A QMIX-based aircraft pulsating assembly line distributed scheduling method

Abstract

The application relates to a QMIX-based aircraft pulsation assembly line distributed scheduling method, which comprises the following steps: setting a distributed decision station agent at each station on the assembly line, wherein the station agent comprises a sequencing sub-agent, an allocation sub-agent, a simulation module and a collaborative decision module; the simulation module simulates an assembly process and updates state data, and the two sub-agents output scheduling rules through interaction with the simulation module; the simulation module saves the state data and the scheduling rules into an experience pool, the collaborative decision module samples the data in the experience pool, and the network parameters of the two sub-modules are trained and updated; finally, the two sub-agents output optimized process processing sequences and personnel allocation schemes after global training. Compared with the prior art, the agents in the application can adjust scheduling decisions according to the production environment through interaction with the environment, greatly improving the autonomy of the entire scheduling process.

Description

Technical Field

[0001] This invention relates to the field of assembly manufacturing, and in particular to a distributed scheduling method for aircraft pulse assembly lines based on QMIX. Background Technology

[0002] Aircraft final assembly is the final and critical stage of aircraft manufacturing, involving a large amount of work such as the docking of major components, laying of conduits and cables, engine installation, and system testing. It directly impacts the final production speed and supply capacity of aircraft. Currently, the main form of aircraft assembly lines is pulse assembly. Pulse assembly means that each assembly station, following the sequence of assembly instructions, begins its assembly task at the current station, completes the designated assembly process within a fixed cycle, and then moves to the next station at the same time to perform the assembly task for the next assembly cycle, ultimately completing all assembly processes. Compared to traditional fixed assembly lines, pulse assembly lines have a more detailed and clear division of labor, with repetitive and repetitive tasks, resulting in higher production efficiency.

[0003] Existing research on pulsed assembly line scheduling mostly focuses on adjusting the assembly teams and assembly process sequences with limited scheduling resources to minimize assembly cycle time and balance the load on assembly stations. The scheduling methods are primarily centralized, employing metaheuristic algorithms for process sequencing. However, real-world aircraft assembly lines involve large-scale processes and numerous assembly teams. Centralized scheduling optimization requires significant computational resources and is prone to getting trapped in local optima. Furthermore, it lacks the ability to effectively respond to dynamic events that frequently occur during assembly, such as changes in orders, team structure, and worker efficiency.

[0004] Therefore, adopting a distributed scheduling approach for large-scale, pulsed assembly lines, classifying and hierarchically scheduling complex production resources, and improving the assembly line's ability to respond to dynamic events has become a new research direction.

[0005] Chinese patent application CN110991056A discloses a scheduling method for aircraft assembly lines. This method includes: constructing a segmented scheduling model for the aircraft assembly line, with the optimization objective of minimizing the total assembly operation duration, while considering predecessor and successor constraints, resource constraints, and spatial constraints; and solving the model using an improved genetic variable neighborhood algorithm. The scheduling process of this method employs a centralized approach, encoding all processes and iteratively optimizing. For large-scale process scheduling, this may lead to local optima. Regarding the scheduling objects, this patent only sorts the process sequence, neglecting the important scheduling object of assembly workers. In terms of the scheduling mechanism, this patent represents a balance optimization during the design phase, meaning the scheduling method cannot be changed after assembly begins on-site. In summary, this existing technology suffers from drawbacks such as a single scheduling object, low autonomy in the scheduling process, and weak adaptability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a distributed scheduling method for aircraft pulse assembly lines based on QMIX.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A distributed scheduling method for aircraft pulsed assembly lines based on QMIX, the method comprising the following steps:

[0009] S1: Distributed decision-making station agents are set up at each station on the assembly line. Each station agent includes two sub-agents: a sorting agent and an allocation agent, as well as a collaborative decision-making module and a simulation module.

[0010] S2: Establish decision networks in the two sub-agents, establish collaborative decision networks in the collaborative decision module, and initialize the parameters of each network;

[0011] S3: Initialize the simulation module parameters and initialize the assembly cycle time based on empirical values;

[0012] S4: The simulation module inputs state information to the decision networks of the two sub-agents, and the two decision networks output the process sequencing rules and worker allocation rules respectively; the simulation module simulates the process assembly process according to the process sequencing rules and worker allocation rules, and updates the process and team state information; after all process allocations are completed, the state data and the rules obtained from the decisions are saved to the experience pool;

[0013] S5: The collaborative decision-making module samples data from the experience pool, performs global training on the two decision networks using the QMIX-based network training and update method, and updates the parameters of the decision networks.

[0014] S6: Every certain number of rounds, update the beat rate to the minimum beat rate of the previous round; when the beat rate no longer changes, the station agent outputs the final process sorting and allocation scheme.

[0015] Furthermore, the scheduling problem of the pulse assembly line is defined as follows: given the number of workstations and personnel information, under the premise of satisfying constraints such as process and team skills, the assembly cycle time and workstation load are optimized by arranging the processing sequence and the allocation scheme from process to personnel.

[0016] The overall optimization objective of the distributed scheduling of the aircraft pulse assembly line is to minimize the assembly cycle time C and the station load variance. During the optimization process, each station independently decides its own process sequence and worker allocation scheme, and the overall decision scheme is completed by the sequential decision of all stations.

[0017] Furthermore, the station agent is an information system with autonomous decision-making and communication capabilities, distributed across various stations, and collaboratively allocates and plans resources within the stations through interaction; the simulation module is used to simulate the information space of the station production environment, including simulating the assembly process and updating assembly status information; the collaborative decision-making module is used to train two sub-agents to enable them to make collaborative decisions.

[0018] Furthermore, in step S4, the two sub-agents cooperate to complete the process sequencing and worker allocation, as detailed below:

[0019] S41. Based on the process predecessor and successor constraints, obtain the set of processes without predecessor processes as the initial set of assemblable processes;

[0020] S42. The simulation module collects the process status and the status of the shift workers and inputs them into the decision network of the two sub-agents;

[0021] S43. The sorting agent decision network outputs sorting rules and calculates the process with the highest priority;

[0022] S44. The sorting agent drives the allocation agent, and the allocation agent's decision network outputs worker allocation rules and selects the worker with the highest priority.

[0023] S45. The simulation module adds the process to the worker's assembly list to simulate the assembly process, assigns the highest priority process to the highest priority worker's assembly list, compares the worker's estimated completion time with the assembly cycle time, and updates the worker's timetable.

[0024] S46. After all processes have been assigned, the decision information and assembly line status information are stored in the experience pool.

[0025] Furthermore, the process sequencing rules include: the more subsequent processes a process has, the higher its priority; the longer the sum of the times of subsequent processes a process has, the higher its priority; and the longer the sum of the times from a process to the point where there are no subsequent processes, the higher its priority.

[0026] The sorting agent decision network outputs a weight vector of a linear combination of the above-mentioned process sorting rules;

[0027] The simulation module inputs process status information into the decision network, calculates the Q value of all output combinations, and takes the rule combination with the largest Q value as the output.

[0028] Furthermore, the worker allocation rules include: workers with shorter actual assembly time are given priority, workers with shorter completion time are given priority, and workers with higher assembly efficiency are given priority.

[0029] The output of the allocation agent decision network is a linear combination of the worker allocation rules described above;

[0030] The simulation module inputs the status information of the workers in the work group into the decision network of the assigned agent, calculates the Q value of all rule combinations, and takes the rule combination with the largest Q value as the output.

[0031] Furthermore, the process of updating the worker's timetable is as follows:

[0032] S451. If the worker's estimated completion time after adding the process does not exceed the predetermined assembly cycle time, the process allocation is successful and input into the simulation module to update the set of assembleable processes and the status data information of the team workers.

[0033] S452. If the worker's estimated completion time exceeds the predetermined assembly cycle time after the process is added, the worker's priority is recalculated and the process is reassigned to other workers, if there are other workers whose completion time does not exceed the assembly cycle time.

[0034] S453. If all workers will exceed the predetermined assembly cycle time for this process, save the process to the assembly list of the next station and delete it from the assembly process set of this station, and reselect a process with higher priority.

[0035] Furthermore, after all assembly processes are completed, the maximum actual completion time of all stations is collected as the actual assembly cycle time C for this round of optimization, and the actual assembly time T and the assembly time t of each shift are collected. b Calculate the reward r t Save to the experience pool;

[0036]

[0037] Where α, β, and γ are the assembly time, the actual assembly cycle time in this round of optimization, and the reward coefficients for the assembly time of each shift, respectively, and B is the total number of shifts.

[0038] Furthermore, the decision-making networks established in the two sub-agents and the collaborative decision-making network established in the collaborative decision-making module respectively include a value network and a target network, wherein the target network is used to delay copying the parameters of the value network.

[0039] Furthermore, the process of performing global training on the two decision networks using the QMIX-based network training and update method in step S5 is as follows:

[0040] S51. Extract the state data at time t from the experience pool. t Scheduling action a at time t-1 t-1 ; respectively Input the value network of the sorting agent. Input the value network for assigning agents. Calculate the Q-values ​​for all rule weight combinations in both value networks, and then take the largest Q-value. and the selected action

[0041] Pick Input to the target network of the sorting agent. The input is fed into the target network of the assigned agent, and the two target networks calculate the target Q-value.

[0042] S52. The maximum Q-value of the value network Q e and global state information s t-1 The value network input to the collaborative decision-making network obtains the global Q-value Q. ae ; set the target Q value Q t and global state information s t The target network of the collaborative decision-making module, input into the collaborative decision-making module, obtains the global objective Q value Q. at ;

[0043] S53. Set the global target Q value Q. at With the reward r in the experience pool t The target value f is obtained by adding them together;

[0044] S54. Based on the objective value f and the global Q value Q ae Calculate the loss function td error ;

[0045] S55. Backpropagate the loss function to update the parameters of the collaborative decision network and the two sub-decision networks;

[0046] S56. After a certain number of rounds, each target network copies the parameters of its corresponding value network after a delay.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1) This invention constructs the various entities of the aircraft pulse assembly line into multi-level agents according to different application levels, classifies and hierarchically schedules the resources of the pulse line, and can efficiently schedule assembly process sets with complex resource relationships and large process scale, which greatly improves the autonomy of the entire scheduling process and improves the efficiency of the scheduling process.

[0049] 2) This invention adopts a network training and update method based on QMIX. Compared with the traditional multi-agent reinforcement learning method that performs value decomposition on the global optimization objective, the global optimization objective in the QMIX method is the value of the output of the collaborative decision network. The parameters of the two sub-decision networks are updated centrally in the collaborative decision network. Through continuous interaction with the environment, the collaborative decision-making is achieved by training each decision link. The solution obtained by this method can effectively improve production efficiency and reduce production costs.

[0050] 3) This invention combines scheduling rules with deep reinforcement learning methods, and is applicable to the scheduling of aircraft pulse assembly lines in the context of large-scale processes.

[0051] 4) This invention uses a periodic scheduling method for scheduling, selects scheduling rules at the scheduling time point, and adaptively adjusts scheduling decisions according to changes in the workshop environment, thereby improving the adaptability of the assembly line to different production environments. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the distributed scheduling method for aircraft pulse assembly lines based on QMIX according to the present invention;

[0053] Figure 2 This is a flowchart illustrating the process sequencing and worker allocation of the present invention.

[0054] Figure 3 This is a schematic diagram of the training of the distributed scheduling network for aircraft pulse assembly lines based on QMIX according to the present invention. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0056] Example:

[0057] This invention provides a distributed scheduling method for aircraft pulsed assembly lines based on QMIX. The method includes the following steps:

[0058] S1: The entities at each station on the assembly line are abstractly modeled as station agents in the information space. Each station agent includes two sub-agents: a sorting agent and an allocation agent, a collaborative decision-making module, and a simulation module. The sorting agent and the allocation agent can interact to make collaborative decisions, and they can also interact with the simulation module to update state information. The collaborative decision-making module is used to train the two sub-agents.

[0059] The pulse assembly line is divided into several stations. The assembly task starts from the current station. After each station completes the specified assembly process within a fixed cycle, it enters the next station at the same time to carry out the assembly task of the next assembly cycle. Finally, all assembly processes are completed. This assembly cycle is generally called the assembly cycle time.

[0060] As attached Figure 1 As shown, the assembly line consists of M stations. Each station entity includes an aircraft entity and a work group entity. The aircraft entity has a set of processes to be assembled, J. The work group entity is the assembly work group divided according to different specialties. Each station has B work groups, and each assembly work group has W workers.

[0061] On large aircraft pulsed assembly lines, the processes are large-scale, and there are complex sequential constraints between different processes. To facilitate process management, all processes can be divided into different process packages according to production relationships, with linear predecessor-successor relationships between each process package. Assembly personnel on the assembly line are organized into several assembly teams according to their specialties. The assembly workers in each team have the same skills, and each team has the same number of workers with a skill level w. s Evenly distributed, the actual assembly time t for worker assembly process j. w =w s *t j , t j The original assembly time is the actual assembly time. If multiple people work together to assemble the assembly, the actual assembly time will be the longest among all workers.

[0062] The aircraft pulse assembly line scheduling optimization problem can be described as follows: Given the number of workstations and personnel information, under the premise of satisfying constraints such as process and team skills, optimize the assembly cycle and balance the workstation load by arranging the processing sequence (including the allocation and sequence of processes between workstations) and personnel assignment.

[0063] Furthermore, the overall optimization objective of the aircraft pulse assembly line scheduling is to minimize the assembly cycle time C and the station load variance SI. During the optimization process, each station independently decides its own process sequence and worker allocation scheme, and the overall decision scheme is completed by the sequential decision-making of all stations.

[0064] Furthermore, the constraints for scheduling aircraft pulsed assembly lines include assembly cycle time constraints, shift constraints, process time constraints, and process predecessor and successor constraints. The assembly cycle time constraint is unique to pulsed assembly lines; the actual assembly time at all stations must not exceed the preset assembly time. Simultaneously, the assembly cycle time is also one of the optimization objectives. Therefore, in the assembly line scheduling optimization process, the assembly cycle time is often preset to a relatively large value, and this value is continuously reduced until it no longer changes during the optimization process.

[0065] To efficiently sequence work processes and allocate personnel, this invention employs a distributed scheduling method to establish a multi-level, distributed scheduling system for assembly line entities in the information space. The distributed system mainly consists of station agents distributed across various workstations. Each station agent includes two sub-agents: a sequencing agent and an allocation agent, a collaborative decision-making module, and a simulation module.

[0066] Among them, the station agent is an information system with autonomous decision-making and communication capabilities, distributed in various stations; the sorting agent and the allocation agent are subordinate to the station agents and can plan and allocate resources in the station through interaction; the simulation module simulates the production environment of the station in the information space and is mainly responsible for updating production status information; the collaborative decision-making module is used to train the two sub-agents so that they can make collaborative decisions.

[0067] S2: Establish decision networks in the two sub-agents, establish collaborative decision networks in the collaborative decision module, and randomly initialize network parameters.

[0068] The decision network employs a neural network. A value network and a target network are established in the two sub-agents and the collaborative decision-making module, respectively, and the network parameters are randomly initialized. The input layer of the decision network is a single linear layer, the hidden layer is a single GRU neural network unit, and the output layer is a single linear layer; the input layer of the collaborative decision-making network is a single linear layer, the hidden layer is a single sequence neural unit, and the output layer is a single linear layer.

[0069] S3: Initialize simulation module parameters and assembly cycle time.

[0070] The assembly cycle time is an important constraint for process allocation and also one of the optimization targets. In common assembly line optimization processes, an assembly cycle time with a margin is often set at the beginning, and this value is gradually reduced during the optimization process until the cycle time no longer changes.

[0071] S4: The simulation module drives the two sub-agents to sort and allocate processes at regular scheduling cycles. The simulation module inputs status information to the decision network of the two sub-agents. The decision network outputs the processes to be assembled and the workers to be assigned to the simulation module. The simulation module simulates the assembly process, updates the process and worker status, and saves the assembly line status data and the rules output by the decision network to the experience pool after all processes are assigned.

[0072] The scheduling cycle is a preset value based on production experience and should exceed the maximum assembly time for a single process. (See attached...) Figure 2 As shown, the specific process of the two sub-agents generating scheduling rules through interaction in step S4 is as follows:

[0073] S41 obtains the initial set of assemblable operations, which is the set of operations without immediate predecessors. Since the operations are divided into several operation packages in this example, the initial set of assemblable operations is the set of operations without immediate predecessors of the first operation package;

[0074] The S42 simulation module collects the process status and the status of the workers in the work group and inputs them into the decision network of the two sub-agents;

[0075] The S43 sorting agent decision network outputs sorting rules. The sorting rules are a weight vector w = [w1, w2, w3], w1 + w2 + w3 = 1, which is a linear combination of the sorting rules for the three processes.

[0076] The three sorting rules include:

[0077] d1. The more subsequent processes a process has, the higher its priority (max subsequent processes numbers, MSPN);

[0078] d2. The longer the sum of the times of subsequent processes, the higher the priority (LSSP-1);

[0079] d3. The longer the sum of the times from one process to the next without subsequent processes, the higher the priority (LSSP-2).

[0080] The calculation method is as follows:

[0081] MSPN: Recursively search for the longest subprocess with no subsequent processing time;

[0082] LSSP-1: Calculates the sum of assembly times for all subsequent processes in a given process;

[0083] LSSP-2: Starting from the current process, recursively search for the longest subprocess with no subsequent processes.

[0084] The process priority is calculated as w·[d1,d2,d3]. The priority of all processes is calculated, and the process with the highest priority is the one to be assembled.

[0085] The specific steps for the ranking agent decision network to output ranking rules are as follows:

[0086] The S431 simulation module inputs process status information into the decision network, calculates the Q-value of all rule combinations, and takes the rule combination with the largest Q-value as the output.

[0087] To better explore the global optimum, S432 employs an ε-greedy strategy when outputting rules. The initial ε is set to 0.4. A random number is generated at each decision stage. If this random number is greater than ε, the decision network's output value is output; otherwise, a random rule is output. Simultaneously, to ensure training stability, ε gradually decreases with increasing training iterations.

[0088] The S44 sorting agent drives the allocation agent to make rule-based decisions. Worker allocation rules include: shorter actual assembly time (min worker process time, MWPT), shorter worker completion time (MWCT), and higher worker assembly efficiency (max workers process efficiency, MWPE). The worker allocation rules here are linear combinations of the aforementioned three rules. Similarly, an ε-greedy strategy is used to generate worker allocation rules, and the highest-priority worker is determined using the same calculation method as the process sorting.

[0089] The S45 simulation module adds the operation to this worker's assembly list and compares the worker's estimated completion time with the assembly cycle time:

[0090] S451 If the worker's estimated completion time does not exceed the assembly cycle time, the process assignment is successful. The process is then entered into the simulation module, removed from the set of assembleable processes, the status information of the team workers is updated, and the process proceeds to S46.

[0091] S452 If the worker's completion time exceeds the assembly cycle time after adding this process, the worker's priority is recalculated and the process is reassigned to other workers. If there are other workers whose completion time does not exceed the assembly cycle time, proceed to S46.

[0092] S453 If all workers would exceed the assembly cycle time for this process, then save the process to the next station's assembleable list and remove it from the current station's assembleable process set, then select a higher priority process.

[0093] S46 Repeat S42-S44 until the set of assembleable processes is empty, then move to the next station until all processes are assembled.

[0094] After all processes in S47 are assembled, the process status information, shift worker status information, and global status information at the scheduling time are saved to the experience pool. The maximum actual completion time of all stations is collected as the actual assembly cycle time C for this round of optimization. The actual assembly time T and the assembly time t of each shift are collected. b Calculate rewards And save it to the experience pool. Where α, β, and γ are the assembly time, the actual assembly cycle time of this round of optimization, and the reward coefficients for the assembly time of each shift, respectively, and B is the total number of shifts.

[0095] S5: The collaborative decision-making module samples the data in the experience pool, trains and updates the parameters of the collaborative decision-making network, and outputs the overall process sequence and personnel allocation plan through global training.

[0096] As attached Figure 3 As shown, the collaborative decision-making network training adopts a network training and update method based on QMIX, and the specific steps include:

[0097] The S51 collaborative decision-making module extracts the state data at time t from the experience pool. Scheduling rules at time t-1 Input sorting agent value network, Input the agent value network, calculate the Q-value for all rule weight combinations in both networks, and take the largest Q-value from each. and the selected action

[0098] S52 take Input to the target network of the sorting agent. Input the data into the target network assigned to the agent, and calculate the target Q-value for each network.

[0099] S53 obtains the value network With global state information s t-1 The value network input to the collaborative decision-making network obtains the total global Q-value Q. ae ; to set two target Q values With global state information s t The input is fed into the cooperative target network to obtain the total global target Q value Q. at .

[0100] S54 calculates the objective value f and the global objective Q value Q. at Then, with the rewards r in the experience pool t Adding them together gives the target value f = r t +γ*Q at .

[0101] S54 calculates the difference loss function td error =Q ae -f

[0102] The S54 differential loss function is backpropagated to update the parameters of the collaborative decision network and the parameters of the two decision networks.

[0103] After a certain number of rounds, each target network copies the parameters of the corresponding decision network after a delay.

[0104] S6: Every certain number of rounds, update the beat rate to the minimum value of the previous round. When the beat rate no longer changes, the station agent outputs the final scheduling scheme, i.e., the process sequencing and allocation scheme.

[0105] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A distributed scheduling method for aircraft pulsed assembly lines based on QMIX, characterized in that, The method steps include: S1: Distributed decision-making station agents are set up at each station on the assembly line. Each station agent includes two sub-agents: a sorting agent and an allocation agent, as well as a collaborative decision-making module and a simulation module. S2: Establish decision networks in the two sub-agents, establish collaborative decision networks in the collaborative decision module, and initialize the parameters of each network; S3: Initialize the simulation module parameters and initialize the assembly cycle time based on empirical values; S4: The simulation module inputs state information to the decision networks of the two sub-agents, and the two decision networks output the process sequencing rules and worker allocation rules respectively; the simulation module simulates the process assembly process according to the process sequencing rules and worker allocation rules, and updates the process and team state information; after all process allocations are completed, the state data and the rules obtained from the decisions are saved to the experience pool; S5: The collaborative decision-making module samples data from the experience pool, performs global training on the two decision networks using the QMIX-based network training and update method, and updates the parameters of the decision networks. The specific process is as follows: S51. Extract the state data at time t from the experience pool. Scheduling actions at time t-1 ; respectively Input the value network of the sorting agent. Input the value network for assigning agents, and use the two value networks to calculate all rule weight combinations. Take the largest value from each. value ， and the selected action ， ; Pick Input to the target network of the sorting agent. The input is fed into the target network of the assigned agent, and the two target networks calculate the target. value ， ; S52. Maximize the value network value and global state information The value network input to the collaborative decision-making network obtains a global value. value ; target value and global state information The target network of the collaborative decision-making module, input into the collaborative decision-making module, obtains the global target. value ; S53. Global Target value Rewards from the experience pool Add them together to get the target value ; S54. Based on target value With the global value Calculate the loss function ; S55. Backpropagate the loss function to update the parameters of the collaborative decision network and the two sub-decision networks; S56. After a certain number of rounds, each target network copies the parameters of its corresponding value network after a delay; S6: Every certain number of rounds, update the beat rate to the minimum beat rate of the previous round; when the beat rate no longer changes, the station agent outputs the final process sorting and allocation scheme.

2. The distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 1, characterized in that, The scheduling problem of the pulse assembly line is defined as follows: given the number of workstations and personnel information, under the premise of meeting the constraints of process and team skills, the assembly cycle time and workstation load are optimized by arranging the processing sequence and the allocation scheme of processes to personnel. The overall optimization objective of the distributed scheduling of the aircraft pulse assembly line is to minimize the assembly cycle time C and the station load variance. During the optimization process, each station independently decides its own process sequence and worker allocation scheme, and the overall decision scheme is completed by the sequential decision of all stations.

3. The distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 1, characterized in that, The station agent is an information system with autonomous decision-making and communication capabilities, distributed in various stations, and coordinates the allocation and planning of resources in the stations through interaction; the simulation module is used to simulate the information space of the station production environment, including simulating the assembly process and updating assembly status information. The collaborative decision-making module is used to train two sub-agents so that they can make collaborative decisions.

4. The distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 3, characterized in that, In step S4, the two sub-agents cooperate to complete the process sequencing and worker allocation, as detailed below: S41. Based on the process predecessor and successor constraints, obtain the set of processes without predecessor processes as the initial set of assemblable processes; S42. The simulation module collects the process status and the status of the shift workers and inputs them into the decision network of the two sub-agents; S43. The sorting agent decision network outputs sorting rules and calculates the process with the highest priority; S44. The sorting agent drives the allocation agent, and the allocation agent's decision network outputs worker allocation rules and selects the worker with the highest priority. S45. The simulation module adds the process to the worker's assembly list to simulate the assembly process, assigns the highest priority process to the highest priority worker's assembly list, compares the worker's estimated completion time with the assembly cycle time, and updates the worker's timetable. S46. After all processes have been assigned, the decision information and assembly line status information are stored in the experience pool.

5. The distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 4, characterized in that, The process sequencing rules include: the more subsequent processes a process has, the higher its priority; the longer the sum of the times of the subsequent processes a process has, the higher its priority; and the longer the sum of the times from a process to having no subsequent processes, the higher its priority. The sorting agent decision network outputs a weight vector of a linear combination of the above-mentioned process sorting rules; The simulation module inputs process status information into the decision network and calculates the output of all possible combinations. Value, take The rule combination with the highest value is output.

6. The distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 4, characterized in that, The worker allocation rules include: workers with shorter actual assembly time are given priority, workers with shorter completion time are given priority, and workers with higher assembly efficiency are given priority. The output of the allocation agent decision network is a linear combination of the worker allocation rules described above; The simulation module inputs the status information of the shift workers into the allocation agent decision network, and calculates all rule combinations. Value, take The rule combination with the highest value is output.

7. A distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 4, characterized in that, The process of updating the worker's timetable is as follows: S451. If the worker's estimated completion time after adding the process does not exceed the predetermined assembly cycle time, the process allocation is successful and input into the simulation module to update the set of assembleable processes and the status data information of the team workers. S452. If the worker's estimated completion time exceeds the predetermined assembly cycle time after the process is added, the worker's priority is recalculated and the process is reassigned to other workers, if there are other workers whose completion time does not exceed the assembly cycle time. S453. If all workers will exceed the predetermined assembly cycle time for this process, save the process to the assembly list of the next station and delete it from the assembly process set of this station, and reselect a process with higher priority.

8. A distributed scheduling method for aircraft pulse assembly lines based on QMIX according to any one of claims 4-7, characterized in that, After all processes are assembled, the maximum actual completion time of all stations is collected as the actual assembly cycle time for this round of optimization. Collect actual assembly time and assembly time for each work group Calculate rewards Save to the experience pool; in, 、 、 These are the assembly time, the actual assembly cycle time in this round of optimization, and the bonus coefficient for the assembly time of each shift. B The total number of work groups This represents the average assembly time.

9. A distributed scheduling method for aircraft pulse assembly lines based on QMIX according to claim 8, characterized in that, The decision-making networks established in the two sub-agents and the collaborative decision-making network established in the collaborative decision-making module respectively include a value network and a target network, wherein the target network is used to delay copying the parameters of the value network.

Images