Hybrid flow shop scheduling method based on dynamic arrival of power battery module workpieces
By adopting a hybrid flow shop scheduling method based on Dueling DQN network, the scheduling problem of power battery module workshop in dynamic environment is solved, real-time adaptive scheduling is realized, the total overdue time of workpieces is reduced, and production efficiency and order delivery on time rate are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUZHOU UNIV
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-08
AI Technical Summary
Existing power battery module workshop scheduling methods suffer from problems such as short-sighted scheduling rules, excessive computation time of heuristic algorithms, and inability to balance real-time performance and global optimization performance when facing complex working conditions such as dynamic arrival of workpieces and random machine failures.
A hybrid flow workshop scheduling method based on the dynamic arrival of power battery module workpieces is adopted. Combined with the Dueling DQN network, state characteristics, scheduling rules and reward functions are designed to perceive environmental changes in real time. The optimal scheduling scheme is solved through the Dueling DQN network to arrange module workpieces to machines and update the scheduling status.
It enables real-time adaptive scheduling in dynamic environments, effectively reducing the total overdue time of workpieces, mitigating the negative impact of machine failures on the overall delivery time, and improving the continuity and efficiency of production activities.
Smart Images
Figure CN121998347A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of dynamic workshop scheduling technology, and in particular to a hybrid flow workshop scheduling method based on the dynamic arrival of power battery module workpieces. Background Technology
[0002] The production of power battery modules is a crucial link in the new energy vehicle industry chain, involving multiple complex processes. Shop floor scheduling, as the core of production management, directly determines production efficiency and resource utilization. Traditional shop floor scheduling research and applications typically assume a static production environment, meaning all processing task information is known and fixed before scheduling begins. Based on this assumption, traditional rule-based scheduling algorithms, mathematical programming methods, and conventional heuristic algorithms can calculate relatively ideal scheduling schemes. However, in actual power battery module production sites, the production environment is highly dynamic and uncertain. This is mainly reflected in the following two aspects: the arrival of workpieces is random, production orders are often dynamically issued, and the time when workpieces arrive at the production line is unpredictable; machine resources are unreliable, and production equipment may experience sudden failures.
[0003] Faced with the aforementioned dynamic workshop environment, existing static scheduling methods have significant limitations: First, they lack real-time adaptive capabilities; the pre-scheduling schemes generated by traditional methods often fail when encountering sudden disturbances, making it impossible to adjust production priorities in real time. Second, rescheduling is lagging; optimization methods relying on global recalculation are time-consuming and cannot meet the needs of rapid response on the production floor. Therefore, there is an urgent need for an efficient scheduling method for power battery module workshops that can perceive environmental changes in real time and respond to machine failures and the arrival of dynamic workpieces. Summary of the Invention
[0004] To address the aforementioned issues, the present invention aims to provide a hybrid flow workshop scheduling method based on the dynamic arrival of power battery module workpieces. This method solves the technical problems of existing power battery module workshop scheduling technologies when facing complex working conditions such as dynamic arrival of workpieces and random machine failures, including short-sighted scheduling rules, excessively long computation time of heuristic algorithms, and the inability to balance real-time performance and global optimization performance.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A hybrid assembly line scheduling method based on the dynamic arrival of power battery module workpieces includes the following steps: Step S1: Based on the uncertainties in the power battery module production workshop, construct an order model for dynamic arrival of workpieces and a machine model that considers faults; Step S2: Based on the production constraints of the power battery module, and combining the order model of dynamic arrival of workpieces and the machine model that considers faults, establish a hybrid flow shop scheduling mathematical model with minimizing overdue time as the optimization objective. Step S3: Based on the mathematical model of hybrid flow shop scheduling, design state characteristics, scheduling rules and reward functions according to the real-time information of the shop; Step S4: Solve for the optimal scheduling scheme based on the Dueling DQN network, output the optimal scheduling rule according to the state characteristics in step S3, arrange the module workpieces to the machine, and update the scheduling status.
[0006] Furthermore, the order model is constructed based on the dynamic arrival characteristics of workpieces in the production of power battery modules, specifically as follows: the arrival of power battery module workpieces in the workshop follows a Poisson process, and each workpiece... The arrival time is a random variable. arrival time interval Follows an exponential distribution: (13) (14) in, It is the average arrival time of the workpiece. It is the workpiece index.
[0007] Furthermore, the machine model is based on the fault start and end times of the machines in the power battery module workshop, as detailed below: The machine malfunction began. and end time For random variables, machine Failure interval and repair time Follows an exponential distribution: (15) (16) (17) (18) in, It's the machine serial number. It is the fault number. It is the machine's mean time between failures. It is the machine's average repair time.
[0008] Furthermore, the production objective is to minimize the total overdue time for all workpieces, as follows: (19) In the formula, The total number of workpieces, and the total delay time for all workpieces is determined by the total delay time for each workpiece. Deadline and completion time calculate; The mathematical model for scheduling the hybrid flow workshop includes the following constraints: Workpiece arrival constraint: Each workpiece arrives at the workshop at a different time, and can only be processed after it arrives at the workshop. (20) in, It is a workpiece First process Start time; Machine fault constraints: For machines The machine must not be in a faulty state during scheduling and processing: (twenty one) in, It is a process In the machine On the processing time, It is a process The start time, It is a decision variable, when the process Assigned to machine The value is 1 if it is true, and 0 otherwise. It is a decision variable, when the machine... The value is 0 when there is a fault, and 1 otherwise.
[0009] Furthermore, the state space includes seven features extracted from the workpieces, machines, and processes in the module workshop, including: workpiece features, such as the average completion rate of the workpiece, the standard deviation of the workpiece completion rate, and the average delay rate of the workpiece; machine features, such as the average machine utilization rate, the standard deviation of the machine utilization rate, and the average machine failure rate; and process features, such as the average delay rate of the process.
[0010] Furthermore, the scheduling rules include five composite scheduling rules, namely: The first step of the composite scheduling rule is to filter the set of workpieces to be scheduled, select the workpieces that are expected to be delayed, select the workpiece with the longest delay time from all the delayed workpieces, and then traverse all the machines to be scheduled and select the machine with the shortest processing time from the current set of machines. The composite scheduling rule 2 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the longest delay time from all the delayed workpieces, and then traverses all the machines to be scheduled, selects the machine with the lowest failure rate from the current set of machines. The composite scheduling rule 3 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the longest delay time from all the delayed workpieces, and then traverses all the machines to be scheduled, selects the machine with the lowest utilization rate from the current set of machines. The composite scheduling rule 4 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the shortest processing time from the current set of machines. The composite scheduling rule 5 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the lowest failure rate from the current set of machines. The composite scheduling rule 6 first filters the set of workpieces to be scheduled, selects the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the lowest utilization rate from the current set of machines.
[0011] Furthermore, the reward function is related to the delay time. To achieve the minimum delay time, a positive reward is given if the delay time at the current moment is lower than that at the next moment; otherwise, a negative reward is given. This encourages the scheduling system to prioritize scheduling tasks with tight delivery deadlines. (twenty two) in, It is the total delay time of all workpieces in the workshop at the current moment.
[0012] Furthermore, the Dueling DQN network includes a sequentially connected common feature extraction layer, a splitting layer, and an aggregation layer; the common feature extraction layer consists of several fully connected layers, used to extract high-order features of the state feature vector; the splitting layer includes parallel state value branches and action advantage branches; the state value branches are used to output the state value function of the current state, and the action advantage branches are used to output the advantage function of each scheduling action; the aggregation layer is used to aggregate the state value function and the advantage function to obtain the Q value of each scheduling action, as shown in formula (13): (twenty three) in, This is the seven-dimensional state feature vector at the current moment. For the selected scheduling action, These are the parameters for the fully connected layer of the network. and These are parameters for the unique layers of value stream and advantage stream, respectively. For the size of the action space, It is a state-value function. It is the action advantage function.
[0013] Furthermore, the parameter update of the target network in the Dueling DQN network adopts a soft update mechanism, specifically including: after each parameter iteration update of the main network, obtaining the current weight parameters of the target network from the current weight parameters of the main network; using a preset soft update coefficient, calculating the linear weighted sum of the weight parameters of the main network and the current weight parameters of the target network; assigning the linear weighted sum to the target network, so that the parameters of the target network smoothly approximate the parameters of the main network. The specific calculation of the soft update is as follows: (twenty four) in, For the updated target network parameters, The preset soft update coefficient has a value range of [value range missing]. .
[0014] Furthermore, the Dueling DQN network employs a decay mechanism. Greedy strategy training specifically includes: setting a preset exploration rate. The initial value, minimum value, and decay step size of the model are determined. In the early stage of model training, a high exploration rate is maintained, allowing the scheduling agent to randomly select a scheduling rule with a high probability to explore the global state space. As the number of training iterations increases, the exploration rate is gradually reduced according to the decay step size until the exploration rate reaches the minimum value. In the later stage of model training, the scheduling agent is made to select the scheduling rule with the largest Q value with a high probability, and the learned optimal strategy is used to make scheduling decisions.
[0015] The present invention has the following beneficial effects: 1. This invention addresses the complex task characteristics of power battery module production workshops, comprehensively considering multiple constraints such as workpiece arrival, machine failure, work sequence, and equipment resources. It designs a highly targeted composite scheduling rule and proposes an improved Dueling DDQN scheduling algorithm, which can perceive the global status of the workshop in real time. This overcomes the short-sightedness of traditional single scheduling rules that only focus on local interests and can effectively reduce the total overdue time of workpieces on a global scale. 2. This invention fully considers uncertainties such as machine failures and workpiece arrival times. When a sudden machine failure occurs in the workshop or the workpiece arrival flow fluctuates, the scheduling agent can capture changes in state characteristics in real time and adaptively select the scheduling action that best matches the current scenario from the composite rule base, ensuring the continuity of production activities in a disturbed environment, thereby effectively reducing the negative impact of machine failures on the overall delivery time. Attached Figure Description
[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a hybrid assembly line layout for power battery module production in one embodiment of the present invention; Figure 3 This is a diagram of the Dueling DQN architecture in one embodiment of the present invention; Figure 4 This refers to the delay time in each training step of Dueling DQN in one embodiment of the present invention. Figure 5 This is a comparison of results based on the delay time index in one embodiment of the present invention; Figure 6 This is a comparison of results based on the average completion time index in one embodiment of the present invention. Detailed Implementation
[0017] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: This embodiment addresses the workpiece-machine collaborative scheduling problem in a power battery module production workshop under conditions of dynamic workpiece arrival and machine failure. The module processing process is as follows: Figure 2 As shown, specifically, there are multiple processing machines and module workpiece orders in the module workshop. Each workpiece arrives at the workshop dynamically over time, and includes multiple sequentially executed processes that can be processed on different equipment that meet the process requirements. Therefore, the scheduling system needs to simultaneously formulate a machine allocation strategy for workpieces and a job sequencing strategy within the machine.
[0018] The processing of power battery module production involves multiple factors: workpiece arrival time (the time when the workpiece arrives at the workshop), which is dynamically determined as production progresses, and only arrived workpieces can be scheduled; workpiece deadline (the set completion time for the workpiece), which imposes requirements on the workshop's processing efficiency, otherwise, penalties for late arrival will apply; machine failure and repair: machines may fail during production, and during a failure, they cannot be assigned to processing tasks and must wait for repairs before production can resume; processing time is the actual time a workpiece is processed on the machine, which is related to the characteristics of the process and the machine used. Throughout the entire production scheduling process, the main optimization objective considered in this method is the total workshop delay time. To achieve the above technical features, this invention is implemented through the following method: refer to Figure 1 In this embodiment, a hybrid assembly line scheduling method based on the dynamic arrival of power battery module workpieces is provided, including the following steps: Step S1: Based on the uncertainties in the power battery module production workshop, construct an order model for the dynamic arrival of workpieces and a machine model for processing failures; Step S2: Based on the production constraints of the power battery module, and combined with the order model and machine model in Step 1, establish a hybrid flow shop scheduling mathematical model with the optimization objective of minimizing overdue time; Step S3: Design state characteristics, scheduling rules, and reward functions based on real-time information from the workshop; Step S4: Establish the Dueling DQN network, output the optimal scheduling rule based on the state characteristics of step S3, arrange the module workpieces to the machine, and update the scheduling state; Preferably, in this embodiment, the training model based on the Dueling DQN network is repeatedly trained by collecting experience and softly updating network parameters until the optimal scheduling performance is achieved.
[0019] Preferably, in this embodiment, step 1 specifically involves the following process: Workpiece Arrival Time: Based on the characteristics of workpieces in the power battery module production workshop, a Poisson process is used to simulate the arrival process of battery module workpieces in the workshop. Specifically, the arrival time of each workpiece is a random variable, and its arrival time interval follows an exponential distribution. (25) (26) Workpiece delivery time: workpiece The delivery time is based on its estimated flow time and arrival time, where the flow time takes into account the average processing load at each stage and the bottleneck process. (27) (28) (29) (30) in, yes Average processing time It is the first Number of machines in each stage The bottleneck process is the processing time. It is flowing time. yes Delivery time.
[0020] Machine Failures and Repair Time: Considering failures in the workshop, construct a failure schedule for each machine. Specifically, the start and end times of each machine's failure are random variables, and the failure interval and repair time of each machine follow an exponential distribution.
[0021] (31) (32) (33) (34) Preferably, in this embodiment, the parameters of the power battery module hybrid production line model in step 2 are shown in Table 1.
[0022] Table 1 Symbol Definitions
[0023] Based on the characteristics and manageability of the power battery module workshop scheduling problem, the following reasonable assumptions are made: (1) When the scheduling starts, the arrival time of the workpieces in the workshop is zero, and the subsequent workpieces arrive in the workshop dynamically one after another; (2) Once the machine malfunctions, it will enter a repair state and cannot process any workpieces; (3) The transportation time of the workpiece is not considered; (4) The preparation time before machine processing is zero; (5) There is no buffer zone in front of any machine. The next workpiece can only be received for processing after the current workpiece has been processed. The optimization objective is to minimize the total overdue time of all artifacts. The objective function is defined as follows: (35) Workpiece arrival constraint: The arrival time of the workpiece in the workshop is dynamic; the workpiece can only be processed after it arrives. (36) Machine fault constraints: For machines If it is in a faulty state, production tasks cannot be carried out. (37) Process constraints: Each workpiece contains multiple processes and must be processed in a specific sequence. (38) Other constraints: (39) (40) Equation (28) indicates that the completion time of each process must be non-negative, and Equation (29) indicates that each process can only be processed on one machine at any given time.
[0024] Preferably, in this embodiment, step 3 specifically involves the following process: Seven features were extracted from the workpieces, machines, and processes in the power battery module workshop, namely: Workpiece feature F1, average completion rate of all workpieces , (41) (42) in, It is a workpiece The completion rate.
[0025] Workpiece characteristic F2, standard deviation of average completion rate of all workpieces , (43) Workpiece feature F3, average delay rate of all workpieces , (44) in, This refers to the number of delayed workpieces. This refers to the number of unfinished workpieces.
[0026] Machine feature F4, average utilization rate of all machines , (45) (46) in, It is a machine Utilization rate.
[0027] Machine characteristic F5, standard deviation of average utilization rate of all machines , (47) Machine feature F6, average failure rate of all machines , (48) in, It is a machine The number of failures.
[0028] Process Feature F7: Delay rate of all machines for all processes , (49) in, It refers to the number of delayed processes. This represents the number of incomplete processes.
[0029] Preferably, in this embodiment, the scheduling method of the module workshop includes five composite scheduling rules, namely: The first step of the composite scheduling rule is to filter the set of workpieces to be scheduled, select the workpieces that are expected to be delayed, select the workpiece with the longest delay time from all the delayed workpieces, and then traverse all the machines to be scheduled and select the machine with the shortest processing time from the current set of machines.
[0030] The composite scheduling rule 2 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the longest delay time from all the delayed workpieces, and then iterates through all the machines to be scheduled, and selects the machine with the lowest failure rate from the current set of machines.
[0031] The composite scheduling rule 3 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the longest delay time from all the delayed workpieces, and then iterates through all the machines to be scheduled, and selects the machine with the lowest utilization rate from the current set of machines.
[0032] The composite scheduling rule 4 first filters the set of workpieces to be scheduled, selects the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the shortest processing time from the current set of machines.
[0033] The composite scheduling rule 5 first filters the set of workpieces to be scheduled, selects the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the lowest failure rate from the current set of machines.
[0034] The composite scheduling rule 6 first filters the set of workpieces to be scheduled, selects the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the lowest utilization rate from the current set of machines.
[0035] To minimize delays in workshop production, a reward function related to delay time is designed to measure the effectiveness of the current scheduling strategy and encourage the scheduling system to prioritize tasks with tight delivery dates. Specifically, a positive reward is given if the delay time decreases; a negative penalty is given if the delay time increases; and no reward is given if the delay time remains unchanged. (50) (51) in, for Cumulative rewards after step scheduling.
[0036] Figure 3 This application provides an embodiment of the Dueling DDQN scheduling architecture, such as... Figure 3 As shown, the scheduling method includes the following procedures: Feature input: The scheduling system first acquires real-time environmental status data of the power battery module production workshop and generates workpiece arrival information. and machine malfunction information The policy network uses seven-dimensional feature vectors. The inputs to the neural network include: the average completion rate of all workpieces in production at the current time (F1), the standard deviation of workpiece completion rate (F2), the average delay rate of workpieces (F3), the average utilization rate of all machines (F4), the standard deviation of machine utilization rate (F5), the machine failure rate (F6), and the process delay rate (F7).
[0037] To improve learning efficiency under complex conditions, the Q-network in this embodiment of the Dueling DQN architecture adopts a competitive network architecture. Unlike traditional DQN which directly outputs action values, this network contains two parallel branch flows: (1) State value flow: evaluating the inherent value of the current workshop state itself. (2) Action advantage flow: Evaluate the relative advantage of adopting a specific composite scheduling rule in the current state compared to the average level. These two branches converge at the network output to calculate the final action value function. .
[0038] (52) in, This is the 7-dimensional feature vector at the current moment. For the selected scheduling action, These are the parameters for the fully connected layer of the network. and These are parameters for the unique layers of value stream and advantage stream, respectively. This refers to the size of the action space.
[0039] Action selection and environmental interaction: Based on the calculated Q-value, the agent adopts... The greedy strategy selects an action from six predefined composite scheduling rules. The selected rule is sent to the shop floor execution system, which assigns the corresponding workpiece to the appropriate processing equipment. After the environment executes the action, it sends back a corresponding reward signal and transitions to a new state.
[0040] In a preferred embodiment of the present invention, a model training based on DDQN is provided, and two networks with the same structure but different parameters are designed: a main network and a target network. During the experience replay training process, when calculating the target Q value, this embodiment separates action selection from value evaluation: the main network is responsible for the next state. Select the optimal action The target network is responsible for calculating the selected action. The Q-value. This avoids biases caused by a single network's self-evaluation, ensuring that the learned scheduling strategy is more reliable in the face of unexpected situations such as machine failures.
[0041] To avoid drastic fluctuations in model parameters and enhance the stability of the training process, a soft update mechanism is employed. After each training iteration, the parameters of the target network are updated. With a small update coefficient To the main network parameters Approach slowly. This ensures a smooth change in the baseline for calculating the target Q value, improving the convergence performance of the entire scheduling system.
[0042] (53) (54) (55) To balance the exploration and utilization of scheduling strategies, and to ensure stable model utilization efficiency while fully exploring the state space, a policy decay mechanism is provided. In the early stages of training, a high exploration rate is set, causing the network to tend to randomly select actions, thus broadly exploring the unknown state space, accumulating diverse empirical data, and avoiding getting trapped in local optima due to premature greed. As the number of training iterations increases, the exploration rate is gradually reduced linearly or exponentially. In the later stages of training, the exploration rate is maintained at a low minimum value, allowing the agent to primarily utilize the learned preferred strategies for decision-making, thereby suppressing random oscillations in action selection, ensuring stable convergence of model parameters, and ultimately generating a high-quality scheduling scheme.
[0043] (56) in, It is the minimum exploration probability. It is the attenuation coefficient. For example... Figure 4 As shown, in the initial training phase, to avoid getting trapped in local optima, the scheduling network uses a pre-defined greedy strategy to explore the solution space with a high probability, resulting in some fluctuations in the latency index. As the training progresses, the strategy decay mechanism begins to take effect, the randomness of action selection gradually decreases, and the target network parameters continuously approach the optimal solution. As can be seen from the figure, after approximately 400 training iterations, the total latency significantly decreases and tends to stabilize. Finally, the curve converges to a low value range, proving that the improved Dueling DDQN algorithm proposed in this invention can effectively adapt to the dynamic workshop environment through autonomous learning and ultimately generate a stable, low-latency, and efficient scheduling strategy. The detailed training process is shown in Figure 2.
[0044] Table 2 Dueling DDQN Training Process
[0045] Example 2: To verify the effectiveness and advancement of the power battery module hybrid production line scheduling optimization method considering workpiece dynamic arrival and machine failure described in this invention, a comparative simulation experiment was conducted in a simulated power battery module production environment. The experiment used total delay time and maximum completion time as core evaluation indicators, and compared the improved Dueling DDQN method proposed in this invention with two other methods: traditional heuristic rules and deep reinforcement learning, including SPT (Shortest Processing Time), LPT (Longest Processing Time), EDD (Earliest Delivery Date), MOR (Most Remaining Operations), FOR (Minimum Remaining Operations), FTM (Fastest Machine), DQN algorithm, and DDQN. Figure 5 The results of various algorithms on latency metrics are presented. The improved Dueling DDQN method proposed in this invention achieves the minimum total latency in all test instances, significantly outperforming traditional heuristic rules and deep reinforcement learning baselines. The deep reinforcement learning agent employed in this invention learns long-term, global scheduling strategies through continuous interaction with the environment. Specifically, regarding machine failures, the Dueling DDQN agent can perceive the failure rate contained in the feature vector and dynamically adjust scheduling actions, thereby effectively avoiding the risk of downtime caused by failures.
[0046] To verify the necessity of the network architecture in this invention, a further comparison was made with various methods in terms of maximum completion time. Figure 6 The completion time distribution of Dueling DDQN and various benchmarks is shown. The Dueling DDQN algorithm has the lowest median, significantly lower than all the compared algorithms. This indicates that in most scheduling scenarios, the scheduling scheme generated by Dueling DDQN enables the workpiece to be processed at the fastest speed, reflecting the optimal average production efficiency.
[0047] In summary, the method proposed in this invention, through an improved Dueling DDQN algorithm combined with targeted state feature design, successfully solves the scheduling challenge of a hybrid production line workshop for power battery modules in a dynamic environment. Compared with existing technologies, this invention not only significantly reduces the total delay time in the production process and improves the on-time delivery rate of orders, but also demonstrates excellent dynamic response capabilities in the face of sudden equipment failures, possessing significant industrial application value. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A hybrid assembly line scheduling method based on the dynamic arrival of power battery module workpieces, characterized in that, Includes the following steps: Step S1: Based on the uncertainties in the power battery module production workshop, construct an order model for dynamic arrival of workpieces and a machine model that considers faults; Step S2: Based on the production constraints of the power battery module, and combining the order model of dynamic arrival of workpieces and the machine model that considers faults, establish a hybrid flow shop scheduling mathematical model with minimizing overdue time as the optimization objective. Step S3: Based on the mathematical model of hybrid flow shop scheduling, design state characteristics, scheduling rules and reward functions according to the real-time information of the shop; Step S4: Solve for the optimal scheduling scheme based on the Dueling DQN network, output the optimal scheduling rule according to the state characteristics in step S3, arrange the module workpieces to the machine, and update the scheduling status.
2. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 1, characterized in that, The order model is constructed based on the dynamic arrival characteristics of workpieces in the production of power battery modules, specifically as follows: the arrival of power battery module workpieces in the workshop follows a Poisson process, and each workpiece... The arrival time is a random variable. arrival time interval Follows an exponential distribution: （1） （2） in, It is the average arrival time of the workpiece. It is the workpiece index.
3. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 2, characterized in that, The machine model is based on the start and end times of faults in the power battery module workshop machines, as detailed below: The machine malfunction began. and end time For random variables, machine Failure interval and repair time Follows an exponential distribution: （3） （4） （5） （6） in, It's the machine serial number. It is the fault number. It is the machine's mean time between failures. It is the machine's average repair time.
4. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 1, characterized in that, The production objective is to minimize the total overdue time of all workpieces, as detailed below: （7） In the formula, The total number of workpieces, and the total delay time for all workpieces is determined by the total delay time for each workpiece. Deadline and completion time calculate; The mathematical model for scheduling the hybrid flow workshop includes the following constraints: Workpiece arrival constraint: Each workpiece arrives at the workshop at a different time, and can only be processed after it arrives at the workshop. （8） in, It is a workpiece First process Start time; Machine fault constraints: For machines The machine must not be in a faulty state during scheduling and processing: （9） in, It is a process In the machine On the processing time, It is a process The start time, It is a decision variable, when the process Assigned to machine The value is 1 if it is true, and 0 otherwise. It is a decision variable, when the machine... The value is 0 when there is a fault, and 1 otherwise.
5. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 1, characterized in that, The state space includes seven features extracted from workpieces, machines, and processes in the module workshop, including: workpiece features, average completion rate of workpieces, standard deviation of workpiece completion rate, and average delay rate of workpieces; machine features, average machine utilization rate, standard deviation of machine utilization rate, and average machine failure rate; and process features, average delay rate of processes.
6. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 5, characterized in that, The scheduling rules include five composite scheduling rules, namely: The first step of the composite scheduling rule is to filter the set of workpieces to be scheduled, select the workpieces that are expected to be delayed, select the workpiece with the longest delay time from all the delayed workpieces, and then traverse all the machines to be scheduled and select the machine with the shortest processing time from the current set of machines. The composite scheduling rule 2 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the longest delay time from all the delayed workpieces, and then traverses all the machines to be scheduled, selects the machine with the lowest failure rate from the current set of machines. The composite scheduling rule 3 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the longest delay time from all the delayed workpieces, and then traverses all the machines to be scheduled, selects the machine with the lowest utilization rate from the current set of machines. The composite scheduling rule 4 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the shortest processing time from the current set of machines. The composite scheduling rule 5 first filters the set of workpieces to be scheduled, filters out the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the lowest failure rate from the current set of machines. The composite scheduling rule 6 first filters the set of workpieces to be scheduled, selects the workpieces expected to be delayed, selects the workpiece with the highest completion rate from all delayed workpieces, and then iterates through all machines to be scheduled, selecting the machine with the lowest utilization rate from the current set of machines.
7. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 6, characterized in that, The reward function is related to the delay time. To achieve the minimum delay time, a positive reward is given if the delay time at the current moment is lower than that at the next moment; otherwise, a negative reward is given. This encourages the scheduling system to prioritize scheduling tasks with tight delivery dates. （10） in, It is the total delay time of all workpieces in the workshop at the current moment.
8. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 1, characterized in that, The Dueling DQN network includes a common feature extraction layer, a splitting layer, and an aggregation layer connected in sequence. The common feature extraction layer consists of several fully connected layers and is used to extract high-order features of the state feature vector. The splitting layer includes parallel state value branches and action advantage branches. The state value branches are used to output the state value function of the current state, and the action advantage branches are used to output the advantage function of each scheduling action. The aggregation layer is used to aggregate the state value function and the advantage function to obtain the Q value of each scheduling action, as shown in formula (13). （11） in, This is the seven-dimensional state feature vector at the current moment. For the selected scheduling action, These are the parameters for the fully connected layer of the network. and These are parameters for the unique layers of value stream and advantage stream, respectively. For the size of the action space, It is a state-value function. It is the action advantage function.
9. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 8, characterized in that, The target network parameter update of the Dueling DQN network adopts a soft update mechanism, specifically including: after each parameter iteration update of the main network, obtaining the current weight parameters of the target network from the current weight parameters of the main network; using a preset soft update coefficient, calculating the linear weighted sum of the weight parameters of the main network and the current weight parameters of the target network; assigning the linear weighted sum to the target network, so that the parameters of the target network smoothly approximate the parameters of the main network. The specific calculation of the soft update is as follows: （12） in, For the updated target network parameters, The preset soft update coefficient has a value range of [value range missing]. .
10. The hybrid assembly line workshop scheduling method based on the dynamic arrival of power battery module workpieces according to claim 1, characterized in that, The Dueling DQN network employs an attenuation mechanism. Greedy strategy training specifically includes: setting a preset exploration rate. The initial value, minimum value, and decay step size of the model are determined. In the early stage of model training, a high exploration rate is maintained, allowing the scheduling agent to randomly select a scheduling rule with a high probability to explore the global state space. As the number of training iterations increases, the exploration rate is gradually reduced according to the decay step size until the exploration rate reaches the minimum value. In the later stage of model training, the scheduling agent is made to select the scheduling rule with the largest Q value with a high probability, and the learned optimal strategy is used to make scheduling decisions.