Multi-agent consistency detection system

Abstract

The invention relates to the technical field of intelligent control, and discloses a multi-agent consistency detection system, which effectively improves the cooperative control capability of a multi-agent system in a dynamic environment. A real-time topology updating mechanism guarantees rapid convergence when a network structure changes, a prediction compensation module significantly reduces control delay caused by environmental interference, and a multi-dimensional health monitoring system improves the accuracy and efficiency of fault recovery. Functional decoupling is achieved through the modular design, system expansion and maintenance are facilitated, autonomous adjustment and optimization of a recovery strategy are achieved through a strategy optimization mechanism driven by reinforcement learning, the fault recovery time is shortened, and the success rate is increased. A secondary health verification mechanism ensures effectiveness of recovery operation, and secondary faults caused by continuous operation of the system in an incomplete recovery state are avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, in particular to a multi-agent consistency detection system. BACKGROUND

[0002] In recent years, with the rapid development of applications such as unmanned aerial vehicle clusters, autonomous vehicle fleets, and industrial multi-robot systems, multi-agent systems need to maintain efficient coordination in complex and uncertain environments, and consistency detection, as a basic link of cooperative control, directly affects the stability and reliability of the system.

[0003] The traditional method has the following problems:

[0004] First, in terms of time-varying topology adaptability, traditional consensus algorithms based on static graph theory models or fixed weight coefficients are difficult to track network structure changes caused by dynamic node joining / leaving in real time. Experimental data shows that when the topology changes more than 3 times per second, the system convergence time will be extended by more than 400%.

[0005] Second, in terms of environmental interference compensation, the current mainstream solution relies on lag analysis of historical state data and cannot effectively predict sudden changes in mobile obstacle trajectories or sudden communication attenuation, resulting in an inherent delay of 150-300ms in compensation instructions.

[0006] Finally, in terms of fault self-healing mechanism, the fixed threshold detection method used cannot adapt to changes in fault characteristics under dynamic load conditions. Actual measurements show that when the calculation load exceeds 70%, the fault misjudgment rate is as high as 35%, and the average time consumed in the recovery process reaches 800ms.

[0007] Therefore, we propose a multi-agent consistency detection system to solve the above problems. SUMMARY

[0008] The purpose of the present application is to provide a multi-agent consistency detection system to solve the problem that traditional consensus algorithms based on static graph theory models or fixed weight coefficients are difficult to track network structure changes caused by dynamic node joining / leaving in real time. Experimental data shows that when the topology changes more than 3 times per second, the system convergence time will be extended by more than 400%.

[0009] To achieve the above purpose, the present application provides the following technical solution: a multi-agent consistency detection system, characterized by comprising a perception fusion module, an event communication module, a time-varying consensus module, a prediction compensation module, a conflict arbitration module, a fault self-healing module, and a monitoring configuration module.

[0010] The perception fusion module is used for collecting the motion state and environmental parameters of the agent in real time through multi-source sensors, and adopting a Kalman filtering algorithm to perform data fusion and noise elimination, and outputting timestamp-aligned standardized state data.

[0011] The event communication module triggers communication based on a preset state change threshold or a link quality event, dynamically maintains a neighbor node list, and optimizes multiple routing paths.

[0012] The time-varying consensus module is used for constructing a time-varying graph network model, updating the Laplacian matrix weight between nodes in real time, and executing a distributed dynamic consensus algorithm to support fast convergence in a topology change scenario.

[0013] The prediction compensation module is used for predicting the environmental dynamic change trend by using a long short-term memory network, and generating a pre-compensation control instruction to offset the expected interference.

[0014] The conflict arbitration module is used for generating a coordination strategy according to real-time resource margins, task priorities, and a preset rule base when detecting a conflict in the behavior of the agent, and retaining a manual operation interface for emergency intervention.

[0015] The fault self-healing module is used for identifying the node fault type through heartbeat detection and behavior analysis, triggering redundant node switching or task migration, and dynamically optimizing fault tolerance parameter configurations based on reinforcement learning.

[0016] The monitoring configuration module is used for visualizing the system consistency indicators, real-time topology, and prediction data, and simultaneously providing a parameter configuration interface for manual tuning.

[0017] Preferably, the perception fusion module comprises a multi-source data acquisition unit and a dynamic fusion calculation unit.

[0018] The multi-source data acquisition unit is used for collecting motion state data of the agent body in real time, synchronously acquiring environmental parameters, applying a unified timestamp to each sensor data, and converting all data to the agent body coordinate system through coordinate transformation.

[0019] The dynamic fusion calculation unit is used for fusing multi-source data by using a Kalman filtering algorithm, and dynamically adjusting the process noise matrix and the observation noise matrix to adapt to different environments.

[0020] Preferably, the event communication module comprises an event detection and triggering unit and an adaptive routing scheduling unit.

[0021] The event detection and triggering unit is used for calculating the state change amount of the agent in real time, evaluating the communication link quality, and adaptively adjusting the triggering threshold according to the network load.

[0022] The adaptive routing scheduling unit is used for maintaining a dynamic neighbor table containing node ID, signal strength, last active time, detecting link state every 200 ms, and automatically removing nodes out of the list if they are lost for more than 500 ms.

[0023] Preferably, the time-varying consensus module includes a topology dynamic modeling unit and a distributed consensus calculation unit.

[0024] The topology dynamic modeling unit is used for constructing a dynamic adjacency matrix according to a real-time neighbor list provided by the communication module, and calculating a time-varying Laplacian matrix.

[0025] The distributed consensus calculation unit is used for executing a dynamic consistency algorithm with time delay compensation.

[0026] Preferably, the prediction compensation module includes a dynamic environment prediction unit and a compensation strategy generation unit.

[0027] The dynamic environment prediction unit processes a historical motion state sequence of an agent and environment perception data through a long short-term memory network, wherein a time window length of the historical state sequence is 510 seconds, and outputs an environment parameter prediction value of an obstacle motion trajectory and a channel attenuation rate within future 23 seconds, and an accompanying confidence evaluation index.

[0028] The compensation strategy generation unit constructs a model predictive control optimization problem with uncertainty constraint based on the environment parameter prediction value, wherein a cost function contains a weighted square sum of state tracking error and control input energy consumption, and solves the problem in a distributed and real-time manner through an alternating direction multiplier method to generate a feedforward compensation control instruction.

[0029] Preferably, the conflict arbitration module includes a multi-objective conflict detection unit and a dynamic priority arbitration unit.

[0030] The multi-objective conflict detection unit continuously monitors behavior conflicts among agents based on a preset rule library, including spatial conflict, resource conflict and logical conflict, adopts a rule engine for conflict matching, and supports dynamic rule loading.

[0031] The dynamic priority arbitration unit is used for calculating an optimal coordination strategy through a multi-objective utility function, wherein function inputs include real-time resource margin, task static and dynamic priority and cooperation history trust degree, and adopts FPGA to accelerate matrix operation and support a distributed negotiation protocol.

[0032] Preferably, the fault self-healing module includes a health state real-time monitoring and elastic recovery decision unit.

[0033] The health state real-time monitoring is used for collecting node operation indexes in real time, including hardware indexes, software indexes and network indexes, adopting a sliding window statistics + isolation forest algorithm to detect abnormal values, and performing fault classification, including node fault, network fault and data fault;

[0034] The elastic recovery decision unit is used for matching a preset strategy library according to the fault type, dynamically optimizing strategy parameters based on reinforcement learning, performing a fault recovery operation, verifying the recovery effect through secondary health checking, and automatically degrading operation if the recovery fails.

[0035] Preferably, the monitoring configuration module includes a multi-dimensional situation visualization unit and a human-computer interaction configuration unit.

[0036] The multi-dimensional situation visualization unit realizes dynamic rendering of two-dimensional data through a framework, and constructs a three-dimensional topology visualization scene in combination with an engine, wherein node states are represented by color coding, and communication link bandwidths are dynamically represented by pipe diameter thickness.

[0037] The human-computer interaction configuration unit implements role-based access control, and defines three levels of permissions, including administrator permission, operator permission and observer permission.

[0038] The administrator permission is used for full parameter modification and emergency intervention, the operator permission is used for parameter modification, and the observer permission is used for limited access.

[0039] The application has the following beneficial effects:

[0040] 1. The application effectively improves the collaborative control ability of the multi-agent system in a dynamic environment. The real-time topology update mechanism ensures fast convergence when the network structure changes, the prediction compensation module significantly reduces the control delay caused by environmental interference, and the multi-dimensional health monitoring system improves the accuracy and efficiency of fault recovery. The modular design realizes function decoupling, facilitating system expansion and maintenance.

[0041] 2. The application can effectively solve the misjudgment problem caused by fault feature drift under dynamic load conditions, and improve the accuracy and environmental adaptability of fault detection. Through the reinforcement learning driven strategy optimization mechanism, the recovery strategy is autonomously optimized, the fault recovery time is shortened, and the success rate is improved. The secondary health checking mechanism ensures the effectiveness of the recovery operation, avoids the secondary fault caused by the system continuing to run in the incomplete recovery state, and provides the last guarantee for the system in the case of serious fault that cannot be repaired. The automatic degradation operation maintains the basic function operation, significantly improves the overall reliability of the multi-agent system. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The figure is a system flowchart of the application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0044] Embodiment one: please refer to Figure 1 A multi-agent consistency detection system, characterized in that: comprising a perception fusion module, an event communication module, a time-varying consensus module, a prediction compensation module, a conflict arbitration module, a fault self-healing module, and a monitoring configuration module.

[0045] The perception fusion module is used to collect the motion state and environmental parameters of the agent in real time through multi-source sensors, and to perform data fusion and noise elimination using Kalman filtering algorithm, and to output timestamp-aligned standardized state data.

[0046] The event communication module triggers communication based on a preset state change threshold or a link quality event, dynamically maintains a neighbor node list and optimizes multiple routing paths.

[0047] The time-varying consensus module is used to construct a time-varying graph network model, to update the Laplacian matrix weight between nodes in real time, and to execute a distributed dynamic consensus algorithm to support fast convergence in a topology change scenario.

[0048] The prediction compensation module is used to predict the trend of environmental dynamic changes using a long short-term memory network, and to generate a pre-compensation control instruction to offset the expected interference.

[0049] The conflict arbitration module is used to generate a coordination strategy according to real-time resource margin, task priority, and a preset rule base when detecting a conflict in agent behavior, and to retain a manual operation interface for emergency intervention.

[0050] The fault self-healing module is used to identify node fault types through heartbeat detection and behavior analysis, to trigger redundant node switching or task migration, and to dynamically optimize fault tolerance parameter configurations based on reinforcement learning.

[0051] The monitoring configuration module is used to visually display system consistency indicators, real-time topology, and prediction data, and to provide a parameter configuration interface for manual tuning.

[0052] In this embodiment: In the prior art, the field of multi-agent cooperative control has long faced the challenge of inadequate adaptability to dynamic environments. Traditional systems use fixed topology models and static parameter configurations, making it difficult to cope with scenarios where nodes frequently join and exit. When the network structure changes more than three times per second, the system's convergence efficiency decreases significantly. The environmental disturbance compensation mechanism relies on historical data analysis and cannot effectively predict sudden changes in the trajectory of moving obstacles, resulting in inherent delays in control instructions. Fault detection uses a fixed threshold strategy, which has a high misjudgment rate under dynamic load conditions and a long recovery process.

[0053] To solve the above problems, the research and development team found that the poor adaptability of dynamic topology is due to the lag in network weight updates, and proposed to build a real-time updated adjacency matrix model. To address the delay in environmental disturbance compensation, it was recognized that a feedforward prediction mechanism was needed to offset anticipated disturbances. In terms of fault recovery, it was found that traditional threshold detection cannot distinguish between transient fluctuations and real faults, and that a multi-dimensional health state assessment needs to be introduced. By integrating dynamic graph theory, time series prediction, and reinforcement learning techniques, a modular solution was formed.

[0054] Therefore, the present application proposes a system architecture including a perception fusion module, an event communication module, a time-varying consensus module, a prediction compensation module, a conflict arbitration module, a fault self-healing module, and a monitoring configuration module. The perception fusion module collects motion states and environmental parameters through multi-source sensors and uses Kalman filtering for data fusion. The event communication module triggers communication based on state change thresholds and link quality events, and dynamically maintains a list of neighbor nodes. The time-varying consensus module builds a time-varying graph network model and updates the Laplacian matrix weights in real time. The prediction compensation module uses a long short-term memory network to generate pre-compensation control instructions. The conflict arbitration module generates coordination strategies based on resource availability and task priority. The fault self-healing module identifies node fault types through heartbeat detection and triggers redundancy switching. The monitoring configuration module provides a visual interface and parameter configuration interface.

[0055] The perception fusion module refers to a data acquisition and processing unit integrating multiple source sensors, which can specifically work cooperatively with an inertial measurement unit, a laser radar and a visual sensor, unify the data coordinate system through coordinate transformation and eliminate the data deviation between sensors. The event communication module refers to a network management unit dynamically adjusting communication triggering, which can specifically evaluate signal strength by using a link quality detection algorithm, trigger route reconstruction when the packet loss rate exceeds a set threshold and maintain network connectivity. The time-varying consensus module refers to a distributed computing unit processing dynamic topology, which can specifically use a weighted adjacency matrix to represent node connection strength and determine the convergence speed through matrix eigenvalue calculation. The prediction compensation module refers to an environmental disturbance prediction unit, which can specifically use a recurrent neural network to process time series data and generate a feedforward control quantity to offset the expected disturbance. The conflict arbitration module refers to a multi-objective decision unit, which can specifically use a game theory model to calculate an optimal coordination strategy and balance task priority and resource consumption. The fault self-healing module refers to a system health management unit, which can specifically use a sliding window to statistically detect node abnormal behavior and optimize the recovery strategy in combination with reinforcement learning. The monitoring configuration module refers to a human-computer interaction interface, which can specifically use WebGL technology to realize three-dimensional topology visualization and guarantee system security through role permission management.

[0056] The multi-source sensor continuously collects the motion state of the agent and the environmental parameters, and outputs standardized data after eliminating noise through Kalman filtering. When the node state changes exceed the preset threshold or the communication link quality decreases, the neighbor list is updated and the routing path is optimized. The time-varying graph network model calculates the Laplacian matrix according to the real-time topology structure, and drives the distributed consensus algorithm to quickly converge. The long short-term memory network analyzes historical data to predict environmental change trends and generates pre-compensation instructions to reduce control delay. The conflict detection unit continuously monitors the agent behavior and generates a coordination strategy based on dynamic priority. The health monitoring unit identifies the fault type through multi-dimensional index analysis and triggers the redundant node to take over the task. The visualization interface displays the system state in real time and supports dynamic parameter configuration.

[0057] The traditional system uses a fixed topology model, which leads to poor adaptability to dynamic scenes. The present scheme realizes real-time updating of topology through a time-varying graph network. The existing method relies on historical data to compensate for environmental disturbances, and the present scheme generates control instructions in advance through a prediction model. The traditional fault detection uses a single threshold, which is easy to misjudge, and the present scheme improves the recovery accuracy in combination with behavior analysis and reinforcement learning.

[0058] The present application effectively improves the cooperative control capability of the multi-agent system in a dynamic environment. The real-time topology updating mechanism guarantees fast convergence when the network structure changes, the prediction compensation module significantly reduces the control delay caused by environmental disturbances, and the multi-dimensional health monitoring system improves the accuracy and efficiency of fault recovery. The modular design realizes function decoupling, facilitating system expansion and maintenance.

[0059] Embodiment Two: Please refer to Figure 1 , the perception fusion module comprises a multi-source data acquisition unit and a dynamic fusion calculation unit;

[0060] The multi-source data acquisition unit is used for collecting the motion state data of the agent body in real time, synchronously acquiring the environment parameters, marking uniform time stamps on the sensor data, and converting all the data to the agent body coordinate system through coordinate transformation;

[0061] The dynamic fusion calculation unit is used for fusing multi-source data by using the Kalman filtering algorithm, and dynamically adjusting the process noise matrix and the observation noise matrix to adapt to different environments.

[0062] In this embodiment, the perception fusion module further comprises a multi-source data acquisition unit and a dynamic fusion calculation unit; the multi-source data acquisition unit is used for collecting the motion state data of the agent body in real time, synchronously acquiring the environment parameters, marking uniform time stamps on the sensor data, and converting all the data to the agent body coordinate system through coordinate transformation; and the dynamic fusion calculation unit is used for fusing multi-source data by using the Kalman filtering algorithm, and dynamically adjusting the process noise matrix and the observation noise matrix to adapt to different environments.

[0063] The multi-source data acquisition unit refers to a hardware component that integrates multiple sensors for synchronous data acquisition, and can be implemented by using a multi-thread acquisition architecture in cooperation with a hardware time synchronization module, realizes microsecond-level time synchronization through GPS or IEEE 1588 protocol, and solves the time domain misalignment problem of multi-source data.

[0064] The dynamic fusion calculation unit refers to a data processing unit based on Kalman filtering, and can be implemented by using an embedded system to carry an adaptive filtering algorithm, estimates the statistical characteristics of sensor noise online, updates the filtering parameters in real time, and eliminates the measurement errors caused by environmental interference.

[0065] The multi-source data acquisition unit collects the motion state data of the agent body and the environment parameters in parallel, adds uniform time stamps to all the sensor data by using a time synchronization mechanism, and converts the multi-source data to the same body coordinate system through coordinate transformation, thereby eliminating the problem of inconsistent space reference caused by differences in sensor installation positions. After receiving the standardized data, the dynamic fusion calculation unit dynamically adjusts the process noise matrix and the observation noise matrix of the Kalman filter according to the current environmental noise level, for example, automatically increases the observation noise weight when the vibration interference is enhanced, thereby suppressing the influence of abnormal data on the fusion result, and outputs stable and reliable state estimation values.

[0066] The traditional method usually adopts a fixed time window for data alignment, cannot eliminate the timing deviation caused by the inherent delay of the sensor, and relies on preset static parameters in noise processing, which is difficult to adapt to dynamic environmental changes. Through hardware-level time synchronization and dynamic noise matrix adjustment, the scheme realizes accurate space-time alignment of multi-source data and environment-adaptive data fusion, effectively improving the accuracy and robustness of state estimation.

[0067] The application solves the problem of reduced fusion accuracy caused by inconsistent space-time reference of sensor data and environmental interference in a multi-agent system, ensures that each agent obtains unified and reliable state information, and provides accurate data basis for subsequent consistency control.

[0068] Embodiment three: please refer to Figure 1 The event communication module includes an event detection and triggering unit and an adaptive routing scheduling unit.

[0069] The event detection and triggering unit is used to calculate the state change of the agent in real time, evaluate the communication link quality, and adaptively adjust the triggering threshold according to the network load.

[0070] The adaptive routing scheduling unit is used to maintain a dynamic neighbor table containing node ID, signal strength, and last active time, detect the link state every 200 ms, and automatically remove nodes that have been disconnected for more than 500 ms from the list.

[0071] In this embodiment, the application further proposes that the event communication module includes an event detection and triggering unit and an adaptive routing scheduling unit; the event detection and triggering unit is used to calculate the state change of the agent in real time, evaluate the communication link quality, and adaptively adjust the triggering threshold according to the network load; and the adaptive routing scheduling unit is used to maintain a dynamic neighbor table containing node ID, signal strength, and last active time, detect the link state every 200 ms, and automatically remove nodes that have been disconnected for more than 500 ms from the list.

[0072] The event detection and triggering unit refers to a logical component that can dynamically adjust the communication triggering condition according to the state change of the agent and the link quality, and can be implemented by using sliding window variance calculation combined with a link error rate monitoring module. By analyzing the state data fluctuation amplitude and channel error rate index in real time, the triggering threshold is dynamically adjusted to balance the communication overhead and information timeliness. The adaptive routing scheduling unit refers to a mechanism that can autonomously maintain the connection state of neighbor nodes and optimize the communication path, and can be implemented by using a neighbor table management algorithm based on link quality prediction. By periodically detecting the link state and automatically removing invalid nodes, the routing information is ensured to be real-time and accurate.

[0073] The event detection and triggering unit continuously monitors the state change amplitude of the agent and the communication link quality index. When the state change exceeds the dynamic threshold corresponding to the current network load or the link quality is lower than the preset standard, the communication event is triggered and the neighbor node information is updated. The adaptive routing scheduling unit scans the communication link state at a fixed period, for example, performs link health degree detection every 200ms, automatically performs list cleaning operation for the node which is continuously out of contact for more than 500ms, and dynamically selects the optimal multi-hop path according to the real-time signal strength. The mechanism effectively solves the communication congestion or information lag problem caused by the traditional fixed threshold communication triggering mechanism under the network load fluctuation through dynamic threshold adjustment and periodic link maintenance.

[0074] The traditional scheme triggers communication by using a fixed state change threshold, which cannot adapt to the dynamic change of bandwidth resources caused by network load fluctuation, and is easy to cause waste of communication resources or delay of information update. The scheme introduces a dynamic threshold adjustment mechanism adaptive to network load, so that the communication triggering condition can be dynamically optimized according to the network state, and at the same time, the real-time performance of the neighbor list and the reliability of the communication path are significantly improved by combining the periodic link state detection and the automatic exclusion mechanism of the invalid node.

[0075] The application can effectively reduce the communication redundancy in the dynamic topology environment, improve the multi-hop routing maintenance efficiency, and ensure the real-time performance and reliability of the communication between agents. The scheme solves the problem that the traditional fixed threshold communication mechanism is easy to cause communication storm when the network load suddenly changes, and avoids the path selection error caused by the retention of invalid nodes in the neighbor list, thereby providing stable communication guarantee for multi-agent cooperation in the time-varying topology scene.

[0076] Embodiment four: please refer to Figure 1 , the time-varying consensus module includes a topology dynamic modeling unit and a distributed consensus computing unit;

[0077] The topology dynamic modeling unit is configured to construct a dynamic adjacency matrix according to the real-time neighbor list provided by the communication module, and calculate a time-varying Laplacian matrix.

[0078] The distributed consensus computing unit is configured to execute a dynamic consistency algorithm with time delay compensation.

[0079] In this embodiment, the application further proposes a technical scheme including a topology dynamic modeling unit and a distributed consensus computing unit. The topology dynamic modeling unit constructs a dynamic adjacency matrix according to the real-time neighbor list provided by the communication module and calculates a time-varying Laplacian matrix, and the distributed consensus computing unit executes a dynamic consistency algorithm with time delay compensation.

[0080] The topology dynamic modeling unit refers to a component for constructing a mathematical model reflecting the connection relationship between nodes in real time. Specifically, an adjacency matrix updating algorithm can be used to achieve this. By periodically receiving the neighbor node information of the communication module, the matrix element value is dynamically adjusted to match the current network topology. Through real-time updating of the adjacency matrix, the unit can effectively track the network structure changes caused by dynamic joining or exiting of nodes.

[0081] The distributed consensus calculation unit refers to a component for performing a multi-agent collaborative calculation process. Specifically, a time delay compensation algorithm can be used to achieve this. By introducing a communication delay estimation value to correct the control instruction, it ensures that each node can reach a state agreement even when there is transmission delay. By compensating for communication delay, the unit can eliminate the problem of decreased convergence speed caused by network fluctuations.

[0082] The topology dynamic modeling unit receives the neighbor node list maintained by the event communication module, such as the dynamic neighbor table updated every 200 milliseconds, to construct an adjacency matrix reflecting the current connection state. The matrix element value is dynamically adjusted according to the communication quality between nodes, such as converting signal strength into weight coefficients. Based on the adjacency matrix, a time-varying Laplace matrix is calculated to provide dynamic topology parameters for subsequent consensus algorithms. After receiving standardized state data, the distributed consensus calculation unit uses an improved consistency algorithm to handle communication delay, such as introducing a time delay compensation term in the state update equation, which eliminates the delay effect by predicting the future state value of adjacent nodes, thereby maintaining the algorithm convergence speed in scenarios with frequent topology changes.

[0083] Traditional solutions usually use fixed topology models or static weight coefficients, which cannot effectively handle network structure fluctuations caused by dynamic changes in nodes. However, this solution can adaptively track topology changes by real-time updating of the adjacency matrix and Laplace matrix, and offset the impact of network transmission delay on algorithm convergence through a time delay compensation mechanism. The consensus algorithm based on static models in existing technologies will have a significant increase in convergence time when the topology changes more than 3 times per second. This solution effectively alleviates this defect through dynamic modeling and delay compensation.

[0084] This application can improve the collaborative control capability of multi-agent systems in dynamic topology scenarios. By accurately reflecting the current connection state through real-time updating of the network model, and combining the time delay compensation mechanism to ensure that each node can quickly reach a state agreement even when there is communication delay. This solves the problem of decreased system convergence speed caused by topology change tracking lag and communication delay accumulation in traditional methods, enhancing the collaborative stability and response real-time performance of multi-agent systems in complex environments.

[0085] Embodiment five: please refer to Figure 1 The prediction compensation module includes a dynamic environment prediction unit and a compensation strategy generation unit.

[0086] The dynamic environment prediction unit processes the agent historical motion state sequence and environment perception data through a long short-term memory network, wherein the time window length of the historical state sequence is 510 seconds, and outputs the environment parameter prediction values of the obstacle motion trajectory and channel attenuation rate within the next 23 seconds, and the confidence evaluation index is attached.

[0087] The compensation strategy generation unit constructs a model predictive control optimization problem with uncertainty constraints based on the environment parameter prediction values, wherein the cost function contains the weighted sum of square of state tracking error and control input energy consumption, and solves it in real time through an alternating direction multiplier method to generate a feedforward compensation control instruction.

[0088] In the embodiment, the application further proposes that the prediction compensation module comprises a dynamic environment prediction unit and a compensation strategy generation unit; the dynamic environment prediction unit processes the agent historical motion state sequence and environment perception data through a long short-term memory network, wherein the time window length of the historical state sequence can be 510 seconds, and outputs the environment parameter prediction values of the obstacle motion trajectory and channel attenuation rate within the next 23 seconds, and the confidence evaluation index is attached; the compensation strategy generation unit constructs a model predictive control optimization problem with uncertainty constraints based on the environment parameter prediction values, wherein the cost function contains the weighted sum of square of state tracking error and control input energy consumption, and solves it in real time through an alternating direction multiplier method to generate a feedforward compensation control instruction.

[0089] The long short-term memory network is a recurrent neural network with a memory gating mechanism, which can be implemented by a network structure containing an input gate, a forgetting gate and an output gate, and is used to capture long-term dependencies in time series data. The time window length refers to the time span of historical data used for model training, which can be sampled using a sliding window mechanism, and the window length can be dynamically adjusted according to the actual scene, for example, set to 510 seconds to balance the completeness of historical information and computational efficiency. The confidence evaluation index refers to a parameter that quantifies the reliability of the prediction result, which can be characterized by the variance of the probability distribution or the prediction error range, and is used for risk assessment of the subsequent control strategy. The model predictive control optimization problem refers to an optimal control problem based on multi-step prediction in the future, which can be constructed by a rolling horizon optimization framework to build the objective function and constraint conditions, and the cost function can contain the weighted combination of state tracking error and control energy consumption. The alternating direction multiplier method is a distributed optimization algorithm, which can be implemented in parallel computing using a decomposition coordination strategy, and is suitable for real-time solving requirements of multi-agent systems.

[0090] The dynamic environment prediction unit performs time series modeling on the historical motion state and environment perception data of the agent by a long short-term memory network, for example, learns the environmental dynamic change rule by using an input sequence with a time window length of 510 seconds. The network outputs the obstacle trajectory prediction value and the channel attenuation rate prediction value in the next 23 seconds, and generates a confidence index to reflect the prediction reliability. After receiving the prediction result, the compensation strategy generation unit constructs an optimization model containing uncertainty constraints, for example, converts the confidence into a constraint boundary, minimizes the tracking error and control energy consumption on the premise of ensuring robustness. Distributed solution is performed by the alternating direction method of multipliers, and each agent only needs to exchange necessary information to cooperatively complete the optimization calculation, and finally generates a control instruction with a feedforward compensation effect.

[0091] The traditional method relies on lag analysis of historical state data and cannot effectively predict sudden environmental changes, resulting in inherent delay of the compensation instruction. The present scheme realizes forward-looking prediction by a long short-term memory network, and combines a distributed real-time optimization algorithm to generate a compensation instruction before the environmental disturbance occurs. At the same time, the introduction of the confidence evaluation index and the uncertainty constraint enhances the robustness of the system in dealing with prediction errors and avoids the problem of control instability caused by prediction deviation.

[0092] The present application effectively reduces the influence of environmental dynamic changes on multi-agent collaborative control, can generate a compensation instruction in advance in the scene of sudden change of the trajectory of a moving obstacle or sudden drop of communication quality, and significantly improves the anti-interference ability and collaborative precision of the system in a complex environment. At the same time, the distributed optimization framework guarantees the real-time demand of the large-scale agent system and avoids the communication bottleneck caused by centralized calculation.

[0093] Embodiment six: please refer to Figure 1 , the conflict arbitration module includes a multi-objective conflict detection unit and a dynamic priority arbitration unit;

[0094] The multi-objective conflict detection unit continuously monitors the behavior conflicts between agents based on a preset rule base, including spatial conflicts, resource conflicts and logical conflicts, uses a rule engine for conflict matching, and supports dynamic rule loading;

[0095] The dynamic priority arbitration unit is used to calculate the optimal coordination strategy by a multi-objective utility function, the function input includes real-time resource margin, task static and dynamic priority and cooperation history trust degree, uses FPGA to accelerate matrix operation, and supports distributed negotiation protocol.

[0096] In the embodiment, the application further proposes a conflict arbitration module including a multi-target conflict detection unit and a dynamic priority arbitration unit. The multi-target conflict detection unit continuously monitors the behavior conflicts among agents based on a preset rule base, including spatial conflicts, resource conflicts, and logical conflicts. A rule engine is used for conflict matching, and dynamic rule loading is supported. The dynamic priority arbitration unit calculates the optimal coordination strategy through a multi-target utility function. The function input includes real-time resource reserves, task static and dynamic priority, and cooperation history trust degree. FPGA is used to accelerate matrix operation, and distributed negotiation protocol is supported.

[0097] The multi-target conflict detection unit refers to a detection mechanism that can identify multiple types of behavior conflicts simultaneously. It can be implemented using a state machine-based event matching engine. The conflict judgment conditions are defined by a preset rule base, and online rule loading is supported to adapt to different scene requirements. This unit is used in the scheme to quickly locate the interaction contradictions among agents and avoid strategy blind spots caused by single-type conflict detection.

[0098] The dynamic priority arbitration unit refers to a decision mechanism that generates coordination strategies based on multi-dimensional parameters. It can be implemented using a mixed integer programming model combined with a hardware accelerator. An utility function is constructed to include resource reserves, task priority, and historical cooperation data. This unit is used in the scheme to balance multi-party interest conflicts and ensure overall system efficiency.

[0099] FPGA accelerated matrix operation refers to a technical means to improve computing performance using programmable logic devices. It can be implemented using a parallel computing architecture to accelerate matrix multiplication in hardware. In the scheme, it is used to shorten the solution time of multi-objective optimization problems and meet real-time decision-making requirements.

[0100] When the system detects behavior conflicts among agents, the multi-target conflict detection unit matches the preset conflict types through the rule engine. For example, when the path planning of multiple agents overlaps, spatial conflict detection is triggered. When shared computing resources are over-allocated, resource conflict detection is triggered. The dynamic priority arbitration unit constructs a multi-objective optimization model based on real-time resource reserve data, preset task priority weights, and historical cooperation success rate indicators. It generates coordination strategies through FPGA accelerated computation, such as reallocating path priority or adjusting resource quotas. Distributed negotiation protocols are supported during arbitration, allowing agents to reach global consensus while maintaining local decision-making autonomy.

[0101] Traditional conflict arbitration schemes usually adopt a single type of conflict detection mechanism, which cannot effectively distinguish the coupling relationship of space, resource and logical conflicts, and relies on a central controller for centralized decision-making, which has the problems of high response delay and poor scalability. The scheme combines multi-dimensional conflict classification detection and distributed negotiation mechanism, improves the system response speed while maintaining detection accuracy, and can adapt to dynamically changing collaborative scenarios.

[0102] The application can effectively solve the real-time coordination problem in complex conflict scenarios in multi-agent systems. Through the cooperative work of multi-type conflict detection and dynamic priority arbitration, the system operation safety is guaranteed while the resource allocation efficiency is optimized, which is suitable for distributed collaborative control systems that need to respond quickly to environmental changes.

[0103] Embodiment seven: please refer to Figure 1 , the fault self-healing module includes a health state real-time monitoring and elastic recovery decision unit;

[0104] The health state real-time monitoring is used to collect node operation indicators in real time, including hardware indicators, software indicators and network indicators, and uses sliding window statistics and isolation forest algorithm to detect abnormal values and perform fault classification, including node failure, network failure and data failure;

[0105] The elastic recovery decision unit is used to match the preset strategy library according to the fault type, dynamically optimize the strategy parameters based on reinforcement learning, execute the fault recovery operation, verify the recovery effect through secondary health check, and automatically degrade the operation if the recovery fails.

[0106] In this embodiment: the application further proposes a fault self-healing module including a health state real-time monitoring and elastic recovery decision unit, the health state real-time monitoring is used to collect node operation indicators in real time, including hardware indicators, software indicators and network indicators, and uses sliding window statistics and isolation forest algorithm to detect abnormal values and perform fault classification, including node failure, network failure and data failure; the elastic recovery decision unit is used to match the preset strategy library according to the fault type, dynamically optimize the strategy parameters based on reinforcement learning, execute the fault recovery operation, verify the recovery effect through secondary health check, and automatically degrade the operation if the recovery fails.

[0107] The sliding window statistics refers to continuously collecting system running index data in time sequence and establishing a dynamic benchmark range, which can be implemented by a fixed time length or a variable time length data window, and abnormal fluctuations are identified by the statistical characteristics of the data in the window. The isolated forest algorithm refers to an unsupervised anomaly detection method based on tree structure, which can calculate the anomaly score of data points by constructing a random partition path, and is suitable for fast anomaly detection of high-dimensional time series data. The secondary health check refers to re-collecting key indicators for verification after performing the recovery operation, which can set a verification period and a verification threshold to confirm whether the fault is truly eliminated. The automatic degradation operation refers to switching to a preset simplified working mode after confirming the recovery failure, which can be implemented by shutting down non-core functional modules or reducing the task execution frequency to ensure the minimum running ability of the system.

[0108] The node running index is continuously collected by multi-dimensional sensors. The hardware index can include processor temperature and memory occupancy rate. The software index can include thread response delay and task queue depth. The network index can include packet loss rate and link delay. When establishing a dynamic benchmark by sliding window statistics, the mean and standard deviation are calculated using a 10-second time window. When the real-time data deviates from the benchmark by more than 3 times the standard deviation, abnormal detection is triggered. When constructing a binary tree structure by the isolated forest algorithm, the data space is divided by randomly selecting features. Abnormal data is isolated at a shallow level due to sparse distribution. In the fault classification process, node faults can be characterized by continuous over-standard hardware indicators, network faults can be characterized by communication interruption across nodes, and data faults can be characterized by error or time sequence disorder. When optimizing strategy parameters by reinforcement learning, a state-action value function is established by the Q-learning algorithm, and the policy selection weight is dynamically adjusted according to the historical recovery effect.

[0109] The traditional fixed threshold detection method cannot accurately distinguish between normal fluctuations and real faults in dynamic load scenarios. The present scheme establishes a dynamic benchmark by sliding window statistics and realizes environment-adaptive anomaly detection by combining the isolated forest algorithm. The existing technology uses a static recovery strategy, which results in long recovery time and low success rate. The present scheme dynamically optimizes parameters by reinforcement learning, so that the recovery strategy can be adjusted automatically according to the real-time system state. The conventional single verification mechanism has a risk of misjudgment. The present scheme forms a closed-loop verification mechanism by secondary health check, effectively avoiding false recovery state determination.

[0110] The application can effectively solve the misjudgment problem caused by fault feature drift under dynamic load conditions, and improve the accuracy and environmental adaptability of fault detection. Through the strategy optimization mechanism driven by reinforcement learning, the autonomous optimization of the recovery strategy is realized, the fault recovery time is shortened and the success rate is improved. The secondary health check mechanism ensures the effectiveness of the recovery operation, avoids the secondary failure caused by the system continuing to run in the state of incomplete recovery, and provides the last safeguard for the automatic degradation operation. When the serious fault cannot be repaired, the basic function operation is maintained, which significantly improves the overall reliability of the multi-agent system.

[0111] Embodiment eight: please refer to Figure 1 The monitoring configuration module includes a multi-dimensional situation visualization unit and a human-computer interaction configuration unit.

[0112] The multi-dimensional situation visualization unit realizes dynamic rendering of two-dimensional data through a framework, and constructs a three-dimensional topology visualization scene combined with an engine, wherein node states are represented by color coding, and communication link bandwidth is dynamically represented by pipe diameter thickness.

[0113] The human-computer interaction configuration unit implements role-based access control, defines three levels of permissions, including administrator permissions, operator permissions and observer permissions.

[0114] The administrator permissions are used for full parameter modification and emergency intervention, the operator permissions are used for parameter modification, and the observer permissions are used for limited access.

[0115] In the embodiment, the monitoring configuration module further includes a multi-dimensional situation visualization unit and a human-computer interaction configuration unit. The multi-dimensional situation visualization unit realizes dynamic rendering of two-dimensional data through a framework, and constructs a three-dimensional topology visualization scene combined with an engine, wherein node states are represented by color coding, and communication link bandwidth is dynamically represented by pipe diameter thickness. The human-computer interaction configuration unit implements role-based access control, defines three levels of permissions, including administrator permissions, operator permissions and observer permissions, wherein the administrator permissions are used for full parameter modification and emergency intervention, the operator permissions are used for parameter modification, and the observer permissions are used for limited access.

[0116] The multi-dimensional situation visualization unit refers to multi-angle display of system state through dynamic rendering technology. Specifically, two-dimensional data rendering can be achieved by using a WebGL framework, a three-dimensional topology scene can be constructed by using a Unity engine, a link bandwidth visualization pipeline can be generated by using a Bezier curve algorithm, and a node health state can be mapped by using an HSV color space. The unit converts complex system parameters into intuitive graphical elements through visual coding, thereby solving the information overload problem of a traditional monitoring interface. The role-based access control refers to dividing permission levels according to operation responsibilities. Specifically, identity authentication can be achieved by using an OAuth2.0 protocol, permission verification can be performed by using a policy engine, the administrator permission can cover all configuration items of the system, the operator permission can be limited within a preset parameter range, and the observer permission can only open a read-only interface. The mechanism prevents misoperation risks by permission isolation, thereby ensuring system configuration security.

[0117] The multi-dimensional situation visualization unit updates a graphical interface by driving the graphical interface with real-time data flow. When a node state changes, a color coding module automatically updates node coloring according to a preset mapping rule. For example, green represents normal operation, and red represents a fault state. A communication link bandwidth monitoring module periodically collects network traffic data, generates dynamic pipe diameter parameters by using an interpolation algorithm, and realizes visual expression of link load. The human-computer interaction configuration unit performs identity verification during user login, loads a corresponding operation interface according to a role type, the administrator can access a full-function panel including an emergency shutdown button, the operator only displays adjustable parameter sliders, and the observer interface locks all input controls. System configuration change requests need to be verified by a policy engine, to ensure that the operation complies with permission constraints.

[0118] Traditional monitoring systems usually use single-dimensional table data display, lack intuitive expression of spatial topology relationships, and usually only distinguish two levels of administrators and ordinary users. The scheme enhances situation awareness capability by using three-dimensional visualization technology, so that an operator can quickly locate a network topology abnormal area. Meanwhile, fine-grained permission division ensures system security and meets operation requirements of different posts.

[0119] The application effectively improves information transmission efficiency of a multi-agent system monitoring interface, so that a complex network state can be quickly identified and understood. Meanwhile, the application reduces configuration error risks by using hierarchical permission control, improves operation convenience on the premise of ensuring system security, and improves operation convenience on the premise of ensuring system security.

[0120] The contents not described in detail in the specification belong to the prior art known to those skilled in the art.

[0121] Although the present application has been described in detail with reference to the foregoing embodiments, the technical solutions recorded in the foregoing embodiments can be modified, or some of the technical features can be replaced by equivalent features, by those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-agent consistency detection system, characterized in that: The perception fusion module, the event communication module, the time-varying consensus module, the prediction compensation module, the conflict arbitration module, the fault self-healing module and the monitoring configuration module are comprised; The perception fusion module is used for collecting the motion state and environmental parameters of the agent in real time through multi-source sensors, and performing data fusion and noise elimination by using a Kalman filtering algorithm, and outputting timestamp-aligned standardized state data; The event communication module triggers communication based on a preset state change threshold or a link quality event, dynamically maintains a neighbor node list and optimizes multiple routing paths; The time-varying consensus module is used for constructing a time-varying graph network model, updating the Laplacian matrix weight between nodes in real time, and executing a distributed dynamic consistency algorithm to support fast convergence in a topology change scenario; The prediction compensation module is used for predicting the environmental dynamic change trend by using a long short-term memory network, and generating a pre-compensation control instruction to offset the expected interference; The conflict arbitration module is used for generating a coordination strategy according to real-time resource margins, task priorities and a preset rule base when detecting a conflict in the agent behavior, and retaining a manual operation interface for emergency intervention. The fault self-healing module is used for identifying the node fault type through heartbeat detection and behavior analysis, triggering redundant node switching or task migration, and dynamically optimizing fault-tolerant parameter configurations based on reinforcement learning. The monitoring configuration module is used for visualizing the system consistency indicators, real-time topology and prediction data, and simultaneously providing a parameter configuration interface for manual tuning.

2. The multi-agent consistency detection system of claim 1, wherein: The perception fusion module comprises a multi-source data acquisition unit and a dynamic fusion calculation unit. The multi-source data acquisition unit is used for collecting motion state data of the agent body in real time, synchronously acquiring environmental parameters, marking all sensor data with a unified timestamp, and converting all data to the agent body coordinate system through coordinate transformation; The dynamic fusion calculation unit is used for fusing multi-source data by using a Kalman filtering algorithm, and dynamically adjusting the process noise matrix and the observation noise matrix to adapt to different environments.

3. The multi-agent consistency detection system of claim 2, wherein: The event communication module comprises an event detection and triggering unit and an adaptive routing scheduling unit. The event detection and triggering unit is used for calculating the state change amount of the agent in real time, evaluating the communication link quality, and adaptively adjusting the triggering threshold according to the network load; The adaptive routing scheduling unit is used for maintaining a dynamic neighbor table containing node ID, signal strength and last active time, detecting the link state every 200 ms, and automatically removing nodes that are out of connection for more than 500 ms from the list.

4. The multi-agent consistency detection system of claim 3, wherein: The time-varying consensus module comprises a topology dynamic modeling unit and a distributed consensus calculation unit. The topology dynamic modeling unit is used for constructing a dynamic adjacency matrix according to the real-time neighbor list provided by the communication module, and calculating a time-varying Laplacian matrix; The distributed consensus calculation unit is used for executing a dynamic consistency algorithm with time delay compensation.

5. The multi-agent consistency detection system of claim 4, wherein: The prediction compensation module comprises a dynamic environment prediction unit and a compensation strategy generation unit; The dynamic environment prediction unit processes the intelligent agent historical motion state sequence and environment perception data through a long short-term memory network, wherein the time window length of the historical state sequence is 510 seconds, and outputs the environment parameter prediction values of the obstacle motion trajectory and channel attenuation rate within the next 23 seconds, and the confidence evaluation index is attached; The compensation strategy generation unit constructs a model predictive control optimization problem with uncertainty constraints based on the environment parameter prediction values, wherein the cost function contains the weighted sum of square of state tracking error and control input energy consumption, and the problem is solved in a distributed and real-time manner through an alternating direction multiplier method to generate a feedforward compensation control instruction.

6. The multi-agent consistency detection system of claim 5, wherein: The conflict arbitration module includes a multi-objective conflict detection unit and a dynamic priority arbitration unit. The multi-objective conflict detection unit continuously monitors the behavior conflicts between intelligent agents based on a preset rule base, including spatial conflicts, resource conflicts, and logical conflicts, and uses a rule engine for conflict matching, supporting dynamic rule loading. The dynamic priority arbitration unit is used to calculate the optimal coordination strategy through a multi-objective utility function, and the function input includes real-time resource margin, task static and dynamic priority, and cooperation history trust degree, and FPGA acceleration matrix operation is adopted to support distributed negotiation protocol.

7. The multi-agent consistency detection system of claim 6, wherein: The fault self-healing module includes a health state real-time monitoring and elastic recovery decision unit. The health state real-time monitoring is used to collect node operation indicators in real time, including hardware indicators, software indicators, and network indicators, and uses sliding window statistics and isolation forest algorithm to detect outliers and perform fault classification, including node fault, network fault, and data fault. The elastic recovery decision unit is used to match the preset strategy library according to the fault type, dynamically optimize the strategy parameters based on reinforcement learning, execute fault recovery operations, and verify the recovery effect through secondary health check. If the recovery fails, it automatically downgrades the operation.

8. The multi-agent consistency detection system of claim 7, wherein: The monitoring configuration module includes a multi-dimensional situation visualization unit and a human-computer interaction configuration unit. The multi-dimensional situation visualization unit realizes dynamic rendering of two-dimensional data through a framework, and constructs a three-dimensional topology visualization scene through an engine, wherein the node state is coded by color, and the communication link bandwidth is dynamically represented by pipe diameter. The human-computer interaction configuration unit implements role-based access control, defines three levels of permissions, including administrator permission, operator permission, and observer permission; The administrator permission is used for full parameter modification and emergency intervention, the operator permission is used for parameter modification, and the observer permission is used for limited access.

Images