Matrix workshop multi-type agv scheduling method based on super heuristic algorithm

Abstract

This invention relates to the field of logistics scheduling in intelligent manufacturing workshops, specifically to a multi-type AGV scheduling method for matrix workshops based on a hyperheuristic algorithm. The method includes: constructing solutions using a type-separated dual-component encoding scheme and its corresponding decoding scheme; generating an initial population based on a mask-based type-interleaved polling initialization method; adaptively selecting low-level heuristic operators through learning; generating child solutions by executing the selected low-level heuristic operators; performing constraint repair and updating the current population; determining the current state based on the logarithmic change in total energy consumption between the parent and child populations' optimal solutions; performing adaptive local search and updating the current population; and determining whether the algorithm termination condition has been met. This invention solves the collaborative scheduling problem of delivery, recycling, and detection AGVs in a matrix workshop, achieving the positive effects of reducing total energy consumption in workshop logistics and improving the collaborative scheduling efficiency of multiple AGV types.

Description

Technical Field

[0001] This invention relates to the field of logistics scheduling in intelligent manufacturing workshops, specifically to a matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm. Background Technology

[0002] Matrix manufacturing systems, with their reconfigurable equipment layout and high production flexibility, have become an important form of flexible manufacturing. Matrix manufacturing systems cluster similar processing equipment and rely on automated guided vehicles (AGVs) to perform logistics tasks such as material delivery, finished product collection, and quality inspection between these clusters, achieving a high degree of automation in the production process. As the scale and variety of products in matrix manufacturing systems continue to expand, the operating efficiency of AGVs directly affects production continuity and workshop energy consumption levels; therefore, efficient AGV scheduling is crucial for the stable operation of matrix manufacturing systems.

[0003] In the actual production of matrix manufacturing systems, the functional requirements of AGVs differ significantly across the three basic stages: material replenishment, product collection, and quality inspection. Differences in weight and cleanliness between raw materials and finished products make it difficult for delivery AGVs and recycling AGVs to share the same vehicle, as this would lead to cross-contamination and load mismatch. Quality inspection tasks in hazardous environments require AGVs equipped with dedicated sensors. Therefore, matrix manufacturing systems need to deploy three different functional types of AGVs simultaneously: delivery, recycling, and inspection, to meet the differentiated transportation needs of different stages. However, most existing AGV scheduling studies in matrix workshops assume that all AGVs perform the same operational functions and have not yet systematically studied the coordinated scheduling problem of AGVs with multiple functional types. Compared to single-type AGV scheduling problems, multi-type AGV scheduling problems involve coordination between different types of AGVs, significantly increasing the complexity in terms of encoding representation, initialization construction, and search optimization.

[0004] For solving the scheduling problem of multiple AGV types, exact methods face unacceptable computational complexity in large-scale instances. While improved metaheuristic algorithms have achieved good results in large-scale scheduling problems, the no-free-lunch theorem indicates that a single metaheuristic algorithm cannot achieve optimal performance on all problem instances. Hyperheuristic algorithms can reconcile the trade-off between search efficiency and solution quality, and are an effective approach to solving such complex scheduling problems. However, existing reinforcement learning-based hyperheuristic algorithms have not yet been applied to scheduling scenarios of multiple AGV types, and there is a lack of encoding schemes, initialization methods, and search operators for multi-type AGV coordination. Therefore, researching a reinforcement learning-based hyperheuristic algorithm to solve the multi-type AGV scheduling problem in a matrix workshop is of great significance for reducing workshop logistics energy consumption and improving the efficiency of multi-type AGV collaborative scheduling. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-type AGV scheduling method for matrix workshops based on a hyperheuristic algorithm, which solves the problem of coordinated scheduling of three types of AGVs—delivery, recycling, and detection—in a matrix workshop, thereby reducing the total energy consumption of workshop logistics and improving the efficiency of coordinated scheduling of multiple types of AGVs.

[0006] The present invention provides a matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm, characterized by comprising the following steps:

[0007] Step 1: Initialize algorithm parameters. Construct the solution using a type-separated two-component encoding scheme and its corresponding decoding scheme, and generate the initial population based on the mask type-interleaved polling initialization method.

[0008] Step 2, using The table stores the state-action pairs. Value, through Learn to adaptively select low-level heuristic operators; use the solutions in the current population as parent solutions, and execute the selected low-level heuristic operators to generate child solutions;

[0009] Step 3: Perform constraint repair on the offspring solutions, and determine whether to accept the offspring solutions to update the current population according to the simulated annealing acceptance criterion; determine the current state based on the logarithmic change in total energy consumption of the optimal solutions of the parent and offspring populations, and update the population using the Bellman equation. The table corresponds to the state-action pairs value;

[0010] Step 4: Perform an adaptive local search on the solutions in the current population and update the current population;

[0011] Step 5: Determine if the algorithm termination condition has been met. If it has, the algorithm ends and the current optimal solution is output; otherwise, return to step 2.

[0012] Furthermore, in step 1, the type-separated two-component encoding scheme represents the solution as a triplet, which consists of three scheduling components corresponding to delivery AGVs, retrieval AGVs, and detection AGVs, respectively; an AGV type index is defined. ,in Indicates delivery type. Indicates recycling type. This indicates a detection-type scheduling component; each scheduling component includes a task sequence. and departure time Task sequence By type The task sequence is arranged by all tasks waiting to be served by the AGV, and divided into several subsequences using 0 as the separator. Each subsequence is a type. The tasks undertaken by each AGV; departure time Record type Departure time and type of each AGV The Middle The departure time of each AGV is expressed as follows: .

[0013] Furthermore, in step 1, the decoding process for the solution represented by the type-separated dual-component encoding scheme is as follows: Each task in the matrix workshop has an allowed earliest arrival time and a allowed latest arrival time. A time window constraint is defined, that is, the allowed earliest arrival time and the allowed latest arrival time together constitute the time window of the corresponding task, and the arrival time of each AGV must be within the time window; a load constraint is defined, that is, the cumulative load of all tasks on each AGV path must not exceed the maximum carrying capacity of the corresponding AGV; for type... AGVs, from the task sequence Get type The task sequence of each AGV, starting from the departure time Get type Departure time of each AGV Starting with the departure time of each AGV, calculate the arrival time, service duration, and cumulative load status of each task sequentially according to the task order in the task sub-sequence; then classify the types... The energy consumption of all AGVs in the system is accumulated, including the energy consumption of driving, starting, and task execution. The total energy consumption of the scheduling scheme is obtained by summing the energy consumption of each of the three types of AGVs.

[0014] Furthermore, in step 1, the mask-based type-interleaved polling initialization method includes a mask mixing construction phase, a type-interleaved polling allocation phase, and a population construction phase, wherein...

[0015] In the mask mixing construction stage, define To determine the population size, a construction operator selection mechanism is used to select one construction operator for each type of AGV from both nearest-neighbor and random construction operators. This mechanism includes a mask sequence consisting of three binary digits: 000, 001, 010, 011, 100, 101, 110, and 111. The first, second, and third digits of the mask sequence correspond to detection-type AGVs, retrieval-type AGVs, and delivery-type AGVs, respectively. When the corresponding bit value in the mask sequence is 1, the nearest-neighbor construction operator is used for that type of AGV; when the corresponding bit value is 0, the random construction operator is used. The mask sequence is traversed to determine the construction operator for each of the detection-type, retrieval-type, and delivery-type AGVs, thus implementing task allocation for these AGVs. Specifically, task allocation is performed once for each of the mask sequences 001 to 111, constructing seven different task sequences. These sequences are then executed independently and repeatedly based on the mask sequence 000. Sub-task allocation, generation Group task sequence;

[0016] During the type-alternating polling allocation phase, the unassigned task sets of delivery, recycling, and detection AGVs and the indices of each type of AGV are processed alternately. The process proceeds sequentially across types with the same index, following the order: Delivery AGV #1, Recycling AGV #1, Detection AGV #1, Delivery AGV #2, Recycling AGV #2, Detection AGV #2, and so on, until all unassigned tasks are allocated. The Middle Departure time of AGV Calculation, where This indicates the departure time of the first AGV in the dispatch process. This indicates the departure interval, thus generating the departure time. ;

[0017] During the population building phase, the mask generated during the construction phase will be mixed with the mask. Task sequence The types generated in the interleaved polling allocation phase are respectively Combine them to construct Find a solution; sort all solutions in ascending order of total energy consumption, and retain the solutions with the lowest total energy consumption. Solution: The population size is The initial population is used to determine the optimal solution, which is the one with the minimum total energy consumption in the current population.

[0018] Furthermore, all unassigned tasks of each type of AGV constitute the unassigned task set of the corresponding type. The specific process of the nearest neighbor construction operator is as follows: taking the location of the AGV currently building a path in the current type as the starting point, calculate the Manhattan distance from each task in the unassigned task set of the current AGV type to the current location; select the task with the smallest Manhattan distance as the next task to be served, and add the selected task to the task sequence of the current AGV; at the same time, remove the selected task from the unassigned task set of the current type, and update the position of the current AGV to the position of the selected task; repeat the process of calculating distance, selecting tasks, and updating positions until the cumulative load of the current AGV reaches the capacity limit of the AGV, or the unassigned task set of the current type of AGV is empty; when the path of an AGV is completed, append a separator 0 to the end of the task sequence of the current AGV to mark the path boundary, and continue to build the path for the next AGV of the same type.

[0019] The specific process of the random task construction operator is as follows: starting from the location of the AGV currently building a path in the current type, a task is randomly selected from the unassigned task set of the current AGV type as the next task to be served, and the selected task is added to the task sequence of the current AGV; at the same time, the selected task is removed from the unassigned task set of the current type of AGV, and the current position of the current AGV is updated to the position of the selected task; the process of randomly selecting tasks and updating positions is repeated until the cumulative load of the current AGV reaches the capacity limit of the AGV, or the unassigned task set of the current type of AGV is empty; when the path of an AGV is completed, a separator 0 is appended to the end of the task sequence of the current AGV to mark the path boundary, and the path is continued to be built for the next AGV of the same type.

[0020] Furthermore, in step 2, the low-level heuristic operators include intra-path insertion operators, inter-path insertion operators, adjacent swap operators, pancake flipping operators, globally greedy insertion operators, destructive reconstruction operators, and two-element optimization operators, among which...

[0021] a. The operation process of the insertion operator within the path is as follows: randomly select a task from the task sequence of an AGV of the same type and remove it from the current position. Then, randomly determine a new position in the task sequence and re-insert the selected task into the new position.

[0022] b. The operation process of the inter-path insertion operator is as follows: task transfer is performed between the task sequences of two different AGVs of the same type. The first AGV with the largest cumulative task load and the second AGV with the smallest cumulative task load are identified. A task is selected from the task sequence of the first AGV and removed. The selected task is then inserted into the task sequence of the second AGV.

[0023] c. The operation process of the adjacent exchange operator is as follows: randomly select the task sequence of an AGV in the same type of AGV, randomly select two adjacent tasks in the task sequence and exchange the positions of the two tasks to perform local perturbation. If the time window constraint or load constraint is violated after the exchange, the exchange operation is cancelled.

[0024] d. The operation process of the pancake flipping operator is as follows: randomly select the task sequence of an AGV in the same type of AGV, randomly select a split position in the task sequence, and rearrange all tasks before the split position in reverse order;

[0025] e. The operation process of the global greedy insertion operator is as follows: randomly remove one task from all task sequences of the same type of AGV as the task to be inserted; traverse each insertion position in all task sequences of the same type of AGV, calculate the total energy consumption after inserting the task to be inserted into each position, and insert the task to be inserted into the position that minimizes the increase in total energy consumption.

[0026] f. The operation process of the destruction and reconstruction operator is as follows: In the destruction phase, the number of tasks to be removed is determined according to the destruction rate, and the tasks are removed according to a random removal rule or a worst-case removal rule. The random removal rule is to randomly select a corresponding number of tasks from all task sequences of the same type of AGV for removal. The worst-case removal rule is to calculate the difference in total energy consumption before and after the removal of each task in all task sequences of the same type of AGV, and prioritize removing the task that reduces the total energy consumption the most. In the reconstruction phase, the difference between the latest and earliest arrival times allowed for each removed task is calculated, and the removed tasks are arranged in ascending order of the difference. The removed tasks are then inserted sequentially into the positions in all task sequences of the same type of AGV that minimize the increase in total energy consumption.

[0027] g. The operation process of the two-element optimization operator is as follows: randomly select an AGV from the same type of AGV, randomly select two non-adjacent positions in the task sequence of the selected AGV, and rearrange all tasks between the two positions in reverse order.

[0028] Furthermore, in step 2, The specific process of learning adaptive selection of low-level heuristic operators involves defining seven optional actions for the low-level heuristic operators: intra-path insertion operator, inter-path insertion operator, adjacent swap operator, pancake flip operator, global greedy insertion operator, destructive reconstruction operator, and two-element optimization operator. The energy consumption change of the population optimal solution is defined as the state. Table storage value, Each row in the table corresponds to a state, each column corresponds to an action, and the intersection of the row and column stores the action selected for that state. value, All in the table The initial values ​​of all values ​​are set to 1; based on In the table Value, using - A greedy strategy is used to select actions. Greedy strategy is based on probability Select the current state The action with the highest value, in probability Randomly select one action from the seven actions. The preset exploration probability parameters are used; the low-level heuristic operator corresponding to the selected action is executed, and the selected low-level heuristic operator is applied to each parent solution in the current population to generate the corresponding child solution.

[0029] Furthermore, in step 3, The specific process of value update is as follows: calculate the total energy consumption of the optimal solution of the parent population in this iteration. Total energy consumption of the optimal solution for the offspring population logarithmic change And the current state is discretized and defined, when When the energy consumption deteriorates, the state value is 0; when When the energy consumption remains unchanged, the state value is 1; when and Values ​​ranging from 1 to 6 represent different degrees of energy efficiency improvement, with the state value being... ;when When the value is 8, it indicates a significant improvement in energy consumption; otherwise, the value is 9. (Based on logarithmic change) For instant rewards Perform hierarchical assignment when hour, ;when hour, ;when and When the value range is 1 to 5, ,Right now The value range is 5 to 9; other cases Using Bellman equations right The corresponding table The value is updated, where, Indicates the current state. This represents the low-level heuristic operator corresponding to the action selected in this iteration. express Status in the table Select action Corresponding value; This indicates the next state after an action is performed. Indicates the candidate action in the next state. Indicates the next state All candidate actions correspond The maximum value; Indicates an immediate reward. Indicates the discount factor. The learning rate is indicated by the number of steps taken to determine the learning rate. according to Dynamic calculation, Indicates the current iteration number. This indicates the maximum number of iterations.

[0030] Furthermore, in step 3, the specific process of the simulated annealing acceptance criterion is as follows: if the total energy consumption of the child solution is less than the total energy consumption of the parent solution, then the child solution is accepted to replace the parent solution; if the total energy consumption of the child solution is less than the total energy consumption of the current optimal solution, then the child solution is updated to the current optimal solution; if the total energy consumption of the child solution is greater than or equal to the total energy consumption of the parent solution, then the child solution is accepted with probability. The decision on whether to accept the offspring solution, among which Indicates child solution Total energy consumption Indicates parent generation solution Total energy consumption

[0031] This indicates the current annealing temperature, used to control the acceptance of offspring solutions with high total energy consumption during population iteration. The larger the value, the greater the probability of accepting a sub-solution with higher total energy consumption. The smaller the value, the lower the probability of accepting a sub-solution with higher total energy consumption; the initial value of the annealing temperature is set to... That is, the initial annealing temperature; the annealing temperature is according to Update, in which This indicates the lower bound of the annealing temperature. Indicates the cooling coefficient. The value range is from 0 to 1.

[0032] Furthermore, in step 3, the specific process of constraint repair is as follows: Detect whether there are any AGVs in the sub-solution whose arrival time violates the time window constraint; if so, identify the AGVs whose arrival time violates the time window constraint as objects to be adjusted; define all AGVs whose departure time is earlier than the departure time of the object to be adjusted and which do not violate the time window constraint as exchangeable objects; swap the departure time of the object to be adjusted with the departure time of the exchangeable object; repeat the operation of swapping the departure time of the object to be adjusted with the exchangeable object until the arrival time of all AGVs satisfies the time window constraint, or there are no exchangeable objects.

[0033] Furthermore, in step 4, the adaptive local search includes three local search operations: knowledge-guided position adjustment, optimal type combination, and AGV merging.

[0034] The specific process of knowledge-guided position adjustment is as follows: sort the solutions in the current population according to their total energy consumption from smallest to largest, and select the solution ranked in the 1st place based on its total energy consumption. Ranked first The solution of each bit is used as the operation object; for each operation object, a random number in the interval [0,1] is generated, and the probability is calculated. If the random number is less than If the position of the object being operated on is adjusted, then the operation is not performed; otherwise, it is not performed. This indicates the time elapsed for normalization. This indicates the time elapsed since the algorithm ran. This represents the total running time. If position adjustment is performed, the difference between the scheduling component energy consumption of the delivery AGV, recovery AGV, and detection AGV in the operation object and the scheduling component energy consumption of the corresponding type of AGV in the current optimal solution is calculated. The AGV type with the largest difference is selected, and the normalized position ratio of all tasks of the selected type of AGV in the task sequence is extracted from the current optimal solution. The normalized position ratio is the ratio of the position of the task in the task sequence to the total length of the sequence. Each task in the task sequence of the corresponding type of AGV in the operation object is adjusted to a position consistent with the normalized position ratio of the same task in the current optimal solution.

[0035] Furthermore, the specific process for the optimal combination of types is as follows: First, identify the scheduling component with the lowest energy consumption among the scheduling components of delivery AGVs, recovery AGVs, and detection AGVs in the current population, and denot it as the current optimal scheduling component for the corresponding type of AGV. Then, construct a new solution from the current optimal scheduling component for the corresponding type of AGV, denoted as the combined solution. Next, regenerate the departure time for the combined solution. If the AGV arrival time in the combined solution violates the time window constraint, then perform constraint repair. Finally, calculate the total energy consumption of the combined solution. If the total energy consumption of the combined solution is less than the total energy consumption of the optimal solution in the current population, then replace the solution with the highest total energy consumption in the current population with the combined solution.

[0036] Furthermore, the specific process of AGV merging is as follows: For each solution in the current population, denoted as the solution to be operated on, an AGV type is randomly selected from the solutions to be operated on. Among the AGVs of the selected type, the AGV with the fewest tasks is identified as the object to be merged. Attempts are made to distribute all tasks in the subsequence of the object to be merged to the subsequences of other AGVs of the selected type. If all AGVs of the selected type satisfy the load constraints after the distribution, the subsequence corresponding to the object to be merged is deleted and the scheduling component of the selected type is updated. The updated scheduling component is combined with the scheduling components of the other two types in the solution to be operated on to construct a new solution, and the departure time is regenerated for the new solution. If there are cases in the new solution where the AGV arrival time violates the time window constraint, constraint repair is performed.

[0037] Furthermore, in step 5, the algorithm termination time is fixed at 5 seconds.

[0038] This invention provides a multi-type AGV scheduling method for matrix workshops based on a hyperheuristic algorithm. It addresses the collaborative scheduling needs of three functionally heterogeneous AGVs—delivery, recycling, and detection—with the optimization objective of minimizing total energy consumption. A type-separated dual-component encoding scheme and a mask-based type-interleaved polling initialization method are employed to generate a high-quality initial population that balances diversity and feasibility. Learning to adaptively select among low-level heuristic operators, through This invention employs adaptive selection among low-level heuristic operators and combines simulated annealing acceptance criteria to balance exploration and development. It utilizes three local search operations—knowledge-guided position adjustment, optimal type combination, and AGV merging—for refined searching, leveraging structural information from the current optimal solution in the population to accelerate convergence. This invention effectively reduces the total energy consumption of heterogeneous AGVs in a matrix workshop, and improves the collaborative scheduling efficiency and solution stability of multiple AGV types. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0040] Figure 2 This is a layout diagram of the matrix manufacturing workshop of the present invention;

[0041] Figure 3 This is a trend graph of the parameter levels of the present invention;

[0042] Figure 4 Box plot comparing the RPD of this invention with four other variant algorithms;

[0043] Figure 5 This is an average scatter plot of the present invention and the comparison algorithm. Detailed Implementation

[0044] like Figure 1 As shown, the matrix workshop multi-type AGV scheduling method based on hyperheuristic algorithm provided by the present invention is mainly implemented through the following steps.

[0045] Step 1: Initialize algorithm parameters. Construct the solution using a type-separated two-component encoding scheme and its corresponding decoding scheme, and generate the initial population based on the mask type-interleaved polling initialization method.

[0046] Specifically, the type-separated two-component encoding scheme represents the solution as a triplet, which consists of three scheduling components corresponding to delivery AGVs, retrieval AGVs, and detection AGVs, respectively, defining an AGV type index. ,in Indicates delivery type. Indicates recycling type. This indicates a detection-type scheduling component. Each scheduling component includes a task sequence. and departure time Task sequence By type The task sequence is arranged by all tasks waiting to be served by the AGV, and divided into several subsequences using 0 as the separator. Each subsequence is a type. The tasks performed by each AGV. Departure time. Record type Departure time and type of each AGV The Middle The departure time of each AGV is expressed as follows: By type index The division separates the three functional types of AGVs at the coding level, making the task allocation and time scheduling of each type of AGV independent of each other, avoiding coding coupling between different types of AGVs, and facilitating subsequent search and optimization operations for specific types.

[0047] The decoding process for the solution represented by the type-separated two-component encoding scheme is as follows: Each task in the matrix workshop has an allowed earliest arrival time and a allowed latest arrival time. A time window constraint is defined, meaning the allowed earliest arrival time and the allowed latest arrival time together constitute the time window for the corresponding task, and the arrival time of each AGV must be within the time window. A load constraint is defined, meaning the cumulative load of all tasks on each AGV path must not exceed the maximum carrying capacity of the corresponding AGV. For type... AGVs, from the task sequence Get type The task sequence of each AGV, starting from the departure time Get type Departure time of each AGV Starting with the departure time of each AGV, the arrival time, service duration, and cumulative load status of each task are calculated sequentially according to the order of tasks in the task sub-sequence. During decoding, if the AGV arrives at a task earlier than the earliest allowed arrival time for that task, the AGV must wait until the earliest arrival time before starting service. The energy consumption of all AGVs in the system is accumulated, including the energy consumption of driving, starting, and task execution. The total energy consumption of the scheduling scheme is obtained by summing the energy consumption of each of the three types of AGVs. The total energy consumption is used as the objective function value for evaluating the quality of the solution. The smaller the total energy consumption, the better the scheduling scheme.

[0048] The mask-based type-interleaved polling initialization method includes a mask hybrid construction phase, a type-interleaved polling allocation phase, and a population construction phase.

[0049] In the mask mixing construction stage, define To determine the population size, a construction operator selection mechanism is used to select one construction operator for each type of AGV from both nearest-neighbor and random construction operators. This mechanism includes a mask sequence of three binary values: 000, 001, 010, 011, 100, 101, 110, and 111. The first, second, and third bits of the mask sequence correspond to detection-type AGVs, retrieval-type AGVs, and delivery-type AGVs, respectively. When the corresponding bit value in the mask sequence is 1, the nearest-neighbor construction operator is used for that type of AGV; when the corresponding bit value is 0, the random construction operator is used. The mask sequence is traversed to determine the construction operator for each of the detection-type, retrieval-type, and delivery-type AGVs, thus implementing task allocation for these AGVs. Specifically, task allocation is performed once for each of the mask sequences 001 to 111, constructing seven different task sequences, which are then independently and repeatedly executed based on mask sequence 000. Sub-task allocation, generation The task sequence is grouped; by traversing the mask sequence, it is ensured that all combinations of the nearest neighbor construction operator and the random construction operator on the three types of AGVs are covered, so that the initial population has both high quality and diversity.

[0050] All unassigned tasks for each type of AGV constitute the corresponding unassigned task set. The specific process of the nearest neighbor construction operator is as follows: taking the location of the AGV currently building a path in the current type as the starting point, calculate the Manhattan distance from each task in the unassigned task set of the current AGV type to the current location; select the task with the smallest Manhattan distance as the next task to be served, and add the selected task to the task sequence of the current AGV; at the same time, remove the selected task from the unassigned task set of the current type, and update the current AGV's location to the location of the selected task. Repeat the process of calculating distance, selecting tasks, and updating locations until the cumulative load of the current AGV reaches the AGV's capacity limit, or the unassigned task set of the current type of AGV is empty; when the path of an AGV is completed, add a separator 0 to the end of the current AGV's task sequence to mark the path boundary, and continue to build a path for the next AGV of the same type. In each step, the nearest neighbor construction operator selects the task with the smallest Manhattan distance from the current location in the unassigned task set and assigns it first, gradually building AGV task paths with shorter travel distances, which helps to improve the overall solution quality of the initial population.

[0051] The specific process of the random task construction operator is as follows: starting from the location of the AGV currently constructing a path, a task is randomly selected from the unassigned task set of the current AGV type as the next task to be served, and the selected task is appended to the task sequence of the current AGV. At the same time, the selected task is removed from the unassigned task set of the current type of AGV, and the current position of the current AGV is updated to the position of the selected task. This process of randomly selecting tasks and updating positions is repeated until the cumulative load of the current AGV reaches the capacity limit of the AGV, or the unassigned task set of the current type of AGV is empty. After the path of an AGV is constructed, a separator 0 is appended to the end of the task sequence of the current AGV to mark the path boundary, and the path is constructed for the next AGV of the same type. By randomly selecting tasks, the random task construction operator makes the solutions in the initial population have differentiated task sequence arrangements, which helps to expand the exploration range of subsequent searches.

[0052] During the type-alternating polling allocation phase, the unassigned task sets of delivery, recycling, and detection AGVs and the indices of each type of AGV are processed alternately. The process proceeds sequentially across types with the same index, following the order: Delivery AGV #1, Recycling AGV #1, Detection AGV #1, Delivery AGV #2, Recycling AGV #2, Detection AGV #2, and so on, until all unassigned tasks are allocated. The Middle Departure time of AGV Calculation, where This indicates the departure time of the first AGV in the dispatch process. This indicates the departure interval, thus generating the departure time. By using a type-alternating polling allocation method, AGVs of different types depart alternately according to their sequence numbers, avoiding the situation where multiple AGVs of the same type depart at the same time, which would cause other types of AGVs to violate the time window constraints and avoid congestion.

[0053] During the population building phase, the mask generated during the construction phase will be mixed with the mask. Task sequence The types generated in the interleaved polling allocation phase are respectively Combine them to construct There are 10 solutions; sort all solutions in ascending order according to their total energy consumption, and retain the top 1 / 10 solutions with the lowest total energy consumption to form a population of 100 solutions. The initial population is used to identify the solution with the lowest total energy consumption in the current population as the current optimal solution. By generating candidate solutions with a number 10 times the population size and selecting the best ones, the overall quality of the initial population is guaranteed.

[0054] Step 2, using The table stores the state-action pairs. Value, through The algorithm learns to adaptively select low-level heuristic operators; it uses solutions in the current population as parent solutions and executes the selected low-level heuristic operators to generate child solutions. The low-level heuristic operators mainly include seven types: intra-path insertion operator, inter-path insertion operator, adjacent swap operator, pancake flip operator, global greedy insertion operator, destructive reconstruction operator, and two-element optimization operator. The specific operation process is as follows:

[0055] a. The operation process of the path insertion operator is as follows: randomly select a task from the task sequence of an AGV of the same type and remove it from its current position. Then, randomly determine a new position in the task sequence and re-insert the selected task into the new position. The path insertion operator achieves local optimization of a single path by adjusting the task order within the task sequence of a single AGV.

[0056] b. The inter-path insertion operator operates as follows: It transfers tasks between the task sequences of two different AGVs of the same type. It identifies the first AGV with the largest cumulative task load and the second AGV with the smallest cumulative task load within the same type. A task is selected from the task sequence of the first AGV, removed, and then inserted into the task sequence of the second AGV. By transferring tasks between different AGVs, the inter-path insertion operator achieves load balancing among AGVs, helping to reduce energy waste caused by load imbalance.

[0057] c. The operation process of the adjacent exchange operator is as follows: randomly select a task sequence of an AGV of the same type, randomly select two adjacent tasks in the task sequence and exchange the positions of the two tasks to perform local perturbation. If the time window constraint or load constraint is violated after the exchange, the exchange operation is cancelled. The perturbation amplitude of the adjacent exchange operator is small and it is suitable for fine local adjustment of the current solution.

[0058] d. The operation process of the pancake flipping operator is as follows: randomly select the task sequence of an AGV in the same type of AGV, randomly select a split position in the task sequence, and rearrange all tasks before the split position in reverse order; the pancake flipping operator reverses part of the task sequence, changes the access order between multiple tasks, and has a large perturbation amplitude, which helps to escape local optima.

[0059] e. The operation process of the global greedy insertion operator is as follows: randomly remove one task from all task sequences of the same type of AGV as the task to be inserted; traverse each insertion position in all task sequences of the same type of AGV, calculate the total energy consumption after inserting the task to be inserted into each position, and insert the task to be inserted into the position that minimizes the increase in total energy consumption; the global greedy insertion operator achieves cross-path task redistribution by traversing all possible insertion positions and selecting the position that minimizes the increase in total energy consumption.

[0060] f. The operation process of the destructive-reconstruction operator is as follows: In the destruction phase, the number of tasks to be removed is determined according to the destruction rate, and removal is carried out according to either a random removal rule or a worst-case removal rule. The random removal rule involves randomly selecting a corresponding number of tasks from all task sequences of the same type of AGV for removal. The worst-case removal rule involves calculating the difference in total energy consumption before and after the removal of each task in all task sequences of the same type of AGV, and prioritizing the removal of tasks that reduce total energy consumption the most. In the reconstruction phase, the difference between the latest and earliest arrival times allowed for each removed task is calculated, and the removed tasks are arranged in ascending order of the difference. The removed tasks are then inserted sequentially into the positions in all task sequences of the same type of AGV that minimize the increase in total energy consumption. By removing some tasks first and then re-inserting them according to the principle of minimizing the increase in total energy consumption, the destructive-reconstruction operator can significantly adjust the structure of the solution and improve the algorithm's ability to escape local optima. It is worth noting that prioritizing the insertion of tasks with tighter time windows in the reconstruction phase can reserve more feasible insertion positions for tasks with looser time windows, thereby improving the feasibility of the reconstructed solution.

[0061] g. The operation process of the two-element optimization operator is as follows: randomly select an AGV from the same type of AGV, randomly select two non-adjacent positions in the task sequence of the selected AGV, and rearrange all tasks between the two positions in reverse order; the two-element optimization operator can effectively reduce the intersection and detour in the path by reversing a subsequence of the task sequence and changing the access order of tasks within the sub-segment.

[0062] The process of learning adaptive selection of low-level heuristic operators involves defining the seven low-level heuristic operators mentioned above as seven optional actions, and defining the energy consumption change of the population optimal solution as a state; using... Table storage value, Each row in the table corresponds to a state, each column corresponds to an action, and the intersection of the row and column stores the action selected for that state. value, All in the table The initial values ​​of all values ​​are set to 1; based on In the table Value, using A greedy strategy is used to select actions. Greedy strategy is based on probability Select the current state The action with the highest value, in probability Randomly select one action from the seven actions. The preset exploration probability parameters are used; the low-level heuristic operator corresponding to the selected action is executed, and the selected low-level heuristic operator is applied to each parent solution in the current population to generate the corresponding child solution; through The learning mechanism allows the algorithm to adaptively adjust its operator selection strategy based on the historical performance of each low-level heuristic operator at different search stages. In the early stages of the search, it tends to select operators with strong exploratory capabilities to expand the search scope, while in the later stages of the search, it tends to select refined operators to accelerate convergence.

[0063] Step 3: Perform constraint repair on the offspring solutions, and determine whether to accept the offspring solutions to update the current population according to the simulated annealing acceptance criterion; determine the current state based on the logarithmic change in total energy consumption of the optimal solutions of the parent and offspring populations, and update the population using the Bellman equation. The table corresponds to the state-action pairs value.

[0064] Specifically, the constraint repair process involves detecting whether any AGVs in the sub-solution violate the time window constraint in terms of arrival time. If so, the AGVs that violate the time window constraint are identified as objects to be adjusted. AGVs whose departure time is earlier than that of the objects to be adjusted and which do not violate the time window constraint themselves are defined as exchangeable objects. The departure time of the objects to be adjusted is swapped with the departure time of the exchangeable objects. This swapping operation is repeated until the arrival time of all AGVs satisfies the time window constraint, or no exchangeable objects exist. Constraint repair, by swapping departure times, corrects time window violations without changing the task sequence structure, thus preserving the optimization effect of the low-level heuristic operators on the solution to the greatest extent.

[0065] The simulated annealing acceptance criterion is as follows: if the total energy consumption of the child solution is less than the total energy consumption of the parent solution, then the child solution is accepted to replace the parent solution; if the total energy consumption of the child solution is less than the total energy consumption of the current optimal solution, then the child solution is updated to the current optimal solution; if the total energy consumption of the child solution is greater than or equal to the total energy consumption of the parent solution, then it is accepted with probability. The decision on whether to accept the offspring solution, among which Indicates child solution Total energy consumption Indicates parent generation solution Total energy consumption This indicates the current annealing temperature, used to control the acceptance of offspring solutions with high total energy consumption during population iteration. The larger the value, the greater the probability of accepting a sub-solution with higher total energy consumption. The smaller the value, the lower the probability of accepting a sub-solution with higher total energy consumption; the initial value of the annealing temperature is set to... That is, the initial annealing temperature; the annealing temperature is according to Update, in which This indicates the lower bound of the annealing temperature. Indicates the cooling coefficient. The value of is between 0 and 1. It is worth noting that in the early stage of the algorithm search, the annealing temperature is relatively high, and the algorithm is more likely to accept the sub-solution with higher total energy consumption, which is conducive to expanding the search space. As the iteration proceeds, the annealing temperature gradually decreases, and the algorithm tends to accept only the sub-solution with lower energy consumption, which is conducive to the algorithm converging towards the optimal solution.

[0066] The value update process involves calculating the total energy consumption of the optimal solution of the parent population in this iteration. Total energy consumption of the optimal solution for the offspring population logarithmic change And define the current state discretized; when When this occurs, it indicates deteriorating energy consumption, and the state value is 0. When the energy consumption remains unchanged, the state value is 1. and Values ​​ranging from 1 to 6 represent different degrees of energy efficiency improvement, with the state value being... ,when When the state represents a significant improvement in energy consumption, its value is 8; otherwise, its value is 9. By discretizing the states using logarithmic changes, it is possible to distinguish the degree of energy consumption improvement at different levels. The learning mechanism can select appropriate operators based on the characteristics of different search stages, based on the logarithmic change. For instant rewards Perform hierarchical assignment when hour, .when hour, .when and When the value range is 1 to 5, ,Right now The value range is from 5 to 9. Other cases... The tiered design of instant rewards makes... Learning enables fine-grained evaluation of the search performance of operators, rather than simple binary judgments.

[0067] Using Bellman equations right The corresponding table The value is updated, where, Indicates the current state. This represents the low-level heuristic operator corresponding to the action selected in this iteration. express Status in the table Select action Corresponding value; This indicates the next state after an action is performed. Indicates the candidate action in the next state. Indicates the next state All candidate actions correspond The maximum value; Indicates an immediate reward. Indicates the discount factor. This represents the learning rate. according to Dynamic calculation, Indicates the current iteration number. This indicates the maximum number of iterations; the learning rate is relatively large in the early stages of the algorithm, making... The table can quickly respond to the search performance of operators; as iterations proceed, the learning rate gradually decreases, making... The table tends to stabilize, avoiding interference from later fluctuations in operator selection.

[0068] Step 4: Perform an adaptive local search on the solutions in the current population and update the current population.

[0069] Specifically, adaptive local search includes three local search operations: knowledge-guided position adjustment, optimal type combination, and AGV merging.

[0070] The specific process of knowledge-guided position adjustment is as follows: sort the solutions in the current population according to their total energy consumption from smallest to largest, and select the solution ranked in the 1st place based on its total energy consumption. Ranked first The solutions of each position are selected as the operation objects, that is, the latter half of the solutions with higher total energy consumption in the population are selected as the operation objects; a random number in the interval [0,1] is generated for each operation object, and the probability is calculated. If the random number is less than If the position of the object being operated on is adjusted, then the operation is not performed; otherwise, it is not performed. This indicates the time elapsed for normalization. This indicates the time elapsed since the algorithm ran. This represents the total running time; according to the probability formula, in the initial stage of the algorithm's operation... Smaller A value close to 1 indicates a higher probability of position adjustment, which is beneficial for quickly improving the quality of poorer solutions by utilizing the structural information of the current optimal solution; as the algorithm's running time increases, Gradually increase The probability of performing position adjustments is gradually reduced, thus preventing the population from prematurely converging to similar solution structures. If position adjustments are performed, the difference between the scheduling component energy consumption of delivery AGVs, recovery AGVs, and detection AGVs in the operation object and the scheduling component energy consumption of the corresponding type of AGV in the current optimal solution is calculated. The type of AGV with the largest difference is selected, and the normalized position ratio of all tasks of the selected type of AGV in the task sequence is extracted from the current optimal solution. The normalized position ratio is the ratio of the position of the task in the task sequence to the total length of the sequence. Each task in the task sequence of the corresponding type of AGV in the operation object is adjusted to a position consistent with the normalized position ratio of the same task in the current optimal solution.

[0071] The specific process of optimal type combination is as follows: First, identify the scheduling component with the lowest energy consumption among the scheduling components of delivery AGVs, recovery AGVs, and detection AGVs in the current population, and denote it as the current optimal scheduling component for the corresponding AGV type. Then, construct a new solution from the current optimal scheduling component for the corresponding AGV type, denoted as the combined solution. Regenerate the departure time for the combined solution; if the AGV arrival time in the combined solution violates the time window constraint, perform constraint repair. Calculate the total energy consumption of the combined solution; if the total energy consumption of the combined solution is less than the total energy consumption of the optimal solution in the current population, replace the solution with the highest total energy consumption in the current population with the combined solution. Optimal type combination utilizes the advantages of different solutions on different AGV types, combining the best-performing scheduling components of each type, which is beneficial for constructing a solution with lower overall energy consumption.

[0072] The specific process of AGV merging is as follows: For each solution in the current population, denoted as the solution to be operated on, an AGV type is randomly selected from the solutions to be operated on. Among the AGVs of the selected type, the AGV with the fewest tasks is identified as the object to be merged. Attempts are made to distribute all tasks in the subsequence of the object to be merged to the subsequences of other AGVs of the selected type. If all AGVs of the selected type satisfy the load constraints after the distribution, the subsequence corresponding to the object to be merged is deleted and the scheduling component of the selected type is updated. The updated scheduling component is combined with the scheduling components of the other two types in the solution to be operated on to construct a new solution, and the departure time is regenerated for the new solution. If there are cases in the new solution where the AGV arrival time violates the time window constraint, constraint repair is performed. AGV merging can reduce the startup energy consumption of AGVs by reducing the number of AGVs used, and further optimize the total energy consumption.

[0073] Step 5: Determine if the algorithm termination condition has been met. If it has, the algorithm ends, outputting the current optimal solution and its corresponding total energy consumption value; otherwise, return to Step 2. Specifically, the algorithm termination time is fixed at 5 seconds.

[0074] This invention presents a multi-type AGV scheduling method for matrix workshops based on a hyperheuristic algorithm, namely RL-HHEA. It addresses the collaborative scheduling needs of three functionally heterogeneous AGVs—delivery, recycling, and detection—in a matrix workshop, with the optimization objective of minimizing total energy consumption. Through a type-separated dual-component encoding scheme and a mask-based type-interleaved polling initialization method, a high-quality initial population that balances diversity and feasibility is generated. The system learns to adaptively select among low-level heuristic operators and coordinates the balance between exploration and exploitation by incorporating simulated annealing acceptance criteria. It employs a fine-grained search by combining three local search operations: knowledge-guided position adjustment, optimal type combination, and AGV merging, leveraging structural information from the current best solution in the population to accelerate convergence.

[0075] The present invention will be further described below by listing specific embodiments.

[0076] This invention runs on a computer equipped with an Intel(R) Core(TM) i7-9700 processor and an NVIDIA GeForce RTX 2060 GPU. The RL-HHEA algorithm is implemented in C++, and the total running time is set to 5N (N is the total number of instances) seconds. All comparison algorithms are executed 10 times on each instance.

[0077] This embodiment applies to a matrix manufacturing workshop with a warehouse and multiple AGVs of various types. The workshop simultaneously deploys AGVs of three functional types: delivery, recycling, and detection. The goal is to determine a scheduling scheme that reduces total energy consumption, which includes AGV driving energy consumption, startup energy consumption, and task execution energy consumption. Figure 2 As shown, the matrix manufacturing workshop consists of a warehouse, production units, tracks, and various types of AGVs. The warehouse is located in the lower left corner of the workshop. Multiple production units are evenly distributed along a track grid, with the tracks intersecting horizontally and vertically to form the AGV's travel paths. The workshop simultaneously deploys three types of AGVs: delivery AGVs, recycling AGVs, and inspection AGVs. Delivery AGVs are responsible for material delivery, recycling AGVs for product retrieval, and inspection AGVs for quality inspection. Each type of AGV travels back and forth between the warehouse and the production units along the tracks.

[0078] The warehouse coordinates are (0,0), and the horizontal and vertical distances between each calling unit are 5.5m and 8.8m, respectively. All three types of AGVs have a speed of 1m / s and a task execution time of 15s. The delivery AGV has a maximum load capacity of 250kg, a basic energy consumption coefficient of 10.0, a startup energy consumption of 50.0, and a task execution energy consumption of 5.0. The recycling AGV has a maximum load capacity of 300kg, a basic energy consumption coefficient of 12.0, a startup energy consumption of 60.0, and a task execution energy consumption of 6.0. The detection AGV has no load requirement, a basic energy consumption coefficient of 8.0, a startup energy consumption of 40.0, and a task execution energy consumption of 7.0. The load influence coefficient is 0.048, the driving energy consumption weight is 1.0, the startup energy consumption weight is 60.0, and the execution energy consumption weight is 1.0. The energy consumption-related parameters mentioned above (basic energy consumption coefficient, startup energy consumption, and task execution energy consumption) are all dimensionless normalized coefficients used to characterize the relative energy consumption levels of different types of AGVs during driving, startup, and task execution. Total energy consumption is calculated by weighted summation of each energy consumption coefficient and its corresponding operating parameter, and is a dimensionless value. The first AGV departs at 365 seconds, and the AGV departure interval is 2 seconds. A type-alternating polling method is used, with delivery, recycling, and detection types departing alternately. The material buffer capacity is 48 blocks, the new material storage area capacity is 50 blocks, the material consumption rate is 1 / 30 block / second, the new material production rate is 1 / 60 block / second, the unit weight of the material is 0.75 kg, and the unit weight of the new material is 1.05 kg. Specifically, the matrix manufacturing workshop example data table of this invention is shown in Table 1:

[0079] Table 1. Calculation Data Table of the Matrix Manufacturing Workshop of the Present Invention

[0080] Call unit number X-axis / m Y-axis / m Task type Task call time / s Material buffer and material storage area inventory / blocks during call Task completion time / s 9 5.5 79.2 1 155 28 755 101 55 8.8 1 329 29 959 91 44 96.8 1 301 30 961 118 60.5 70.4 1 220 28 820 86 44 52.8 1 288 30 948 82 44 17.6 1 188 28 788 90 44 88 1 142 28 742 23 11 26.4 1 318 29 948 87 44 61.6 1 78 29 708 112 60.5 17.6 1 228 30 888 15 5.5 44 1 153 30 813 88 44 70.4 1 122 29 752 101 49.5 96.8 1 205 28 805 50 22 88 1 272 28 872 29 11 79.2 1 347 30 1007 34 16.5 35.2 1 181 30 841 105 55 44 1 178 30 838 32 16.5 17.6 1 211 28 811 113 60.5 26.4 1 132 29 762 14 5.5 35.2 1 258 28 858 16 5.5 52.8 2 173 34 893 54 27.5 35.2 2 229 36 829 28 11 70.4 2 227 36 827 27 11 61.6 2 131 35 791 51 27.5 8.8 2 22 34 742 110 55 88 2 218 36 818 97 49.5 61.6 2 115 36 715 115 60.5 44 2 310 36 910 22 11 17.6 2 55 34 775 39 16.5 79.2 2 340 34 1060 66 33 52.8 3 127 45 727 63 33 26.4 3 144 46 774 104 55 35.2 3 274 45 874 67 33 61.6 3 12 46 642 60 27.5 88 3 149 46 779 75 38.5 44 3 157 46 787 45 22 44 3 12 46 642 59 27.5 79.2 3 264 46 894 72 38.5 17.6 3 280 45 880 55 27.5 44 3 280 46 910

[0081] Table 1 records the relevant data of task requests sent by each calling unit in the matrix manufacturing workshop to the control system within a certain time period, including the calling unit number, location coordinates (X-axis / m, Y-axis / m), task type, task call time, buffer inventory at the time of call, and latest task completion time. Task type 1 represents a delivery task, executed by a delivery AGV; task type 2 represents a retrieval task, executed by a retrieval AGV; and task type 3 represents a detection task, executed by a detection AGV. The unit for both the task call time and the latest task completion time is seconds, which together constitute the time window constraint for the corresponding task.

[0082] For the RL-HHEA algorithm of this invention, there are 5 key parameters: population size , - Exploration probability parameter in greedy strategies , Discount Factors in Learning Initial temperature in simulated annealing and the destruction rate in the destruction reconstruction operator Population size The unit for is 1; the other parameters are dimensionless values ​​and have no units. For each parameter, select one of the following three levels: , , , , A full factorial experimental design was used to analyze parameter effects and interactions, with 3 levels for each of the 5 factors, resulting in 3... 5 =243 parameter combinations. Each parameter combination is executed independently 10 times on the calibration set, and the average RPD is calculated as the response variable. The formula for RPD is RPD. ,in This represents the total energy consumption value obtained from the current algorithm. This represents the optimal total energy consumption among all compared algorithms. A smaller RPD value indicates higher solution quality. Optimal parameters are determined using the main effects plot, as shown below. Figure 3 As shown in the main effects diagram, analysis reveals that... The inflection point is reached at 120, indicating that a moderate population size achieves the optimal balance between solution diversity and computational efficiency. It shows a monotonically increasing trend. The fact that the minimum RPD is achieved at this time indicates that a smaller exploration probability is advantageous. Learning mechanisms make better use of already learned knowledge; Achieving an inflection point at 0.3 allows for a proper balance between immediate and long-term rewards; The differences between the three levels are relatively small. The time efficiency is slightly better, indicating that the algorithm has strong robustness to the initial temperature setting; An inflection point of 0.2 is achieved, providing an effective balance between preserving a good solution structure and achieving sufficient neighborhood exploration. Based on the above scores, the optimal parameter configuration is finally determined as follows: indivual, , , , In subsequent experiments, the optimal parameter configurations were used for each task size.

[0083] To comprehensively evaluate the performance of the RL-HHEA algorithm of this invention, test case sets with four different task sizes were generated based on the actual production scenario of a matrix manufacturing workshop. The four task sizes are 40, 60, 80, and 100 tasks, with the number of delivery, retrieval, and inspection tasks under each size being (20, 10, 10), (30, 20, 10), (30, 30, 20), and (40, 40, 20), respectively. The initial available number of each type of AGV is configured according to the number of tasks of the corresponding type. During the optimization process, the algorithm dynamically reduces the actual number of AGVs used through AGV merging operations, ensuring that the actual number of AGVs used for each type in the final scheduling scheme is determined as part of the optimization result. Each size contains 20 test case instances, for a total of 80 test cases. The naming convention for the test cases is as follows: ,in Indicates the number of tasks. Indicates the example number.

[0084] To demonstrate the effectiveness of each component strategy of this invention, an ablation experiment is conducted to compare the performance differences between the RL-HHEA algorithm of this invention and four variant algorithms that remove specific components. The four variant algorithms are: RL-HHEA_NI represents a variant algorithm that removes the mask-based type interleaved polling initialization method of this invention and replaces it with random initialization; RL-HHEA_NQ represents a variant algorithm that removes the mask-based type interleaved polling initialization method of this invention and replaces it with random initialization; The algorithm employs an adaptive operator selection mechanism and a variant algorithm that randomly selects seven low-level heuristic operators for replacement; RL-HHEA_NK represents a variant algorithm that removes the knowledge-guided position adjustment of this invention; RL-HHEA_NT represents a variant algorithm that removes the optimal type combination of this invention. The comparison results of the RL-HHEA algorithm of this invention with the four variant algorithms on a task size of 40 are shown in Table 2.

[0085] Table 2. Comparison results of the RL-HHEA algorithm of the present invention with four variant algorithms on a task size of 40.

[0086] Calculation example RL-HHEA (RPD) RL-HHEA_NI (RPD) RL-HHEA_NQ (RPD) RL-HHEA_NK (RPD) RL-HHEA_NT (RPD) J40I1 2.01% 13.57% 3.42% 1.20% 1.95% J40I2 0.88% 17.21% 2.46% 0.69% 0.82% J40I3 15.97% 19.44% 14.99% 11.23% 9.17% J40I4 2.88% 19.83% 10.46% 8.11% 6.43% J40I5 10.53% 18.53% 13.21% 9.12% 8.19% J40I6 2.63% 16.07% 11.47% 7.22% 5.63% J40I7 2.12% 17.04% 10.04% 5.06% 1.83% J40I8 1.55% 13.74% 4.18% 1.14% 1.09% J40I9 0.59% 18.12% 13.14% 9.54% 9.59% J40I10 2.57% 18.19% 2.50% 0.96% 2.87% J40I11 0.67% 14.94% 10.97% 5.78% 7.06% J40I12 3.27% 15.04% 9.30% 7.65% 6.85% J40I13 1.19% 17.78% 13.14% 9.19% 8.25% J40I14 3.24% 16.44% 10.19% 8.67% 7.39% J40I15 7.70% 14.13% 9.33% 6.98% 7.86% J40I16 2.57% 16.73% 3.35% 1.44% 1.94% J40I17 2.78% 19.95% 13.22% 10.66% 8.34% J40I18 0.43% 15.30% 10.07% 8.08% 8.57% J40I19 2.09% 17.21% 9.22% 7.92% 7.40% J40I20 4.45% 20.77% 4.52% 1.62% 3.82% average value 3.51% 17.00% 8.96% 6.11% 5.75%

[0087] The comparison results of the RL-HHEA algorithm of this invention with four variant algorithms on a task size of 60 are shown in Table 3:

[0088] Table 3. Comparison results of the RL-HHEA algorithm of the present invention with four variant algorithms on a task size of 60.

[0089] Calculation example RL-HHEA (RPD) RL-HHEA_NI (RPD) RL-HHEA_NQ (RPD) RL-HHEA_NK (RPD) RL-HHEA_NT (RPD) J60I1 7.32% 14.92% 17.78% 7.14% 8.35% J60I2 14.54% 20.40% 22.21% 15.02% 12.23% J60I3 5.22% 14.03% 14.33% 5.32% 5.59% J60I4 0.50% 21.57% 6.57% 0.72% 1.08% J60I5 1.19% 10.81% 6.44% 1.43% 3.07% J60I6 8.54% 14.18% 17.09% 8.90% 7.82% J60I7 8.94% 15.66% 16.93% 9.53% 8.68% J60I8 3.43% 14.40% 14.65% 3.87% 4.82% J60I9 6.42% 14.33% 19.04% 7.82% 11.81% J60I10 6.92% 13.14% 14.29% 7.06% 7.09% J60I11 4.54% 14.25% 7.31% 4.58% 4.13% J60I12 0.86% 11.57% 2.16% 1.11% 0.67% J60I13 3.55% 25.29% 18.05% 4.28% 11.79% J60I14 3.47% 15.01% 14.24% 3.31% 5.19% J60I15 5.71% 24.04% 7.01% 5.71% 5.84% J60I16 7.00% 17.88% 15.12% 6.95% 7.22% J60I17 17.99% 24.40% 25.46% 18.85% 16.59% J60I18 9.55% 18.09% 18.60% 10.16% 10.24% J60I19 6.62% 23.15% 16.06% 7.97% 9.69% J60I20 10.34% 21.90% 25.55% 11.98% 13.09% average value 6.63% 17.45% 14.94% 7.09% 7.75%

[0090] The comparison results of the RL-HHEA algorithm of this invention with four variant algorithms on a task size of 80 are shown in Table 4:

[0091] Table 4. Comparison results of the RL-HHEA algorithm of the present invention with four variant algorithms on a task size of 80.

[0092] Calculation example RL-HHEA (RPD) RL-HHEA_NI (RPD) RL-HHEA_NQ (RPD) RL-HHEA_NK (RPD) RL-HHEA_NT (RPD) J80I1 2.94% 8.76% 9.98% 3.19% 4.86% J80I2 3.71% 11.80% 12.43% 3.87% 5.29% J80I3 5.72% 11.47% 16.56% 6.00% 6.67% J80I4 2.52% 10.35% 8.67% 2.77% 4.90% J80I5 2.86% 11.52% 10.54% 3.08% 4.50% J80I6 2.56% 11.80% 12.25% 2.86% 4.23% J80I7 3.18% 8.96% 13.34% 3.34% 4.24% J80I8 2.88% 9.04% 11.43% 3.34% 5.02% J80I9 2.48% 10.92% 9.86% 2.94% 4.90% J80I10 2.65% 10.66% 9.71% 2.86% 3.85% J80I11 8.93% 20.00% 18.90% 9.30% 12.57% J80I12 10.18% 17.04% 19.47% 9.85% 11.63% J80I13 6.97% 12.28% 15.02% 7.60% 6.30% J80I14 2.88% 12.03% 11.46% 3.36% 5.14% J80I15 6.34% 15.00% 16.98% 6.48% 5.34% J80I16 2.19% 9.39% 11.66% 2.68% 4.27% J80I17 3.63% 20.50% 13.01% 4.25% 3.95% J80I18 1.93% 10.62% 11.53% 2.58% 4.63% J80I19 2.92% 9.58% 12.96% 3.40% 4.02% J80I20 0.61% 8.58% 8.43% 0.81% 3.16% average value 3.90% 12.02% 12.71% 4.23% 5.47%

[0093] The comparison results of the RL-HHEA algorithm of this invention with four variant algorithms on a task size of 100 are shown in Table 5:

[0094] Table 5. Comparison results of the RL-HHEA algorithm of the present invention with four variant algorithms on a task size of 100.

[0095] Calculation example RL-HHEA (RPD) RL-HHEA_NI (RPD) RL-HHEA_NQ (RPD) RL-HHEA_NK (RPD) RL-HHEA_NT (RPD) J100I1 4.21% 17.99% 16.87% 4.48% 6.96% J100I2 2.99% 15.44% 12.33% 3.93% 4.92% J100I3 2.50% 10.14% 16.34% 3.56% 5.91% J100I4 1.96 18.89% 13.60% 2.88% 6.66% J100I5 3.81% 17.52% 12.77% 3.71% 4.85% J100I6 6.26% 15.81% 16.25% 6.15% 6.87% J100I7 2.49% 14.13% 10.92% 2.69% 4.04% J100I8 2.92% 20.70% 14.78% 3.64% 5.36% J100I9 3.08% 22.11% 8.59% 3.72% 4.39% J100I10 2.70% 17.43% 13.31% 4.08% 5.81% J100I11 1.72% 16.18% 13.29% 2.43% 4.38% J100I12 7.85% 21.14% 20.91% 8.72% 9.98% J100I13 4.37% 16.19% 17.23% 4.63% 5.66% J100I14 2.85% 19.93% 14.01% 3.62% 5.70% J100I15 4.51% 21.34% 14.68% 4.97% 7.29% J100I16 1.98% 20.84% 10.14% 2.28% 4.04% J100I17 2.77% 16.88% 11.16% 3.10% 4.23% J100I18 3.81% 20.32% 19.73% 4.08% 5.37% J100I19 4.68% 15.95% 17.21% 3.65% 6.82% J100I20 3.88% 19.64% 13.97% 5.76% 5.67% average value 3.57% 17.93% 14.40% 4.10% 5.75%

[0096] The values ​​corresponding to each example in the table represent the average RPD obtained by repeating the example 30 times. The last row of average values ​​represents the average RPD of the RL-HHEA algorithm and its variants across all 20 test examples. The data in the table shows that in 80 test examples across four task sizes (40, 60, 80, and 100), the average RPD of the RL-HHEA algorithm of this invention is 3.51%, 6.63%, 3.90%, and 3.57%, respectively, all lower than the corresponding values ​​of the four variant algorithms. Specifically, the average RPD of the RL-HHEA_NI variant is 17.00%, 17.45%, 12.02%, and 17.93%, respectively, performing the worst among all variants. Compared to the RL-HHEA algorithm of this invention, the average RPD increases by 13.49%, 10.82%, 8.12%, and 14.36% across the four task sizes, indicating that the mask-based type-interleaved polling initialization method plays a crucial role in generating a high-quality initial population. The average RPD of the RL-HHEA_NQ variants were 8.96%, 14.94%, 12.71%, and 14.40%, respectively. Compared with the RL-HHEA algorithm of this invention, the average RPD increased by 5.45%, 8.31%, 8.81%, and 10.83% for the four task sizes, respectively, and the difference widened with increasing task size, indicating that the RPD based on... The adaptive operator selection mechanism developed through learning contributes more significantly to large-scale problems. The average RPD of the RL-HHEA_NK variant is 6.11%, 7.09%, 4.23%, and 4.10%, respectively, while that of the RL-HHEA_NT variant is 5.75%, 7.75%, 5.47%, and 5.75%, respectively. These values ​​are relatively close to the RL-HHEA algorithm of this invention, but they are still inferior to the RL-HHEA algorithm of this invention across all task scales, indicating that knowledge-guided position adjustment and optimal type combination are equally indispensable for improving solution quality. These results demonstrate that each component proposed in this invention is indispensable for achieving optimal performance, and the synergistic effect among these components enables the RL-HHEA algorithm of this invention to efficiently solve the total energy consumption minimization scheduling problem for multiple types of AGVs in a matrix workshop.

[0097] Box plots comparing the RPD of the RL-HHEA algorithm of this invention with four variant algorithms are shown below. Figure 4 As shown in the box plot, RPD represents Relative Percentage Deviation. The box plot reveals that the RL-HHEA algorithm has the most compact box. The upper boundary of the data distribution range represents the maximum value (upper whisker), and the lower boundary represents the minimum value (lower whisker). Both whiskers are positioned relatively low. The horizontal line inside the box represents the median. The RL-HHEA algorithm exhibits a concentrated RPD data distribution and small extreme value differences, indicating that the RL-HHEA algorithm of this invention significantly outperforms the other four variant algorithms in terms of solution quality and robustness, demonstrating superior median performance and lower outlier performance, resulting in outstanding overall stability. The RL-HHEA_NI variant has the highest median and the largest distribution range, confirming that the mask-based type-interleaved polling initialization strategy is crucial for generating high-quality initial solutions. The RL-HHEA_NQ variant has a moderate median but high variability, indicating that the random operator selection leads to unstable performance of the algorithm on different problem instances.

[0098] Ablation experiments show that each component proposed in this invention is indispensable for achieving optimal performance, and the synergistic effect between the components enables the algorithm to efficiently solve the scheduling problem of minimizing the total energy consumption of multiple types of AGVs in the matrix workshop.

[0099] To further verify the overall superiority of the RL-HHEA algorithm proposed in this invention, it was compared with four existing comparative algorithms: Multi-Type Discrete Artificial Bee Colony (MTDABC), Parallel Evolutionary Algorithm (PEA), Variable Neighborhood Search (VNS), and Adaptive Iterative Greedy Algorithm (SAIG). All comparative algorithms were run under the same experimental conditions, with a maximum single-run time of 5 seconds for each algorithm on each example. The comparison results are shown in Tables 6-9. Table 6 shows the comparison results of the RL-HHEA algorithm of this invention with the MTDABC, PEA, VNS, and SAIG ​​algorithms on a task size of 40.

[0100] Table 6. Comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 40.

[0101] Calculation example RL-HHEA (RPD) MTDABC (RPD) PEA (RPD) VNS (RPD) SAIG (RPD) J40I1 1.72% 6.11% 11.54% 15.64% 1.10% J40I2 0.87% 5.10% 13.14% 16.11% 0.59% J40I3 14.70% 8.17% 16.48% 26.60% 9.23% J40I4 2.88% 9.91% 11.18% 11.76% 7.08% J40I5 9.80% 10.57% 15.68% 28.26% 8.48% J40I6 2.51% 6.44% 16.31% 25.14% 7.12% J40I7 2.03% 2.77% 10.69% 12.87% 5.97% J40I8 1.43% 2.48% 4.12% 4.28% 1.13% J40I9 0.59% 10.79% 25.45% 27.82% 9.42% J40I10 2.57% 9.14% 14.46% 17.30% 0.98% J40I11 0.48% 1.58% 12.91% 13.90% 5.23% J40I12 3.51% 6.69% 19.67% 24.79% 7.91% J40I13 0.35% 2.03% 10.85% 11.78% 8.40% J40I14 2.68% 2.59% 7.58% 10.54% 7.97% J40I15 7.22% 9.06% 10.16% 11.30% 6.54% J40I16 1.97% 3.57% 14.67% 18.03% 1.02% J40I17 1.51% 6.89% 10.87% 4.29% 9.74% J40I18 0.43% 9.11% 10.12% 10.96% 8.08% J40I19 2.09% 9.09% 10.67% 11.19% 6.76% J40I20 4.45% 18.06% 13.95% 19.27% 1.55% average value 3.19% 7.01% 13.02% 16.09% 5.72%

[0102] The comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 60 are shown in Table 7:

[0103] Table 7. Comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 60.

[0104] Calculation example RL-HHEA (RPD) MTDABC (RPD) PEA (RPD) VNS (RPD) SAIG (RPD) J60I1 6.81% 14.85% 21.24% 23.11% 4.92% J60I2 14.54% 13.26% 18.24% 15.96% 9.40% J60I3 5.22% 10.34% 20.21% 22.44% 3.90% J60I4 0.50% 10.51% 10.76% 10.74% 9.48% J60I5 5.48% 10.53% 12.04% 12.31% 5.75% J60I6 8.54% 16.21% 18.69% 22.19% 5.25% J60I7 7.58% 14.59% 17.58% 22.82% 2.98% J60I8 3.89% 16.45% 19.04% 21.75% 5.52% J60I9 5.78% 16.76% 19.45% 18.54% 14.79% J60I10 6.92% 9.02% 14.09% 14.59% 5.43% J60I11 3.36% 5.27% 6.09% 5.92% 1.85% J60I12 0.86% 9.55% 11.54% 5.12% 2.94% J60I13 3.55% 18.29% 19.88% 19.19% 11.58% J60I14 4.32% 10.44% 12.77% 11.29% 0.93% J60I15 5.71% 19.32% 21.29% 21.33% 13.71% J60I16 5.32% 14.75% 16.29% 20.66% 9.83% J60I17 5.73% 11.03% 14.26% 8.66% 2.37% J60I18 9.37% 13.18% 23.92% 14.46% 6.91% J60I19 6.62% 24.07% 22.44% 26.74% 13.64% J60I20 10.09% 18.63% 22.60% 23.38% 15.06% average value 6.01% 13.85% 17.12% 17.06% 7.31%

[0105] The comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 80 are shown in Table 8:

[0106] Table 8. Comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 80.

[0107] Calculation example RL-HHEA (RPD) MTDABC (RPD) PEA (RPD) VNS (RPD) SAIG (RPD) J80I1 0.30% 1.70% 2.25% 1.52 2.25% J80I2 2.40% 7.70% 7.97% 8.83% 1.31% J80I3 3.52% 8.65% 6.59% 13.09% 3.26% J80I4 2.48% 8.54% 8.55% 9.70% 1.79% J80I5 2.43% 5.09% 6.34% 9.18% 1.30% J80I6 1.40% 8.06% 8.33% 10.02% 1.49% J80I7 2.27% 1.04% 2.54% 2.38% 1.71% J80I8 2.88% 8.74% 6.47% 3.15% 1.72% J80I9 2.48% 7.15% 9.37% 5.19% 2.02% J80I10 1.81% 5.01% 6.01% 7.96% 2.00% J80I11 5.14% 16.86% 16.85% 19.72% 10.78% J80I12 2.79% 6.88% 5.35% 4.06% 2.96% J80I13 1.30% 5.01% 6.94% 7.24% 1.21% J80I14 2.88% 2.71% 4.13% 3.96% 2.42% J80I15 6.34% 7.12% 8.12% 8.06% 7.54% J80I16 2.00% 6.97% 6.56% 8.71% 1.80% J80I17 3.63% 17.35% 18.62% 20.14% 13.04% J80I18 1.22% 2.35% 5.51% 7.98% 1.28% J80I19 2.14% 7.41% 7.83% 10.46% 2.56% J80I20 2.91% 4.69% 8.26% 8.34% 3.23% average value 2.62% 6.95% 7.63% 8.48% 3.28%

[0108] The comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 100 are shown in Table 9:

[0109] Table 9. Comparison results of the RL-HHEA algorithm of the present invention with the MTDABC algorithm, PEA algorithm, VNS algorithm, and SAIG ​​algorithm on a task size of 100.

[0110] Calculation example RL-HHEA (RPD) MTDABC (RPD) PEA (RPD) VNS (RPD) SAIG (RPD) J100I1 4.21% 7.45% 8.23% 8.22% 6.96% J100I2 2.99% 14.33% 9.73% 17.26% 6.55% J100I3 2.50% 10.17% 8.27% 12.88% 0.47% J100I4 1.96% 5.99% 7.10% 6.25% 6.00% J100I5 3.81% 12.24% 13.09% 15.95% 6.82% J100I6 3.46% 9.05% 7.63% 9.75% 3.77% J100I7 2.15% 12.03% 8.31% 12.02% 6.63% J100I8 2.92% 14.47% 12.58% 17.86% 10.24% J100I9 3.08% 15.05% 16.81% 14.88% 13.81% J100I10 2.70% 9.70% 9.89% 13.66% 6.79% J100I11 1.72% 16.43% 13.94% 18.73% 8.10% J100I12 7.85% 13.37% 12.37% 10.94% 8.30% J100I13 4.37% 12.19% 11.62% 12.44% 10.48% J100I14 2.85% 6.41% 7.04% 6.46% 5.72% J100I15 4.51% 15.51% 12.36% 19.22% 8.55% J100I16 1.98% 8.87% 9.31% 9.13% 8.27% J100I17 2.77% 10.22% 10.33% 9.02% 7.56% J100I18 3.81% 11.23% 11.74% 11.70% 10.01% J100I19 4.44% 6.72% 9.07% 8.53% 4.13% J100I20 3.88% 14.83% 13.18% 18.66% 9.27% average value 3.40% 11.31% 10.63% 12.68% 7.42%

[0111] The values ​​corresponding to each example in the table are the average RPD values ​​obtained by repeating the example 30 times. The last row of average values ​​is the average RPD value of the corresponding algorithm across all 20 test examples. According to Tables 6 to 9, in 80 test examples across four task sizes (40, 60, 80, and 100), the average RPD values ​​of the RL-HHEA algorithm of this invention are 3.19%, 6.01%, 2.62%, and 3.40%, respectively. It achieves the best results across all task sizes, significantly outperforming the MTDABC algorithm (7.01%, 13.85%, 6.95%, 11.31%), PEA algorithm (13.02%, 17.12%, 7.63%, 10.63%), VNS algorithm (16.09%, 17.06%, 8.48%, 12.68%), and SAIG ​​algorithm (5.72%, 7.31%, 3.28%, 7.42%). Among them, the RL-HHEA algorithm of this invention, compared with the SAIG ​​algorithm which has the closest performance, reduces the average RPD by 2.53%, 1.30%, 0.66%, and 4.02% across four task sizes, respectively. As the task size increases from 40 to 100, the advantage of the RL-HHEA algorithm of this invention compared with the comparative algorithms generally shows an increasing trend, indicating that the proposed algorithm based on… The adaptive operator selection mechanism learned, along with the synergistic effect of three local search operations—knowledge-guided position adjustment, optimal type combination, and AGV merging—enables the RL-HHEA algorithm of this invention to have stronger problem-solving capabilities for problems of different scales.

[0112] The scatter plot comparing the average RPD and 95% confidence interval of the RL-HHEA algorithm of this invention with those of the MTDABC, PEA, VNS, and SAIG ​​algorithms is shown below. Figure 5 As shown in the figure, each scatter point represents the average RPD of the corresponding algorithm across all test cases, and the error bars represent the 95% confidence interval. The scatter plot shows that the average RPD of the RL-HHEA algorithm is significantly lower than that of the SAIG, MTDABC, PEA, and VNS algorithms, and the 95% confidence interval for the RL-HHEA algorithm is the most compact. This indicates that the RL-HHEA algorithm of this invention significantly outperforms the other four comparative algorithms in terms of solution quality, while also exhibiting the highest stability and consistency.

[0113] In summary, the RL-HHEA algorithm of this invention achieves its goals by incorporating a dual-component coding scheme with type separation, a mask-based type-interleaved polling initialization method, and a... The adaptive low-level heuristic operator selection mechanism, simulated annealing acceptance criterion, and three local search strategies—knowledge-guided position adjustment, optimal type combination, and AGV merging—ensure the diversity and quality of the initial population, the balance between exploration and development, and the fine-grained search capability of the solution. This effectively reduces the total energy consumption of heterogeneous AGVs in the matrix workshop and improves the collaborative scheduling efficiency and solution stability of multiple types of AGVs.

Claims

1. A matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm, characterized in that, Includes the following steps, Step 1: Initialize algorithm parameters. Construct the solution using a type-separated two-component encoding scheme and its corresponding decoding scheme, and generate the initial population based on the mask type-interleaved polling initialization method. Step 2, using The table stores the state-action pairs. Value, through Learn to adaptively select low-level heuristic operators; use the solutions in the current population as parent solutions, and execute the selected low-level heuristic operators to generate child solutions; Step 3: Perform constraint repair on the offspring solutions, and determine whether to accept the offspring solutions to update the current population according to the simulated annealing acceptance criterion; determine the current state based on the logarithmic change in total energy consumption of the optimal solutions of the parent and offspring populations, and update the population using the Bellman equation. The table corresponds to the state-action pairs value; Step 4: Perform an adaptive local search on the solutions in the current population and update the current population; Step 5: Determine if the algorithm termination condition has been met. If it has, the algorithm ends and the current optimal solution is output; otherwise, return to step 2.

2. The matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 1, characterized in that, In step 1, the type-separated two-component encoding scheme represents the solution as a triple, which consists of three scheduling components corresponding to delivery AGVs, retrieval AGVs, and detection AGVs, respectively; an AGV type index is defined. ,in Indicates delivery type. Indicates recycling type. This indicates a detection-type scheduling component; each scheduling component includes a task sequence. and departure time Task sequence By type The task sequence is arranged by all tasks waiting to be served by the AGV, and divided into several subsequences using 0 as the separator. Each subsequence is a type. The tasks undertaken by each AGV; departure time Record type Departure time and type of each AGV The Middle The departure time of each AGV is expressed as follows: .

3. The matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 2, characterized in that, In step 1, the decoding process for the solution represented by the type-separated dual-component encoding scheme is as follows: Each task in the matrix workshop has an allowed earliest arrival time and a allowed latest arrival time. A time window constraint is defined, meaning the allowed earliest arrival time and the allowed latest arrival time together constitute the time window for the corresponding task. The arrival time of each AGV must be within the time window. A load constraint is defined, meaning the cumulative load of all tasks on each AGV path must not exceed the maximum carrying capacity of the corresponding AGV. For type... AGVs, from the task sequence Get type The task sequence of each AGV, starting from the departure time Get type Departure time of each AGV Starting with the departure time of each AGV, calculate the arrival time, service duration, and cumulative load status of each task sequentially according to the task order in the task sub-sequence; then classify the types... The energy consumption of all AGVs in the system is accumulated, including the energy consumption of driving, starting, and task execution. The total energy consumption of the scheduling scheme is obtained by summing the energy consumption of each of the three types of AGVs.

4. The matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 3, characterized in that, In step 1, the mask-based type-interleaved polling initialization method includes a mask mixing construction phase, a type-interleaved polling allocation phase, and a population construction phase, wherein... In the mask mixing construction stage, define To determine the population size, a construction operator selection mechanism is used to select one construction operator for each type of AGV from both nearest-neighbor and random construction operators. This mechanism includes a mask sequence consisting of three binary digits: 000, 001, 010, 011, 100, 101, 110, and 111. The first, second, and third digits of the mask sequence correspond to detection-type AGVs, retrieval-type AGVs, and delivery-type AGVs, respectively. When the corresponding bit value in the mask sequence is 1, the nearest-neighbor construction operator is used for that type of AGV; when the corresponding bit value is 0, the random construction operator is used. The mask sequence is traversed to determine the construction operator for each of the detection-type, retrieval-type, and delivery-type AGVs, thus implementing task allocation for these AGVs. Specifically, task allocation is performed once for each of the mask sequences 001 to 111, constructing seven different task sequences. These sequences are then executed independently and repeatedly based on the mask sequence 000. Sub-task allocation, generation Group task sequence; During the type-alternating polling allocation phase, the unassigned task sets of delivery, recycling, and detection AGVs and the indices of each type of AGV are processed alternately. The process proceeds sequentially across types with the same index, following the order: Delivery AGV #1, Recycling AGV #1, Detection AGV #1, Delivery AGV #2, Recycling AGV #2, Detection AGV #2, and so on, until all unassigned tasks are allocated. The Middle Departure time of AGV Calculation, where This indicates the departure time of the first AGV in the dispatch process. This indicates the departure interval, thus generating the departure time. ; During the population building phase, the mask generated during the construction phase will be mixed with the mask. Task sequence The types generated in the interleaved polling allocation phase are respectively Combine them to construct Find a solution; sort all solutions in ascending order of total energy consumption, and retain the solutions with the lowest total energy consumption. Solution: The population size is The initial population is used to determine the optimal solution, which is the one with the minimum total energy consumption in the current population.

5. A matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 4, characterized in that, All unassigned tasks for each type of AGV constitute the unassigned task set for that type. The specific process of the nearest neighbor construction operator is as follows: taking the location of the AGV currently building a path in the current type as the starting point, calculate the Manhattan distance from each task in the unassigned task set of the current AGV type to the current location; select the task with the smallest Manhattan distance as the next task to be served, and add the selected task to the task sequence of the current AGV; at the same time, remove the selected task from the unassigned task set of the current type, and update the position of the current AGV to the position of the selected task; repeat the process of calculating distance, selecting tasks, and updating positions until the cumulative load of the current AGV reaches the capacity limit of the AGV, or the unassigned task set of the current type of AGV is empty; when the path of an AGV is completed, append a separator 0 to the end of the task sequence of the current AGV to mark the path boundary, and continue to build the path for the next AGV of the same type. The specific process of the random task construction operator is as follows: starting from the location of the AGV currently building a path in the current type, a task is randomly selected from the unassigned task set of the current AGV type as the next task to be served, and the selected task is added to the task sequence of the current AGV; at the same time, the selected task is removed from the unassigned task set of the current type of AGV, and the current position of the current AGV is updated to the position of the selected task; the process of randomly selecting tasks and updating positions is repeated until the cumulative load of the current AGV reaches the capacity limit of the AGV, or the unassigned task set of the current type of AGV is empty; when the path of an AGV is completed, a separator 0 is appended to the end of the task sequence of the current AGV to mark the path boundary, and the path is continued to be built for the next AGV of the same type.

6. The matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 5, characterized in that, In step 2, the low-level heuristic operators include intra-path insertion operators, inter-path insertion operators, adjacent swap operators, pancake flipping operators, globally greedy insertion operators, destructive reconstruction operators, and two-element optimization operators. a. The operation process of the insertion operator within the path is as follows: randomly select a task from the task sequence of an AGV of the same type and remove it from the current position. Then, randomly determine a new position in the task sequence and re-insert the selected task into the new position. b. The operation process of the inter-path insertion operator is as follows: task transfer is performed between the task sequences of two different AGVs of the same type. The first AGV with the largest cumulative task load and the second AGV with the smallest cumulative task load are identified. A task is selected from the task sequence of the first AGV and removed. The selected task is then inserted into the task sequence of the second AGV. c. The operation process of the adjacent exchange operator is as follows: randomly select the task sequence of an AGV in the same type of AGV, randomly select two adjacent tasks in the task sequence and exchange the positions of the two tasks to perform local perturbation. If the time window constraint or load constraint is violated after the exchange, the exchange operation is cancelled. d. The operation process of the pancake flipping operator is as follows: randomly select the task sequence of an AGV in the same type of AGV, randomly select a split position in the task sequence, and rearrange all tasks before the split position in reverse order; e. The operation process of the global greedy insertion operator is as follows: randomly remove one task from all task sequences of the same type of AGV as the task to be inserted; traverse each insertion position in all task sequences of the same type of AGV, calculate the total energy consumption after inserting the task to be inserted into each position, and insert the task to be inserted into the position that minimizes the increase in total energy consumption. f. The operation process of the destruction and reconstruction operator is as follows: In the destruction phase, the number of tasks to be removed is determined according to the destruction rate, and the tasks are removed according to a random removal rule or a worst-case removal rule. The random removal rule is to randomly select a corresponding number of tasks from all task sequences of the same type of AGV for removal. The worst-case removal rule is to calculate the difference in total energy consumption before and after the removal of each task in all task sequences of the same type of AGV, and prioritize removing the task that reduces the total energy consumption the most. In the reconstruction phase, the difference between the latest and earliest arrival times allowed for each removed task is calculated, and the removed tasks are arranged in ascending order of the difference. The removed tasks are then inserted sequentially into the positions in all task sequences of the same type of AGV that minimize the increase in total energy consumption. g. The operation process of the two-element optimization operator is as follows: randomly select an AGV from the same type of AGV, randomly select two non-adjacent positions in the task sequence of the selected AGV, and rearrange all tasks between the two positions in reverse order.

7. A matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 6, characterized in that, In step 2, The specific process of learning adaptive selection of low-level heuristic operators involves defining seven optional actions for the low-level heuristic operators: intra-path insertion operator, inter-path insertion operator, adjacent swap operator, pancake flip operator, global greedy insertion operator, destructive reconstruction operator, and two-element optimization operator. The energy consumption change of the population optimal solution is defined as the state. Table storage value, Each row in the table corresponds to a state, each column corresponds to an action, and the intersection of the row and column stores the action selected for that state. value, All in the table The initial values ​​of all values ​​are set to 1; based on In the table Value, using - A greedy strategy is used to select actions. Greedy strategy is based on probability Select the current state The action with the highest value, in probability Randomly select one action from the seven actions. The preset exploration probability parameters are used; the low-level heuristic operator corresponding to the selected action is executed, and the selected low-level heuristic operator is applied to each parent solution in the current population to generate the corresponding child solution.

8. The matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 7, characterized in that, In step 3, The specific process of value update is as follows: calculate the total energy consumption of the optimal solution of the parent population in this iteration. Total energy consumption of the optimal solution for the offspring population logarithmic change And the current state is discretized and defined, when When the energy consumption deteriorates, the state value is 0; when When the energy consumption remains unchanged, the state value is 1; when and Values ​​ranging from 1 to 6 represent different degrees of energy efficiency improvement, with the state value being... ;when When the value is 8, it indicates a significant improvement in energy consumption; otherwise, the value is 9. (Based on logarithmic change) For instant rewards Perform hierarchical assignment when hour, ;when hour, ;when and When the value range is 1 to 5, ,Right now The value range is 5 to 9; other cases Using Bellman equations right The corresponding table The value is updated, where, Indicates the current state. This represents the low-level heuristic operator corresponding to the action selected in this iteration. express Status in the table Select action Corresponding value; This indicates the next state after an action is performed. Indicates the candidate action in the next state. Indicates the next state All candidate actions correspond The maximum value; Indicates an immediate reward. Indicates the discount factor. The learning rate is indicated by the number of steps taken to determine the learning rate. according to Dynamic calculation, Indicates the current iteration number. This indicates the maximum number of iterations.

9. A matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 8, characterized in that, In step 3, the specific process of constraint repair is as follows: detect whether there are any AGVs in the sub-solution whose arrival time violates the time window constraint; if so, identify the AGVs whose arrival time violates the time window constraint as objects to be adjusted; define all AGVs whose departure time is earlier than the departure time of the objects to be adjusted and which do not violate the time window constraint as exchangeable objects; and swap the departure time of the objects to be adjusted with the departure time of the exchangeable objects. Repeat the operation of exchanging departure times between the object to be adjusted and the exchangeable object until the arrival times of all AGVs meet the time window constraints, or there are no exchangeable objects.

10. A matrix workshop multi-type AGV scheduling method based on a hyperheuristic algorithm according to claim 9, characterized in that, In step 4, the adaptive local search includes three local search operations: knowledge-guided position adjustment, optimal type combination, and AGV merging. The specific process of knowledge-guided position adjustment is as follows: sort the solutions in the current population according to their total energy consumption from smallest to largest, and select the solution ranked in the 1st place based on its total energy consumption. Ranked first The solution of each bit is used as the operation object; for each operation object, a random number in the interval [0,1] is generated, and the probability is calculated. If the random number is less than If the position of the object being operated on is adjusted, then the operation is not performed; otherwise, it is not performed. This indicates the time elapsed for normalization. This indicates the time elapsed since the algorithm ran. This represents the total running time. If position adjustment is performed, the difference between the scheduling component energy consumption of the delivery AGV, recovery AGV, and detection AGV in the operation object and the scheduling component energy consumption of the corresponding type of AGV in the current optimal solution is calculated. The AGV type with the largest difference is selected, and the normalized position ratio of all tasks of the selected type of AGV in the task sequence is extracted from the current optimal solution. The normalized position ratio is the ratio of the position of the task in the task sequence to the total length of the sequence. Each task in the task sequence of the corresponding type of AGV in the operation object is adjusted to a position consistent with the normalized position ratio of the same task in the current optimal solution. The specific process for optimal type combination is as follows: First, identify the scheduling component with the lowest energy consumption among the scheduling components of delivery AGVs, recovery AGVs, and detection AGVs in the current population, and denote it as the current optimal scheduling component for the corresponding type of AGV. Then, construct a new solution from the current optimal scheduling component for the corresponding type of AGV, denoted as the combined solution. Next, regenerate the departure time for the combined solution. If the AGV arrival time in the combined solution violates the time window constraint, then perform constraint repair. Finally, calculate the total energy consumption of the combined solution. If the total energy consumption of the combined solution is less than the total energy consumption of the optimal solution in the current population, then replace the solution with the highest total energy consumption in the current population with the combined solution. The specific process of AGV merging is as follows: For each solution in the current population, denoted as the solution to be operated on, an AGV type is randomly selected from the solutions to be operated on. Among the AGVs of the selected type, the AGV with the fewest tasks is identified as the object to be merged. Attempts are made to distribute all tasks in the subsequence of the object to be merged to the subsequences of other AGVs of the selected type. If all AGVs of the selected type satisfy the load constraints after the distribution, the subsequence corresponding to the object to be merged is deleted and the scheduling component of the selected type is updated. The updated scheduling component is combined with the scheduling components of the other two types in the solution to be operated on to construct a new solution, and the departure time is regenerated for the new solution. If there are cases in the new solution where the AGV arrival time violates the time window constraint, constraint repair is performed.

Images