Aircraft pulsation assembly line balance optimization method and equipment based on graph neural network and deep reinforcement learning, and medium

Abstract

The invention discloses an aircraft pulsation assembly line balance optimization method and device based on a graph neural network and deep reinforcement learning and a medium, and relates to the field of intelligent manufacturing, and the method comprises the following steps: S1, obtaining the initial information of an assembly line, constructing a disjunction graph model, and building a Markov decision process; s2, performing feature embedding extraction on the disjunction graph model by using a graph neural network to obtain node embedding and global embedding; s3, the nodes and the global embedding are input into an Actor strategy network, probability distribution is calculated to select a task to be assembled, and parameters of the graph nerve and the strategy network are dynamically adjusted; and S4, judging whether the current assembly line balance scheme meets a preset condition, if so, outputting the scheme, and otherwise, returning to the step S2 to continue training. The scheduling deviation caused by incomplete model expression is avoided, the modeling mode provides a solid foundation for real-time decision making, and the response speed of the assembly line to the dynamic production environment is increased.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to aircraft pulse assembly line balancing optimization technology, specifically a method, equipment, and medium for aircraft pulse assembly line balancing optimization based on graph neural networks and deep reinforcement learning. Background Technology

[0002] With the rapid development of intelligent manufacturing technology, the aircraft assembly industry faces increasing complexity and efficiency requirements. Aircraft Pulse Assembly Lines (APALBPs), as a highly efficient production mode, have become a mainstream solution for improving production efficiency through the cyclical (takt) movement of fixed workstations. In practical applications, effectively solving the Aircraft Pulse Assembly Line Balancing Problem (APALBP)—that is, rationally allocating assembly tasks to each workstation while meeting various constraints, optimizing takt time, balancing load distribution, and improving resource utilization—has become a core bottleneck restricting its performance.

[0003] Current research on APALBP is relatively limited, and its optimization algorithms are mainly divided into two categories: exact solutions and approximate solutions. Exact solutions, such as branch and bound and mathematical programming, have high accuracy for small-scale problems, but as the problem size increases, they are prone to the "curse of dimensionality," resulting in extremely low computational efficiency in practical applications. Approximate solutions, represented by metaheuristic methods such as genetic algorithms, are currently the mainstream methods in APALBP. Their advantage lies in finding the optimal approximate scheduling solution relatively quickly, but they still require a long solution time for complex scheduling problems and generally rely on static environment assumptions, resulting in poor adaptability.

[0004] Chinese invention patent CN118551468A discloses a resource-constrained aircraft assembly line balancing optimization method, proposing a genetic algorithm to solve the aircraft assembly line balancing problem. This method evenly distributes assembly tasks to each workstation, improving the uneven workstation load to some extent. Chinese invention patent CN118607854A discloses a multi-agent-based aircraft assembly line scheduling method, utilizing a genetic algorithm for optimization to achieve joint optimization of the number of workers and cycle time. While traditional metaheuristic algorithms have achieved some success in solving the APALBP problem, these methods typically rely on empirically setting the initial cycle time and then assigning tasks to each workstation. After each state update, the assembly time of each workstation needs to be repeatedly compared to adjust the actual cycle time, resulting in low efficiency, reduced task allocation flexibility, difficulty in self-learning, and inability to quickly generate solutions in scenarios with high real-time requirements. APALBP is an NP-hard problem involving complex task station allocation and task sorting. The computational difficulty increases dramatically with the number of tasks and workstations.

[0005] In recent years, reinforcement learning methods have demonstrated excellent policy learning capabilities in workshops, possessing advantages such as interaction with the environment and autonomous optimization. However, traditional reinforcement learning heavily relies on human experience to select and design multiple scheduling rules as the action space to choose the current task. When applied to the aircraft assembly line balancing problem, multiple scheduling rules need to be designed according to the characteristics of the aircraft assembly line. When the manufacturing system environment changes, a significant amount of effort is required to re-determine the scheduling rules.

[0006] Therefore, current research on APALBP has not fully combined the dual requirements of the complexity of assembly task structure and real-time intelligent decision-making, and lacks an efficient optimization method that simultaneously takes into account the ability to express task dependencies and self-learning scheduling capabilities. Summary of the Invention

[0007] This invention overcomes the shortcomings of existing technologies and provides a method, device, and medium for optimizing the balancing of aircraft pulsed assembly lines based on graph neural networks and deep reinforcement learning. Addressing the balancing problem of aircraft pulsed assembly lines, with the goal of optimizing cycle time, it leverages the flexibility of graph models to design a novel solution representation based on disjunctive graphs. APALBP is expressed as a Markov decision process embedding heuristic station allocation rules, ensuring rapid station allocation for task selection without requiring additional learning of station allocation knowledge, thus guaranteeing performance and computational efficiency. Furthermore, an integrated graph neural network and actor-critic structure is designed to capture the complex relationships between different tasks and achieve end-to-end real-time decision-making without explicitly designing additional scheduling rules. The entire structure is size-independent, meaning it maintains a certain level of generalization across instances of different sizes.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: First aspect: The present invention provides a method for optimizing the balancing of an aircraft pulsed assembly line based on graph neural networks and deep reinforcement learning, comprising the following steps:

[0009] S1. Obtain the initial information of the aircraft pulse assembly line, construct the disjunctive graph model, and establish a Markov decision process based on the disjunctive graph model, in which heuristic station allocation rules are embedded.

[0010] S2. Use graph neural networks to extract feature embeddings from the disjunctive graph model to obtain node embeddings and global embeddings;

[0011] S3. Input the node embedding and global embedding into the Actor policy network, calculate the probability distribution in the action space to select the task to be assembled, and dynamically adjust the parameters of the graph neural network and the policy network.

[0012] S4. Determine whether the current assembly line balancing scheme meets the preset conditions. If it does, output the scheme; otherwise, return to step S2 to continue training.

[0013] In a preferred embodiment of the present invention, in step S1, the initial information includes the sequential constraints of assembly tasks, assembly time, the type of worker shifts required for the task, and the number of workstations.

[0014] In a preferred embodiment of the present invention, the construction of the disjunctive graph model in step S1 specifically includes:

[0015] The disjunction graph model is represented as follows: ;in, This represents the set of all assembly tasks for the aircraft, including all assembly task nodes and virtual start and end nodes. S and T represent virtual nodes; C represents the set of conjunctive edges, which represent priority constraints between assembly tasks; D represents the set of disjunctive edges, which represent shared constraints of assembly tasks on the same worker shift type.

[0016] Each node contains characteristic information, including node status, assembly time, estimated completion time, and number of successor tasks.

[0017] In a preferred embodiment of the present invention, in step S1, the embedded heuristic station allocation rule specifically includes:

[0018] Once the start time or estimated start time of all assembly tasks is determined, the initial cycle time is the ratio of the total assembly time to the number of stations.

[0019] Based on the start and end times of the tasks, tasks are distributed to workstations with load balancing.

[0020] The final cycle time is determined by the maximum assembly time of all workstations in the work group.

[0021] In a preferred embodiment of the present invention, in step S1, the state of the Markov decision process is represented by the current disjunctive graph model, the action space consists of the currently feasible assembly tasks, the state transition includes updating the disjunctive graph direction and recalculating the task allocation and completion time, and the reward function is set as the difference between the tick time of the previous state and the tick time of the current state.

[0022] In a preferred embodiment of the present invention, step S2, which involves using the graph neural network to perform feature embedding extraction on the disjunctive graph model, specifically includes:

[0023] The node embedding is iteratively updated using a multilayer perceptron, and the calculation formula is as follows:

[0024] ;in, This represents the l-th generation embedding information of node n; Indicates that the lth generation has parameters Multilayer perceptron; Represents the set of neighboring nodes of node n; It is any number that can be learned;

[0025] The global embedding h is calculated using the average pooling function. G .

[0026] In a preferred embodiment of the present invention, in step S3, calculating the probability distribution in the action space specifically includes: using node embedding and global embedding together as input to the Actor policy network; calculating the score of each action through a multilayer perceptron; and deriving the probability distribution of the action using a softmax function.

[0027] The parameter adjustment of the graph neural network and policy network specifically includes: using the Actor network as the action selection network to learn a policy that maximizes cumulative reward; using the Critic network to evaluate state value and sharing the multilayer perceptron structure with the Actor network; and updating the parameters of the graph neural network, Actor network, and Critic network through a proximal policy optimization algorithm.

[0028] In a preferred embodiment of the present invention, in step S4, the preset condition is that the cycle time reaches a minimum value or the number of training iterations reaches a threshold.

[0029] Second aspect: The present invention provides an electronic device, comprising: at least one processor; and a memory communicatively connected to at least one of the processors;

[0030] The memory stores a computer program that is executed by at least one of the processors, such that the at least one processor is able to execute the aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning as described above.

[0031] Thirdly, the present invention provides a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning as described above.

[0032] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0033] (1) This invention provides a method, equipment and medium for balancing optimization of aircraft pulsed assembly lines based on graph neural networks and deep reinforcement learning. By constructing a disjunctive graph model, it accurately represents the task dependencies and workstation constraints in the aircraft pulsed assembly line. It can fully capture the priority of assembly tasks and shared resource information. By transforming the problem into a graph structure and combining it with graph neural networks for feature extraction, it can deeply explore the complex relationships between tasks, thereby making the model more accurate and comprehensive in expressing the problem, improving the accuracy and reliability of the optimization scheme. Compared with the existing technology that relies on static assumptions and manual rules, this invention avoids scheduling deviations caused by incomplete model expression. Its modeling method provides a solid foundation for real-time decision-making, thereby enhancing the model's decision-making ability under different data distributions and improving the assembly line's response speed to dynamic production environments.

[0034] (2) In this invention, a heuristic station allocation rule is embedded into the Markov decision process. This rule dynamically allocates workstations while sorting tasks to ensure load balance. By integrating the rule into the reinforcement learning framework, the agent can focus on task sorting decisions without having to learn additional station allocation knowledge. This enables automatic load balancing between workstations, effectively reducing resource idleness and overcrowding. Compared with traditional metaheuristic methods that require repeated rule adjustments and reliance on manual intervention, this invention can significantly improve computational efficiency and scheduling flexibility, thereby maintaining stable production line operation while minimizing cycle time.

[0035] (3) The present invention adopts a structure that combines graph neural network and deep reinforcement learning, and achieves self-learning optimization through ActorCritic network and near-end policy optimization algorithm. It can learn policies autonomously from interactive data and dynamically adjust network parameters to adapt to different problem scales. Its self-learning mechanism can bring efficient real-time decision-making capabilities without relying on pre-designed scheduling rules, which can significantly shorten the solution generation time and improve versatility, thereby generalizing to assembly line environments of different scales and providing sustainable optimization support for intelligent manufacturing systems. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to Embodiment 1 of the present invention;

[0038] Figure 2 This is a disjunctive diagram of the aircraft pulse assembly line balancing problem in Embodiment 1 of the present invention, wherein (a) is a disjunctive diagram containing initial information, and (b) is a disjunctive diagram of a feasible solution, where workers assemble tasks sequentially and tasks are also assigned to different workstations.

[0039] Figure 3 This is a schematic diagram of the state transition in Embodiment 1 of the present invention;

[0040] Figure 4 This is the assembly task Gantt chart of Embodiment 1 of the present invention;

[0041] Figure 5 A schematic diagram of an electronic device structure that can be used to implement Embodiment 1 of the present invention is shown. Detailed Implementation

[0042] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0044] Example 1: As Figure 1 As shown, the aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning includes the following steps:

[0045] S1. Obtain the initial information of the aircraft pulse assembly line, construct the disjunctive graph model, and establish a Markov decision process based on the disjunctive graph model, in which heuristic station allocation rules are embedded.

[0046] S2. Use graph neural networks to extract feature embeddings from the disjunctive graph model to obtain node embeddings and global embeddings;

[0047] S3. Input the node embedding and global embedding into the Actor policy network, calculate the probability distribution in the action space to select the task to be assembled, and dynamically adjust the parameters of the graph neural network and the policy network.

[0048] S4. Determine whether the current assembly line balancing scheme meets the preset conditions. If it does, output the scheme; otherwise, return to step S2 to continue training.

[0049] In some specific implementations, in step S1, in order to construct the disjunctive graph model and establish the Markov decision process, it is necessary to first obtain the initial information of the aircraft pulse assembly line. The initial information includes: the sequential constraints of the assembly tasks, the assembly time, the type of worker shifts required for the tasks, and the number of workstations.

[0050] In this embodiment, the constructed aircraft pulse assembly line dissociation diagram model is represented as follows: ;

[0051] in, The set represents all assembly tasks for the aircraft. These tasks are represented as nodes in the graph. The node set contains information about all tasks that need to be assembled, as well as two empty nodes representing the start and end of assembly. S and T represent virtual nodes. C represents the set of conjunctive edges, which represent priority constraints between assembly tasks. D represents the set of disjunctive edges, which represent shared constraints between two assembly tasks and worker teams. When two assembly tasks are assembled by the same team type, the corresponding task nodes are connected by disjunctive edges.

[0052] Furthermore, the node feature information design includes: node status, assembly time, estimated completion time, and number of subsequent tasks; the node status includes 0 and 1, which represent that the task has been arranged and the task has not been arranged, respectively; the assembly time represents the assembly time required for the task; the number of subsequent tasks is the number of tasks immediately following the specific assembly task; at time t, if task n has not been arranged, the completion time of its predecessor node is recursively calculated to obtain its estimated completion time; if task n has been arranged, the estimated completion time is equal to its actual completion time.

[0053] For example, Figure 2 A simplified example is shown, comprising 12 tasks, 3 workstations, and 3 worker types. Figure 2 (a) describes the assembly task organization for a specific model, with different colors representing different worker types. Figure 2 (b) is a feasible solution to this example; it can be seen that assembly tasks 1, 2, 4, and 5 are assigned to station 1, assembly tasks 3, 6, 7, and 8 are assigned to station 2, and assembly tasks 9, 10, 11, and 12 are assigned to station 3. At station 1, assembly tasks 1 and 4 are assembled sequentially by worker shift type 2. This disjunctive graph clearly and effectively expresses the solution to the APALBP problem. Furthermore, the graph structure of the disjunctive graph provides an important foundation for the deep integration of graph neural networks and APALBP.

[0054] Specifically, for the APALBP problem being addressed, two key decisions must be made: assigning tasks to workstations and determining task start times (task sequencing). An effective task assignment strategy can ensure workload balance among workstations, while an optimized task sequencing method can significantly improve productivity while maintaining workstation load balance.

[0055] Specifically, a heuristic workstation allocation rule is designed to assign tasks to workstations to maintain load balance and calculate the cycle time. The designed balancing heuristic algorithm is further integrated into the Markov decision process, enabling the agent to focus on ranking decisions without ignoring workstation load balance.

[0056] The allocation rules are detailed below:

[0057] S11. Once the start time / estimated start time of all assembly tasks is determined, initialize the cycle time to makespan / K, where K is the number of stations.

[0058] S12. For each task, obtain the start time T. s and end time T e ;

[0059] S13, if Then the task is assigned to the station. Otherwise, they will be assigned to a station. ;

[0060] S14. Calculate the cycle time. For each station, each assembly task is determined, and the start and end times are also determined. Therefore, the total assembly time for different shift types is determined. Thus, the final cycle time can be determined by calculating the longest assembly time for a shift among all stations.

[0061] Furthermore, the representation of APALBP in this invention is based on disjunctive graphs, and its Markov decision process is as follows: Here, denoted in order, are state, action, state transition probability, reward, and discount factor, as follows:

[0062] S111, State: At decision time t, disjunctive graph This indicates the current state, where, Including all disjunction arcs with specified directions, For disjunctive arcs whose direction is yet to be determined, the nodes in set V contain four key pieces of information:

[0063] a. If the node has already been assigned, the value is 1; otherwise, it is 0.

[0064] b. The assembly time corresponding to this node;

[0065] c. Lower bound of the estimated completion time of the assembly task corresponding to this node At decision time t, if the task has already been assembled, then the estimated lower bound of its completion time is equal to its actual completion time; otherwise, it is calculated... PT n Let n be the set of all preceding tasks. The estimated lower bound of task n is determined by the node with the slowest completion time among its preceding nodes. If a task has no preceding tasks, its estimated start time is 0. This allows us to obtain the estimated lower bounds of the completion times for all tasks. Using the allocation rules, we can calculate the estimated allocation and tempo of the current decision point. Once all tasks are completed, we can obtain the final allocation and tempo.

[0066] d. The number of immediate successor nodes of a node reflects the node's influence;

[0067] S112, Action: An action is an assembly task that meets the conditions at decision time t, where the action space A t It is the set of tasks that meet the conditions at the current time step;

[0068] S113, State Transition: Once a task is determined to be assigned, its completion time is fixed, and the direction of the disjunction graph is updated; the estimated completion times of other unplanned tasks are updated, and based on the allocation rules and the calculation method for estimated completion times, the new workstation allocation and completion time for each task can be obtained; such as Figure 3 As shown, in state S6, task 7 was selected as the next action, therefore the disjunction arc was updated to point from 8 to 7, and the current position needs to be adjusted according to the allocation rules; Figure 3 In the middle, Task 4 was reassigned to position 2 to ensure the balance of the assembly line, and then the current beat CT was updated;

[0069] S114, Reward: The goal of this method is to learn a strategy to optimize sorting and allocation, minimizing the tick time. This example combines the algorithm's optimization objective with the characteristics of the APALBP problem, setting the reward function as... That is, the difference between the beat of the previous state and the beat of the current state, when the discount factor... Accumulated maximum reward Therefore, maximizing the reward is consistent with minimizing the beat.

[0070] In some specific implementations, step S2 uses a graph neural network to process the graph structure data of the extracted graph, deeply mining information in the network topology to generate graph embeddings as input to the policy network. This specifically includes the following sub-steps:

[0071] S21. Node encoding, also known as graph embedding, is a compressed representation of graph data. This invention uses a Graph Isomorphism Network (GIN) to extract disjunctive graph information. For each node n corresponding to an assembly task in the disjunctive graph, its state is updated as follows:

[0072] ;

[0073] in, This represents the l-th generation embedding information of node n; This represents the initial input information for node n; Indicates that the lth generation has parameters Multilayer perceptron; Represents the set of neighboring nodes of node n; It is any learnable number. The parameters of the input, hidden, and output layers of the graph neural network are trained using a multilayer perceptron (MLP). Batch normalization is performed after each MLP layer to enhance the stability of training.

[0074] S22. After L iterations, the final graph-level global information h is obtained. G Calculated using the average pooling function: ;

[0075] The multilayer perceptron of this invention comprises two hidden layers, each consisting of 64 neurons; each node contains 4 features, and the input layer dimension is 4.

[0076] In some specific implementations, in step S3, the task to be assembled is directly selected through action selection; after obtaining the node embedding... and global embedded h G Then, the action selection network (task decision-making) calculates the action space A. t The probability distribution on the t-th step is used to determine the action at step t.

[0077] Specifically, feature processing is performed, combining node embeddings and global embeddings to obtain the input for the action selection network. The policy network uses an MLP to calculate the score for each feasible action, denoted as . .

[0078] Furthermore, action selection is performed, and the probability distribution of the actions is derived using the softmax function:

[0079] ;

[0080] Furthermore, by calculating the probability of each action in the action space, sampling is performed based on the probability during model training to select the current action; during model validation and testing, a greedy strategy is adopted to select the action with the highest probability.

[0081] Furthermore, the graph neural network and action selection network are optimized by using the Proximal Policy Optimization (PPO) algorithm for training. The Actor network, as the action selection network, learns a policy that maximizes the cumulative reward. The Critic network shares the same multilayer perceptron (MLP) network architecture as the Actor network, takes global embedding as input, and outputs a scalar value to evaluate the decision-making effect in the state.

[0082] In some specific implementations, in step S4, the model is evaluated based on the obtained results to determine whether the results meet expectations. If they do, the current solution is output; otherwise, the parameters of the graph neural network, Actor network, and Critic network are trained using the proximal policy optimization algorithm, and the process returns to step S2 for retraining.

[0083] To further simplify and make the objectives and effects of the present invention easier to understand, the present invention will be further illustrated in conjunction with the following specific embodiments.

[0084] To verify the effectiveness of the proposed model, instance parameters were set based on the characteristics of APALBP-2. Table 1 shows the parameters, including the number of assembly tasks (N), the number of worker shift types (M), the number of workstations (K), and the task assembly time (t). n The duration of the experiments ranged from 2 to 9 hours; all experiments were run on machines with Intel(R) Core(TM) i5-13400F CPUs and NVIDIA GeForce RTX 3060 Ti GPUs.

[0085] Table 1: Parameter settings for generating instances

[0086] parameter scope N {20,50,80,100} M 3 K 3,4 <![CDATA[t n ]]> [2,9]

[0087] To train and evaluate the proposed model, 50 instances of each size were generated for training, 40 for validation, and 40 for testing. For each instance size, the model was trained for 1000 iterations, with the policy network updated at each iteration.

[0088] The GNN used for feature extraction has 3 layers (including the input layer). The MLP in the feature extraction network has 2 layers and a hidden dimension of 64. The MLPs in the actor and critic networks have 3 layers and a hidden dimension of 32. The pruning parameter is set to 0.2, and the coefficients of the value function, policy loss, and entropy are set to 1.0, 2.0, and 0.01, respectively. The model is trained using the Adam optimizer, and the learning rate is dynamically adjusted using StepLR provided by PyTorch. The discount factor is set to... Learning rate lr = 2 × 10 -4 .

[0089] This method is compared with scheduling rules, and the selected rules are: longest processing time, shortest processing time, most succeeding tasks, and fewest succeeding tasks.

[0090] Compared with metaheuristic algorithms, this method was selected as the genetic algorithm (GA).

[0091] Figure 4 A Gantt chart visualization of the solution results for one example is shown.

[0092] Expected experimental results: The algorithm demonstrates competitiveness with genetic algorithms in cycle time optimization, but achieves near real-time solution times across different scales; it exhibits good generalization performance, adapting to aircraft with varying assembly sizes. For example... Figure 4 As shown, the present invention can effectively achieve task sorting and station allocation, while ensuring that the load on each station is relatively even.

[0093] Example 2: Figure 5A schematic diagram of an electronic device structure that can be used to implement Embodiment 1 of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0094] like Figure 5 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0095] Furthermore, the processor 11 can perform various appropriate actions and processes based on a computer program stored in the read-only memory (ROM) 12 or a computer program loaded from the storage unit 18 into the random access memory (RAM) 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0096] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0097] Furthermore, processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as methods for resource management of a database.

[0098] In some specific embodiments, the method for managing database resources can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the method for managing database resources described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the method for managing database resources by any other suitable means (e.g., by means of firmware).

[0099] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0100] Computer programs used to implement the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0101] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method for resource management of a database provided by this invention. The computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0102] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0103] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0104] Optionally, the computing system may include clients and servers. Clients and servers are generally geographically separated and typically interact via a communication network. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. This addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0105] The above description is based on the preferred embodiments of the present invention. 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 exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and 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.

[0106] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for balancing and optimizing aircraft pulsed assembly lines based on graph neural networks and deep reinforcement learning, characterized in that, Includes the following steps: S1. Obtain the initial information of the aircraft pulse assembly line, construct the disjunctive graph model, and establish a Markov decision process based on the disjunctive graph model, in which heuristic station allocation rules are embedded. S2. Use graph neural networks to extract feature embeddings from the disjunctive graph model to obtain node embeddings and global embeddings; S3. Input the node embedding and global embedding into the Actor policy network, calculate the probability distribution in the action space to select the task to be assembled, and dynamically adjust the parameters of the graph neural network and the policy network. S4. Determine whether the current assembly line balancing scheme meets the preset conditions. If it does, output the scheme; otherwise, return to step S2 to continue training.

2. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S1, the initial information includes the sequential constraints of the assembly tasks, the assembly time, the type of worker shifts required for the task, and the number of workstations.

3. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S1, constructing the disjunctive graph model specifically includes: The disjunction graph model is represented as follows: ;in, This represents the set of all assembly tasks for the aircraft, including all assembly task nodes and virtual start and end nodes. S and T represent virtual nodes; C represents the set of conjunctive edges, which represent priority constraints between assembly tasks; D represents the set of disjunctive edges, which represent shared constraints of assembly tasks on the same worker shift type. Each node contains characteristic information, including node status, assembly time, estimated completion time, and number of successor tasks.

4. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S1, the embedded heuristic station allocation rule specifically includes: Once the start time or estimated start time of all assembly tasks is determined, the initial cycle time is the ratio of the total assembly time to the number of stations. Based on the start and end times of the tasks, tasks are distributed to workstations with load balancing. The final cycle time is determined by the maximum assembly time of all workstations in the work group.

5. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S1, the state of the Markov decision process is represented by the current disjunctive graph model, the action space consists of the currently feasible assembly tasks, the state transition includes updating the disjunctive graph direction and recalculating task allocation and completion time, and the reward function is set as the difference between the tick time of the previous state and the tick time of the current state.

6. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S2, the feature embedding extraction of the disjunctive graph model using the graph neural network specifically includes: The node embedding is iteratively updated using a multilayer perceptron, and the calculation formula is as follows: ;in, This represents the l-th generation embedding information of node n; Indicates that the lth generation has parameters Multilayer perceptron; Represents the set of neighboring nodes of node n; It is any number that can be learned; The global embedding h is calculated using the average pooling function. G .

7. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S3, calculating the probability distribution in the action space specifically includes: using node embeddings and global embeddings together as input to the Actor policy network; calculating the score of each action through a multilayer perceptron; and deriving the probability distribution of the action using a softmax function. The parameter adjustment of the graph neural network and policy network specifically includes: using the Actor network as the action selection network to learn a policy that maximizes cumulative reward; using the Critic network to evaluate state value and sharing the multilayer perceptron structure with the Actor network; and updating the parameters of the graph neural network, Actor network, and Critic network through a proximal policy optimization algorithm.

8. The aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning according to claim 1, characterized in that: In step S4, the preset condition is that the cycle time reaches a minimum value or the number of training iterations reaches a threshold.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores a computer program executed by at least one of the processors, which enables the at least one processor to perform the aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the aircraft pulse assembly line balancing optimization method based on graph neural networks and deep reinforcement learning as described in any one of claims 1-8.

Images