Intelligent scheduling and resource allocation method and system for flexible AGV

Abstract

The invention relates to the technical field of AGV intelligent scheduling, and discloses an intelligent scheduling and resource allocation method and system for a flexible AGV, and the method comprises the steps: carrying out the grid processing of an AGV operation region, and obtaining a dynamic scheduling network; modeling an AGV state space in the dynamic scheduling network to obtain a multi-agent scheduling model; calculating a task priority score of the AGV based on a multi-agent scheduling model, discretizing an operation time axis into time slices, and obtaining a time slice resource allocation result; according to the time slice resource allocation result, the AGV scheduling decision is optimized to obtain a distributed scheduling strategy; dividing the distributed scheduling strategy to obtain a master control AGV and a slave AGV, and generating a cluster collaborative decision result; and resource allocation and dynamic load balancing are carried out based on a cluster collaborative decision result to obtain a scheduling instruction of the AGV cluster, collaborative decision of the AGV cluster is realized, and consistency of global task decomposition and local resource allocation is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AGV intelligent scheduling, and particularly relates to an intelligent scheduling and resource allocation method and system for flexible AGV. BACKGROUND

[0002] AGV (Automatic Guided Vehicle) is widely used in modern factories and warehouse centers as the core equipment of intelligent logistics system. Traditional AGV scheduling system adopts centralized control architecture, which has problems such as slow system response speed, high computing load and poor scalability when facing large-scale and dynamically changing working environment.

[0003] Current AGV intelligent scheduling research mainly focuses on single AGV path planning and static task allocation, and lacks in-depth research on multi-AGV collaborative scheduling and dynamic resource allocation. Especially in complex working conditions, the task coordination and resource competition among multiple AGVs are increasingly prominent, and the traditional scheduling method is difficult to effectively balance system efficiency and resource utilization. In addition, the existing AGV scheduling system has insufficient processing capacity for sudden events and poor system robustness. When AGV failure, task change or environmental interference occurs, there is a lack of effective adaptive adjustment mechanism, which easily causes local or global performance degradation, affecting the stable operation of the entire logistics system. SUMMARY

[0004] The present application provides an intelligent scheduling and resource allocation method and system for flexible AGV, which realizes collaborative decision-making of AGV cluster and ensures the consistency of global task decomposition and local resource allocation.

[0005] In a first aspect, the present application provides an intelligent scheduling and resource allocation method for flexible AGV, which comprises: performing grid processing on the AGV working area to obtain a dynamic scheduling network; modeling the AGV state space in the dynamic scheduling network to obtain a multi-agent scheduling model; calculating the task priority score of AGV based on the multi-agent scheduling model, and discretizing the working time axis into time slices to obtain time slice resource allocation results; optimizing AGV scheduling decisions according to the time slice resource allocation results to obtain a distributed scheduling strategy; dividing the distributed scheduling strategy to obtain master AGV and slave AGV, and generating cluster collaborative decision-making results; performing resource allocation and dynamic load balancing based on the cluster collaborative decision-making results to obtain scheduling instructions for AGV cluster.

[0006] In a second aspect, the present application provides a flexible AGV intelligent scheduling and resource allocation system, comprising: a meshing module for meshing the AGV operation area to obtain a dynamic scheduling network; a modeling module for modeling the AGV state space in the dynamic scheduling network to obtain a multi-agent scheduling model; a calculation module for calculating the task priority score of the AGV based on the multi-agent scheduling model, and discretizing the operation time axis into time slices to obtain a time slice resource allocation result; an optimization module for optimizing the AGV scheduling decision according to the time slice resource allocation result to obtain a distributed scheduling strategy; a division module for dividing the distributed scheduling strategy to obtain master AGVs and slave AGVs, and generating a cluster collaborative decision result; an allocation module for resource allocation and dynamic load balancing based on the cluster collaborative decision result to obtain the scheduling instruction of the AGV cluster.

[0007] In the technical scheme provided by the present application, the continuous operation space is discretized by establishing a meshed dynamic scheduling network, the AGV motion state and resource distribution are accurately modeled, the system has real-time perception and dynamic updating capability. A multi-agent scheduling model is used for distributed decision making, each AGV is regarded as an independent agent for autonomous decision making, the calculation load of the central controller is reduced, and the scalability of the system is enhanced. The priority time window mechanism is introduced, resources are dynamically allocated according to the task urgency, remaining power and historical completion rate, and the resource competition problem of multiple AGVs in the time and space dimensions is solved. A self-adaptive scheduling optimization framework is constructed by combining deep neural network and asynchronous advantage actor critic algorithm, and the learning ability of the system to complex environment is improved. Through the master-slave hierarchical structure design, the collaborative decision of the AGV cluster is realized, and the consistency of global task decomposition and local resource allocation is guaranteed. A dynamic load balancing mechanism based on multi-objective optimization is established, and the system performance is balanced in multiple dimensions such as resource utilization, task completion rate and energy efficiency. A distributed consistency protocol is used to handle emergencies, and the robustness of the system in the face of uncertainty is enhanced. A unified scheduling instruction format is designed, the control information is standardized and standardized, and the system maintenance and expansion are facilitated. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0009] Figure 1 The step schematic diagram of the intelligent scheduling and resource allocation method of the flexible AGV in the embodiments of the present application. Figure 2 The structure schematic diagram of the intelligent scheduling and resource allocation system of the flexible AGV in the embodiments of the present application. DETAILED DESCRIPTION

[0010] The embodiments of the present application provide an intelligent scheduling and resource allocation method and system of a flexible AGV. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0011] For the convenience of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1 One embodiment of the intelligent scheduling and resource allocation method of the flexible AGV in the embodiments of the present application includes: Step S1, performing grid processing on the AGV operation area to obtain a dynamic scheduling network; It can be understood that the execution subject of the present application can be an intelligent scheduling and resource allocation system of a flexible AGV, and can also be a terminal or a server, and the specific place is not limited. The embodiments of the present application take a server as an execution subject for example.

[0012] Specifically, the total area of the AGV operation area is divided into equal-interval grids for calculation. By dividing the entire operation area into a plurality of regular grid cells, each grid cell represents a specific spatial area, and a local coordinate system is established with the center point of each grid cell as the coordinate origin, ensuring that each grid cell can be located in space and at the center position through the coordinate system after gridding processing. In order to facilitate subsequent identification and operation, each grid cell is assigned a unique identification code to form a grid cell identification matrix, representing the distribution and identification information of all grid cells in the operation area. The key resource nodes in the AGV operation area, including the coordinates of charging piles, workstation coordinates, and storage device coordinates, are input into the above grid cell identification matrix, and the positions of these resource nodes are associated with the grid cells through mapping calculation to generate a resource node distribution map, which represents the specific position distribution of key resources in the entire operation area, ensuring that the spatial information of the resource nodes corresponds to the identification of the grid cells. The adjacent grid cells in the resource node distribution map are detected for connectivity state, to determine whether there is a passable path between adjacent grids, such as no obstacles, no device occupation, etc. under the condition of passability, and a network topology connection matrix is constructed based on the detection results to describe the connectivity between grid cells. In order to optimize the actual scheduling path of the AGV, the network topology connection matrix is input into a multi-objective optimization model. The optimization objectives of the model include path distance minimization, turn number minimization, and travel time minimization, which optimize the scheduling performance of the AGV from the aspects of path length, navigation complexity, and time efficiency. To achieve these objectives, a Pareto optimal solution is used for solving, generating a target network topology structure that considers multiple optimization objectives. The passable state of each grid cell in the target network topology structure is detected in real time to generate a passable state matrix. This matrix dynamically reflects whether each grid cell is passable, for example, whether there are obstacles or congestion. At the same time, real-time monitoring of the resource occupation of key resource nodes (such as charging piles, workstations, and storage devices) in the target network topology structure is performed to generate a resource occupancy rate matrix, which dynamically describes the current availability of each resource node. Based on the grid cell identification matrix, the target network topology structure, the passable state matrix, and the resource occupancy rate matrix, a dynamic scheduling network is generated. This network integrates the spatial information, resource distribution, and real-time state data of the operation area, and dynamically responds to environmental and task changes.

[0013] Step S2, model the AGV state space in the dynamic scheduling network to obtain a multi-agent scheduling model; Specifically, the position coordinate data, the remaining power value, the current task number being executed, and the real-time load weight of each AGV are collected to obtain an original state data set. The state space of the AGV is modeled as a four-dimensional vector using the original state data set, where each dimension corresponds to a different attribute. The position coordinates are mapped to the spatial dimension to describe the specific position of the AGV in the work area; the remaining power is mapped to the energy dimension to represent the endurance of the AGV; the current task number being executed is mapped to the task dimension to reflect the task category and its priority being processed by the AGV; and the real-time load weight is mapped to the load dimension to describe the current transportation capacity and working state of the AGV. In this way, the state space vector of all AGVs is systematically defined, providing multi-dimensional information support for the scheduling decision of each AGV. After constructing the state space vector, a unique agent identifier is assigned to each AGV, so that each AGV is regarded as an independent agent, and a communication link matrix between agents is established through a communication protocol. The communication link matrix describes the information interaction relationship between each agent, forming an initial agent network. In order to depict the interaction relationship between agents, the adjacent agents in the initial agent network are quantitatively calculated. By introducing a task correlation coefficient and a resource competition coefficient, the cooperation degree and the competition degree between each pair of agents are quantified, an interaction weight matrix is constructed, the cooperation mode of the agents in the dynamic scheduling network is effectively reflected, and an agent interaction network is obtained. According to the agent interaction network, the historical scheduling data is segmented and sampled to generate a training sample set, and the dynamic changes of the agent state space are modeled based on these sample data. In order to depict the trend of the state space vector between adjacent time points, Markov property analysis is performed on the state space vector. Through this analysis, the transition probability of the state space vector between different time points is calculated, a state transition probability matrix is constructed, and the possible behavior change path of the agent in the dynamic environment is described. Based on the state transition probability matrix, a reward function is designed to evaluate the pros and cons of the scheduling scheme. In the reward function, the task completion time, energy consumption, and resource utilization are set as key evaluation indexes, so as to comprehensively consider the scheduling efficiency, energy consumption cost, and resource use effect. By optimizing this reward function, a multi-agent scheduling model that can dynamically adapt to complex environments is constructed.

[0014] Step S3, calculating the task priority score of the AGV based on the multi-agent scheduling model, and discretizing the work time axis into time slices to obtain a time slice resource allocation result; Specifically, the deadline of the AGV task is subtracted from the current time, and the task urgency coefficient is set as the normalized reciprocal of the difference value to obtain the task urgency score, reflecting the urgency of the task. The higher the value, the more urgent the task, and the higher the priority in subsequent scheduling. At the same time, based on the energy dimension in the multi-agent scheduling model, the ratio of the remaining power of the AGV to the power required by the task is calculated to obtain the power sufficiency score, so as to evaluate whether each AGV has enough energy to complete the specified task and avoid the situation that the task is interrupted due to insufficient power. In order to quantify the scheduling history performance of the AGV, the historical scheduling data in the multi-agent scheduling model is analyzed. The ratio of the number of tasks completed by the AGV in the past K time windows to the total number of tasks is calculated to obtain the historical task completion rate score. Reflects the task execution efficiency of the AGV in a period of time. Based on the task urgency score, the power sufficiency score, and the historical task completion rate score, the priority score is calculated. Through weighted calculation or other comprehensive scoring methods, the three types of scores are integrated into a task priority comprehensive score. The operation time axis is equally divided according to a fixed length to generate a set of time slices. Each time slice represents a time window of the scheduling system, and the division of the time axis ensures the refinement of the scheduling operation. After generating the set of time slices, a time window resource matrix is established to record the available resource information in each time slice, including the number of charging piles, the number of workstations, and the number of currently schedulable AGVs. These resource data are stored in the time slice resource capacity table in the form of resource capacity vectors. In the resource allocation process, all AGVs are arranged in descending order according to the task priority comprehensive score, and resource allocation is performed from the AGV with the highest priority. Each AGV is allocated to an available time slice in the time slice resource capacity table, and is matched according to its task demand and resource availability. If the resources of the current time slice cannot meet the task demand of a certain AGV, it needs to be allocated to the next time slice until all task resource allocations are completed. Through this process, the time slice resource allocation result is obtained, which records the scheduling arrangement and resource usage of each AGV in different time slices.

[0015] Step S4, according to the time slice resource allocation result, optimizing the AGV scheduling decision to obtain a distributed scheduling strategy; Specifically, the current system state is feature extracted based on the time slice resource allocation result. The core features of each AGV are extracted, including position coordinates, remaining power, task status, and load conditions, and these data are constructed into a state feature vector. Based on the extracted state feature vector, a deep neural network with a dual network architecture is designed, which is divided into a policy network and a value network. The policy network generates a scheduling policy, while the value network is used to evaluate the value of the current state, providing feedback support for policy optimization. In this design, the Actor learning rate and policy gradient threshold are set for the policy network to ensure that it can stably generate high-quality scheduling policies; the Critic learning rate and time difference threshold are set for the value network to accurately calculate the time difference value of the state, ensuring the reliability of its evaluation results. Through this dual network architecture, a deep learning model with policy generation and value evaluation capabilities is constructed. On the basis of the dual network model, an asynchronous advantage function is introduced to optimize the scheduling policy of the AGV. The construction of the advantage function is based on the difference between the actual reward value of each AGV and the estimated value of the value network, which quantifies the degree of superiority of the current policy compared to the baseline policy. In order to improve the comprehensiveness and reliability of the evaluation, the advantage function accumulates the data over time to reflect the dynamic trend of the advantage. Through this process, the advantage evaluation result of each AGV in the current environment is obtained. The advantage evaluation result is input into the actor-critic algorithm, which combines the characteristics of the policy network and the value network to update the policy gradient and value loss of each AGV. In the policy network, the policy gradient is optimized to improve the adaptability and efficiency of the scheduling policy; in the value network, the value loss is reduced to improve the accuracy of state evaluation. Through this training process, the deep neural network is further optimized to better adapt to complex dynamic scheduling environments, and a trained neural network is obtained, which has the ability to generate efficient scheduling policies. In order to apply the trained neural network to the actual scheduling scenario, a dynamic model is established. By mapping the speed and acceleration of the AGV to discrete state variables, a discrete event model of the AGV is constructed. This model can discretize the continuous physical motion characteristics to match the time slice division in the scheduling optimization process. At the same time, this model can effectively describe the dynamic behavior of the AGV. Based on the discrete event model of the AGV, the system is optimized. The mixed optimization combines global scheduling goals and local adjustment needs, and can balance the global optimization and local response ability of distributed scheduling in dynamic environments. Through this optimization process, a distributed scheduling strategy is generated, which fully utilizes the advantages of each AGV while considering the coordination of global resource allocation. The distributed scheduling strategy can dynamically respond to changes in task demand and effectively solve resource competition and path conflict problems, thereby improving the operation efficiency and task completion ability of the entire AGV system.

[0016] Step S5, the distributed scheduling strategy is divided to obtain master AGV and slave AGV, and a cluster cooperative decision result is generated; Specifically, the performance of the AGV cluster in the distributed scheduling strategy is evaluated. By quantitatively analyzing the task processing capacity, remaining power level and communication bandwidth of each AGV, an AGV performance score table is constructed. In the score table, the overall running ability of the AGV is sorted and classified by comprehensively considering these performance indicators. Based on the AGV performance score table, the AGVs are divided into a master AGV set and a slave AGV set according to the preset score range. The AGVs whose performance scores reach or exceed the preset range are classified into the master AGV set, while the other AGVs are classified into the slave AGV set, forming a hierarchical AGV cluster hierarchical model with hierarchical structure. The hierarchical structure ensures that in task scheduling, the master AGV assumes the main responsibility of coordination and decision-making, while the slave AGV mainly executes specific tasks, optimizing the cooperative efficiency of the cluster. The hierarchical AGV cluster structure is mapped to the communication topology network, a fully connected network structure is constructed for the master AGV set to ensure efficient information sharing and coordination among master AGVs. At the same time, a star network structure is constructed for the slave AGV set and its corresponding master AGV, so that each slave AGV can indirectly communicate with other nodes through its master AGV, forming a double-layer communication topology graph. The double-layer network structure reduces the communication complexity while maintaining the efficiency of information transmission, adapting to the needs of the distributed scheduling system. The global task matrix in the double-layer communication topology graph is regionally segmented to generate a task decomposition sequence. The task decomposition process analyzes the spatial and temporal distribution of the global task to decompose the complex task into multiple executable sub-task sequences, and assigns a time window requirement to each sub-task. On this basis, the scheduling period of the AGV cluster is synchronously divided to generate a task scheduling time sequence table. At the same time of generating the task scheduling time sequence table, the resource allocation problem is processed in detail. By assigning weights to the charging pile usage, workstation occupation and warehouse equipment usage, a resource allocation matrix is constructed. This matrix describes the allocation of different types of resources in each time window. In order to ensure the consistency of resource allocation, a consistency constraint condition is imposed on the resource allocation matrix, and a distributed average consistency protocol is constructed using the Laplacian matrix. This protocol iteratively updates the decision variables of AGVs, so that all AGVs gradually reach an agreement on resource use and task scheduling. After multiple iterations, a consistency decision vector is generated. The consistency decision vector is input into the cooperative optimizer, and the decision variables of the AGVs are iteratively optimized based on the actual operation of the cluster. The optimization process combines multiple constraints of resource allocation, task decomposition and scheduling time sequence to adjust the variables and ensure further improvement of the overall performance of the system. The cluster cooperative decision result after optimization is generated.

[0017] Step S6, resource allocation and dynamic load balancing based on the cluster collaborative decision result, to obtain the scheduling instruction of the AGV cluster.

[0018] Specifically, the cluster collaborative decision result is input into the scheduler. In the scheduler, key data in the execution period is sampled in real time, including the position coordinates of each AGV, the task state, and the occupation of the current resource. Through the dynamic collection of these data, the running state of the AGV cluster is reflected in real time. Based on the AGV cluster state data, the charging pile occupation number is divided by the total number to obtain the charging device utilization rate; the number of tasks being processed by the workstation is divided by the total processing capacity of the workstation to calculate the utilization rate of the workstation; at the same time, the access frequency of the warehouse equipment is counted and divided by the operation time to obtain the warehouse utilization rate. Through weighted summation, these resource utilization rate indicators are integrated into a comprehensive resource utilization rate indicator, which is used to evaluate the overall use efficiency and load of the resources in the system. The efficiency of task execution is quantitatively analyzed. Also based on the state data of the AGV cluster, the task completion rate is obtained by calculating the ratio of the number of tasks completed to the total number of tasks, representing the completion degree of the current task; the number of tasks being executed is divided by the number of online AGVs to obtain the task execution rate, which is used to measure the balance of the current task distribution; at the same time, the average response time of the task is divided by the standard response time to calculate the time efficiency, reflecting the timeliness of task processing. These indicators are integrated into a task completion rate indicator through linear combination, describing the performance of the AGV cluster in task processing. The energy efficiency is analyzed, the remaining power of each AGV is divided by its full power to calculate the power level, reflecting the current endurance of the AGV; the driving mileage is divided by the total energy consumption to obtain the driving efficiency, indicating the energy use efficiency; at the same time, the charging frequency is calculated by dividing the number of charging times by the number of completed tasks, describing the charging demand status of the AGV. After normalization processing, these indicators are integrated into an energy efficiency indicator, representing the overall state of the AGV cluster in energy utilization. Based on the resource utilization rate indicator, the task completion rate indicator and the energy efficiency indicator, comprehensive load analysis is carried out to generate load evaluation results. By quantifying the contribution weight of different indicators to the system load, the system load distribution is dynamically adjusted in combination with the current real-time data. On the basis of load evaluation, the scheduler generates a robust control strategy to ensure that the system can still maintain stable performance when facing load fluctuations and uncertain factors. The robust control strategy dynamically adjusts the task allocation, path selection and resource use plan of the AGV according to the load evaluation results, so as to achieve the goal of load balancing. Based on the robust control strategy, a set of scheduling instructions is generated. The instruction set integrates the motion parameters (such as speed, direction, acceleration) of the AGV, the task execution sequence (including task priority and task allocation), and the resource use permission (such as the allocation of charging piles and workstations) into a unified control message. The generated control message is directly issued to the AGV cluster as a scheduling instruction to guide the actual operation of the AGV.

[0019] In the embodiment of the present application, by establishing a grid dynamic scheduling network, the continuous job space is discretely processed, the accurate modeling of AGV motion state and resource distribution is realized, the system has real-time perception and dynamic updating capability. A multi-agent scheduling model is used for distributed decision making, each AGV is regarded as an independent agent for autonomous decision making, the computing load of the central controller is reduced, and the scalability of the system is enhanced. The priority time window mechanism is introduced, resources are dynamically allocated according to the task urgency, the remaining power and the historical completion rate, and the resource competition problem of multiple AGVs in the time and space dimensions is solved. Combined with deep neural network and asynchronous advantage actor critic algorithm, an adaptive scheduling optimization framework is constructed, and the learning ability of the system to complex environment is improved. Through the design of master-slave hierarchical structure, the cooperative decision making of AGV cluster is realized, and the consistency of global task decomposition and local resource allocation is guaranteed. A dynamic load balancing mechanism based on multi-objective optimization is established, and the balance of system performance in multiple dimensions such as resource utilization, task completion rate and energy efficiency is realized. A distributed consistency protocol is used to handle emergencies, and the robustness of the system in the face of uncertainty is enhanced. A unified scheduling instruction format is designed, the control information is standardized and normalized, and the maintenance and expansion of the system are facilitated.

[0020] In a specific embodiment, the process of step S1 can specifically include the following steps: The total area of the AGV job area is equally spaced grid division calculation, a plurality of grid units are obtained, and the center point of each grid unit is taken as the coordinate origin to establish a local coordinate system. Each grid unit is given a unique identification code to obtain a grid unit identification matrix; The charging pile position coordinates, workstation position coordinates and storage device position coordinates in the AGV job area are input into the grid unit identification matrix for mapping calculation to obtain a resource node distribution map; The adjacent grid units in the resource node distribution map are detected for connectivity state to obtain a network topology connection matrix, and the network topology connection matrix is input into a multi-objective optimization model. Three optimization objectives of minimizing path distance, minimizing turning number and minimizing passing time are set, and a Pareto optimal solution is used for solving to obtain a target network topology structure; The passing state of each grid unit in the target network topology structure is detected in real time to obtain a passing state matrix, and the resource occupation of the charging pile, workstation and storage device in the target network topology structure is monitored in real time to obtain a resource occupation rate matrix; The dynamic scheduling network is generated based on the grid unit identification matrix, the target network topology structure, the passing state matrix and the resource occupation rate matrix.

[0021] Specifically, the total area of the entire job area is equally spaced grid division. Let the length and width of the job area be L and W respectively, and the grid spacing be S. The total area of the job area is A = L * W, and the number of grid units is N = A / S2. The total area of the entire job area is equally spaced grid division. Let the length and width of the job area be L and W respectively, and the grid spacing be S. The total area of the job area is A = L * W, and the number of grid units is N = A / S2. and , the side length of the grid is . The positions of resource nodes (such as charging piles, workstations, storage devices) in the job area are input into the grid identification matrix for mapping calculation. In order to establish the connectivity relationship between the grid units, the adjacent grid units in the resource node distribution diagram are detected for connectivity state. Let the connectivity state of the grid units and be a binary variable , if there is a passable path between the two grids, then , otherwise . This process is completed by detecting obstacles or other restrictions between the two grids to generate a network topology connection matrix . The network topology connection matrix is input into the multi-objective optimization model for path planning optimization. The optimization objectives include the minimization of path distance , the minimization of the number of turns , and the minimization of travel time . By using the Pareto optimal solution method to comprehensively solve the above objectives, the optimal path planning result, i.e., the target network topology structure , is obtained. The passable state of each grid unit in the target network topology structure is detected in real time to generate a passable state matrix . If the grid unit is passable, then , otherwise . At the same time, the occupation of resource nodes such as charging piles, workstations, and storage devices is monitored in real time. Based on the grid unit identification matrix , the target network topology structure , the passable state matrix , and the resource occupation rate matrix , a dynamic scheduling network is comprehensively generated. The dynamic scheduling network can dynamically respond to environmental changes.

[0022] The embodiment also includes a self-organizing cooperative scheduling strategy based on dynamic topology reconstruction, which includes: hierarchical topology decomposition of the dynamic scheduling network, construction of core task subgraph, standby path subgraph and resource competition subgraph, extraction of multi-dimensional topology features, and obtaining of network topology feature tensor; inputting the network topology feature tensor into a dynamic graph neural network to model the time sequence evolution of node embedding vectors and edge attribute vectors, constructing a time-varying topology predictor, and obtaining a topology evolution model; based on the topology evolution model, reconstructing the local topology structure of the AGV cluster in real time, mapping the reconstructed topology structure to the resource allocation space, dividing the dynamic resource partition using a spectral clustering method, and obtaining an adaptive resource allocation scheme; performing robustness analysis on the adaptive resource allocation scheme, constructing constraint conditions based on Lyapunov stability theory, designing a distributed consistency protocol, and obtaining stability criteria for topology reconstruction; based on the stability criteria, dynamically recombining the local communication topology of the AGV cluster, constructing a self-organizing cooperative mechanism, and obtaining a topology reconstruction decision; coupling and optimizing the topology reconstruction decision and the task allocation strategy, constructing a multi-level mixed integer programming model, solving the optimal scheduling scheme, and obtaining a cooperative scheduling strategy; evaluating the performance of the cooperative scheduling strategy, calculating the redundancy, robustness and adaptability of the topology structure, and obtaining a strategy evaluation result; and according to the strategy evaluation result, adaptively adjusting the topology reconstruction parameters and the scheduling optimization parameters to obtain a self-organizing cooperative scheduling strategy.

[0023] In a specific embodiment, the process of performing step S2 can specifically include the following steps: Collecting the position coordinate data, the remaining power value, the current executed task number, and the real-time load weight of each AGV to obtain an original state data set; According to the original state data set, constructing a four-dimensional vector for the AGV state space, mapping the position coordinates to the spatial dimension, the power to the energy dimension, the task number to the task dimension, and the load weight to the load dimension to obtain a state space vector; Based on the state space vector, assigning a unique agent identifier to each AGV, establishing a communication link matrix between agents, and obtaining an initial agent network; Quantitatively calculating the interaction relationship between adjacent agents in the initial agent network, constructing an interaction weight matrix based on the task correlation coefficient and the resource competition coefficient, and obtaining an agent interaction network; According to the agent interaction network, segmenting and sampling historical scheduling data to obtain a training sample set, and performing Markov property analysis on the training sample set to calculate the transition probability of the state space vector between adjacent time instants, and constructing a state transition probability matrix; Based on the state transition probability matrix, constructing a reward function, setting the task completion time, energy consumption, and resource utilization as evaluation indexes, and obtaining a multi-agent scheduling model.

[0024] Specifically, the key data of each AGV is collected in real time. Assuming that the position information of a certain AGV is , the remaining power is , the current task number is , and the real-time load is . These data are aggregated into an original state data set , where represents the current position of the AGV, represents the remaining capacity of the battery, represents the current task identifier, and represents the current load. By collecting the states of all AGVs, a global original state data set is formed. According to the original state data set, the state space of the AGV is modeled as a four-dimensional vector, where each dimension corresponds to a different physical attribute. The position information is mapped to the spatial dimension to describe the specific position of the AGV in the work area; the power is mapped to the energy dimension to represent the current endurance of the AGV. On the basis of the state space vector, a unique agent identifier is assigned to each AGV, and a communication link matrix is established according to the communication ability between agents. The element of the communication link matrix represents the communication state between agent and agent , if the two can communicate directly, then otherwise . Through this process, an initial agent network is formed, which describes the communication topology between all AGVs. In the initial agent network, the interaction relationship between adjacent agents is quantitatively calculated to construct an agent interaction network. The interaction relationship is measured by a task correlation coefficient and a resource competition coefficient. According to the agent interaction network, historical scheduling data is segmented and sampled to obtain a training sample set. Assuming that the historical data contains time periods, the state vector of each period is , then the training sample set is represented as . Through Markov property analysis of these samples, the transition probability of the state vector at adjacent times is calculated. Based on the state transition probability matrix, a reward function is designed to evaluate the pros and cons of the scheduling scheme. By optimizing the reward function, a multi-agent scheduling model is constructed, which can dynamically adapt to complex environments and optimize the scheduling strategy.

[0025] In a specific embodiment, the process of performing step S3 can specifically include the following steps: The difference between the deadline and the current time of the AGV task is calculated, and the task urgency coefficient is set as the normalized reciprocal of the difference to obtain the task urgency score. Based on the energy dimension in the multi-agent scheduling model, the ratio of the remaining power of the AGV to the power required for the task is calculated to obtain the power sufficiency score. Based on the historical scheduling data in the multi-agent scheduling model, the ratio of the number of tasks completed by the AGV in the past K time windows to the total number of tasks is calculated to obtain the historical task completion rate score. Priority scores are calculated based on task urgency, battery adequacy, and historical task completion rate, resulting in a comprehensive task priority score. The operation time axis is divided into equal intervals according to a fixed length to generate a set of time slices. A time window resource matrix is ​​established based on the set of time slices. The number of available charging piles, workstations, and AGVs in each time slice is recorded as a resource capacity vector to obtain a time slice resource capacity table. The AGVs are sorted in descending order based on their overall task priority score, and then assigned to the available time slices in the time slice resource capacity table to obtain the time slice resource allocation result.

[0026] Specifically, the difference between the AGV task's deadline and the current time is calculated, and the task urgency coefficient is set as the normalized reciprocal of this difference to obtain a task urgency score. For example, if the remaining time for a task is close to zero, then... When the value is close to 1, and the remaining time is relatively long, The value is relatively small. Based on the energy dimension in the multi-agent scheduling model, the power adequacy score of the AGV is calculated. The higher the score, the more sufficient the energy reserve of the AGV, and the greater the probability of completing the task. The task completion status of the AGV is analyzed, and the task completion rate score is calculated using historical data in the multi-agent scheduling model. After obtaining the urgency, power adequacy, and task completion rate scores, the comprehensive task priority score is calculated by weighted summation. The operation time axis is divided into several equally spaced time slices of fixed length. All tasks are sorted in descending order according to the comprehensive task priority score. Let the task set be... Task priorities are sorted by score as follows: If the current time slice If the resource capacity meets the task requirements, then the task will be... Assigned to time slice If the conditions are not met, the next time slice is checked until a suitable time slice is found. Through the above process, the time slice resource allocation result is obtained.

[0027] In one specific embodiment, the process of performing step S4 may specifically include the following steps: According to the time slice resource allocation result, the current system state is characterized, and the AGV position coordinates, remaining power, task state, and load condition are constructed into a state feature vector; Based on the state feature vector, the deep neural network architecture is divided into a policy network and a value network, and the policy network is set with an Actor learning rate and a policy gradient threshold, and the value network is set with a Critic learning rate and a time difference threshold, to obtain a double-network training model; According to the double-network training model, an asynchronous advantage function is constructed, the difference between the actual reward value of each AGV and the value network estimation value is taken as the advantage measure, and the advantage function is calculated in time series to obtain an advantage evaluation result; The advantage evaluation result is input into the Actor-Critic algorithm, the policy gradient and value loss of each AGV are updated, and a trained neural network is obtained; The trained neural network is subjected to dynamic modeling, the AGV speed and acceleration are mapped into discrete state variables, an AGV discrete event model is obtained, and hybrid optimization is performed based on the AGV discrete event model to obtain a distributed scheduling strategy.

[0028] Specifically, the current system state is characterized. The system state is composed of key parameters of AGVs, including position coordinates, remaining power, task state, and load. Assuming that the current position of an AGV is , the remaining power is , the current task number is , and the load is . Based on the state feature vector, a deep neural network model is constructed, which is composed of a policy network and a value network. The policy network is used to generate a scheduling policy, and the value network is used to estimate the value function of a given state. The learning rate and the policy gradient threshold of the policy network are set, and the learning rate and the time difference threshold of the value network are set. These hyperparameters control the convergence speed and training stability of the model. Based on the double-network model, an asynchronous advantage function is constructed to measure the relative superiority of the current policy compared to the baseline policy. The advantage evaluation result is input into the Actor-Critic algorithm, and the gradient of the policy network and the loss function of the value network are optimized. The optimized deep neural network is trained. On the trained neural network, dynamic modeling is performed, the speed and acceleration of the AGV are mapped into discrete state variables, and the discrete event model of the AGV is constructed. Based on the discrete event model, hybrid optimization is performed, considering the goals of path planning, task allocation, and resource scheduling. Through mixed integer linear programming or other optimization algorithms, the optimal scheduling strategy is solved. The distributed scheduling strategy combines the motion parameters of each AGV, the task priority, and the resource usage permission to generate an efficient global scheduling scheme.

[0029] In a specific embodiment, the process of performing step S5 can specifically include the following steps: An AGV performance score table is constructed for the AGV cluster in the distributed scheduling strategy according to task processing capacity, remaining power level, and communication bandwidth; AGVs in the AGV performance score table that meet a preset score range are divided into a master AGV set, and the remaining AGVs are divided into a slave AGV set, to obtain an AGV cluster hierarchical structure; The AGV cluster hierarchical structure is mapped to a communication topology network, a full-connection network structure is constructed for the master AGV set, and a star network structure is constructed for the slave AGV set and the corresponding master AGV, to obtain a two-layer communication topology graph; The global task matrix in the two-layer communication topology graph is regionally segmented to obtain a task decomposition sequence, and the scheduling period of the AGV cluster is synchronously divided according to the time window requirement of each subtask in the task decomposition sequence, to obtain a task scheduling time sequence table; Based on the task scheduling time sequence table, the charging pile usage weight, the workstation occupation weight, and the storage device usage weight are calculated, to obtain a resource allocation matrix; A consistency constraint condition is set for the resource allocation matrix, a distributed average consistency protocol based on a Laplacian matrix is constructed, and the AGV decision variable is iteratively updated to obtain a consistency decision vector; The consistency decision vector is input into a collaborative optimizer, the AGV decision variable is iteratively optimized, and a cluster collaborative decision result is obtained.

[0030] Specifically, an AGV performance score table is constructed for the AGV cluster in the distributed scheduling strategy according to task processing capacity, remaining power level, and communication bandwidth. All indexes are normalized to [0, 1] to ensure that data of different dimensions can be compared. According to the performance score table, AGVs with scores meeting a preset range are divided into a master AGV set, and the remaining AGVs are divided into a slave AGV set. A hierarchical structure of the AGV cluster is obtained, in which the master AGVs are responsible for high-level decision-making, and the slave AGVs are responsible for the execution of specific tasks. The hierarchical AGV cluster is mapped to a communication topology network, a full-connection network structure is constructed for the master AGV set to ensure efficient information exchange and collaboration between the master AGVs, and a star network structure is constructed for the slave AGV set and the corresponding master AGV. The full-connection network and the star network are integrated to form a two-layer communication topology graph that describes the communication structure of the entire AGV cluster. On the basis of the two-layer communication topology graph, the global task matrix is regionally segmented. It is assumed that the global task matrix is wherein represents the execution requirement of task and grid cell the degree of association. The region division divides the task into multiple sub-tasks, denoted as , each of which corresponds to a specific region and is divided according to the spatial distribution of the task. The time window requirement for each sub-task is defined, and the time window is ensured to complete the task within the corresponding time slice. Based on the task decomposition sequence and the time window, the scheduling cycle of the AGV cluster is synchronously divided to form a task scheduling time sequence table. At the same time, the weight of resource usage is calculated for each task, including the charging pile usage weight , workstation occupation weight , and warehouse equipment usage weight , forming a resource allocation matrix , where represents the resource demand of task in time slice . The resource allocation matrix is subjected to a consistency constraint condition, and a distributed average consistency protocol based on the Laplacian matrix is constructed to update the decision variables of AGVs. Through iterative updating formula, the decision variables of AGVs gradually tend to be consistent, and a consistent decision vector is obtained. The consistent decision vector is input into the collaborative optimizer, and the optimization objectives include minimizing the task completion time, maximizing the resource utilization rate, and optimizing the path planning. Through iterative optimization, the cluster collaborative decision result is finally obtained, which allocates motion parameters, task sequence and resource authority to each AGV, thereby realizing intelligent scheduling.

[0031] In a specific embodiment, the process of performing step S6 can specifically include the following steps: Input the cluster collaborative decision result into the scheduler, and sample the position coordinate data, task state data, and resource occupation data in the system execution cycle to obtain AGV cluster state data; Based on the AGV cluster state data, the charging device utilization rate is obtained by dividing the number of charging pile occupation by the total number, the workstation utilization rate is obtained by dividing the number of tasks by the processing capacity, the warehouse utilization rate is obtained by dividing the number of storage and retrieval times by the time, and the resource utilization rate index is calculated by weighted summation; Based on the AGV cluster state data, the completion rate is obtained by dividing the number of completed tasks by the total number of tasks, the execution rate is obtained by dividing the number of tasks being executed by the number of online AGVs, the timeliness efficiency is obtained by dividing the average response time of the task by the standard response time, and the task completion rate index is calculated by linear combination; According to the AGV cluster state data, the energy efficiency index is obtained by normalizing the energy efficiency index, which is obtained by dividing the remaining power by the full power, dividing the driving mileage by the energy consumption, and dividing the number of charging times by the number of tasks; Based on the resource utilization index, task completion rate index and energy efficiency index, comprehensive load analysis is performed to obtain load evaluation results, and a robust control strategy is generated according to the load evaluation results; According to the robust control strategy, a set of scheduling instructions is generated, AGV motion parameters, task execution sequences and resource usage permissions are packaged into unified control messages, and scheduling instructions for AGV clusters are generated.

[0032] Specifically, the cluster collaborative decision result is input into the scheduler, and the position coordinates , task state , resource occupation and other data of each AGV are collected for the dynamic running state within the system execution cycle. Through real-time sampling, a dynamic data set describing the overall state of the AGV cluster is formed , wherein , represents the total number of AGVs. Based on the collected AGV state data, the resource utilization index is calculated. The charging device utilization rate is defined as the ratio of the number of occupied charging piles to the total number of charging piles , the workstation utilization rate is obtained by dividing the number of tasks of the workstation by its maximum processing capacity , the storage device utilization rate is the ratio of the number of storage access times per unit time to the operation time ; the comprehensive resource utilization rate is obtained by weighted sum of the above three indexes. The task completion rate index is calculated based on the AGV state data. The completion rate is the ratio of the number of completed tasks to the total number of tasks . The execution rate is the ratio of the number of tasks being executed to the number of online AGVs ; the task time efficiency is the ratio of the average response time of the task to the standard response time ; the task completion rate index is obtained by linear combination. The energy efficiency index is calculated. The power level is defined as the ratio of the remaining power of the AGV to its full power . The travel efficiency is the ratio of the travel mileage to the total energy consumption . The charging frequency is the ratio of the number of charges to the number of tasks The ratio of the above indicators is normalized, and the comprehensive energy efficiency index is obtained by weighted summation . Based on resource utilization , task completion rate and energy efficiency , comprehensive load analysis is carried out to generate load evaluation results , reflecting the importance of each indicator. According to the load evaluation results, a robust control strategy is generated to adjust task allocation, path selection and resource usage priority. The process of control strategy generation includes dynamically adjusting the motion parameters of AGV, task execution sequence , resource allocation authority , etc. These information is packaged into a unified control message to generate a set of scheduling instructions.

[0033] The above describes the intelligent scheduling and resource allocation method of flexible AGV in the embodiment of the present application. Next, the intelligent scheduling and resource allocation system of flexible AGV in the embodiment of the present application is described. Please refer to Figure 2 , an embodiment of the intelligent scheduling and resource allocation system of flexible AGV in the embodiment of the present application includes: Grid module, for grid processing of AGV operation area to obtain dynamic scheduling network; Modeling module, for modeling AGV state space in dynamic scheduling network to obtain multi-agent scheduling model; Calculation module, for calculating AGV task priority score based on multi-agent scheduling model, and discretizing job time axis into time slices to obtain time slice resource allocation result; Optimization module, for optimizing AGV scheduling decision according to time slice resource allocation result to obtain distributed scheduling strategy; Division module, for dividing distributed scheduling strategy to obtain master AGV and slave AGV, and generating cluster collaborative decision result; Allocation module, for resource allocation and dynamic load balancing based on cluster collaborative decision result to obtain AGV cluster scheduling instructions.

[0034] Through the synergistic cooperation of each component, the continuous job space is discretized by establishing a grid-based dynamic scheduling network, the precise modeling of AGV motion state and resource distribution is realized, and the system has real-time perception and dynamic updating capabilities. A multi-agent scheduling model is used for distributed decision-making, each AGV is an independent agent for autonomous decision-making, which reduces the computational load of the central controller and enhances the scalability of the system. The priority time window mechanism is introduced, and resources are dynamically allocated according to the task urgency, remaining power and historical completion rate, solving the resource competition problem of multiple AGVs in the time and space dimensions. Combined with deep neural networks and asynchronous advantage actor critic algorithm, an adaptive scheduling optimization framework is constructed to improve the system's learning ability in complex environments. Through the design of master-slave hierarchical structure, the cooperative decision-making of AGV cluster is realized, ensuring the consistency of global task decomposition and local resource allocation. A dynamic load balancing mechanism based on multi-objective optimization is established to balance the system performance in multiple dimensions such as resource utilization, task completion rate and energy efficiency. A distributed consistency protocol is used to handle unexpected events, enhancing the robustness of the system in the face of uncertainty. A unified scheduling instruction format is designed to standardize and normalize control information, facilitating system maintenance and expansion.

[0035] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, system and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0036] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0037] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalent replacements; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent scheduling and resource allocation of flexible AGVs, characterized in that, The method includes: The AGV operating area is gridded to obtain a dynamic scheduling network; The state space of the AGV in the dynamic scheduling network is modeled to obtain a multi-agent scheduling model; Based on the multi-agent scheduling model, the task priority score of the AGV is calculated, and the operation time axis is discretized into time slices to obtain the time slice resource allocation result. Based on the time slice resource allocation results, the AGV scheduling decision is optimized to obtain a distributed scheduling strategy; The distributed scheduling strategy is divided into master AGVs and slave AGVs, and cluster collaborative decision-making results are generated. Based on the cluster collaborative decision-making results, resource allocation and dynamic load balancing are performed to obtain the scheduling instructions for the AGV cluster.

2. The intelligent scheduling and resource allocation method for flexible AGVs according to claim 1, characterized in that, The process of gridding the AGV operating area to obtain a dynamic scheduling network includes: The total area of ​​the AGV operating area is divided into equally spaced grids to obtain multiple grid cells. A local coordinate system is established with the center point of each grid cell as the origin. Each grid cell is assigned a unique identifier to obtain the grid cell identifier matrix. The coordinates of the charging piles, workstations, and storage equipment within the AGV's operating area are input into the grid cell identifier matrix for mapping calculation to obtain a resource node distribution map. Connectivity state detection is performed on adjacent grid cells in the resource node distribution map to obtain the network topology connection matrix. The network topology connection matrix is ​​then input into a multi-objective optimization model. Three optimization objectives are set: minimizing path distance, minimizing the number of turns, and minimizing travel time. Pareto optimal solution is used to solve the model to obtain the target network topology. Real-time traffic status detection is performed on each grid cell in the target network topology to obtain a traffic status matrix. Real-time monitoring of resource occupancy of charging piles, workstations, and warehousing equipment in the target network topology is also performed to obtain a resource occupancy matrix. A dynamic scheduling network is generated based on the grid cell identifier matrix, the target network topology, the traffic status matrix, and the resource occupancy matrix.

3. The intelligent scheduling and resource allocation method for flexible AGVs according to claim 2, characterized in that, The process of modeling the AGV state space in the dynamic scheduling network to obtain a multi-agent scheduling model includes: The location coordinates, remaining power, currently executing task number, and real-time load capacity of each AGV are collected to obtain the raw status dataset; Based on the original state dataset, a four-dimensional vector is constructed for the AGV state space. The position coordinates are mapped to the spatial dimension, the power is mapped to the energy dimension, the task number is mapped to the task dimension, and the load weight is mapped to the load dimension, thus obtaining the state space vector. Based on the state space vector, a unique agent identifier is assigned to each AGV, and a communication link matrix between agents is established to obtain the initial agent network. The interaction relationships between adjacent agents in the initial agent network are quantified and calculated. An interaction weight matrix is ​​constructed based on the task relevance coefficient and resource competition coefficient to obtain the agent interaction network. Based on the intelligent agent interaction network, the historical scheduling data is sampled in segments to obtain a training sample set, and Markov characteristic analysis is performed on the training sample set to calculate the transition probability of the state space vector between adjacent time moments and construct a state transition probability matrix. A reward function is constructed based on the state transition probability matrix, and task completion time, energy consumption, and resource utilization are set as evaluation indicators to obtain a multi-agent scheduling model.

4. The intelligent scheduling and resource allocation method for flexible AGVs according to claim 3, characterized in that, The process of calculating the task priority score of the AGV based on the multi-agent scheduling model and discretizing the operation time axis into time slices to obtain the time slice resource allocation result includes: The difference between the deadline and the current time of the AGV task is calculated, and the task urgency coefficient is set as the normalized reciprocal of the difference to obtain the task urgency score. Based on the energy dimension in the multi-agent scheduling model, the ratio of the remaining power of the AGV to the power required for the task is calculated to obtain the power sufficiency score. Based on the historical scheduling data in the multi-agent scheduling model, the ratio of the number of tasks completed by the AGV in the past K time windows to the total number of tasks is calculated to obtain the historical task completion rate score. Based on the task urgency score, the battery adequacy score, and the historical task completion rate score, a priority score is obtained to obtain a comprehensive task priority score. The operation time axis is divided into equal intervals according to a fixed length to generate a time slice set. A time window resource matrix is ​​established based on the time slice set. The number of available charging piles, workstations, and AGVs in each time slice is recorded as a resource capacity vector to obtain a time slice resource capacity table. The AGVs are sorted in descending order based on the comprehensive score of task priority, and then the AGVs are allocated to the available time slices in the time slice resource capacity table in sequence to obtain the time slice resource allocation result.

5. The intelligent scheduling and resource allocation method for flexible AGVs according to claim 4, characterized in that, The step of optimizing the AGV scheduling decision based on the time slice resource allocation result to obtain a distributed scheduling strategy includes: Based on the time slice resource allocation results, feature extraction is performed on the current system state, and the AGV position coordinates, remaining power, task status, and load status are constructed into a state feature vector. Based on the state feature vector, the deep neural network architecture is divided into a policy network and a value network. The Actor learning rate and policy gradient threshold are set for the policy network, and the Critic learning rate and temporal difference threshold are set for the value network to obtain a dual-network training model. An asynchronous advantage function is constructed based on the dual-network training model. The difference between the actual reward value of each AGV and the value network estimate is used as the advantage metric. The advantage function is then calculated by time-series accumulation to obtain the advantage evaluation result. The advantage evaluation results are input into the actor commentator algorithm to update the policy gradient and value loss of each AGV, resulting in a trained neural network. The trained neural network is dynamically modeled, and the AGV speed and acceleration are mapped to discrete state variables to obtain an AGV discrete event model. Based on the AGV discrete event model, hybrid optimization is performed to obtain a distributed scheduling strategy.

6. The intelligent scheduling and resource allocation method for flexible AGVs according to claim 5, characterized in that, The process of dividing the distributed scheduling strategy into master AGVs and slave AGVs, and generating cluster collaborative decision-making results, includes: An AGV performance scoring table is constructed for the AGV cluster in the distributed scheduling strategy based on task processing capacity, remaining power level, and communication bandwidth. The AGVs that meet the preset scoring range in the AGV performance scoring table are divided into a set of master AGVs, and the remaining AGVs are divided into a set of subordinate AGVs, thus obtaining a hierarchical structure of the AGV cluster. The hierarchical structure of the AGV cluster is mapped to a communication topology network. A fully connected network structure is constructed for the master control AGV set, and a star network structure is constructed for the subordinate AGV set and the corresponding master control AGV, resulting in a two-layer communication topology diagram. The global task matrix in the two-layer communication topology is divided into regions to obtain a task decomposition sequence. Based on the time window requirements of each subtask in the task decomposition sequence, the scheduling cycle of the AGV cluster is synchronously divided to obtain a task scheduling time sequence table. Based on the task scheduling timeline, the weights of charging pile usage, workstation occupancy, and warehousing equipment usage are assigned and calculated to obtain a resource allocation matrix. Consistency constraints are set on the resource allocation matrix, a distributed average consistency protocol based on the Laplace matrix is ​​constructed, and the AGV decision variables are iteratively updated to obtain a consistent decision vector. The consensus decision vector is input into the collaborative optimizer to iteratively optimize the AGV decision variables and obtain the cluster collaborative decision result.

7. The intelligent scheduling and resource allocation method for flexible AGVs according to claim 6, characterized in that, The process of allocating resources and dynamically balancing loads based on the cluster collaborative decision-making results to obtain scheduling instructions for the AGV cluster includes: The cluster collaborative decision-making results are input into the scheduler, and the location coordinate data, task status data, and resource usage data within the system execution cycle are sampled in real time to obtain AGV cluster status data. Based on the AGV cluster status data, the utilization rate of charging equipment is obtained by dividing the number of charging piles occupied by the total number, the utilization rate of workstations is obtained by dividing the number of workstation tasks by the processing capacity, and the utilization rate of warehouses is obtained by dividing the number of times warehouse equipment is accessed by the time. The resource utilization rate index is obtained by weighted summation. Based on the AGV cluster status data, the completion rate is obtained by dividing the number of completed tasks by the total number of tasks, the execution rate is obtained by dividing the number of tasks currently being executed by the number of online AGVs, and the time efficiency is obtained by dividing the average task response time by the standard response time. The task completion rate index is obtained by linear combination calculation. Based on the AGV cluster status data, the remaining power of the AGV is divided by the full power to obtain the power level, the driving distance is divided by the energy consumption to obtain the driving efficiency, the number of charging times is divided by the number of tasks to obtain the charging frequency, and the energy efficiency index is obtained through normalization. A comprehensive load analysis is performed based on the resource utilization rate index, the task completion rate index, and the energy efficiency index to obtain the load assessment result, and a robust control strategy is generated based on the load assessment result. Based on the robust control strategy, a scheduling instruction set is generated, and AGV motion parameters, task execution sequences, and resource usage permissions are packaged into a unified control message to generate scheduling instructions for the AGV cluster.

8. An intelligent scheduling and resource allocation system for flexible AGVs, characterized in that, The intelligent scheduling and resource allocation method for executing the flexible AGV as described in any one of claims 1-7, wherein the intelligent scheduling and resource allocation system for the flexible AGV comprises: The gridding module is used to perform grid processing on the AGV operating area to obtain a dynamic scheduling network; The modeling module is used to model the state space of AGVs in the dynamic scheduling network to obtain a multi-agent scheduling model. The calculation module is used to calculate the task priority score of the AGV based on the multi-agent scheduling model, and discretize the operation time axis into time slices to obtain the time slice resource allocation result; The optimization module is used to optimize the AGV scheduling decision based on the time slice resource allocation result to obtain a distributed scheduling strategy. The partitioning module is used to partition the distributed scheduling strategy to obtain the master AGV and slave AGV, and generate cluster collaborative decision results; The allocation module is used to allocate resources and dynamically balance loads based on the cluster collaborative decision results, and to obtain scheduling instructions for the AGV cluster.

Images