A distributed flexible scheduling system based on double-buffer synchronization tuning mechanism

Abstract

This invention belongs to the field of intelligent manufacturing and production scheduling optimization technology, specifically a distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism, including a system coding and initialization stage, an evolution stage, and an optimization stage. The system coding adopts a three-layer one-dimensional vector: operation sequence (OS) layer, machine allocation (MA) layer, and factory allocation (FA) layer. A hybrid initialization strategy generates an initial population of size N. In the first 50% of individuals, FA adopts the factory load balancing rule; OS adopts the highest remaining operation number priority rule; MA adopts the shortest processing time rule; in the last 50% of individuals, FA, OS, and MA are all randomly generated. The evolution stage and the optimization stage are connected in series. The evolution stage is responsible for exploring the global solution space, using an elite retention strategy and a Dual-RL mechanism to dynamically adjust the crossover and mutation probabilities. The optimization stage executes seven local search operators on the current best elite individual and performs fine-tuning through DSTM.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and production scheduling optimization technology. Specifically, it is a distributed flexible scheduling system based on a dual-buffer synchronous optimization mechanism, which is applicable to the flexible job shop scheduling problem in a distributed production environment, and is especially suitable for large-scale, highly complex distributed manufacturing scenarios. Background Technology

[0002] With increasing global competition and volatile market demands, distributed manufacturing has become the mainstream production paradigm in modern industrial systems. The Distributed Flexible Job Shop Scheduling Problem (DFJSP) is an extension of the traditional flexible job shop scheduling problem in a distributed environment, significantly increasing its complexity. DFJSP comprises three subproblems: allocating jobs to different factories, arranging the processing sequence within each factory, and selecting processing machines for each process. This problem has been proven to be an NP-hard combinatorial optimization problem, with the solution space growing exponentially with the number of jobs and factories.

[0003] Existing optimization methods for DFJSP are mainly divided into three categories: exact methods, heuristic methods, and intelligent optimization methods. Exact methods perform well on small-scale problems, but are difficult to apply on large-scale problems due to excessive computation time. Heuristic methods quickly construct feasible solutions based on scheduling rules, with high computational efficiency, but the quality of the solutions is difficult to guarantee, and the problem dependence is strong. Intelligent optimization methods improve the quality of solutions through iterative search and have become the mainstream research direction. However, existing intelligent optimization methods still face the following key technical bottlenecks: (1) Insufficient dynamic balance between exploration and development: In traditional genetic algorithms, the crossover probability and mutation probability are usually fixed and difficult to adapt to the needs of different evolutionary stages. In the early stage of population evolution, it is necessary to strengthen the exploration ability to avoid getting trapped in local optima, and in the later stage, it is necessary to strengthen the development ability for fine search. Fixed parameter settings cannot dynamically adjust this balance, resulting in slow convergence speed or premature convergence. (2) Difficulty in maintaining population diversity: In the evolutionary process, superior individuals quickly dominate the population, resulting in the loss of genetic diversity, and the algorithm is prone to getting trapped in local optima. Existing elite retention strategies often overprotect the best individuals and inhibit population diversity. (3) Insufficient utilization of coding coupling characteristics: There is an inherent coupling relationship between the three coding layers of DFJSP. Existing methods usually optimize each layer independently, ignoring the synergistic gains that may be brought about by joint adjustment.

[0004] To address the aforementioned issues, there is an urgent need to design a novel optimization method that can dynamically balance exploration and development, effectively maintain population diversity, and fully utilize the coding coupling characteristics, in order to improve the solution efficiency and quality of DFJSP. Summary of the Invention

[0005] To address the processing problems existing in the prior art, this invention provides a distributed flexible scheduling system based on a dual-buffer synchronous optimization mechanism.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism includes a system coding and initialization phase, an evolution phase (dual-reinforcement learning (Dual-RL) module), and an optimization phase (dual-buffer synchronous tuning (DSTM) module).

[0008] In the system coding and initialization phase, the system coding adopts a three-layer one-dimensional vector: the Operation Sequence (OS) layer, with a length equal to the total number of operations for all workpieces, where each element is a workpiece number, and the frequency of the workpiece number indicates which operation for that workpiece; the Machine Assignment (MA) layer, with the same length as the OS layer, where each element represents the machine index selected for the corresponding operation; and the Factory Assignment (FA) layer, with a length equal to the total number of workpieces n, where each element represents the factory number assigned to the corresponding workpiece. A hybrid initialization strategy generates an initial population of size N. In the first 50% of individuals, FA adopts the factory load balancing rule; OS adopts the highest remaining operation number priority rule; and MA adopts the shortest processing time rule. In the last 50% of individuals, FA, OS, and MA are all randomly generated.

[0009] The evolutionary stage and the optimization stage are linked. The evolutionary stage is responsible for exploring the global solution space and uses an elite retention strategy and a Dual-RL mechanism to dynamically adjust the crossover and mutation probabilities. The optimization stage executes seven local search operators on the current best elite individuals and performs fine-tuning through DSTM. The two stages work together to balance global search capability and local development accuracy.

[0010] Preferably, the evolutionary phase includes elite-preserving crossover and Dual-RL parameter tuning;

[0011] In the elite-preserving crossover, each generation randomly and evenly divides the population into subpopulations P and Q. Individuals in P are sequentially crossovered with the current global best individual, and the better offspring is selected to enter the next generation. Individuals in Q are randomly paired and crossovered, and the better offspring is selected. Different crossover operators are used for different coding layers. For the FA layer, masked uniform crossover (UX) is used, and a 0 / 1 mask is randomly generated. When the mask is 1, it inherits from parent generation 1, and when it is 0, it inherits from parent generation 2. For the OS layer, priority-preserving order-based crossover (POX) is used, and the workpiece set is randomly divided. The offspring inherits the process position of the workpiece set from the parent generation, and the remaining processes are filled according to the order of the other parent generation. For the MA layer, a combination of two-point crossover and uniform crossover is used, and the two methods are selected with a certain probability.

[0012] In Dual-RL parameter tuning, two reinforcement learning algorithms, Q-learning and SARSA, are used to adjust the selection of parameters during the evolutionary process. The Dual-RL mechanism uses a switching factor Λ to alternate between SARSA and Q-learning algorithms. The state is defined as a weighted combination of average fitness, population diversity, and optimal individual fitness. Actions are used to select the range of crossover and mutation probabilities. Rewards are calculated based on the iterative improvement amount between optimal and average fitness. The switching factor Λ compares the reward difference between the current algorithm and the other algorithm in the iteration; if Λ > 0, the algorithm is switched. This mechanism achieves adaptive adjustment of evolutionary parameters, effectively balancing exploration and development. Specifically:

[0013] State: Divided into 20 discrete intervals, determined by the average fitness of the population, the diversity of the population, and the fitness of the best individual, as shown in formulas (1) to (4):

[0014]

[0015] Among them, formula (4) will include the population state value Defined as a weighted combination of formulas (1) to (3); formula (1) represents the average fitness of the population. The calculation method, formula (2), is expressed as the population diversity. The calculation method, formula (3) is expressed as the fitness of the optimal individual. The calculation method; Let represent the fitness of the i-th individual in generation t. Let N represent the fitness of the i-th individual in the initial generation, and N represent the population size. Let represent the fitness of the optimal individual in generation t. The fitness of the best individual in the initial generation is represented by w1, w2, and w3, which are weight values ​​that make w1+w2+w3=1, used to quantify the relative importance of these three components.

[0016] Action: P c and P m The adaptation constitutes the execution of the agent's actions; the agent uses different actions to obtain the appropriate P. c and P m This includes 10 actions in the action set; after selecting an action, the agent from P... c Sets and P m A random value is selected from each set and set as the true P. c Value and P m Value; the ε-greedy strategy is used in Dual-RL to select appropriate actions that balance exploration and development, as described in Equation (5):

[0017]

[0018] in, The policy function represents the state S. t Choose action a t Decision-making rules. Indicates that in state S t The action that takes the maximum Q value among all possible actions 'a' is given below. ε represents the greed rate, and r is a random value between 0 and 1.

[0019] Rewards: The reward function is designed based on the individual's optimal fitness and the population's average fitness. The calculation method is shown in formulas (6) to (8).

[0020]

[0021] Among them, reward c Indicates adjustment P c The reward value after that. m Indicates adjustment P m The reward value after each iteration; the reward calculation for each iteration is shown in formula (8).

[0022] The conversion factor Λ dynamically determines whether the RL algorithm needs to be switched. In each iteration, both RL algorithms are used to calculate the reward, while the current algorithm is used to update the Q value. The agent determines which RL algorithm to use in the next iteration based on the conversion factor Λ. The calculation method of Λ is shown in formula (9):

[0023]

[0024] Among them, a reward e represents the actual reward value of the algorithm used. reward This represents the reward value of another algorithm.

[0025] Preferably, during the optimization phase, one of the following seven operators is randomly selected and executed for the current best individual:

[0026] LS1: OS is divided into m segments, and the member processes and the head process are exchanged sequentially within each segment;

[0027] LS2: The OS is divided into m equal segments. After each segment is removed, it is inserted into all possible positions of the remaining segments.

[0028] LS3: A random position in the OS is swapped with the other n-1 positions in sequence;

[0029] LS4: Randomly select an element in the OS and insert it into n-1 possible positions in sequence;

[0030] LS5: Exchange two critical processes within a critical plant;

[0031] LS6: Randomly select a workpiece within the critical plant and reassign it to another plant;

[0032] LS7: Randomly select a critical process within the critical plant and move it to another available machine.

[0033] Preferably, during the optimization phase, DSTM utilizes the coupling characteristics between the operation sequence (OS) and factory assignment (FA) in the DFJSP encoding scheme to maintain the independence of machine assignment (MA) while setting up two candidate solution record stacks. When a single-level adjustment of OS or FA to the current best individual cannot directly improve fitness, the candidate solution is still stored in the stack. Subsequently, by synchronously searching the candidate combinations in the two stacks, a solution that can produce better fitness after joint adjustment of OS and FA is found. This mechanism breaks through the limitation of traditional methods that only accept single-level improvements, can capture ignored cooperative gains, accelerate convergence, and avoid premature convergence.

[0034] This invention proposes a hierarchical optimization algorithm with a dual-cache synced tuning mechanism (HOA-DSTM). This algorithm dynamically adjusts evolutionary parameters through dual-reinforcement learning (Dual-RL), maintains population diversity through an elite preservation strategy, enhances local exploitation capabilities through seven local search operators, and fully utilizes the encoding coupling characteristics to optimize the optimal individual via the dual-cache synced tuning mechanism (DSTM), thereby efficiently solving the DFJSP problem.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] (1) This invention innovatively proposes DSTM: by recording and jointly adjusting the candidate solutions of OS and FA, the coding coupling characteristics are fully utilized, which significantly improves the optimization quality, especially in large-scale problems.

[0037] (2) This invention introduces Dual-RL parameter adaptive adjustment based on the conversion factor Λ: dynamically switching SARSA and Q-learning, overcoming the limitations of fixed parameters or single RL strategies, and achieving a better balance between exploration and development.

[0038] (3) The present invention designs a differentiated elite retention strategy: by dividing the subgroups and stratifying the crossover, it retains the knowledge of elite solutions and maintains the diversity of the population, effectively suppressing premature convergence.

[0039] (4) This invention integrates seven complementary local search operators: covering multiple dimensions such as process, machine, and factory, which enhances the local development capability of the algorithm and improves the quality of the optimal solution.

[0040] (5) The overall framework of the present invention is clear and modular, and it is easy to extend to other distributed scheduling problems, with broad application prospects. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0042] Figure 1 This is a schematic diagram of the DFJSP structure in this invention;

[0043] Figure 2 This is an example diagram of the three-layer coding scheme in this invention;

[0044] Figure 3 This is a flowchart of the HOA-DSTM algorithm in this invention;

[0045] Figure 4 This is a schematic diagram of the encoding coupling characteristics in DFJSP of the present invention;

[0046] Figure 5 This is a schematic diagram of the DSTM in this invention;

[0047] Figure 6 This is an example Gantt chart from the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] A distributed flexible scheduling system for the distributed flexible job shop scheduling problem, the overall system framework is as follows: Figure 3 As shown, its core modules are as follows:

[0051] (1) Two-stage hierarchical optimization framework: The framework consists of an evolutionary stage and an optimization stage connected in series. The evolutionary stage is responsible for exploring the global solution space, and adopts an elite retention strategy and a Dual-RL mechanism to dynamically adjust the crossover and mutation probabilities; the optimization stage executes seven local search operators on the current best elite individuals and performs fine-tuning through DSTM. The two stages work together to balance global search capability and local development accuracy.

[0052] (2) Dual-RL Module: The Dual-RL mechanism uses a switching factor Λ to alternate between two reinforcement learning algorithms, SARSA and Q-learning. The state is defined as a weighted combination of the population's average fitness, population diversity, and optimal individual fitness; actions are used to select the range of crossover and mutation probabilities; rewards are calculated based on the iterative improvement of the optimal and average fitness. The switching factor Λ compares the reward difference between the current algorithm and the other algorithm in that iteration; if Λ > 0, the algorithm is switched. This mechanism achieves adaptive adjustment of evolutionary parameters, effectively balancing exploration and development.

[0053] (3) Dual-buffered synchronous tuning (DSTM) module: such as Figure 4 As shown, DSTM utilizes the coupling between Operation Sequence (OS) and Factory Assignment (FA) in the DFJSP encoding scheme to maintain the independence of Machine Assignment (MA) while setting up two candidate solution record stacks, such as... Figure 5 As shown, when a single-level adjustment of OS or FA to the current best individual cannot directly improve fitness, candidate solutions are still stored in the stack. Subsequently, by synchronously searching the candidate combinations in both stacks, a solution that can produce a better fitness after joint adjustment of OS and FA is found. This mechanism breaks through the limitation of traditional methods that only accept single-level improvements, can capture ignored cooperative gains, accelerate convergence, and avoid premature convergence.

[0054] Example 2

[0055] A distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism, such as Figure 1 As shown, it includes:

[0056] 1. System coding and initialization, such as Figure 2 As shown,

[0057] (1) The encoding adopts a three-layer one-dimensional vector: the process sequence (OS) layer, the length of which is the total number of processes of all workpieces, each element is the workpiece number, and the number of times the workpiece number appears indicates which process of the workpiece; the machine allocation (MA) layer, the same length as the OS layer, each element represents the machine index selected for the corresponding process; the factory allocation (FA) layer, the length of which is the total number of workpieces n, each element represents the factory number assigned to the corresponding workpiece.

[0058] (2) Initialization: A hybrid initialization strategy is used to generate an initial population of size N. In the first 50% of individuals, FA adopts the factory load balancing rule; OS adopts the highest remaining number of processes priority rule; and MA adopts the shortest processing time rule. In the last 50% of individuals, FA, OS, and MA are all randomly generated.

[0059] 2. Algorithm Flow Description

[0060] The proposed HOA-DSTM flowchart is as follows: Figure 3 As shown, its termination condition is set to cputime=n∗m∗f.

[0061] The entire algorithm begins with mixed partitioning initialization to generate a diverse and high-quality initial population, while simultaneously establishing a reinforcement learning state-action set. Before iterative search, the current optimal individual is determined. During the evolutionary phase, individuals are selected from the population, which is then divided into two parts, and heterogeneous crossover and mutation strategies are implemented. P is adjusted online in real-time via a Dual-RL mechanism. c and P m Dual-RL alternates between SARSA and Q-learning through a transition factor Λ, using reward feedback in each generation to balance exploration and development, preserving genetic diversity while retaining elite individuals. During the optimization phase, a local search operator is used to explore the neighborhood of each individual. After evaluating the population, based on the three-layer encoding coupling characteristic, DSTM is used to further optimize the current best individual. By simultaneously maintaining two stacks recording candidate individuals and only adopting coordinated work sequence (OS) and factory assignment (FA) collaborative improvement schemes, DSTM can capture high-quality individuals missed by adjusting work sequence (OS) or factory assignment (FA) alone, accelerating algorithm convergence without affecting machine assignment (MA). The HOA-DSTM algorithm runs iteratively until the termination condition is met, finally outputting the best individual as the final scheduling solution for the problem.

[0062] 3. Evolutionary Stages

[0063] (1) Elite retention cross

[0064] In each generation, the population is randomly and evenly divided into subpopulations P and Q. Individuals in P are sequentially crossbred with the current global best individual, and the better offspring is selected to enter the next generation. Individuals in Q are randomly paired and crossbred, and the better offspring is selected again. Different crossbringing operators are used for different coding layers. For the FA layer, masked uniform crossbringing (UX) is used, and a 0 / 1 mask is randomly generated. When the mask is 1, it inherits from parent generation 1, and when it is 0, it inherits from parent generation 2. For the OS layer, priority-preserving order-based crossbringing (POX) is used. The job set is randomly divided, and the offspring inherits the operation position of the job set belonging to a certain job set from the parent generation. The remaining operations are filled according to the order of the other parent generation. For the MA layer, a combination of two-point crossbringing and uniform crossbringing is used, and the two methods are selected with a certain probability.

[0065] (2) Adjustment of Dual-RL parameters

[0066] Two reinforcement learning algorithms, Q-learning and SARSA, are used to adjust the selection of parameters during the evolution process.

[0067] State: Divided into 20 discrete intervals, determined by the average fitness of the population, the diversity of the population, and the fitness of the best individual, as shown in formulas (1) to (4):

[0068]

[0069] Formula (4) will use the population state value Defined as a weighted combination of formulas (1) to (3). Formula (1) represents the average fitness of the population. The calculation method, formula (2), is expressed as the population diversity. The calculation method, formula (3) is expressed as the fitness of the optimal individual. The calculation method. Let represent the fitness of the i-th individual in generation t. Let N represent the fitness of the i-th individual in the initial generation, and N represent the population size. Let represent the fitness of the optimal individual in generation t. The fitness of the best individual in the initial generation is represented by w1, w2, and w3, which are weight values ​​that make w1+w2+w3=1, used to quantify the relative importance of these three components.

[0070] Action: P c and P m The adaptation constitutes the execution of the agent's actions. The agent uses different actions to obtain the appropriate P. c and P m This includes 10 actions in the action set. After selecting an action, the agent from P... c Sets and P mA random value is selected from each set and set as the true P. c Value and P m Value. The ε-greedy strategy is used to select appropriate actions that balance exploration and development in Dual-RL, as described in Equation (5):

[0071]

[0072] in The policy function represents the state S. t Choose action a t Decision-making rules. Indicates that in state S t The action that maximizes the Q value among all possible actions 'a' is defined below. ε represents the greed rate, and r is a random value between 0 and 1.

[0073] Rewards: The reward function is designed based on the individual's optimal fitness and the population's average fitness. The calculation method is shown in formulas (6) to (8).

[0074]

[0075] Among them, reward c Indicates adjustment P c The reward value after that. m Indicates adjustment P m The reward value after each iteration is calculated as shown in formula (8).

[0076] This invention designs a switching factor Λ to dynamically determine whether an RL algorithm needs to be switched. In each iteration, both RL algorithms are used to calculate the reward, while the current algorithm is used to update the Q value. The agent determines which RL algorithm to use in the next iteration based on the switching factor Λ. The calculation method of Λ is shown in formula (9):

[0077]

[0078] Where a reward e represents the actual reward value of the algorithm used. reward This represents the reward value of another algorithm.

[0079] 4. Optimization Phase

[0080] (1) Local search operator

[0081] For the current best individual, randomly select one of the following seven operators to execute:

[0082] LS1: OS is divided into m segments, and the member processes and the head process are exchanged sequentially within each segment.

[0083] LS2: The OS is divided into m segments. After each segment is removed, it is inserted into all possible positions of the remaining segments.

[0084] LS3: A random position in the OS is swapped with the other n-1 positions in turn.

[0085] LS4: Randomly select an element in the OS and insert it into n-1 possible positions in sequence.

[0086] LS5: Exchange two critical processes within a critical plant.

[0087] LS6: Randomly select a workpiece within the critical plant and reassign it to another plant.

[0088] LS7: Randomly select a critical process within the critical plant and move it to another available machine.

[0089] (2) Dual-buffered synchronous tuning mechanism

[0090] During the optimization phase, local search operations on the OS sequence sometimes fail to improve the overall fitness of an individual, leading to skipping such neighboring individuals and evaluating other individuals instead. Even if an individual's OS sequence is suboptimal, there are still cases where adjusting the FA sequence simultaneously produces neighboring individuals with better fitness than the original individual. Retaining certain candidate solutions requires dynamically synchronizing adjustments to both OS and FA. A balance between local optimization and global coordination is achieved through dynamically coordinating OS and FA adjustments; synchronized optimization of OS and FA reflects the finer-grained interaction within the decoupled framework. Two candidate solution record stacks are used in DSTM to record the set of candidate solutions during the adjustment process.

[0091] Example 3

[0092] like Figure 6 As shown, four workpieces are assigned to two isomorphic factories, each with three production machines. Based on the proposed coding scheme, a Gantt chart is given with the following codes: OS=[1,3,2,2,4,3,4,1,2,4], MA=[1,2,3,2,1,2,3,1,3,2], FA=[2,1,1,2]. First, workpieces 1 and 4 are assigned to factory 2, and workpieces 2 and 3 are assigned to factory 1. For workpiece 1, the first operation is performed on machine 1, and the second operation on machine 2. For workpiece 2, the first operation is performed on machine 3, the second operation on machine 2, and the third operation on machine 1. According to OS, the first operation of workpiece 1 is performed first, followed by the first operation of workpiece 3, then the first and second operations of workpiece 2. This process continues until all workpiece operations are completed. The completion time of the last factory to complete all operations is the total completion time.

[0093] Industrial Applications: This invention can be widely applied to distributed production scenarios such as automobile manufacturing, electronic assembly, and mold processing, and is especially suitable for group manufacturing enterprises that require multi-factory collaborative scheduling. The system can be deployed in an enterprise-level scheduling center, and connects with the Manufacturing Execution System (MES) of each factory through API to receive real-time tasks and equipment status, dynamically generate optimized scheduling schemes, significantly shorten production cycles, improve equipment utilization, and reduce production costs.

[0094] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism, comprising a system coding and initialization phase, an evolution phase, and an optimization phase; In the system coding and initialization phase, the system coding adopts a three-layer one-dimensional vector: the Operation Sequence (OS) layer, with a length equal to the total number of operations for all workpieces, where each element is a workpiece number, and the frequency of the workpiece number indicates which operation for that workpiece; the Machine Assignment (MA) layer, with the same length as the OS layer, where each element represents the machine index selected for the corresponding operation; and the Factory Assignment (FA) layer, with a length equal to the total number of workpieces n, where each element represents the factory number assigned to the corresponding workpiece. A hybrid initialization strategy generates an initial population of size N. In the first 50% of individuals, FA adopts the factory load balancing rule; OS adopts the highest remaining operation number priority rule; and MA adopts the shortest processing time rule. In the last 50% of individuals, FA, OS, and MA are all randomly generated. The evolutionary stage and the optimization stage are linked. The evolutionary stage is responsible for exploring the global solution space and uses an elite retention strategy and a Dual-RL mechanism to dynamically adjust the crossover and mutation probabilities. The optimization stage executes seven local search operators on the current best elite individuals and performs fine-tuning through DSTM. The two stages work together to balance global search capability and local development accuracy.

2. The distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 1, characterized in that, The evolutionary phases include elite-preserving crossover and Dual-RL parameter tuning.

3. The distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 2, characterized in that, In the elite-preserving crossover, each generation randomly and evenly divides the population into subpopulations P and Q. Individuals in P are sequentially crossovered with the current global best individual, and the better offspring is selected to enter the next generation. Individuals in Q are randomly paired and crossovered, and the better offspring is selected. Different crossover operators are used for different coding layers. For the FA layer, masked uniform crossover is used, and 0 / 1 masks are randomly generated. When the mask is 1, it inherits from parent generation 1, and when it is 0, it inherits from parent generation 2. For the OS layer, priority-preserving order-based crossover is used, and the workpiece set is randomly divided. The offspring inherit the process position of the workpiece set in the parent generation, and the remaining processes are filled according to the order of the other parent generation. For the MA layer, a combination of two-point crossover and uniform crossover is used, and the two methods are selected with a certain probability.

4. The distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 2, characterized in that, In Dual-RL parameter tuning, two reinforcement learning algorithms, Q-learning and SARSA, are used to adjust the selection of parameters during the evolutionary process. The Dual-RL mechanism is based on the alternation of SARSA and Q-learning reinforcement learning algorithms using a switching factor Λ. Here, the state is defined as a weighted combination of the population's average fitness, population diversity, and optimal individual fitness; the action is used to select the value range of crossover probability and mutation probability; the reward is calculated based on the iterative improvement amount of the optimal fitness and the average fitness; the switching factor Λ compares the reward difference between the current algorithm and the other algorithm in the iteration, and if Λ > 0, the algorithm is switched. This mechanism realizes the adaptive adjustment of evolutionary parameters and effectively balances exploration and development.

5. The distributed flexible scheduling system based on a dual-buffer synchronization optimization mechanism according to claim 4, characterized in that, The state is divided into 20 discrete intervals, which are determined by the average fitness of the population, the diversity of the population, and the fitness of the best individual, as shown in formulas (1) to (4): Among them, formula (4) will include the population state value Defined as a weighted combination of formulas (1) to (3); formula (1) represents the average fitness of the population. The calculation method, formula (2), is expressed as the population diversity. The calculation method, formula (3) is expressed as the fitness of the optimal individual. The calculation method; Let represent the fitness of the i-th individual in generation t. Let N represent the fitness of the i-th individual in the initial generation, and N represent the population size. Let represent the fitness of the optimal individual in generation t. The fitness of the best individual in the initial generation is represented by w1, w2, and w3, which are weight values ​​that make w1+w2+w3=1, used to quantify the relative importance of these three components.

6. The distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 4, characterized in that, The action is: P c and P m The adaptation constitutes the execution of the agent's actions; the agent uses different actions to obtain the appropriate P. c and P m This includes 10 actions in the action set; after selecting an action, the agent from P... c Sets and P m A random value is selected from each set and set as the true P. c Value and P m Value; the ε-greedy strategy is used in Dual-RL to select appropriate actions that balance exploration and development, as described in Equation (5): in, The policy function represents the state S. t Choose action a t Decision-making rules; Indicates that in state S t The action that maximizes the Q value among all possible actions 'a' is defined below. ε represents the greed rate, and r is a random value between 0 and 1.

7. A distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 4, characterized in that, The reward is designed based on the individual's optimal fitness and the population's average fitness, and the calculation method is shown in formulas (6) to (8): Among them, reward c Indicates adjustment P c The reward value after that. m Indicates adjustment P m The reward value after each iteration; the reward calculation for each iteration is shown in formula (8).

8. The distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 4, characterized in that, The conversion factor Λ dynamically determines whether the RL algorithm needs to be switched. In each iteration, both RL algorithms are used to calculate the reward, while the current algorithm is used to update the Q value. The agent determines which RL algorithm to use in the next iteration based on the conversion factor Λ. The calculation method of Λ is shown in formula (9): #imgpt13# Among them, a reward e represents the actual reward value of the algorithm used. reward This represents the reward value of another algorithm.

9. A distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 1, characterized in that, During the optimization phase, one of the following seven operators is randomly selected and executed for the current best individual: LS1: OS is divided into m segments, and the member processes and the head process are exchanged sequentially within each segment; LS2: The OS is divided into m equal segments. After each segment is removed, it is inserted into all possible positions of the remaining segments. LS3: A random position in the OS is swapped with the other n-1 positions in sequence; LS4: Randomly select an element in the OS and insert it into n-1 possible positions in sequence; LS5: Exchange two critical processes within a critical plant; LS6: Randomly select a workpiece within the critical plant and reassign it to another plant; LS7: Randomly select a critical process within the critical plant and move it to another available machine.

10. A distributed flexible scheduling system based on a dual-buffer synchronous tuning mechanism according to claim 9, characterized in that, During the optimization phase, DSTM utilizes the coupling characteristics between the operation sequence (OS) and factory assignment (FA) in the DFJSP coding scheme to maintain the independence of machine assignment (MA) while setting up two candidate solution record stacks. When a single-level adjustment of OS or FA to the current best individual cannot directly improve fitness, the candidate solution is still stored in the stack. Subsequently, by synchronously searching the candidate combinations in the two stacks, a scheme that can produce better fitness after joint adjustment of OS and FA is found.

Images