Multi-agv intelligent scheduling optimization method

Abstract

The present application relates to the field of automation logistics, aiming at the problems of high empty load rate, poor coordination and easy to fall into local optimum of algorithm of multi-distribution center AGV scheduling in the prior art, a multi-AGV intelligent scheduling optimization method is disclosed, a multi-distribution center semi-open heterogeneous AGV simultaneous collection and distribution path planning model MDHOHVRPSPDTW with mixed time window is constructed. The method uses mixed time window constraints to adapt to the time efficiency demand of different stations, uses the simultaneous collection and distribution mode to reduce the empty load rate of the vehicle, and allows the AGV to return nearby through the semi-open strategy to reduce the empty running. The solving process adopts the multi-subgroup velocity pause particle swarm optimization algorithm MSVPPSO-LNS combined with large neighborhood search, and through the introduction of chaos initialization, velocity pause mechanism and multi-subgroup cooperation strategy, the global optimization ability is enhanced. The experimental results show that the present application can effectively reduce the transportation cost and penalty cost, and significantly improve the flexibility and resource utilization rate of the scheduling system.

Description

Technical Field

[0001] This invention relates to the field of automated logistics technology, and in particular to an intelligent scheduling and optimization method for multiple AGVs. Background Technology

[0002] With the rapid development of intelligent manufacturing and automated logistics technologies, Automated Guided Vehicles (AGVs) have been widely used in material distribution in large production workshops, especially in complex logistics systems with multiple distribution centers. However, existing AGV scheduling technologies still have significant limitations when dealing with large-scale, multi-constraint production scenarios.

[0003] First, in a single distribution center scenario, traditional scheduling methods often employ a single time window constraint (either a hard time window or a soft time window). While hard time windows ensure timeliness, they lack flexibility, while soft time windows, though flexible, can easily lead to delays at critical workstations and are difficult to adapt to the diverse needs of both critical and general workstations within the workshop. Furthermore, existing solutions often employ a separate collection and distribution operation mode (e.g., delivering first and then picking up), resulting in high AGV single-trip empty load rates and low transportation efficiency.

[0004] Secondly, in multi-distribution-center collaborative scenarios, the existing mainstream scheduling models are mostly closed-loop, requiring AGVs to return to their original departure center after completing their tasks. This model leads to a significant amount of wasted empty return trips in workshops with large distances between multiple centers. Furthermore, existing technologies often fail to adequately consider the performance and cost differences between heterogeneous fleets (such as light-load and heavy-load AGVs), resulting in low vehicle-to-task matching and insufficient resource utilization.

[0005] Finally, at the algorithm level, faced with multiple complex constraints such as mixed time windows, simultaneous collection and distribution of goods, and heterogeneous fleet coordination, traditional path planning algorithms (such as standard genetic algorithms and particle swarm algorithms) are prone to getting trapped in local optima, resulting in low solution efficiency and difficulty in meeting the quality requirements of real-time production scheduling. Summary of the Invention

[0006] This invention provides an intelligent scheduling optimization method for multiple AGVs. Through comprehensive optimization of hybrid time window management, simultaneous collection and distribution operations, and semi-open multi-center collaborative scheduling, it significantly reduces logistics and transportation costs and vehicle empty load rate, and effectively improves the scheduling flexibility and overall operating efficiency of heterogeneous AGV fleets in complex production environments.

[0007] This invention provides an intelligent scheduling and optimization method for multiple AGVs, comprising:

[0008] S1. Obtain multi-dimensional logistics basic data of the production workshop, including workstation task parameters, layout parameters, and fleet parameters; wherein, the workstation task parameters include workstation coordinates and operation time windows, the layout parameters are the location coordinates of each distribution center, and the fleet parameters include the load limit and running speed of each heterogeneous AGV.

[0009] S2. Establish an objective function for the vehicle route planning model based on the multi-dimensional logistics basic data. The objective function aims to minimize the total delivery cost, which consists of vehicle fixed and driving costs and time window penalty costs. The vehicle fixed and driving costs are calculated based on the fleet parameters, and the time window penalty costs are calculated based on the workstation task parameters.

[0010] S3. Based on the multidimensional logistics basic data, establish the constraints of the vehicle route planning model, including semi-open route and vehicle scheduling constraints, heterogeneous vehicle capacity and simultaneous collection and distribution constraints, and mixed time window constraints.

[0011] S4. The vehicle path planning model is solved by using a multi-subgroup speed-pause particle swarm optimization algorithm that combines large neighborhood search; wherein, the multi-subgroup speed-pause particle swarm optimization algorithm is an improvement on the standard particle swarm algorithm, including using Logistic chaotic mapping to initialize the population distribution, introducing a speed-pause mechanism in the iterative update, using a multi-subgroup cooperative strategy to share global optimal information, and incorporating a large neighborhood search mechanism.

[0012] S5. Obtain the global optimal solution obtained by the multi-subgroup velocity-pause particle swarm optimization algorithm, decode the global optimal solution to obtain a scheduling scheme and convert it into a multi-AGV scheduling instruction to be issued to each heterogeneous AGV; wherein, the scheduling scheme includes the optimal workstation access order corresponding to the workstation task parameters, the workstation assignment center determined from the layout parameters, and the vehicle type allocation matched from the fleet parameters.

[0013] Furthermore, S1 specifically includes:

[0014] S101. Determine all workstations in the production workshop that require material delivery or recycling to establish a workstation set. Obtain the planar coordinates, delivery demand, pickup demand, loading and unloading service time, and optimal service time window of each workstation as workstation task parameters. The optimal service time window is divided into hard time windows and soft time windows based on the urgency of the workstation task.

[0015] S102. Determine the layout of multiple distribution centers in the production workshop to establish a distribution center set, obtain the planar coordinates of each distribution center as layout parameters, and calculate the Euclidean distance between any two points based on the planar coordinates of the work site set and the distribution center set.

[0016] S103. Determine the different vehicle types in the heterogeneous AGV fleet to establish a heterogeneous AGV set, and obtain the maximum load, average driving speed, total number available in the workshop, fixed usage cost per trip, and transportation cost per unit driving distance for each type of AGV as fleet parameters; wherein, the heterogeneous AGV set includes at least two types: light-load AGV and heavy-load AGV.

[0017] Furthermore, S2 specifically includes:

[0018] S201. Construct a fixed usage cost function for vehicles, and calculate the sum of the fixed departure costs of all scheduled and activated heterogeneous AGVs. The calculation formula is as follows:

[0019]

[0020] in, R represents the fixed operating cost of the vehicle; R is the set of heterogeneous AGV types. Let A be the set of vehicles of type r AGV, A be the set of arc segments formed by all nodes, and C be the set of work points; This represents the fixed dispatch cost of the r-th type AGV; The variable is a decision variable; it takes the value 1 if the kth AGV in the rth class is activated, and 0 otherwise.

[0021] S202. Construct a vehicle transportation cost function to calculate the sum of the travel distance costs incurred by all scheduled and activated heterogeneous AGVs during the delivery task. The calculation formula is as follows:

[0022]

[0023] in, For vehicle transportation costs; This represents the unit distance transportation cost of the r-th type of AGV; This represents the Euclidean distance between node i and node j; The value of the variable is 1 if the kth AGV in the rth class travels from node i to node j, and 0 otherwise.

[0024] S203. Construct a time window penalty cost function. For workstations with a set soft time window, calculate the sum of delay penalty costs caused by the AGV arrival time exceeding the upper bound of the optimal service time window. The calculation formula is as follows:

[0025]

[0026] in, Penalty costs for time windows; This represents the penalty coefficient for delay per unit time. This represents the time when the k-th AGV in class r arrives at workstation i and begins service; This indicates the latest end time of the optimal service time window for workstation i; This indicates a violation of the time window; penalties are calculated only if the arrival time is later than the end time of the time window.

[0027] S204. Add the vehicle fixed usage cost function, vehicle travel cost function, and time window penalty cost function to establish a vehicle route planning model objective function with the goal of minimizing total delivery cost. .

[0028] Furthermore, S3 specifically includes:

[0029] S301. Establish semi-open path and vehicle scheduling constraints, specifically including:

[0030] Vehicle activation and route association constraints: ;

[0031] Available vehicle quantity constraints: ;

[0032] Unique service constraint: ;

[0033] Flow balance constraints: ;

[0034] Semi-open start and end constraints: Departure constraints Return constraints No-idle running constraint: ;

[0035] Eliminate sub-loop constraints:

[0036] Where D is the set of distribution centers and C is the set of work sites; For decision variables, the value is 1 if the kth AGV of the rth class travels from the distribution center d to the workstation j, and 0 otherwise; This is a decision variable; it is 1 if the vehicle is activated, and 0 otherwise. Let be the total number of available AGVs of type r; V be the set of all nodes. This indicates that the vehicle travels from node i to j; S is any subset of the workstation set C, and |S| is the number of elements in that subset;

[0037] S302. Establish heterogeneous vehicle capacity and simultaneous cargo collection and distribution constraints, specifically including:

[0038] Initial load definition: ;

[0039] Dynamic balance constraints on vehicle load capacity: ;

[0040] Capacity limit constraint: ; ;

[0041] in, The load of the vehicle when it leaves the distribution center d. The delivery demand for workstation j; , These represent the cargo loads when leaving nodes j and i, respectively. Let M be the required quantity of goods to be picked up at workstation j; M is a maximum positive number; A is the set of all arc segments; This represents the maximum rated load capacity of the r-th type AGV.

[0042] S303. Establish mixed time window constraints and variable definitions, specifically including:

[0043] Service time continuity constraints: ;

[0044] Soft time window delay constraint: ;

[0045] Initial time constraints: ;

[0046] Definition of decision variables and nonnegativity constraints: ;

[0047] in, , These represent the times when the AGV arrives at workstations j and i, respectively. Service duration for workstation i; Let M be the travel time from node i to j; M is a maximum positive number. The upper bound of the optimal service time window for workstation i; if ,but It is at least the difference between the two.

[0048] Furthermore, S4 specifically includes:

[0049] S401. Construct a three-segment particle coding structure and use Logistic chaotic mapping to generate the position and velocity of the initial population; wherein, the three-segment particle coding structure includes a first segment code representing the workstation assignment center, a second segment code representing the workstation vehicle type allocation, and a third segment code representing the workstation access order; the initialization formula of the Logistic chaotic mapping is:

[0050]

[0051] in, This represents the normalized variable generated in the nth iteration, with a value range of (0,1); This is a chaos control parameter with a value range of [3.57, 4], used to ensure that the sequence is in a completely chaotic state;

[0052] S402. Divide the particle population into multiple independent search subgroups, decode the position vector of each particle into a path scheme and calculate the fitness value, and initialize the local optimal solution of each subgroup and the global optimal solution of the population; wherein, the fitness value is calculated by the formula: f(x)=1 / Z, where f(x) represents the fitness value of particle x; Z is the comprehensive cost of the particle, including fixed cost, travel distance, and penalty term generated by time window constraint;

[0053] S403. Perform multi-subgroup collaborative update with the introduction of velocity pause mechanism. During the iteration process, determine whether the particle should maintain the current velocity or perform standard velocity update based on probability, and update the particle position.

[0054] S404. At the end of each iteration, perform a large neighborhood search destruction operation on the current global optimal solution, and remove several workstation nodes in the path according to a preset strategy; perform a large neighborhood search repair operation on the destroyed solution, and reinsert the removed workstation nodes into the path to generate a new solution. If the new solution is better than the original global optimal solution, then update it.

[0055] S405. Determine whether the maximum number of iterations or the convergence accuracy requirement has been reached. If the requirements are met, output the global optimal solution; otherwise, return to step S403 to continue iterating.

[0056] Furthermore, in S403, the specific update rules for the speed pause mechanism are as follows:

[0057] Set speed pause probability A random number is generated before each particle velocity update. ;like If the particle's velocity is paused, it will maintain its velocity from the previous moment. ;

[0058] like Then the particle performs a standard velocity update operation, that is:

[0059]

[0060] In the formula, and Let i represent the velocity and position of particle i in the t-th iteration, respectively. For inertial weights, As a learning factor, A random number between [0, 1]; Let be the individual historical best position of particle i. This represents the globally optimal position for the population.

[0061] Furthermore, in S404, the destructive operation adopts a hybrid removal strategy, including random removal, worst-case removal, and similarity removal; wherein, random removal randomly selects several workstations from the current solution for elimination; worst-case removal calculates the reduction in objective function cost after removing each workstation, and prioritizes removing the workstations that contribute the most to system cost; similarity removal removes a group of workstations with high similarity based on the geographical proximity or material requirement similarity between workstations;

[0062] The repair operation employs an optimal incremental insertion strategy, which sequentially calculates the objective function increment generated by the workstation to be inserted at all feasible insertion positions, and selects the position with the smallest increment for insertion.

[0063] Furthermore, S5 specifically includes:

[0064] S501. Obtain the global optimal particle vector output by the multi-subgroup velocity pause particle swarm optimization algorithm, and parse it into a three-segment structure of workstation assignment center vector, workstation vehicle allocation vector, and workstation access order vector.

[0065] S502. Extract the values ​​from the workstation access sequence vector and sort them. Establish the access priority sequence of the workstations based on the sorting results. The sorting rule is as follows: arrange the random real numbers in the workstation access sequence vector in ascending order. The workstation with the smaller value has a higher access priority in the scheduling sequence.

[0066] S503. Traverse the workstation assignment center vector and the workstation vehicle allocation vector, and determine the distribution center to which each workstation belongs and the matched AGV vehicle through numerical interval mapping rules; wherein, the numerical interval mapping rules include:

[0067] For the workstation assignment center vector, the numerical range [0,1] is divided into several equal intervals corresponding to the number of distribution centers, and the distribution center number assigned to the workstation is determined according to the interval into which the particle value falls.

[0068] For the vehicle type allocation vector at the workstation, the numerical range [0,1] is divided into several intervals corresponding to the number of heterogeneous vehicle types. The light-load AGV or heavy-load AGV allocated to the workstation is determined according to the interval into which the particle value falls.

[0069] S504. Based on the established workstation resource ownership and access priority sequence, workstations are allocated to specific AGVs, and a semi-open strategy is used to construct the specific driving path for each AGV; wherein, the process of constructing the specific driving path using the semi-open strategy includes:

[0070] Workstations assigned to the same AGV are sorted according to access priority to form intermediate task path segments. The first and last service workstations of each path segment are determined. The distances between the first service workstation and all distribution centers are calculated, and the nearest distribution center is selected as the AGV's departure node. The distances between the last service workstation and all distribution centers are calculated, and the nearest distribution center is selected as the AGV's return node. The departure node, intermediate task path segments, and return node are connected sequentially to form a complete semi-open delivery path.

[0071] S505 converts the constructed driving path into a multi-AGV scheduling instruction containing navigation coordinates, action type and time parameters, and sends it to each heterogeneous AGV for execution via wireless network.

[0072] The present invention also provides a multi-AGV intelligent scheduling optimization device, based on the multi-AGV intelligent scheduling optimization method described above, the device comprising:

[0073] The acquisition module is used to acquire multi-dimensional logistics basic data of the production workshop, including workstation task parameters, layout parameters, and fleet parameters. Among them, the workstation task parameters include workstation coordinates and operation time windows, the layout parameters are the location coordinates of each distribution center, and the fleet parameters include the load limit and running speed of each heterogeneous AGV.

[0074] The first module is used to establish an objective function for a vehicle route planning model based on the multi-dimensional logistics basic data. The objective function aims to minimize the total delivery cost, which consists of vehicle fixed and driving costs and time window penalty costs. The vehicle fixed and driving costs are calculated based on the fleet parameters, and the time window penalty costs are calculated based on the workstation task parameters.

[0075] The second module is used to establish the constraints of the vehicle route planning model based on the multidimensional logistics basic data, including semi-open route and vehicle scheduling constraints, heterogeneous vehicle capacity and simultaneous collection and distribution constraints, and mixed time window constraints.

[0076] The solution module is used to solve the vehicle path planning model using a multi-subgroup speed-pause particle swarm optimization algorithm that combines large neighborhood search. The multi-subgroup speed-pause particle swarm optimization algorithm is an improvement on the standard particle swarm algorithm, including using Logistic chaotic mapping to initialize the population distribution, introducing a speed-pause mechanism in the iterative update, adopting a multi-subgroup cooperative strategy to share global optimal information, and incorporating a large neighborhood search mechanism.

[0077] The output module is used to obtain the global optimal solution obtained by the multi-subgroup velocity-pause particle swarm optimization algorithm, decode the global optimal solution to obtain a scheduling scheme, and convert it into a multi-AGV scheduling instruction to be issued to each heterogeneous AGV; wherein, the scheduling scheme includes the optimal workstation access order corresponding to the workstation task parameters, the workstation assignment center determined from the layout parameters, and the vehicle type allocation matched from the fleet parameters.

[0078] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0079] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0080] The beneficial effects of this invention are as follows:

[0081] This invention first introduces a hybrid time window constraint, setting hard time windows for critical workstations such as precision machining to ensure stringent timeliness, and setting soft time windows for general workstations such as material buffering to improve scheduling flexibility, effectively balancing production rhythm and resource utilization. Secondly, it adopts a simultaneous delivery and distribution mechanism, enabling AGVs to simultaneously complete raw material delivery and finished product recovery along the same path, dynamically balancing vehicle load and significantly reducing single-trip empty load rate and total travel distance compared to the traditional separate mode. Furthermore, this invention innovatively designs a semi-open multi-center scheduling strategy, allowing AGVs to return to the nearest distribution center after completing their tasks instead of necessarily returning to the origin, greatly reducing invalid return trips and lowering transportation costs. Finally, it employs a multi-subgroup velocity-pause particle swarm optimization algorithm combined with large neighborhood search (MSVPPSO-LNS). Through the velocity-pause mechanism, multi-subgroup cooperation, and large neighborhood destruction-repair strategy, it effectively overcomes the premature convergence defects of traditional algorithms, achieving globally optimal scheduling of heterogeneous vehicle fleets under complex constraints, significantly improving the system's solution quality and response speed. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of AGV logistics transportation in the production workshop of multiple distribution centers in this invention.

[0083] Figure 2 This is a flowchart illustrating the intelligent scheduling and optimization method for multiple AGVs of the present invention.

[0084] Figure 3 This is a schematic diagram of the particle coding structure in this invention.

[0085] Figure 4 This is an example diagram of particle encoding in this invention.

[0086] Figure 5 This is a schematic diagram of the decoding process in this invention.

[0087] Figure 6 This is a spatial point distribution diagram of the Logistic model in this invention.

[0088] Figure 7 This is the Logistic frequency histogram in this invention.

[0089] Figure 8 This is a schematic diagram of the damage-repair operation in this invention.

[0090] Figure 9 This is a flowchart of the MSVPPSO-LNS algorithm in this invention.

[0091] Figure 10 This is a graph showing the algorithm iteration curve in this invention.

[0092] Figure 11 This is the optimal delivery route map in this invention.

[0093] Figure 12 This is a schematic diagram of the intelligent scheduling and optimization device for AGVs of the present invention.

[0094] Figure 13 This is a schematic diagram of the internal structure of the computer device in this invention.

[0095] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0096] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0097] This invention introduces simultaneous cargo collection and distribution in a multi-center scheduling mode to reduce empty loads, adopts heterogeneous fleets to improve resource adaptability, and combines a semi-open collaborative delivery strategy to enhance scheduling flexibility, thereby improving the efficiency and resource utilization of the AGV system. The MDHOHVRPSPDTW studied in this invention is a comprehensive extension of the classic VRP (route planning) problem. It organically combines multiple distribution centers, heterogeneous vehicle structures, time windows, and simultaneous pickup and delivery constraints. It is mainly developed from the following issues: (1) MDVRP refers to vehicles starting from multiple distribution centers, completing delivery tasks for customers, and returning to the starting center. In the semi-open mode, some vehicles do not need to return to the starting center and can directly stay at the nearest dispatch center after completing the task; (2) MDVRPTW introduces time window constraints on the basis of multiple centers, requiring vehicles to complete services according to the customer's preset time, thereby improving the timeliness and satisfaction of the service; (3) The HVRP model considers that the fleet consists of various vehicle types, and the vehicles have differences in capacity, cost, etc., making the dispatch system closer to actual applications. For example, in the production workshop, light AGVs are suitable for high-frequency short-distance transportation, while heavy AGVs are suitable for long-distance or heavy-load tasks. Therefore, in route planning, it is necessary to reasonably allocate vehicle types to improve resource utilization and service quality. Based on MDVRP, MDVRPTW and HVRP, this invention considers the path decision under the constraints of vehicle execution of pickup and delivery tasks and time window service, and constructs the path optimization problem in the scenario of semi-open collaborative delivery and heterogeneous fleets picking up and delivering goods simultaneously with time windows, namely the MDHOHVRPSPDTW model, which provides theoretical support and methodological basis for AGV scheduling and path planning in multi-center production workshops.

[0098] Multi-distribution-center production workshops typically consist of multiple distribution centers, multiple workstations, different types of AGVs, and a central dispatching system. The dispatching center formulates daily delivery task plans based on the material requirements of each workstation and rationally allocates tasks to different distribution centers. Simultaneously, it needs to coordinate the operation plans of various AGV types and optimize vehicle scheduling to improve resource utilization efficiency. This invention focuses on a semi-open delivery mode under the collaborative operation of multiple distribution centers and heterogeneous AGV fleets. In this mode, AGV vehicles no longer return to the original distribution center; instead, they can choose the nearest distribution center to complete replenishment, charging, or task completion operations based on actual operating status and delivery needs. This scheduling strategy can effectively adapt to dynamic and changing production logistics environments, reduce vehicle empty-running rates, and improve system response flexibility, thereby significantly reducing total operating costs while ensuring task completion rates.

[0099] Based on the aforementioned characteristics of logistics and distribution, this invention establishes a path planning mathematical model (MDHOHVRPSPDTW) with the objective of minimizing the sum of vehicle fixed costs, transportation costs, and late arrival time window penalty costs at workstations. The key to the model lies in achieving collaborative delivery among multiple distribution centers, while simultaneously considering the unified allocation of heterogeneous AGV vehicle resources to meet the constraints of workstations simultaneously having pickup and delivery needs and strictly adhering to time windows. Specifically, it can be described as follows: given the clear delivery needs of multiple distribution centers and each workstation, the distribution centers cooperate and flexibly allocate vehicle resources to optimize and determine a series of efficient and reasonable delivery routes while meeting vehicle load and workstation requirements, thereby minimizing the total delivery cost. Furthermore, to ensure the correct model construction and the smoothness and controllability of delivery operations, it is stipulated that each workstation demand point can only be served by one AGV, and each delivery route can also be completed by only one AGV. Specific delivery modes can be found in [reference needed]. Figure 1 As shown.

[0100] This invention addresses the problem of simultaneous distribution of goods by semi-open heterogeneous AGVs in a multi-center production workshop. Considering multi-center joint distribution, vehicle load capacity, simultaneous pickup and delivery, and time windows, a reasonable distribution scheme is planned for manufacturing enterprises. Because there are too many practical factors to consider in actual distribution, the following assumptions are made for the convenience of model construction: (1) No downtime, malfunctions, or AGV collisions will occur in the workshop; (2) Multiple distribution centers will distribute goods simultaneously. AGVs will start from one distribution center, complete the delivery and pickup tasks at all workstations along the path, and then return to the nearest distribution center; (3) The geographical location, workstation demand, workstation pickup volume, time window, and service time of each workstation are known and will not change; the demand for each workstation cannot be divided and can only be served by one AGV, but each AGV can serve multiple workstations; (4) The maximum load capacity, speed, and vehicle type of different types of AGVs are known and remain constant; (5) Time-varying load and energy consumption due to overcoming frictional resistance, air resistance, and slope resistance are ignored. (6) All vehicles originate and terminate at the distribution center, and vehicles must not be overloaded during delivery. (7) The speed of the AGV when turning and going straight is a constant.

[0101] like Figure 2 As shown, this invention provides an intelligent scheduling and optimization method for multiple AGVs, including:

[0102] S1. Obtain multi-dimensional logistics basic data of the production workshop, including workstation task parameters, layout parameters, and fleet parameters; wherein, the workstation task parameters include workstation coordinates and operation time windows, the layout parameters are the location coordinates of each distribution center, and the fleet parameters include the load limit and running speed of each heterogeneous AGV.

[0103] In this embodiment, the multidimensional logistics basic data is obtained through a shop floor manufacturing execution system (MES) or a warehouse management system (WMS). Step S1 specifically includes the following sub-steps:

[0104] S101. First, determine all workstations within the production workshop that require material delivery or recycling, denoted as the set of workstations C = {1, 2, ..., n}. For each workstation i (i ∈ C) in the set, obtain the workstation coordinates and the operation time window, specifically:

[0105] Workstation coordinates: Obtain the coordinates of workstation i in the workshop plane coordinate system. This is used for subsequent distance calculations. Since a simultaneous delivery and collection model is adopted, the delivery demand for workstation i needs to be obtained separately. (i.e., raw material demand) and pickup demand (i.e., finished product recycling volume). According to the model assumptions, the demand at each workstation is indivisible and can only be served by one AGV. Simultaneously, the service time required for the AGV to perform loading and unloading operations at workstation i is obtained. .

[0106] Operation Time Window: Set the optimal service time window for work station i. ,in To be the earliest permitted start time for service, This is the latest allowed start time for service. Based on the urgency of the production process, time windows are divided into hard time windows and soft time windows. Hard time windows target critical workstations (such as precision machining equipment), requiring the AGV to arrive within the window; otherwise, the task is considered a failure, with no penalty cost calculation, and it is directly treated as a constraint. Soft time windows target general workstations (such as material buffer zones), allowing the AGV to arrive outside the window, but the time window penalty cost must be calculated. Therefore, a delay cost penalty coefficient per unit time also needs to be set. .

[0107] S102. Determine the layout of multiple distribution centers within the production workshop, denoted as D. For each distribution center d (d∈D) in the set, obtain its coordinates in the workshop's planar coordinate system. Based on the coordinate information of the work site set C and the distribution center set D, calculate the Euclidean distance between the nodes. , Let represent the straight-line distance between node i and node j, where i,j∈{C∪D}. Based on the model assumptions, the speed difference of the AGV when turning and traveling straight is ignored, and the path distance is directly used as the basis for calculating transportation costs and time.

[0108] S103. This embodiment involves heterogeneous fleet scheduling, where the fleet contains different types of AGVs (e.g., light-duty AGVs and heavy-duty AGVs). Let R be the set of AGV types, and the specific set of available AGVs be... For each type r (r∈R) of AGV, the maximum load capacity and operating speed are obtained as follows:

[0109] Maximum load capacity: Get the maximum load capacity of this type of AGV. This serves as an important basis for subsequent capacity constraints on heterogeneous vehicles, ensuring that the vehicle load (initial load + pickup quantity - delivery quantity) does not exceed this upper limit during the delivery process.

[0110] Operating speed: Obtain the average travel speed v of this type of AGV, assuming the AGV travels at a constant speed, to calculate the travel time between nodes. .

[0111] In addition, obtain the total number of this type of AGV available in the workshop. And cost parameters, available total quantity Used to limit the maximum usage of this type of vehicle in the scheduling scheme; cost parameters include fixed usage costs. and transportation costs Fixed usage cost This refers to the startup / depreciation costs incurred when this type of AGV is activated (i.e., the vehicle is assigned a task); and the operating and transportation costs. This refers to the variable cost (such as energy consumption cost) generated per unit travel distance of this type of AGV. Generally, the unit travel cost of heavy-duty AGVs is higher than that of light-duty AGVs.

[0112] S2. Establish an objective function for the vehicle route planning model based on the multi-dimensional logistics basic data. The objective function aims to minimize the total delivery cost, which consists of vehicle fixed and driving costs and time window penalty costs. The vehicle fixed and driving costs are calculated based on the fleet parameters, and the time window penalty costs are calculated based on the workstation task parameters.

[0113] In this example, to evaluate the merits of different AGV scheduling schemes, a mathematical model (MDHOHVRPSPDTW) is constructed with the goal of minimizing the total delivery cost. This objective function minZ mainly consists of three cost components: vehicle fixed usage cost, vehicle transportation cost, and time window penalty cost. Step S2 specifically includes the following sub-steps:

[0114] S201. Construct the vehicle fixed-use cost function.

[0115] Vehicle fixed costs refer to the inherent expenses (such as startup energy consumption, depreciation, etc.) incurred once a particular AGV is put into operation. These costs are independent of the driving distance but related to the vehicle model. Based on the fleet parameters obtained in step S1, the following definition is established. For the fixed dispatch cost of the r-th type AGV, Let $\frac{ ... This represents the sum of the fixed costs of all activated AGVs, calculated using the following formula:

[0116]

[0117] Where R is the set of AGV types, Let r be the set of AGVs of class r. This part aims to guide the algorithm to minimize unnecessary vehicle deployment and improve vehicle loading rate.

[0118] S202. Construct the vehicle transportation cost function.

[0119] Vehicle operating costs are the variable costs (such as electricity consumption) incurred by AGVs during task execution, directly related to the travel distance and the energy consumption characteristics of the vehicle model. Based on the fleet parameters and layout parameters obtained in step S1, the following is defined: Let r be the unit distance travel cost of the r-th type AGV. Let be the Euclidean distance from node i to node j. Let $\mathbf{k}$ be the decision variable (equivalent to 1 if the $k$-th AGV in class $r$ travels from node $i$ to node $j$, otherwise 0). Then, the vehicle travel cost is... It is expressed as the sum of the products of the total mileage traveled by all AGVs and the corresponding unit cost. The specific calculation formula is as follows:

[0120]

[0121] Here, A represents the set of arcs formed by all nodes (workstations and distribution centers). This part aims to guide the algorithm in planning the shortest or most efficient travel path.

[0122] S203. Construct the time window penalty cost function.

[0123] In terms of penalty term design, MDHOHVRPSPDTW is a typical multi-constraint, multi-objective combinatorial optimization problem. Therefore, a penalty term mechanism is introduced to impose additional costs on delivery schemes that violate time window constraints, thereby increasing their total delivery cost. The time window penalty cost is set for general workstations (soft time windows) and is used to quantify the potential production losses caused by the AGV not arriving within the optimal time window. Based on the workstation task parameters obtained in step S1, the following is defined: The penalty coefficient for delay per unit time. The time when the AGV arrives at workstation i and begins service. (or Let be the latest allowed start time for workstation i (i.e., the upper bound of the soft time window). In this embodiment, penalties are only imposed on late arrivals (allowing vehicles to arrive early and wait, with penalties only imposed on late arrivals). The time window penalty cost is... It is expressed as the sum of the products of the delay time of all workstations and the penalty coefficient. The specific calculation formula is as follows:

[0124]

[0125] Where C is the set of work sites, This represents the time window violation amount for workstation i. If the AGV arrives on time or early, this item is 0; if it arrives late, the difference is calculated. This section aims to meet production timeliness requirements and balance delivery efficiency with on-time performance.

[0126] S204. Establish the objective function Z that minimizes the total delivery cost.

[0127] Combining the costs from the three components mentioned above, a final objective function for the vehicle routing model is established. This function is expressed as a weighted sum, and the goal of model optimization is to find a set of decision variables (vehicle allocation scheme y and route selection scheme x) that minimizes the total cost Z.

[0128]

[0129] Right now:

[0130]

[0131] Through this objective function, the model can find the optimal balance between reducing vehicle input, shortening travel paths, and reducing delay risks, thereby achieving global optimization of multi-AGV scheduling.

[0132] S3. Based on the multidimensional logistics basic data, establish the constraints of the vehicle route planning model, including semi-open route and vehicle scheduling constraints, heterogeneous vehicle capacity and simultaneous collection and distribution constraints, and mixed time window constraints.

[0133] In this example, to ensure that the generated scheduling scheme meets the physical constraints and process requirements of the production workshop, it is necessary to set a strict set of constraints in the model based on the basic data obtained in step S1. Step S3 specifically includes the following sub-steps:

[0134] S301. Establish semi-open path and vehicle scheduling constraints.

[0135] This constraint aims to enable flexible vehicle scheduling in multi-distribution-center scenarios, ensuring route connectivity and rational resource allocation.

[0136] 1) Vehicle activation and route association constraints: Ensure that AGVs can only be assigned routes from the distribution center after they are activated.

[0137]

[0138] Where D is the set of distribution centers and C is the set of work sites; For decision variables, the value is 1 if the kth AGV of the rth class travels from the distribution center d to the workstation j, and 0 otherwise; This is a decision variable; it is 1 if the vehicle is enabled, and 0 otherwise.

[0139] 2) Available vehicle quantity constraint: Ensure that the number of each type of AGV used does not exceed the maximum available quantity in the workshop.

[0140]

[0141] in, Let be the total number of available AGVs of type r.

[0142] 3) Unique service constraint: Ensure that each work site is served by only one AGV once.

[0143]

[0144] Where V represents the set of all nodes (including workstations and distribution centers). This indicates that the vehicle travels from node i to node j.

[0145] 4) Flow balance constraint: Ensure that after the AGV arrives at a certain work point, it must leave from that point to ensure the continuity of the path.

[0146]

[0147] 5) Semi-open start and stop constraints

[0148] Departure constraint: Each AGV must depart from a specific distribution center.

[0149]

[0150] Return constraint: Each AGV must return to a specific distribution center after completing its task (returns to different locations are allowed).

[0151]

[0152] No-empty-running constraint: AGVs are prohibited from traveling directly between distribution centers.

[0153]

[0154] 6) Eliminate sub-loop constraints: Prevent isolated loops in the path that do not pass through the distribution center.

[0155]

[0156] Where S is any subset of the workstation set C, and |S| is the number of elements in the subset.

[0157] S302. Establish heterogeneous vehicle capacity and simultaneous cargo collection and distribution constraints.

[0158] 1) Initial load definition: Define the load capacity of the AGV when it departs from the distribution center.

[0159]

[0160] in, The load of the vehicle when it leaves the distribution center d. The delivery demand for workstation j.

[0161] 2) Dynamic balance constraint of vehicle load: Update the load changes between nodes according to delivery and pickup needs.

[0162]

[0163] in, , These represent the cargo loads when leaving nodes j and i, respectively. Let M be the required quantity of goods to be picked up at workstation j; M is a maximum positive number; and A is the set of all arc segments.

[0164] 3) Capacity limit constraint: Ensure that the load does not exceed the limit and is non-negative.

[0165]

[0166]

[0167] in, This represents the maximum rated load capacity of the r-th type AGV.

[0168] S303. Establish mixed time window constraints and variable definitions.

[0169] 1) Service time continuity constraint: Ensure that the time when the AGV arrives at the next workstation conforms to the physical logic.

[0170]

[0171] in, , These represent the times when the AGV arrives at workstations j and i, respectively. Service duration for workstation i; Let M be the travel time from node i to j; M is a maximum positive number.

[0172] 2) Soft time window delay constraint: Define the time window delay variable This is used for subsequent calculations of penalty costs.

[0173]

[0174] in, (or ) is the upper bound of the optimal service time window (soft time window cutoff time) for workstation i; if ,but It is at least the difference between the two.

[0175] Hard time window constraint: For nodes defined as critical workstations, the AGV is required to be within the time window. Arrive inside, if If the path is invalid (infeasible solution), then the path is invalid, i.e.:

[0176]

[0177] in, This refers to the set of nodes defined as critical workstations.

[0178] 3) Initial time constraint: The time when the AGV departs from the distribution center is set to 0.

[0179]

[0180] 4) Definition of decision variables and non-negativity constraints: Clearly define the range of values ​​for the variables, and combine them with... Ensure that delay time is non-negative (i.e., there is no penalty for arriving early).

[0181]

[0182] S4. The vehicle path planning model is solved using a multi-subgroup speed-pausing particle swarm optimization algorithm that combines large neighborhood search. The multi-subgroup speed-pausing particle swarm optimization algorithm is an improvement on the standard particle swarm algorithm, including using Logistic chaotic mapping to initialize the population distribution, introducing a speed-pausing mechanism in the iterative update, adopting a multi-subgroup cooperative strategy to share global optimal information, and incorporating a large neighborhood search mechanism.

[0183] In this example, MDHOHVRPSPDTW is a typical NP-hard problem (complex and computationally difficult problem). Due to its large scale and multiple constraints, a metaheuristic algorithm must be used to obtain a high-quality approximate optimal solution within a finite time. Metaheuristic algorithms, as an effective method for solving complex optimization problems, have the advantage of not depending on the characteristics of the problem. Particle Swarm Optimization (PSO) has become one of the most widely used metaheuristic algorithms due to its strong robustness, efficiency, and simplicity. PSO performs well in optimization problems in many fields such as wireless communication, artificial intelligence, prestressed design, and image segmentation. However, PSO often faces the problem of premature convergence during the solution process. To overcome this problem, a velocity pause mechanism and a multi-subgroup cooperative strategy are introduced, proposing an improved particle swarm algorithm. Furthermore, a large neighborhood search algorithm is combined in the local search phase to further enhance the algorithm's ability to escape local optima. Finally, a multi-subgroup velocity pause particle swarm optimization algorithm combining large neighborhood search (MSVPPSO-LNS) is proposed. MSVPPSO-LNS, based on the characteristics of MDHOHVRPSPDTW, has made the following three optimizations and improvements to the standard PSO algorithm:

[0184] (1) Velocity Pause Mechanism: To address the shortcoming of traditional PSO algorithms that are prone to getting trapped in local optima as the number of iterations increases, a velocity pause strategy is adopted in the algorithm design. When updating the particle velocity, the particle is allowed to retain the velocity of the previous iteration with a certain probability. Each particle can not only move towards its own optimal position and the group's optimal position as in traditional PSO, but also explore unknown areas at its original velocity, which effectively improves the algorithm's exploration ability in the search space and its global search efficiency.

[0185] (2) Multi-subgroup cooperative strategy: Drawing on the idea of ​​multi-subgroup cooperative particle swarm optimization, the entire population is divided into multiple subgroups. Each subgroup performs optimization search independently, but in each iteration, the subgroups share the optimal particle information, forming effective cooperation. This strategy effectively improves the algorithm's search space coverage and global exploration capability, and reduces the possibility of the algorithm getting stuck in local optima.

[0186] (3) Chaotic mapping initialization and large neighborhood search: The algorithm uses Logistic chaotic mapping to generate a diverse initial population to enhance the diversity of the initial population and make the particle distribution more uniform; in the large neighborhood search part, three neighborhood operators are used to perform "destruction-repair" operation to increase the quality of the solution and avoid premature convergence.

[0187] Step S4 specifically includes the following sub-steps:

[0188] S401, encoding, decoding and population initialization.

[0189] (1) Encoding and Decoding

[0190] The PSO algorithm is primarily used to solve continuous linear optimization problems and cannot be directly applied to discrete optimization models. Therefore, it is necessary to construct a mapping relationship from random values ​​between [0,1] to the solution space of discrete optimization problems to meet the needs of problem solving. To better accommodate complex constraints such as pickup and delivery scenarios, heterogeneous AGVs, and time windows, a three-segment coding structure is adopted to represent the workstation access order, workstation assignment center, and vehicle type allocation scheme using random real numbers.

[0191] like Figure 3 As shown, for each particle Define a 3×N dimension (N is the number of workstations), where the first N dimensions represent the workstation assignment center (first segment of code), with a value range of [0,1], and the distribution center to which the workstation belongs is determined through interval mapping; the [N+1,2N] dimension represents the workstation AGV allocation (workstation vehicle type allocation, second segment of code), with a value range of [0,1], and the type of AGV serving the workstation is determined through interval mapping; the last N dimensions represent the workstation access order (third segment of code), with a value range of [0,1], and are sorted by value (e.g., ascending order), with smaller values ​​representing higher access priority, thus determining the service order of the workstations.

[0192] ① Encoding method

[0193] There are N workstations and M distribution centers, each particle having 3N dimensions. The first dimension [1, N] represents the distribution center assigned to the workstation. Initially, a real number is randomly generated within [0, 1]. When the dimension value is between 0 and 0.25, the workstation is assigned to distribution center 1, and so on. The second dimension [N+1, 2N] represents the vehicle type assigned to the workstation, also taking a random real number between [0, 1]. When the value is between 0 and 0.5, a light AGV is selected for that workstation; when the value is between 0.5 and 1.0, a heavy AGV is selected. The last N dimensions determine the access order of the workstations. Each dimension is also randomly generated between 0 and 1, with smaller values ​​indicating higher access priority. These dimension values ​​are sorted from smallest to largest to determine the final access order. Example (N=3, M=2) is shown below. Figure 4 As shown.

[0194] ②Decoding method

[0195] The decoding process mainly includes three steps: a. determining the workstation access order; b. assigning the workstation to the distribution center and vehicle type; c. constructing the delivery route and verifying constraints. First, the access order of the workstations is determined by sorting the random numbers [0,1] generated in the last N dimensions of the particle encoding in ascending order. Since the problem involves simultaneous order fulfillment, each workstation only needs to be accessed once. Next, based on the N dimensions of the particle encoding to determine the assigned distribution center of the workstation and the middle N dimensions to determine the assigned vehicle type, the distribution center and assigned AGV vehicle type of each workstation are determined. This forms the initial allocation structure of "workstation-distribution center-vehicle type". Finally, the workstations are inserted one by one into the path corresponding to their assigned distribution center and vehicle type according to the access order, while simultaneously verifying constraints such as vehicle load, time window, and fleet size. If the constraints cannot be met after insertion, a local reconstruction mechanism is used, such as switching vehicle types, adjusting distribution centers, or rearranging the access order, to generate an efficient delivery plan that meets the feasibility conditions. The specific route construction is as follows: Figure 5 As shown.

[0196] like Figure 5 As shown, firstly, the set of workstations served by each AGV is determined based on the vehicle information. The workstation set served by heavy-duty AGV1 is {1,4,5}, the workstation set served by light-duty AGV2 is {2,3,6}, and the workstation set served by heavy-duty AGV3 is {7,8}. Then, the workstation sets of each AGV are sorted according to the service order to generate the final service order. After sorting, the service order of heavy-duty AGV1 is 5-1-4, the service order of light-duty AGV2 is 3-2-6, and the service order of heavy-duty AGV3 is 8-7. This completes the first stage of decoding. In the second stage of decoding, based on the positions of the first and last workstations of the AGV, the nearest distribution center is selected as the start and end node for judgment. First, the nearest distribution center A is selected for the first workstation of vehicle 1, and it is determined whether the center has available delivery capacity. If the condition is met, the starting point of the AGV's delivery path is determined. Subsequently, the final destination delivery center B corresponding to the last workstation of AGV1 is determined using the same method and incorporated into the complete path, ultimately forming the travel path A-5-1-4-B for the heavy-duty AGV1. Similarly, the starting and ending delivery centers for the light-duty AGV2 and AGV3 are determined sequentially according to the nearest principle, resulting in travel paths B-3-2-6-A and C-8-7-C, respectively. Finally, the delivery paths for all AGVs are fully generated, ensuring that workstation requirements are met. This process continues until all workstations have completed delivery allocation, ultimately determining the most cost-effective delivery solution.

[0197] (2) Population initialization

[0198] To improve the quality of initial solutions and search diversity in the PSO algorithm, a Logistic chaotic mapping mechanism is introduced based on random initialization and the nearest neighbor greedy method to enhance the distribution characteristics and coverage of the initial population. After random initialization of vehicle location and service order vectors, the solution vectors are first normalized to map them to [0,1]. Then, a perturbation is introduced through the Logistic mapping to improve the uniformity of the initial population distribution and increase the coverage of the search space. The mathematical expression of the Logistic mapping is as follows:

[0199]

[0200] in, Let X(0) represent the normalized variable of the nth iteration, which is a chaos control parameter. Its value is typically in the range [3, 57, 4] to maintain the ergodicity of the chaotic sequence. The generated chaotic sequence is mapped to the particle's value space to obtain uniformly distributed initial positions X(0) and velocities V(0). By generating a uniformly distributed sequence with chaotic characteristics within the interval [0, 1] and converting the mapping result back to the actual range of vehicle positions and service order vectors, the search coverage can be significantly expanded during the PSO initialization phase, reducing the risk of getting trapped in local optima, thereby improving the algorithm's global search efficiency and solution quality.

[0201] In the multi-center heterogeneous AGV path optimization problem, the solution space is irregular and the path distribution is complex. Improper initialization can easily cause the population to become trapped in concentrated regions, affecting the algorithm's convergence speed and the efficiency of exploring feasible solutions. This method not only overcomes the problem of decreased search quality caused by uneven initial population distribution but also provides more diverse and stable initial solutions for subsequent speed updates and global optimal positioning. This provides a solid foundation for improving the overall performance of the MSVPPSO-LNS algorithm. The visualization of the relevant initialization results is shown below. Figure 6 and Figure 7 As shown.

[0202] S402, Divide the population into subgroups and calculate the initial fitness.

[0203] The initial particle population is randomly divided into K subgroups. Each subgroup performs independent search and evolution, but shares the global optimal solution gbest.

[0204] Since the decoded delivery plan may not be feasible, a delivery plan that heavily penalizes violations of constraints is used to provide the driving path. This avoids getting trapped in local optima and increases the optimization capability of the MSVPPSO-LNS algorithm. Therefore, the starting cost of the delivery plan consists of three parts: fixed cost, travel cost, and penalty cost. The specific fitness value formula is: f(x) = 1 / Z, where f(x) represents the fitness value of particle x, reflecting the quality of its solution; Z is the particle's overall cost, including fixed cost, travel distance, and penalties imposed by constraints such as time window. During the algorithm's iteration process, a larger fitness value f(x) indicates a smaller overall cost Z for the particle, and its corresponding delivery plan is more advantageous in the population, thus more likely to be selected as the global optimum.

[0205] According to the encoding rules in step S401, the position vector of each particle is decoded into a specific vehicle path scheme. The path scheme is then substituted into the objective function constructed in step S2 and the constraints constructed in step S3. If the hard time window or load constraint is violated, a large penalty value is applied; otherwise, the total delivery cost Z is calculated. The fitness function f(x) = 1 / Z is defined, where higher fitness indicates a better solution. The individual historical optimal solution Pbest of each particle, the local optimal solution of each subgroup, and the global optimal solution gbest of the entire population are updated.

[0206] S403, Execute speed update strategy.

[0207] Traditional particle swarm optimization (PSO) algorithms often suffer from insufficient population diversity and susceptibility to local optima when dealing with complex optimization problems. To improve the overall optimization capability of the algorithm in complex path scheduling scenarios, this paper proposes an improved PSO algorithm that integrates a speed pause update mechanism and a multi-subgroup cooperative strategy. The former enhances the local search depth and the ability to escape local optima traps by introducing a probabilistic pause update and local perturbation mechanism; the latter improves the overall search breadth and diversity of the population by dividing the population into multiple cooperative subgroups, which, while maintaining their independent search capabilities, achieve global cooperative evolution through information sharing.

[0208] The velocity pause mechanism enhances the search capability of the PSO algorithm. Based on the traditional PSO velocity update rule, this mechanism introduces the option of maintaining the velocity unchanged from the previous time step with probability 'a', rather than being limited to acceleration or deceleration updates. Its mathematical expression is as follows:

[0209]

[0210] Where 'a' is the probability of speed stopping. A random number taking values ​​in [0,1], where Pbest and gbest represent the historical best and global best positions of the particle, respectively; Used to balance the update magnitude of particles at different times, it is defined as follows:

[0211]

[0212] Where t represents the current iteration number, and T is the maximum iteration number. Through the aforementioned velocity pause mechanism, in each iteration, the particles move with probability... By maintaining the original speed, a "constant speed" mode has been added in addition to the acceleration and deceleration modes, improving the algorithm's adaptability to complex search spaces.

[0213] Furthermore, the velocity-pause particle swarm algorithm employs a two-group structure of "main group-subgroup" to enhance global exploration capabilities: the main group particles update their velocity according to the following formula, while the subgroup particles undergo random perturbation around the global optimum, as shown in the following equation:

[0214]

[0215] in, Using random numbers in the [0,1] range, this method can guide particles to continuously try to escape local traps and promote global search in the later stages of the search.

[0216] To enhance the algorithm's global exploration capability for complex optimization problems over a wider range, the entire particle swarm is divided into several subgroups. Each subgroup independently searches within a local area, seeking its own local optimum. This allows different subgroups to explore multiple regions simultaneously, increasing the overall chance of escaping local optima. Each subgroup independently executes a velocity pause update mechanism. In the multi-subgroup cooperative strategy, particles in each subgroup evolve independently during the update process, but all subgroups share a unique global optimum (gbest). The specific update formula is defined as follows:

[0217]

[0218]

[0219] In the formula, the inertial weight w is defined as:

[0220]

[0221] The inertia weight ranges from [0.4, 0.9]. The above multi-subgroup cooperation strategy enables the population to maintain diversity during the overall search process. At the same time, through the sharing mechanism of the global optimal solution, information exchange and cooperation between subgroups are realized, which effectively avoids the algorithm from getting stuck in local optima and enhances the overall optimization efficiency of the algorithm.

[0222] In summary, during the iteration process, the velocity and position of particles in each subgroup are updated. To prevent premature convergence of the algorithm, a velocity pause mechanism is introduced. Specifically,

[0223] Set pause probability Before each update, generate a random number rand. The particles perform a "pause" operation, that is, retain the velocity of the previous generation. This allows the particles to maintain their current inertial direction of exploration, increasing the diversity of the search. If The particle performs a standard PSO update, which involves learning from the individual best (Pbest) and the globally best (gbest); the new particle position is calculated based on the updated velocity. .

[0224] S404. Perform a large neighborhood search on the global optimal solution.

[0225] Large Neighborhood Search (LNS) is employed to enhance the algorithm's local search capabilities. LNS is based on the core idea of ​​"Destroy and Repair." Its fundamental principle is to selectively remove some workstation nodes through destruction operations, thereby disrupting the original path structure and introducing significant disturbances. Subsequently, repair operations are used to reinsert the removed workstation nodes into the solution, minimizing the insertion cost and achieving solution reconstruction and optimization. This process not only expands the search space but also guides the solution towards better regions while preserving some high-quality path structures, effectively improving the algorithm's ability to escape local optima and enhancing its global optimization performance.

[0226] For path construction considering simultaneous pickup and delivery by multi-center heterogeneous AGVs, given that AGVs need to perform pickup and delivery tasks simultaneously, LNS dynamically adjusts the path structure through a "destruction-repair" approach, effectively addressing the complexity arising from task allocation and path coupling. Under the premise of satisfying constraints such as capacity and time windows, LNS allows for the reconstruction of part of the path structure of the current solution, thereby exploring more high-quality feasible solutions in the solution space. This mechanism not only enhances the algorithm's ability to escape local optima but also provides an effective supplement to overcoming the premature convergence problem faced by differential evolution or other global search strategies in complex environments.

[0227] The large neighborhood algorithm consists of three neighborhood operators, each of which includes two parts: destruction and repair.

[0228] (1) Destruction process

[0229] For the AGV path solution set of the MDHOHVRPTW problem in the production workshop, at the end of each iteration, a large neighborhood search "destruction" operation is performed on the current global optimal solution gbest, removing q workstation nodes from the current path. Three removal strategies are adopted: random removal, worst removal, and similarity removal.

[0230] Random removal refers to randomly selecting several workstations from the current solution for elimination, regardless of workstation location or demand attributes, in order to maintain the diversity of search solutions; worst-case removal sorts workstations according to their "contribution" to the total cost of the path, prioritizing the removal of workstations with the greatest impact on system cost, thereby quickly weakening the disadvantageous parts of the current solution; similarity removal strategy removes workstations in groups based on their "geographical proximity" or "material demand similarity", which helps to break the local clustering structure in the original path and provides more space for the re-optimization and scheduling of similar workstation groups.

[0231] (2) Repair process

[0232] During the reconstruction phase, the best insertion strategy is employed to process the workstations to be inserted sequentially. For the set of workstations removed during the destruction phase, without violating constraints such as time windows and load limits, each workstation is attempted to be inserted into all possible positions in the current solution, and the incremental change in the objective function caused by different insertion schemes is calculated. Then, the position with the smallest increment in the objective function is selected as the optimal insertion point for that workstation, and it is inserted at that position to minimize the impact of this round of reconstruction on the overall solution quality. This operation is applied sequentially to all workstations to be inserted until the entire reconstruction is completed, forming a new set of feasible path solutions. This strategy can effectively repair the structural integrity of the solution while controlling the disturbance amplitude, guiding the search to gradually converge towards a better solution region. A possible destruction-repair mechanism operation process is as follows: Figure 8 As shown. That is, for the set of workstations removed in S404, a "repair" operation is performed to re-insert them into the path. This involves traversing all removed workstations, attempting to insert them into all feasible positions in the current path, and calculating the increment of the objective function resulting from the insertion position. ;choose Insert the workstation at the smallest position until all workstations are rearranged to form a new complete solution; if the new solution is better than the original gbest, then update the global optimal solution.

[0233] S405, Iteration terminated.

[0234] The MSVPPSO-LNS algorithm sets two types of stopping criteria: First, a maximum iteration limit, meaning the algorithm terminates iteration and outputs the current optimal solution after reaching a preset maximum number of iterations (MaxIter). Second, a convergence accuracy limit, meaning the algorithm is considered to have converged prematurely and terminates iteration if the global optimal solution has not significantly improved within a certain number of consecutive iterations (the improvement is less than a set threshold). After termination, the global optimal solution is selected as the final solution output; otherwise, the algorithm returns to step S403 to continue iteration.

[0235] Based on step S4, to address the bottlenecks frequently encountered by traditional PSO algorithms in complex path planning problems, such as insufficient search space exploration and getting trapped in local optima, an improved particle swarm optimization algorithm is proposed: the improved particle swarm optimization algorithm combining large neighborhood search (MSVPPSO-LNS). The limitations of traditional PSO mainly stem from the uneven particle distribution during population initialization and the excessively rapid convergence of the particle swarm to local optima in the later stages of iteration, lacking an effective information exchange and cooperation mechanism, thus reducing global search capability and overall algorithm performance.

[0236] To address this issue, the MSVPPSO-LNS algorithm proposed in this embodiment is improved in the following aspects: First, a chaotic mapping strategy is adopted to improve the initialization process of the population, making the initial particles more evenly distributed in the solution space, which is conducive to early extensive exploration; second, a velocity pause mechanism is introduced in the particle position update stage to avoid premature particle convergence, thereby maintaining the diversity of the population; at the same time, an information sharing and cooperation mechanism among multiple subgroups is introduced to enhance the global search capability of the algorithm; finally, by combining a large neighborhood search strategy, the optimization efficiency of the algorithm in the local space is improved. The synergistic effect of these strategies effectively improves the global optimization accuracy and convergence speed of the algorithm. The specific implementation process of the MSVPPSO-LNS algorithm is as follows: Figure 9 As shown.

[0237] S5. Obtain the global optimal solution obtained by the multi-subgroup velocity-pause particle swarm optimization algorithm, decode the global optimal solution to obtain a scheduling scheme and convert it into a multi-AGV scheduling instruction to be issued to each heterogeneous AGV; wherein, the scheduling scheme includes the optimal workstation access order corresponding to the workstation task parameters, the workstation assignment center determined from the layout parameters, and the vehicle type allocation matched from the fleet parameters.

[0238] In this example, the globally optimal solution gbest output by the algorithm is a high-dimensional numerical vector, which needs to be converted into a physical scheduling scheme executable in the production environment through specific decoding rules. Step S5 specifically includes the following sub-steps:

[0239] S501. Obtain the converged global optimal particle gbest from step S4. This particle consists of real values ​​in 3N dimensions (N is the number of workstations). Divide it into three logical segments. The first segment (dimensions 1 to N) corresponds to the workstation assignment center vector, which is used to determine the delivery area to which each workstation belongs. The second segment (dimensions N+1 to 2N) corresponds to the workstation vehicle type allocation vector, which is used to determine the AGV type (such as light load or heavy load) serving each workstation. The third segment (dimensions 2N+1 to 3N) corresponds to the workstation access order vector, which is used to determine the service priority of the workstation.

[0240] S502. Extract the random real values ​​of the third segment of the particle (dimensions 2N+1 to 3N), sort them in ascending order, and the sorted workstation index sequence represents the globally optimal workstation access order. Workstations with smaller values ​​have higher service priority in the scheduling scheme and should be accessed first.

[0241] S503. Traverse the first segment of the vector (dimensions 1 to N), and determine the assigned delivery center for each workstation based on the range of the values ​​(e.g., [0, 0.25] corresponds to center 1). Traverse the second segment of the vector (dimensions N+1 to 2N), and determine the AGV model matched for each workstation based on the range of the values ​​(e.g., [0, 0.5] corresponds to light-load AGV, (0.5, 1.0] corresponds to heavy-load AGV). This forms the initial allocation structure of "Workstation ID - Assigned Center - Matched Model".

[0242] S504. Based on the results of S502 and S503, generate the complete travel path for each AGV, that is, group the workstations assigned to the same distribution center and matching the same vehicle type together and assign them to specific AGVs (such as heavy-duty AGV1, light-duty AGV2). Within the task set of each AGV, sort the workstations according to the access priority determined in S502 to form intermediate path segments (such as: workstation 5 → workstation 1 → workstation 4).

[0243] Starting point determination: Calculate the distance between the first service station of the AGV (e.g., station 5) and each distribution center, and select the nearest distribution center (e.g., center A) as the starting point of the AGV.

[0244] Destination determination: Calculate the distance between the last service station of the AGV (e.g., station 4) and each distribution center, and select the nearest distribution center (e.g., center B) as the return point of the AGV.

[0245] By connecting the starting point and the ending point to the intermediate path segment, a complete semi-open delivery path is formed (e.g., Center A → Workstation 5 → Workstation 1 → Workstation 4 → Center B).

[0246] S505: Convert the AGV path schemes generated in S504 into machine-readable instruction code, that is: convert the path node sequence into specific navigation instructions (such as from coordinates). Drive to Action instructions (such as performing a picking operation at workstation 5, increasing the load) ) and time instructions (such as waiting until a certain time) (Start operation). Instruction packets are sent in parallel to the onboard controllers of each heterogeneous AGV via the workshop's wireless local area network (WLAN) or 5G network. Upon receiving the instructions, each heterogeneous AGV's drive motor executes driving and loading / unloading actions, realizing the physical flow of materials.

[0247] To verify the effectiveness and adaptability of the MDHVRP, MDHVRP, and MDHVRP PDTW models and algorithms constructed in this embodiment, simulation experiments were conducted on three typical examples, covering MDVRP, MDHVRP, and MDHVRP PDTW problems. Two sets of sensitivity experiments were also designed to evaluate the impact of multi-center collaborative delivery strategies and vehicle type configurations on scheduling results. Since MDVRP and MDHVRP are crucial foundations of this research model and core components of its construction, algorithm performance was first tested on these two problems. All algorithms were implemented in Python 3.11 using the PyCharm integrated development platform, running on a Windows 10 operating system equipped with an AMD Ryzen 95900HX processor (3.3GHz) and 32GB of memory. The key parameter settings for the improved particle swarm optimization algorithm (MSVPPSO-LNS) combined with large neighborhood search are shown in Table 1 below.

[0248] Table 1 Parameter Settings

[0249] parameter describe Value Maximum number of iterations of the algorithm 500 Subgroup number 4 Number of principal particles in each subgroup 15 Number of subparticles in each subgroup 15 Particle pause probability 0.3 Large domain algorithm search count 10

[0250] To verify and analyze the algorithm's performance, a Multi-Center Vehicle Routing Problem (MDVRP) example was selected, comprising 3 distribution centers (numbered 1-3) and 30 customer points (numbered 4-33). Each vehicle has a maximum load capacity of 10 tons and a maximum travel distance of 50 kilometers. Each distribution center can use a maximum of 4 vehicles, and the known optimal solution is 116.01 kilometers. In the experiment, the MSVPPSO-LNS algorithm was run 20 times, and the best, worst, and average values ​​(Aug) of the 20 runs were recorded. Table 2 below compares the results of the MSVPPSO-LNS algorithm with those of Wolf Pack Algorithm (WPA), Competitive Decision Algorithm (CDA), Tabu Search Algorithm (TS), Quantum Genetic Algorithm (QGA), and Cloud Quantum Genetic Algorithm (COGA). The evaluation metrics in the table include the known optimal value (KOS), the best value, the worst value, the average value (Aug), the deviation of the optimal value (%Dev), and the average running time (CPU / s). The optimal vehicle delivery route is also provided, and the results demonstrate the performance differences of various algorithms on the MDVRP problem.

[0251] Table 2 Algorithm Performance Comparison Results

[0252] algorithm KOS BEST %Dev Worst %Dev Avg %Dev CPU / s CDA 123.33 6.31 WPA 122.42 5.53 TS 116.01 0 343.31 195.93 187.34 61.49 243 QGA 116.01 116.01 0 326.48 181.42 176.37 52.03 216 CQGA 116.01 0 162.38 39.97 156.32 34.75 168 MSVPPSO-LNS -2.06 147.10 26.81% 127.70 10.1% 4.814

[0253] Table 2 shows that, in terms of solution quality, the CDA algorithm simplifies the multi-center path problem into solving multiple single-center problems separately, failing to effectively coordinate global resources, resulting in a significant gap in overall solution quality. While the introduction of a joint delivery strategy achieves the currently known optimal solution in some cases, it still falls short in terms of solution stability and computational efficiency. In contrast, the MSVPPSO-LNS algorithm demonstrates relatively better solution performance across multiple performance metrics. Its optimal solution reaches 113.6176 km, slightly lower than the known optimal value of 116.01 km, showing some improvement in optimization accuracy. Regarding the worst-case solution, MSVPPSO-LNS keeps it below 147.10 km, significantly lower than the worst results of other comparative algorithms. Its QGA and TS are 326.48 km and 343.31 km respectively, demonstrating its advantage in controlling solution quality. The average solution is 127.70 km, with a deviation of 10.1%, showing greater stability compared to methods like CDA and WPA under the same performance metrics. Overall results show that the proposed algorithm has potential in improving the quality of solutions and maintaining stability.

[0254] In terms of solution speed, such as Figure 10As shown, the MSVPPSO-LNS algorithm can quickly find the optimal solution in a fewer iterations and has strong global search capabilities, avoiding getting trapped in local optima. Whether considering the quality of the optimal solution, the worst solution, or the average solution, the MSVPPSO-LNS algorithm outperforms the compared algorithms. Furthermore, in terms of runtime, the algorithm in this embodiment also demonstrates a significant advantage, with an average search time of only 4.814 seconds. Figure 11 The corresponding vehicle delivery routes are displayed.

[0255] In summary, the data fully demonstrate that the algorithm proposed in this embodiment exhibits superior performance in solving the MDVRP example, verifying the superiority of the MSVPPSO-LNS algorithm in terms of solution quality, computational efficiency, and stability.

[0256] This invention first focuses on the scheduling requirements of multi-distribution center collaborative scheduling, heterogeneous fleet configuration, and the coexistence of pickup and delivery tasks, systematically analyzing and describing its key characteristics. Then, a path optimization model is established with the objective of minimizing the sum of fixed cost, variable cost, and time window penalty cost. Finally, an improved particle swarm optimization algorithm, MSVPPSO-LNS (Multi-Subgroup Velocity Pause Particle Swarm Optimization Algorithm Combined with Large Neighborhood Search), is designed to address the specific problem characteristics. This algorithm employs a multi-subgroup collaborative strategy and a velocity pause update mechanism, and combines a large neighborhood search algorithm to improve local search capabilities, thereby enhancing the overall performance of the algorithm. The algorithm of this invention demonstrates superior performance compared to other algorithms, reducing costs and travel distances by approximately 13.8% and 15.1%, respectively, compared to the traditional closed-loop multi-depot model.

[0257] like Figure 12 As shown, the present invention also provides a multi-AGV intelligent scheduling optimization device, based on the multi-AGV intelligent scheduling optimization method described above, the device comprising:

[0258] The acquisition module 1 is used to acquire multi-dimensional logistics basic data of the production workshop, including workstation task parameters, layout parameters, and fleet parameters; wherein, the workstation task parameters include workstation coordinates and operation time windows, the layout parameters are the location coordinates of each distribution center, and the fleet parameters include the load limit and running speed of each heterogeneous AGV.

[0259] The first module 2 is used to establish an objective function for a vehicle route planning model based on the multi-dimensional logistics basic data. The objective function aims to minimize the total delivery cost, which consists of vehicle fixed and driving costs and time window penalty costs. The vehicle fixed and driving costs are calculated based on the fleet parameters, and the time window penalty costs are calculated based on the workstation task parameters.

[0260] The second module 3 is used to establish the constraints of the vehicle route planning model based on the multidimensional logistics basic data, including semi-open route and vehicle scheduling constraints, heterogeneous vehicle capacity and simultaneous collection and distribution constraints, and mixed time window constraints.

[0261] The solution module 4 is used to solve the vehicle path planning model using a multi-subgroup speed-pause particle swarm optimization algorithm that combines large neighborhood search. The multi-subgroup speed-pause particle swarm optimization algorithm is an improvement on the standard particle swarm algorithm, including using Logistic chaotic mapping to initialize the population distribution, introducing a speed-pause mechanism in the iterative update, adopting a multi-subgroup cooperative strategy to share global optimal information, and incorporating a large neighborhood search mechanism.

[0262] Output module 5 is used to obtain the global optimal solution obtained by the multi-subgroup velocity pause particle swarm optimization algorithm, decode the global optimal solution to obtain a scheduling scheme and convert it into a multi-AGV scheduling instruction to be issued to each heterogeneous AGV; wherein, the scheduling scheme includes the optimal workstation access order corresponding to the workstation task parameters, the workstation assignment center determined from the layout parameters, and the vehicle type allocation matched from the fleet parameters.

[0263] Each of the above modules is used to execute the corresponding steps in the above-mentioned intelligent scheduling and optimization method for multiple AGVs. The specific implementation method is as described in the above-mentioned method embodiment, and will not be repeated here.

[0264] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores all data required for the intelligent scheduling and optimization method for multiple AGVs. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the intelligent scheduling and optimization method for multiple AGVs.

[0265] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0266] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any one of the above-described intelligent scheduling and optimization methods for multiple AGVs.

[0267] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0268] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0269] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent scheduling and optimization of multiple AGVs, characterized in that, include: S1. Obtain multi-dimensional logistics basic data of the production workshop, including workstation task parameters, layout parameters, and fleet parameters; wherein, the workstation task parameters include workstation coordinates and operation time windows, the layout parameters are the location coordinates of each distribution center, and the fleet parameters include the load limit and running speed of each heterogeneous AGV. S2. Establish an objective function for the vehicle route planning model based on the multi-dimensional logistics basic data. The objective function aims to minimize the total delivery cost, which consists of vehicle fixed and driving costs and time window penalty costs. The vehicle fixed and driving costs are calculated based on the fleet parameters, and the time window penalty costs are calculated based on the workstation task parameters. S3. Based on the multidimensional logistics basic data, establish the constraints of the vehicle route planning model, including semi-open route and vehicle scheduling constraints, heterogeneous vehicle capacity and simultaneous collection and distribution constraints, and mixed time window constraints. S4. The vehicle path planning model is solved by using a multi-subgroup speed-pause particle swarm optimization algorithm that combines large neighborhood search; wherein, the multi-subgroup speed-pause particle swarm optimization algorithm is an improvement on the standard particle swarm algorithm, including using Logistic chaotic mapping to initialize the population distribution, introducing a speed-pause mechanism in the iterative update, using a multi-subgroup cooperative strategy to share global optimal information, and incorporating a large neighborhood search mechanism. S5. Obtain the global optimal solution obtained by the multi-subgroup velocity-pause particle swarm optimization algorithm, decode the global optimal solution to obtain a scheduling scheme and convert it into a multi-AGV scheduling instruction to be issued to each heterogeneous AGV; wherein, the scheduling scheme includes the optimal workstation access order corresponding to the workstation task parameters, the workstation assignment center determined from the layout parameters, and the vehicle type allocation matched from the fleet parameters.

2. The intelligent scheduling and optimization method for multiple AGVs according to claim 1, characterized in that, S1 specifically includes: S101. Determine all workstations in the production workshop that require material delivery or recycling to establish a workstation set. Obtain the planar coordinates, delivery demand, pickup demand, loading and unloading service time, and optimal service time window of each workstation as workstation task parameters. The optimal service time window is divided into hard time windows and soft time windows based on the urgency of the workstation task. S102. Determine the layout of multiple distribution centers in the production workshop to establish a distribution center set, obtain the planar coordinates of each distribution center as layout parameters, and calculate the Euclidean distance between any two points based on the planar coordinates of the work site set and the distribution center set. S103. Determine the different vehicle types in the heterogeneous AGV fleet to establish a heterogeneous AGV set, and obtain the maximum load, average driving speed, total number available in the workshop, fixed usage cost per trip, and transportation cost per unit driving distance for each type of AGV as fleet parameters; wherein, the heterogeneous AGV set includes at least two types: light-load AGV and heavy-load AGV.

3. The intelligent scheduling and optimization method for multiple AGVs according to claim 1, characterized in that, S2 specifically includes: S201. Construct a fixed usage cost function for vehicles, and calculate the sum of the fixed departure costs of all scheduled and activated heterogeneous AGVs. The calculation formula is as follows: in, R represents the fixed operating cost of the vehicle; R is the set of heterogeneous AGV types. Let A be the set of vehicles of type r AGV, A be the set of arc segments formed by all nodes, and C be the set of work points; This represents the fixed dispatch cost of the r-th type AGV; The variable is a decision variable; it takes the value 1 if the kth AGV in the rth class is activated, and 0 otherwise. S202. Construct a vehicle transportation cost function to calculate the sum of the travel distance costs incurred by all scheduled and activated heterogeneous AGVs during the delivery task. The calculation formula is as follows: in, For vehicle transportation costs; This represents the unit distance transportation cost of the r-th type of AGV; This represents the Euclidean distance between node i and node j; The value of the variable is 1 if the kth AGV in the rth class travels from node i to node j, and 0 otherwise. S203. Construct a time window penalty cost function. For workstations with a set soft time window, calculate the sum of delay penalty costs caused by the AGV arrival time exceeding the upper bound of the optimal service time window. The calculation formula is as follows: in, Penalty costs for time windows; This represents the penalty coefficient for delay per unit time. This represents the time when the k-th AGV in class r arrives at workstation i and begins service; This indicates the latest end time of the optimal service time window for workstation i; This indicates a violation of the time window; penalties are calculated only if the arrival time is later than the end time of the time window. S204. Add the vehicle fixed usage cost function, vehicle transportation cost function, and time window penalty cost function to establish a vehicle route planning model objective function with the goal of minimizing total delivery cost. .

4. The intelligent scheduling and optimization method for multiple AGVs according to claim 1, characterized in that, S3 specifically includes: S301. Establish semi-open path and vehicle scheduling constraints, specifically including: Vehicle activation and route association constraints: ; Available vehicle quantity constraints: ; Unique service constraint: ; Flow balance constraints: ; Semi-open start and end constraints: Departure constraints Return constraints No-idle running constraint: ; Eliminate sub-loop constraints: Where D is the set of distribution centers and C is the set of work sites; For decision variables, the value is 1 if the kth AGV of the rth class travels from the distribution center d to the workstation j, and 0 otherwise; This is a decision variable; it is 1 if enabled, and 0 otherwise. Let be the total number of available AGVs of type r; V be the set of all nodes. This indicates that the vehicle travels from node i to j; S is any subset of the workstation set C, and |S| is the number of elements in that subset; S302. Establish heterogeneous vehicle capacity and simultaneous cargo collection and distribution constraints, specifically including: Initial load definition: ; Dynamic balance constraints on vehicle load capacity: ; Capacity limit constraint: ; ; in, The load of the vehicle when it leaves the distribution center d. The delivery demand for workstation j; , These represent the cargo loads when leaving nodes j and i, respectively. Let M be the required quantity of goods to be picked up at workstation j; M is a maximum positive number; A is the set of all arc segments; This represents the maximum rated load capacity of the r-th type AGV. S303. Establish mixed time window constraints and variable definitions, specifically including: Service time continuity constraints: ; Soft time window delay constraint: ; Initial time constraints: ; Definition of decision variables and nonnegativity constraints: ; in, , These represent the times when the AGV arrives at workstations j and i, respectively. Service duration for workstation i; Let M be the travel time from node i to j; M is a maximum positive number. The upper bound of the optimal service time window for workstation i; if ,but It is at least the difference between the two.

5. The intelligent scheduling and optimization method for multiple AGVs according to claim 1, characterized in that, S4 specifically includes: S401. Construct a three-segment particle coding structure and use Logistic chaotic mapping to generate the position and velocity of the initial population; wherein, the three-segment particle coding structure includes a first segment code representing the workstation assignment center, a second segment code representing the workstation vehicle type allocation, and a third segment code representing the workstation access order; the initialization formula of the Logistic chaotic mapping is: in, This represents the normalized variable generated in the nth iteration, with a value range of (0,1); This is a chaos control parameter with a value range of [3.57, 4], used to ensure that the sequence is in a completely chaotic state; S402. Divide the particle population into multiple independent search subgroups, decode the position vector of each particle into a path scheme and calculate the fitness value, and initialize the local optimal solution of each subgroup and the global optimal solution of the population; wherein, the fitness value is calculated by the formula: f(x)=1 / Z, where f(x) represents the fitness value of particle x; Z is the comprehensive cost of the particle, including fixed cost, travel distance, and penalty term generated by time window constraint; S403. Perform multi-subgroup collaborative update with the introduction of velocity pause mechanism. During the iteration process, determine whether the particle should maintain the current velocity or perform standard velocity update based on probability, and update the particle position. S404. At the end of each iteration, perform a large neighborhood search destruction operation on the current global optimal solution, and remove several workstation nodes in the path according to a preset strategy; perform a large neighborhood search repair operation on the destroyed solution, and reinsert the removed workstation nodes into the path to generate a new solution. If the new solution is better than the original global optimal solution, then update it. S405. Determine whether the maximum number of iterations or the convergence accuracy requirement has been reached. If the requirements are met, output the global optimal solution; otherwise, return to step S403 to continue iterating.

6. The intelligent scheduling and optimization method for multiple AGVs according to claim 5, characterized in that, In S403, the specific update rules for the speed pause mechanism are as follows: Set speed pause probability A random number is generated before each particle velocity update. ;like If the particle's velocity is paused, it will maintain its velocity from the previous moment. ; like Then the particle performs a standard velocity update operation, that is: In the formula, and Let i represent the velocity and position of particle i in the t-th iteration, respectively. For inertial weights, As a learning factor, A random number between [0, 1]; Let be the individual historical best position of particle i. This represents the globally optimal position for the population.

7. The intelligent scheduling and optimization method for multiple AGVs according to claim 5, characterized in that, In S404, the destructive operation adopts a hybrid removal strategy, including random removal, worst-case removal, and similarity removal. Random removal randomly selects several workstations from the current solution for elimination. Worst-case removal calculates the reduction in objective function cost after removing each workstation and prioritizes removing the workstations that contribute the most to system cost. Similarity removal removes a group of workstations with high similarity based on their geographical proximity or material requirement similarity. The repair operation employs an optimal incremental insertion strategy, which sequentially calculates the objective function increment generated by the workstation to be inserted at all feasible insertion positions, and selects the position with the smallest increment for insertion.

8. The intelligent scheduling and optimization method for multiple AGVs according to claim 1, characterized in that, S5 specifically includes: S501. Obtain the global optimal particle vector output by the multi-subgroup velocity pause particle swarm optimization algorithm, and parse it into a three-segment structure of workstation assignment center vector, workstation vehicle allocation vector, and workstation access order vector. S502. Extract the values ​​from the workstation access sequence vector and sort them. Establish the access priority sequence of the workstations based on the sorting results. The sorting rule is as follows: arrange the random real numbers in the workstation access sequence vector in ascending order. The workstation with the smaller value has a higher access priority in the scheduling sequence. S503. Traverse the workstation assignment center vector and the workstation vehicle allocation vector, and determine the distribution center to which each workstation belongs and the matched AGV vehicle through numerical interval mapping rules; wherein, the numerical interval mapping rules include: For the workstation assignment center vector, the numerical range [0,1] is divided into several equal intervals corresponding to the number of distribution centers, and the distribution center number assigned to the workstation is determined according to the interval into which the particle value falls. For the vehicle type allocation vector at the workstation, the numerical range [0,1] is divided into several intervals corresponding to the number of heterogeneous vehicle types. The light-load AGV or heavy-load AGV allocated to the workstation is determined according to the interval into which the particle value falls. S504. Based on the established workstation resource ownership and access priority sequence, workstations are allocated to specific AGVs, and a semi-open strategy is used to construct the specific driving path for each AGV; wherein, the process of constructing the specific driving path using the semi-open strategy includes: Workstations assigned to the same AGV are sorted according to access priority to form intermediate task path segments. The first and last service workstations of each path segment are determined. The distances between the first service workstation and all distribution centers are calculated, and the nearest distribution center is selected as the AGV's departure node. The distances between the last service workstation and all distribution centers are calculated, and the nearest distribution center is selected as the AGV's return node. The departure node, intermediate task path segments, and return node are connected sequentially to form a complete semi-open delivery path. S505 converts the constructed driving path into a multi-AGV scheduling instruction containing navigation coordinates, action type and time parameters, and sends it to each heterogeneous AGV for execution via wireless network.

9. A multi-AGV intelligent scheduling and optimization device, based on the multi-AGV intelligent scheduling and optimization method according to any one of claims 1 to 8, characterized in that, The device includes: The acquisition module is used to acquire multi-dimensional logistics basic data of the production workshop, including workstation task parameters, layout parameters, and fleet parameters. Among them, the workstation task parameters include workstation coordinates and operation time windows, the layout parameters are the location coordinates of each distribution center, and the fleet parameters include the load limit and running speed of each heterogeneous AGV. The first module is used to establish an objective function for a vehicle route planning model based on the multi-dimensional logistics basic data. The objective function aims to minimize the total delivery cost, which consists of vehicle fixed and driving costs and time window penalty costs. The vehicle fixed and driving costs are calculated based on the fleet parameters, and the time window penalty costs are calculated based on the workstation task parameters. The second module is used to establish the constraints of the vehicle route planning model based on the multidimensional logistics basic data, including semi-open route and vehicle scheduling constraints, heterogeneous vehicle capacity and simultaneous collection and distribution constraints, and mixed time window constraints. The solution module is used to solve the vehicle path planning model using a multi-subgroup speed-pause particle swarm optimization algorithm that combines large neighborhood search. The multi-subgroup speed-pause particle swarm optimization algorithm is an improvement on the standard particle swarm algorithm, including using Logistic chaotic mapping to initialize the population distribution, introducing a speed-pause mechanism in the iterative update, adopting a multi-subgroup cooperative strategy to share global optimal information, and incorporating a large neighborhood search mechanism. The output module is used to obtain the global optimal solution obtained by the multi-subgroup velocity-pause particle swarm optimization algorithm, decode the global optimal solution to obtain a scheduling scheme, and convert it into a multi-AGV scheduling instruction to be issued to each heterogeneous AGV; wherein, the scheduling scheme includes the optimal workstation access order corresponding to the workstation task parameters, the workstation assignment center determined from the layout parameters, and the vehicle type allocation matched from the fleet parameters.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images