Multi-agent dynamic scheduling method and system based on efficiency and energy multi-objective optimization

Abstract

The invention provides a multi-agent dynamic scheduling method and system based on efficiency and energy multi-objective optimization. The method belongs to the cross technical field of intelligent manufacturing, multi-agent systems and production scheduling optimization. The method comprises the following steps: constructing a global coordination and local execution double-layer intelligent agent architecture, generating a layered scheduling network topology, carrying out robot cluster dynamic modeling, generating a hundred-level robot state data set, initializing a task pool and a real-time environment map, and forming initial scheduling scene data; through joint optimization of task allocation and path planning, the robot can complete tasks in a more reasonable mode, and the production throughput rate is remarkably improved.

Description

Technical Field

[0001] This invention proposes a multi-agent dynamic scheduling method and system based on multi-objective optimization of efficiency and energy, belonging to the interdisciplinary fields of intelligent manufacturing, multi-agent systems and production scheduling optimization. Background Technology

[0002] In the field of intelligent manufacturing, as production scale continues to expand, scenarios involving hundreds of robots working collaboratively are becoming increasingly common. However, current multi-robot scheduling methods have many shortcomings in dealing with such complex scenarios.

[0003] Most existing systems focus only on single-objective optimization, either simply pursuing the shortest task completion time or only focusing on reducing task latency, while ignoring the cumulative energy consumption effect caused by the robot's movement distance. Although this single-objective optimization approach can improve local efficiency, it often leads to high efficiency accompanied by high energy consumption, failing to achieve efficient utilization of global resources.

[0004] Meanwhile, traditional scheduling methods often employ static allocation strategies, such as the Hungarian algorithm and auction mechanisms. These methods allocate tasks all at once at the beginning of the task, making it difficult to cope with real-time disturbances such as dynamic task insertion, robot malfunctions, and path congestion. Once an emergency occurs, the entire scheduling system may fall into chaos and be unable to adjust in time, severely impacting production efficiency.

[0005] Furthermore, the disconnect between the decision-making levels of task allocation and path planning is a significant problem. The lack of coordination between high-level task allocation and low-level path planning often results in robots that appear to be closest to the task being assigned, but actually require detours, increasing energy consumption and task completion time. Moreover, centralized solvers time out when dealing with hundreds of robots, while pure reinforcement learning methods suffer from convergence difficulties due to state space explosion, exhibiting poor scalability. Therefore, a novel scheduling method is urgently needed that can achieve multi-objective optimization of efficiency and energy, possess dynamic rescheduling capabilities, and is suitable for hundreds of robots. Summary of the Invention

[0006] This invention provides a multi-agent dynamic scheduling method and system based on efficiency and energy multi-objective optimization to solve the problems mentioned in the background art above:

[0007] This invention proposes a multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization, the method comprising:

[0008] S1. Construct a two-layer intelligent agent architecture for global coordination and local execution, generate a hierarchical scheduling network topology, perform dynamic modeling of robot clusters, generate a state dataset of hundreds of robots, initialize the task pool and real-time environment map, and form initial scheduling scenario data.

[0009] S2. Input the initial scheduling scenario data into the improved NSGA-II algorithm module, perform multi-objective optimization modeling, generate Pareto front solution set, and use the spatiotemporal A* algorithm to perform spatiotemporal collision detection and dynamic obstacle avoidance on the candidate paths in the Pareto front solution set to generate a joint optimization path scheme set.

[0010] S3. Based on the joint optimization path scheme set, evaluate the matching degree between tasks and robots, generate a dual-objective weighted fitness function value, screen the optimal task allocation scheme, and simultaneously generate a robot individual path planning instruction set, which is then sent to the local execution layer agent to form initial scheduling instruction data.

[0011] S4. Real-time collection of robot motion state data and task execution progress data by local execution layer intelligent agents, generation of dynamic scheduling feedback data stream, detection of real-time disturbance events, generation of disturbance type identification and impact range assessment report;

[0012] S5. Trigger the online rescheduling mechanism based on the disturbance type identifier and impact range assessment report. Solve the updated joint optimization problem through the improved NSGA-II algorithm module and the spatiotemporal A* algorithm, generate a set of rescheduling path correction schemes, and fuse them with the initial scheduling instruction data to generate dynamic scheduling update instructions.

[0013] S6. Based on the dynamic scheduling update instructions, execute the real-time scheduling operation of the hundred-unit robot system, generate scheduling performance evaluation data, adjust the multi-objective optimization weight parameters of the global coordination layer, form a closed-loop adaptive optimization loop, and finally generate optimized scheduling result data.

[0014] This invention proposes a system for implementing the multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization as described above, the system comprising:

[0015] Dynamic Modeling Module: Constructs a two-layer intelligent agent architecture for global coordination and local execution, generates a hierarchical scheduling network topology, performs dynamic modeling of robot clusters, generates a state dataset of hundreds of robots, initializes the task pool and real-time environment map, and forms initial scheduling scenario data;

[0016] Dynamic obstacle avoidance module: Input the initial scheduling scenario data into the improved NSGA-II algorithm module, perform multi-objective optimization modeling, generate Pareto front solution set, and use the spatiotemporal A* algorithm to perform spatiotemporal collision detection and dynamic obstacle avoidance processing on the candidate paths in the Pareto front solution set to generate a joint optimized path scheme set;

[0017] Path planning module: Based on the joint optimization path scheme set, the task and robot matching degree is evaluated, a dual-objective weighted fitness function value is generated, the optimal task allocation scheme is selected, and the individual robot path planning instruction set is generated simultaneously; the individual robot path planning instruction set is sent to the local execution layer agent to form the initial scheduling instruction data;

[0018] Dynamic scheduling module: Collects robot motion state data and task execution progress data in real time through local execution layer intelligent agents, generates dynamic scheduling feedback data stream, detects real-time disturbance events, and generates disturbance type identification and impact range assessment report;

[0019] Joint optimization module: Based on the disturbance type identifier and impact range assessment report, the online rescheduling mechanism is triggered. The improved NSGA-II algorithm module and the spatiotemporal A* algorithm are used to solve the updated joint optimization problem, generate a set of rescheduling path correction schemes, and fuse them with the initial scheduling instruction data to generate dynamic scheduling update instructions.

[0020] Performance evaluation module: Based on dynamic scheduling update instructions, executes real-time scheduling operations for a system of hundreds of robots, generates scheduling performance evaluation data, adjusts the multi-objective optimization weight parameters of the global coordination layer, forms a closed-loop adaptive optimization loop, and finally generates optimized scheduling result data.

[0021] The beneficial effects of this invention are as follows: By jointly optimizing task allocation and path planning, robots can complete tasks in a more rational manner, significantly improving production throughput. Simultaneously, it reduces the cumulative energy consumption during robot operation, balancing efficiency and energy efficiency to achieve green production. It enhances the system's ability to cope with real-time disturbances, responding within milliseconds and completing online rescheduling in the face of sudden situations such as dynamic task insertion, robot malfunctions, and path congestion, ensuring production continuity. It reduces invalid path planning caused by fragmented decision-making levels, avoiding increased energy consumption and time costs due to robot detours. It avoids the rigidity problem of traditional static allocation methods in complex scenarios, preventing production stoppages caused by the inability to adjust in a timely manner. It meets the stringent efficiency requirements of large-scale intelligent manufacturing while aligning with green development principles to reduce energy consumption, providing an efficient, green, and robust dynamic scheduling infrastructure for intelligent manufacturing. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the steps of the method described in this invention;

[0023] Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0024] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] One embodiment of the present invention, such as Figure 1 As shown, a multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization is proposed, the method comprising:

[0026] S1. Construct a two-layer intelligent agent architecture with global coordination and local execution, and generate a hierarchical scheduling network topology; perform dynamic modeling of the robot cluster based on the hierarchical scheduling network topology to generate a state dataset of hundreds of robots; initialize the task pool and real-time environment map based on the state dataset of hundreds of robots to form initial scheduling scenario data;

[0027] S2. Input the initial scheduling scenario data into the improved NSGA-II algorithm module to perform multi-objective optimization modeling and generate a Pareto front solution set, which includes task completion time and energy consumption accumulation; perform spatiotemporal collision detection and dynamic obstacle avoidance processing on the candidate paths in the Pareto front solution set using the spatiotemporal A* algorithm to generate a joint optimization path scheme set.

[0028] S3. Evaluate the task and robot matching degree based on the joint optimization path scheme set, and generate a dual-objective weighted fitness function value; select the optimal task allocation scheme based on the dual-objective weighted fitness function value, and simultaneously generate a robot individual path planning instruction set; send the robot individual path planning instruction set to the local execution layer agent to form initial scheduling instruction data;

[0029] S4. Collect robot motion state data and task execution progress data in real time through the local execution layer intelligent agent to generate a dynamic scheduling feedback data stream; detect real-time disturbance events based on the dynamic scheduling feedback data stream, including task insertion, robot failure, path blocking, etc., and generate a disturbance type identifier and impact range assessment report.

[0030] S5. Trigger the online rescheduling mechanism based on the disturbance type identifier and impact range assessment report. Solve the updated joint optimization problem through the improved NSGA-II algorithm module and the spatiotemporal A* algorithm to generate a set of rescheduling path correction schemes. Then, fuse the set of rescheduling path correction schemes with the initial scheduling instruction data to generate a dynamic scheduling update instruction with a millisecond-level response.

[0031] S6. Based on the dynamic scheduling update instruction, execute the real-time scheduling operation of the hundred-robot system to generate scheduling performance evaluation data, which includes throughput and energy consumption indicators; adjust the multi-objective optimization weight parameters of the global coordination layer according to the scheduling performance evaluation data to form a closed-loop adaptive optimization loop, and finally generate optimized scheduling result data.

[0032] The working principle and effects of the above technical solution are as follows: By combining a two-layer intelligent agent architecture with an improved multi-objective optimization algorithm, the scheduling efficiency of a 100-unit robot swarm is significantly improved, and the overall operating energy consumption is reduced. Spatiotemporal obstacle avoidance effectively avoids the risk of collisions during robot movement, reducing downtime losses caused by collisions. The dynamic rescheduling mechanism can respond promptly to disturbances such as task insertion and robot malfunctions, avoiding scheduling chaos or task delays caused by disturbances. The closed-loop adaptive optimization loop continuously adjusts the optimization weights, enhancing the adaptability of the scheduling scheme. It can ensure high throughput while stably controlling energy consumption, significantly improving the stability and robustness of robot swarm scheduling, and ensuring the efficient and orderly completion of scheduling tasks in complex scenarios.

[0033] In one embodiment of the present invention, S1 includes:

[0034] S11. Integrating distributed communication protocols and hierarchical decision-making mechanisms, a two-layer intelligent agent architecture consisting of a global coordination layer and a local execution layer is constructed, generating a hierarchical scheduling network topology. The global coordination layer is used for task allocation and target optimization; the local execution layer is used for action execution and status feedback. The hierarchical scheduling network topology includes node connection relationships and data interaction interfaces.

[0035] S12. Based on the node function division of the hierarchical scheduling network topology, the kinematic model, energy consumption model, and task execution capability model of the hundred-unit robot are parametrically modeled to generate a state dataset of the hundred-unit robot. The state dataset of the hundred-unit robot covers position coordinates, remaining power, load status, and task adaptability.

[0036] S13. Based on the capability boundary parameters of the state dataset of hundreds of robots, initialize a multi-type task pool. The multi-type task pool includes different priorities, time window constraints and resource requirements. Simultaneously, collect environmental obstacle information and passable path network through LiDAR and visual SLAM technology to generate initial scheduling scene data that integrates task information and environmental information.

[0037] The working principle and effects of the above technical solution are as follows: By integrating distributed communication protocols and hierarchical decision-making mechanisms to construct a two-layer intelligent agent architecture, the efficiency of global and local collaboration can be improved, avoiding chaotic or delayed transmission of scheduling instructions. Parametric modeling of hundreds of robots generates state datasets, improving the accuracy of understanding core states such as robot position and battery level, reducing mismatches between subsequent task allocation and robot capabilities. Integrating multiple task pools and environmental information to generate initial scheduling scenario data makes the initial scheduling basis more comprehensive, avoiding deviations in the initial scheduling plan due to missing task or environmental information, while enhancing the reliability of subsequent optimization modeling, laying a solid foundation for efficient scheduling.

[0038] In one embodiment of the present invention, step S12 includes:

[0039] Based on the functional division of nodes in the hierarchical scheduling network topology, the robot performance adaptation requirements corresponding to each node are extracted to generate a set of robot performance requirement indicators.

[0040] Based on the robot performance requirement index set, the core parameters of the kinematic model, energy consumption model and task execution capability model of hundreds of robots are defined, and a model parameter configuration table is generated.

[0041] Based on the model parameter configuration table, the motion characteristics, energy consumption level, and task processing capabilities of each robot are quantitatively modeled one by one to generate single robot state data.

[0042] All individual robot status data are aggregated and integrated to form a status dataset of hundreds of robots, covering location coordinates, remaining power, load status, and task adaptability.

[0043] The working principle and effects of the above technical solution are as follows: Based on node functions, robot performance adaptation requirements are extracted to generate an indicator set, making subsequent parameter definition more targeted and avoiding parameter deviations caused by blind modeling. The indicator set defines core parameters to form a configuration table, improving the uniformity and standardization of model parameters and reducing the problem of inconsistent modeling standards for different robots. Individual robot state data is generated through quantified modeling, accurately capturing the individual differences in motion characteristics, energy consumption levels, etc., of each robot, avoiding overall data distortion caused by the masking of individual characteristics. The resulting dataset of state data for hundreds of robots comprehensively and completely reflects the overall state of the cluster, reducing the mismatch between subsequent task allocation and robot capabilities, and providing reliable data support for the initialization of subsequent scheduling scenarios.

[0044] In one embodiment of the present invention, S2 includes:

[0045] S21. Input the task constraints, robot performance parameters, and environmental constraint information from the initial scheduling scenario data into the improved NSGA-II algorithm module. The task constraints include the deadline and execution accuracy. The robot performance parameters include the maximum speed and energy consumption coefficient. The environmental constraint information includes the path width and slope resistance. Construct a multi-objective optimization model that minimizes the task completion time and energy consumption accumulation through an adaptive weight encoding mechanism to generate a Pareto front solution set that satisfies Pareto optimality.

[0046] S22. Extract the spatiotemporal coordinate sequence of each candidate path in the Pareto front solution set, construct a four-dimensional collision detection model through the spatiotemporal A* algorithm, the four dimensions including three-dimensional space + one-dimensional time, divide the robot's motion trajectory into millisecond-level time slices and calculate the spatial overlap, and perform real-time planning of dynamic obstacle avoidance path.

[0047] S23. Combine the obstacle avoidance results to perform secondary energy consumption calibration on the candidate paths. The secondary calibration includes energy consumption correction for uphill sections and energy consumption compensation for idling, generating a set of jointly optimized path schemes that take into account path optimality, obstacle avoidance safety and energy economy.

[0048] The working principle and effects of the above technical solution are as follows: Multiple constraint information is input into the improved algorithm to construct a multi-objective optimization model, generating a Pareto front solution set. This accurately balances task completion time and energy consumption accumulation, avoiding the unintended consequences of single-objective optimization. Through a four-dimensional collision detection model and millisecond-level time slice partitioning, the timeliness and accuracy of dynamic obstacle avoidance are significantly improved, preventing spatiotemporal collisions during robot movement and reducing downtime losses. Secondary energy consumption calibration is performed based on the obstacle avoidance results, further reducing ineffective energy consumption while ensuring the optimality and safety of the path scheme. The final jointly optimized path scheme set improves the overall efficiency of path planning and enhances the practicality and reliability of the scheme, providing a high-quality path basis for subsequent task allocation.

[0049] In one embodiment of the present invention, step S22 includes:

[0050] Extract the spatiotemporal coordinate sequences of each candidate path from the Pareto front solution set to generate a set of candidate path spatiotemporal coordinate sequences;

[0051] Based on the set of candidate path spatiotemporal coordinate sequences, a four-dimensional collision detection model is constructed using the spatiotemporal A* algorithm, and the four-dimensional collision detection model structure is generated.

[0052] Based on the four-dimensional collision detection model structure, the robot's motion trajectory is divided into millisecond-level time slices to generate trajectory time slice division results.

[0053] Based on the trajectory time slice division results, the spatial overlap of the robot's motion trajectory within each time slice is calculated, and spatial overlap data is generated.

[0054] Based on spatial overlap data, potential collision risks are identified, dynamic obstacle avoidance path planning is performed in real time, and dynamic obstacle avoidance path planning results are generated.

[0055] The working principle and effects of the above technical solution are as follows: A set of spatiotemporal coordinate sequences of candidate paths is extracted to provide a precise data foundation for subsequent collision detection, avoiding detection bias caused by fragmented coordinate information. A four-dimensional collision detection model is constructed using the spatiotemporal A* algorithm, taking into account both spatial location and temporal dimensions, enhancing the comprehensiveness of collision risk identification and reducing potential dangers missed by single spatial dimensions. Millisecond-level time slice division allows for more refined trajectory analysis, improving the timeliness of collision prediction and preventing missed optimal avoidance opportunities due to insufficient time precision. Calculating the spatial overlap of each time slice accurately locates risk points, making obstacle avoidance planning more targeted and reducing ineffective obstacle avoidance operations. Finally, through real-time obstacle avoidance path planning, spatiotemporal collisions during robot movement are effectively avoided, reducing downtime losses and ensuring the safety and orderliness of the movement of a swarm of hundreds of robots.

[0056] In one embodiment of the present invention, S3 includes:

[0057] S31. Extract key features from the set of joint optimization path schemes. The key features include path length, energy consumption cost, and task matching timeliness. Construct a task and robot dual-objective matching degree evaluation matrix. Determine the weight of each feature using the entropy weight method. Generate dual-objective weighted fitness function value.

[0058] S32. Based on the bi-objective weighted fitness function value, perform non-dominated sorting and crowding screening on all task allocation combinations, eliminate invalid schemes that exceed the robot's load capacity and violate task time constraints, and select the globally optimal task allocation scheme.

[0059] S33. For each robot in the optimal task allocation scheme, generate a robot individual path planning instruction set. The robot individual path planning instruction set includes path node coordinates, motion speed instructions, and energy consumption control thresholds. The instruction format is adapted and the integrity is verified through the communication interface between the global coordination layer and the local execution layer.

[0060] S34. The verified robot individual path planning instruction set is sent to the local execution layer agent. The instruction is parsed and prepared for execution in combination with the robot hardware driver protocol to form the initial scheduling instruction data that can directly drive the robot's actions.

[0061] The working principle and effects of the above technical solution are as follows: Multi-dimensional key features are extracted to construct a dual-objective matching degree evaluation matrix. Entropy weighting is used to determine weights and generate fitness function values, making the task-robot matching evaluation more accurate and avoiding evaluation bias caused by subjective weight settings. Invalid solutions are eliminated through non-dominated sorting and crowding screening, improving the reliability of the globally optimal task allocation scheme and reducing situations where tasks exceed robot load or violate time constraints. Individual path planning instruction sets are generated and their format adaptation and integrity verification are performed to ensure the compatibility and accuracy of instruction transmission, avoiding execution failures caused by inconsistent instruction formats or missing content. The verified instructions are then sent, parsed, and prepared for execution, allowing initial scheduling instructions to directly drive robot actions, improving scheduling startup efficiency, avoiding time delays caused by secondary instruction processing, and laying a solid foundation for subsequent efficient scheduling execution.

[0062] In one embodiment of the present invention, S33 includes:

[0063] For the optimal task allocation scheme, extract the task allocation information and associated path parameters corresponding to each robot to generate single robot task path association data;

[0064] Based on the single robot task path association data, a draft of the robot individual path planning instruction is generated. The draft of the robot individual path planning instruction includes path node coordinates, motion speed instructions and energy consumption control thresholds.

[0065] The communication interface specifications between the global coordination layer and the local execution layer are retrieved, and the initial draft of the robot's individual path planning instructions is adapted and adjusted to generate a set of instructions with adapted format.

[0066] The completeness of the instruction set after format adaptation is checked to verify the completeness and accuracy of the core content, which includes path node coordinates, motion speed commands, and energy consumption control thresholds, and a robot individual path planning instruction set that passes the verification is generated.

[0067] The working principle and effects of the above technical solution are as follows: Task allocation information and path parameters corresponding to a single robot are extracted to generate associated data, making subsequent instruction generation more tailored to individual needs and avoiding the problem of instructions being disconnected from the robot's actual tasks and paths. Based on the associated data, a draft instruction containing core parameters is generated, ensuring the integrity of the core instruction information and reducing repetitive work in supplementing and perfecting subsequent instructions. The instruction format is adjusted according to the communication interface specification to improve the transmission compatibility of instructions between the global and local execution layers, avoiding instruction recognition failures or transmission failures due to format inconsistencies. A completeness check is performed on the adapted instruction set to accurately control the accuracy of core parameters, preventing robot execution deviations or even malfunctions caused by missing or incorrect parameters. The final generated verified instruction set not only improves the reliability of the instructions but also provides a guarantee for the smooth issuance and execution of subsequent instructions.

[0068] In one embodiment of the present invention, step S4 includes:

[0069] S41. Through the multi-sensor fusion module carried by the local execution layer intelligent agent, the multi-sensor includes an odometer, gyroscope, energy consumption sensor and vision sensor, to collect data such as the robot's real-time position, movement speed, remaining power, and task execution progress in real time. After Kalman filtering noise reduction and data format standardization processing, a dynamic scheduling feedback data stream is generated.

[0070] S42. Based on the dynamic scheduling feedback data stream, a disturbance event feature library is constructed. The disturbance event feature library is used for task insertion features, robot fault features, path blocking features, etc. Real-time data matching and anomaly detection are performed through a deep learning model to identify the disturbance event type and generate a disturbance type identifier.

[0071] S43. For the identified disturbance events, combine the task dependencies and robot cluster topology in the initial scheduling scenario data to conduct a quantitative analysis of the impact range. The quantitative analysis of the impact range includes the number of tasks involved in the faulty robot and the passage area affected by path blockage, and generates a disturbance impact range assessment report.

[0072] The working principle and effects of the above technical solution are as follows: By fusing multi-sensor data to collect multi-dimensional operational data of the robot, and combining this with Kalman filtering for noise reduction and format standardization, the accuracy and consistency of the feedback data are significantly improved, reducing data distortion caused by noise interference and avoiding deviations in subsequent anomaly detection. Constructing a disturbance event feature library and combining it with a deep learning model for matching and detection enables rapid and accurate identification of disturbance types such as task insertion and robot malfunctions, improving the timeliness of disturbance event detection and preventing the escalation of disturbances that could lead to more serious scheduling chaos. Combining task dependencies and cluster topology to quantitatively analyze the impact range allows for a more comprehensive and accurate assessment of disturbance impact, reducing resource waste caused by blind responses and providing a reliable basis for triggering targeted rescheduling mechanisms, ensuring the stable operation of the scheduling system in the face of emergencies.

[0073] In one embodiment of the present invention, step S5 includes:

[0074] S51. Based on the disturbance type identifier and impact range assessment report, set a rescheduling trigger threshold. The rescheduling trigger threshold is triggered when a single robot failure affects more than 3 tasks, and the online rescheduling mechanism is started to synchronously update the task pool status (add / cancel tasks), robot status (fault / recovery), and environment map (blocked / unblocked).

[0075] S52. Input the updated scheduling scenario data into the improved NSGA-II algorithm module, re-model the multi-objective optimization, combine the spatiotemporal A* algorithm to perform collision detection and obstacle avoidance optimization on the corrected path, collaboratively solve the updated joint optimization problem, and generate a set of rescheduling path correction schemes.

[0076] S53. Using a timing alignment and conflict resolution algorithm, the set of rescheduling path correction schemes is compared and analyzed with the initial scheduling instruction data. Valid instructions that are not affected by disturbances are retained, and the content of affected instructions is replaced to generate intermediate instruction fusion data.

[0077] S54. Perform millisecond-level compression and transmission delay compensation on the intermediate data of instruction fusion, adapt to the robot's real-time control protocol, and generate dynamic scheduling update instructions that meet the millisecond-level response requirements.

[0078] The working principle and effects of the above technical solution are as follows: Setting a rescheduling trigger threshold based on disturbance information accurately determines the necessity of rescheduling, avoiding resource waste and scheduling fluctuations caused by blindly initiating rescheduling. Synchronously updating the task pool, robot status, and environmental map ensures the accuracy and freshness of the data source for subsequent optimization modeling, avoiding the generation of invalid correction schemes using outdated data. The improved algorithm, in collaboration with the spatiotemporal A* algorithm, generates rescheduling path correction schemes that can adapt to the updated scheduling scenario, reducing task delays and collision risks caused by disturbances. The timing alignment and conflict resolution algorithm retains valid instructions and replaces affected instructions, improving instruction fusion efficiency and avoiding time losses caused by full instruction reconstruction. Millisecond-level compression and latency compensation processing enable scheduling update instructions to quickly adapt to the control protocol, achieving millisecond-level response, avoiding robot execution chaos caused by response lag, and ensuring the stable and efficient operation of the scheduling system under disturbance scenarios.

[0079] In one embodiment of the present invention, S53 includes:

[0080] A time-series alignment and conflict resolution algorithm is used to align the rescheduling path correction scheme set and the initial scheduling instruction data in the time dimension, generating a time-series aligned dual dataset.

[0081] Based on the time-aligned dual datasets, an item-by-item comparative analysis of the instruction content is conducted to identify the instruction items affected by disturbances and the valid instruction items that are not affected by disturbances, and to generate instruction impact identification results.

[0082] Based on the results of instruction impact identification, valid instruction entries that are not affected by disturbances are extracted and summarized to form a subset of valid instructions;

[0083] For the affected instruction entries marked in the instruction impact identification results, the corresponding corrected instruction content in the rescheduling path correction scheme set is called to replace them, and a replacement instruction subset is generated.

[0084] The effective instruction subset and the replaced instruction subset are merged to form instruction fusion intermediate data.

[0085] The working principle and effects of the above technical solution are as follows: Timing alignment ensures that the rescheduling correction scheme set and the initial instruction data are consistent in the time dimension, avoiding instruction identification errors caused by time misalignment in subsequent comparative analysis. Item-by-item comparative analysis accurately distinguishes between instruction items affected by disturbances and those unaffected, improving the accuracy of instruction impact identification and reducing the possibility of misjudging valid instructions as affected or omitting affected instructions. Extracting a subset of valid instructions retains the undisturbed valid content, avoiding resource waste and efficiency loss caused by full instruction reconstruction. Targeted replacement of affected instruction items ensures that the corrected instructions can adapt to the scheduling scenario after disturbance, reducing execution conflicts caused by the reuse of old instructions. Merging the two subsets to generate intermediate instruction fusion data ensures both the integrity and continuity of instructions, improves the efficiency of instruction fusion, and avoids gaps in subsequent instruction generation due to incomplete fusion, providing a reliable foundation for millisecond-level response scheduling update instructions.

[0086] In one embodiment of the present invention, step S6 includes:

[0087] S61. Based on dynamic scheduling update instructions, the robot actuator is driven by a local execution layer intelligent agent to perform motion adjustment and task execution, and data is collected synchronously during the scheduling process. The data during the scheduling process includes throughput (number of tasks completed per unit time), energy consumption index (average energy consumption per unit task), task delay rate, and equipment failure rate, generating multi-dimensional scheduling performance evaluation data.

[0088] S62. Normalize and perform weight sensitivity analysis on the scheduling performance evaluation data. Adjust the task completion time weight and energy consumption weight parameters of the global coordination layer through the analytic hierarchy process. For example, increase the time weight when the throughput is insufficient and increase the energy consumption weight when the energy consumption exceeds the standard.

[0089] S63. Feed the adjusted weight parameters back to the multi-objective optimization modeling stage, update the objective function weight configuration of the improved NSGA-II algorithm, and form a closed-loop adaptive optimization cycle of scheduling execution, performance evaluation, parameter adjustment and optimization modeling.

[0090] S64. Through multiple rounds of closed-loop iterative optimization, the multi-objective optimization performance of the scheduling scheme is continuously improved, and finally optimized scheduling result data that takes into account high throughput, low operating energy consumption and high robustness is generated, realizing efficient and low-consumption dynamic scheduling of a cluster of hundreds of robots.

[0091] The working principle and effects of the above technical solution are as follows: Based on dynamic scheduling update instructions, the robot performs motion adjustments and task processing, simultaneously collecting multi-dimensional scheduling performance data to make the scheduling effect evaluation more comprehensive and objective, avoiding optimization bias caused by single-indicator judgments. The evaluation data is normalized and subjected to weight sensitivity analysis, adjusting task completion time and energy consumption weights as needed to make the weight configuration more suitable for actual scheduling scenarios, avoiding the problem of fixed weights being unable to adapt to different operating states. The adjusted weights are fed back to the optimization modeling stage to form a closed loop, allowing the scheduling scheme to continuously iterate and improve, reducing the performance limitations caused by one-time modeling. After multiple rounds of closed-loop optimization, the multi-objective adaptability of the scheduling scheme is significantly improved, ensuring high throughput, stable control of operating energy consumption, enhanced scheduling robustness, reduced task delay rate and equipment failure rate, and avoiding large fluctuations in scheduling performance under complex scenarios, ultimately achieving efficient and low-consumption dynamic scheduling of hundreds of robot clusters.

[0092] One embodiment of the present invention, such as Figure 2 As shown, a system for implementing the multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization as described above is provided, the system comprising:

[0093] Dynamic modeling module: Constructs a two-layer intelligent agent architecture for global coordination and local execution, and generates a hierarchical scheduling network topology; performs dynamic modeling of the robot cluster based on the hierarchical scheduling network topology, and generates a state dataset of hundreds of robots; initializes the task pool and real-time environment map based on the state dataset of hundreds of robots to form initial scheduling scenario data;

[0094] Dynamic obstacle avoidance module: Input the initial scheduling scenario data into the improved NSGA-II algorithm module, perform multi-objective optimization modeling, and generate a Pareto front solution set, which includes task completion time and energy consumption accumulation; perform spatiotemporal collision detection and dynamic obstacle avoidance processing on the candidate paths in the Pareto front solution set through the spatiotemporal A* algorithm to generate a joint optimized path scheme set.

[0095] Path planning module: Based on the joint optimization path scheme set, the task and robot matching degree is evaluated, and a dual-objective weighted fitness function value is generated; based on the dual-objective weighted fitness function value, the optimal task allocation scheme is selected, and a robot individual path planning instruction set is generated simultaneously; the robot individual path planning instruction set is sent to the local execution layer agent to form initial scheduling instruction data;

[0096] Dynamic scheduling module: Collects robot motion state data and task execution progress data in real time through local execution layer intelligent agents, and generates dynamic scheduling feedback data stream; detects real-time disturbance events based on the dynamic scheduling feedback data stream, including task insertion, robot failure, path blocking, etc., and generates disturbance type identifier and impact range assessment report;

[0097] Joint optimization module: Based on the disturbance type identifier and impact range assessment report, the online rescheduling mechanism is triggered. The improved NSGA-II algorithm module and the spatiotemporal A* algorithm are used to solve the updated joint optimization problem in collaboration, generating a set of rescheduling path correction schemes. The set of rescheduling path correction schemes is then fused with the initial scheduling instruction data to generate a dynamic scheduling update instruction with a millisecond-level response.

[0098] Performance evaluation module: Based on dynamic scheduling update instructions, executes real-time scheduling operations for a system of hundreds of robots, generates scheduling performance evaluation data, which includes throughput and energy consumption indicators; adjusts the multi-objective optimization weight parameters of the global coordination layer according to the scheduling performance evaluation data, forming a closed-loop adaptive optimization loop, and finally generates optimized scheduling result data.

[0099] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization, characterized in that, The method includes: S1. Construct a two-layer intelligent agent architecture for global coordination and local execution, generate a hierarchical scheduling network topology, perform dynamic modeling of robot clusters, generate a state dataset of hundreds of robots, initialize the task pool and real-time environment map, and form initial scheduling scenario data. S2. Input the initial scheduling scenario data into the improved NSGA-II algorithm module, perform multi-objective optimization modeling, generate Pareto front solution set, and use the spatiotemporal A* algorithm to perform spatiotemporal collision detection and dynamic obstacle avoidance on the candidate paths in the Pareto front solution set to generate a joint optimization path scheme set. S3. Based on the joint optimization path scheme set, evaluate the matching degree between tasks and robots, generate a dual-objective weighted fitness function value, screen the optimal task allocation scheme, and simultaneously generate a robot individual path planning instruction set, which is then sent to the local execution layer agent to form initial scheduling instruction data. S4. Real-time collection of robot motion state data and task execution progress data by local execution layer intelligent agents, generation of dynamic scheduling feedback data stream, detection of real-time disturbance events, generation of disturbance type identification and impact range assessment report; S5. Trigger the online rescheduling mechanism based on the disturbance type identifier and impact range assessment report. Solve the updated joint optimization problem through the improved NSGA-II algorithm module and the spatiotemporal A* algorithm, generate a set of rescheduling path correction schemes, and fuse them with the initial scheduling instruction data to generate dynamic scheduling update instructions. S6. Based on the dynamic scheduling update instructions, execute the real-time scheduling operation of the hundred-unit robot system, generate scheduling performance evaluation data, adjust the multi-objective optimization weight parameters of the global coordination layer, form a closed-loop adaptive optimization loop, and finally generate optimized scheduling result data.

2. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 1, characterized in that, S1 includes: S11. Integrate distributed communication protocols and hierarchical decision-making mechanisms to construct a two-layer intelligent agent architecture with a global coordination layer and a local execution layer, generating a hierarchical scheduling network topology; S12. Based on the node function division of the hierarchical scheduling network topology, perform parametric modeling of the kinematic model, energy consumption model, and task execution capability model of the hundred-unit robot, and generate a state dataset of the hundred-unit robot. S13. Based on the capability boundary parameters of the state dataset of hundreds of robots, initialize multi-type task pools, and simultaneously collect environmental obstacle information and passable path network through LiDAR and visual SLAM technology to generate initial scheduling scenario data that integrates task information and environmental information.

3. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 1, characterized in that, The S2 includes: S21. Input the task constraints, robot performance parameters and environmental constraints from the initial scheduling scenario data into the improved NSGA-II algorithm module. Construct a multi-objective optimization model that minimizes task completion time and energy consumption accumulation through an adaptive weight encoding mechanism, and generate a Pareto front solution set that satisfies Pareto optimality. S22. Extract the spatiotemporal coordinate sequence of each candidate path in the Pareto front solution set, construct a four-dimensional collision detection model through the spatiotemporal A* algorithm, perform millisecond-level time slice division and spatial overlap calculation on the robot's motion trajectory, and execute real-time planning of dynamic obstacle avoidance path. S23. Combine the obstacle avoidance results to perform secondary energy consumption calibration on the candidate paths and generate a joint optimized path scheme set.

4. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 3, characterized in that, S22 includes: Extract the spatiotemporal coordinate sequences of each candidate path from the Pareto front solution set to generate a set of candidate path spatiotemporal coordinate sequences; Based on the set of candidate path spatiotemporal coordinate sequences, a four-dimensional collision detection model is constructed using the spatiotemporal A* algorithm, and the four-dimensional collision detection model structure is generated. Based on the four-dimensional collision detection model structure, the robot's motion trajectory is divided into millisecond-level time slices to generate trajectory time slice division results. Based on the trajectory time slice division results, the spatial overlap of the robot's motion trajectory within each time slice is calculated, and spatial overlap data is generated. Based on spatial overlap data, potential collision risks are identified, dynamic obstacle avoidance path planning is performed in real time, and dynamic obstacle avoidance path planning results are generated.

5. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 1, characterized in that, The S3 includes: S31. Extract key features from the set of joint optimization path schemes, construct a task and robot dual-objective matching degree evaluation matrix, determine the weight of each feature through the entropy weight method, and generate dual-objective weighted fitness function values. S32. Based on the bi-objective weighted fitness function value, perform non-dominated sorting and crowding screening on all task allocation combinations, eliminate invalid schemes that exceed the robot's load capacity and violate task time constraints, and select the globally optimal task allocation scheme. S33. For each robot in the optimal task allocation scheme, generate a path planning instruction set for the robot and perform instruction format adaptation and integrity verification through the communication interface between the global coordination layer and the local execution layer. S34. The verified robot individual path planning instruction set is sent to the local execution layer intelligent agent, and instruction parsing and execution preparation are carried out in combination with the robot hardware driver protocol to form the initial scheduling instruction data.

6. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 1, characterized in that, The S4 includes: S41. Through the multi-sensor fusion module carried by the local execution layer intelligent agent, robot data is collected in real time, and after Kalman filtering noise reduction and data format standardization processing, dynamic scheduling feedback data stream is generated. S42. Based on the dynamic scheduling feedback data stream, construct a disturbance event feature library, perform real-time data matching and anomaly detection through a deep learning model, identify disturbance event types, and generate disturbance type identifiers; S43. For the identified disturbance events, combine the task dependencies and robot cluster topology in the initial scheduling scenario data to conduct a quantitative analysis of the impact range and generate a disturbance impact range assessment report.

7. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 1, characterized in that, The S5 includes: S51. Based on the disturbance type identifier and impact range assessment report, set the rescheduling trigger threshold, start the online rescheduling mechanism, and synchronously update the task pool status, robot status, and environmental map. S52. Input the updated scheduling scenario data into the improved NSGA-II algorithm module, re-model the multi-objective optimization, and combine the spatiotemporal A* algorithm to perform collision detection and obstacle avoidance optimization on the corrected path to generate a set of rescheduling path correction schemes. S53. Using a timing alignment and conflict resolution algorithm, the set of rescheduling path correction schemes and the initial scheduling instruction data are compared and analyzed to generate instruction fusion intermediate data. S54. Perform millisecond-level compression and transmission delay compensation on the intermediate data of instruction fusion, adapt to the robot's real-time control protocol, and generate dynamic scheduling update instructions that meet the millisecond-level response requirements.

8. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 7, characterized in that, S53 includes: A time-series alignment and conflict resolution algorithm is used to align the rescheduling path correction scheme set and the initial scheduling instruction data in the time dimension, generating a time-series aligned dual dataset. Based on the time-aligned dual datasets, an item-by-item comparative analysis of the instruction content is conducted to identify the instruction items affected by disturbances and the valid instruction items that are not affected by disturbances, and to generate instruction impact identification results. Based on the results of instruction impact identification, valid instruction entries that are not affected by disturbances are extracted and summarized to form a subset of valid instructions; For the affected instruction entries marked in the instruction impact identification results, the corresponding corrected instruction content in the rescheduling path correction scheme set is called to replace them, and a replacement instruction subset is generated. The effective instruction subset and the replaced instruction subset are merged to form instruction fusion intermediate data.

9. The multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization according to claim 1, characterized in that, The S6 includes: S61. Based on dynamic scheduling update instructions, the robot actuator is driven by a local execution layer intelligent agent to perform motion adjustment and task execution, and data is collected synchronously during the scheduling process to generate multi-dimensional scheduling performance evaluation data. S62. Normalize the scheduling performance evaluation data and perform weight sensitivity analysis. Adjust the task completion time weight and energy consumption weight parameters of the global coordination layer through the analytic hierarchy process. S63. Feed the adjusted weight parameters back to the multi-objective optimization modeling stage, update the objective function weight configuration of the improved NSGA-II algorithm, and form a closed-loop adaptive optimization loop; S64. Through multiple rounds of closed-loop iterative optimization, the multi-objective optimization performance of the scheduling scheme is continuously improved, and finally optimized scheduling result data is generated.

10. A system for implementing the multi-agent dynamic scheduling method based on efficiency and energy multi-objective optimization as described in claim 1, characterized in that, The system includes: Dynamic Modeling Module: Constructs a two-layer intelligent agent architecture for global coordination and local execution, generates a hierarchical scheduling network topology, performs dynamic modeling of robot clusters, generates a state dataset of hundreds of robots, initializes the task pool and real-time environment map, and forms initial scheduling scenario data; Dynamic obstacle avoidance module: Input the initial scheduling scenario data into the improved NSGA-II algorithm module, perform multi-objective optimization modeling, generate Pareto front solution set, and use the spatiotemporal A* algorithm to perform spatiotemporal collision detection and dynamic obstacle avoidance processing on the candidate paths in the Pareto front solution set to generate a joint optimized path scheme set; Path planning module: Based on the joint optimization path scheme set, the task and robot matching degree is evaluated, a dual-objective weighted fitness function value is generated, the optimal task allocation scheme is selected, and the individual robot path planning instruction set is generated simultaneously; the individual robot path planning instruction set is sent to the local execution layer agent to form the initial scheduling instruction data; Dynamic scheduling module: Collects robot motion state data and task execution progress data in real time through local execution layer intelligent agents, generates dynamic scheduling feedback data stream, detects real-time disturbance events, and generates disturbance type identification and impact range assessment report; Joint optimization module: Based on the disturbance type identifier and impact range assessment report, the online rescheduling mechanism is triggered. The improved NSGA-II algorithm module and the spatiotemporal A* algorithm are used to solve the updated joint optimization problem, generate a set of rescheduling path correction schemes, and fuse them with the initial scheduling instruction data to generate dynamic scheduling update instructions. Performance evaluation module: Based on dynamic scheduling update instructions, executes real-time scheduling operations for a system of hundreds of robots, generates scheduling performance evaluation data, adjusts the multi-objective optimization weight parameters of the global coordination layer, forms a closed-loop adaptive optimization loop, and finally generates optimized scheduling result data.

Images