Electronic atomization production line scheduling simulation method based on deep learning

Abstract

This invention discloses a deep learning-based simulation method for scheduling electronic atomization production lines, comprising: S1, collecting and preprocessing historical production data from the electronic atomization production line to construct a standardized dataset; S2, constructing and training an electronic atomization production line scheduling model using a multi-dimensional spatiotemporal collaborative Transformer network; S3, inputting real-time production line status information into the trained electronic atomization production line scheduling model and evaluating production line operation indicators to generate an optimized set of candidate scheduling strategies; S4, constructing a multi-objective simulation environment for the electronic atomization production line and simulating the execution of the optimized set of candidate scheduling strategies; S5, using an improved multi-objective ant colony optimization algorithm to select the optimal scheduling strategy; S6, applying the optimal scheduling strategy to the actual production line and collecting feedback information in real time; S7, continuously iteratively optimizing the scheduling model based on the feedback data. This invention improves production line operating efficiency, resource utilization, and scheduling response capabilities.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production line scheduling optimization technology, and in particular to a deep learning-based simulation method for electronic atomization production line scheduling. Background Technology

[0002] In the manufacturing process of modern electronic atomization products, production line scheduling is a core element, directly impacting overall production efficiency, resource utilization, and the timeliness of order delivery. With increasingly diversified market demands and continuously expanding production scale, traditional scheduling methods are struggling to meet the complex and ever-changing production scenarios. Currently, flexible manufacturing systems and intelligent production lines are gradually being adopted. These production lines involve a large number of heterogeneous devices, concurrent tasks, and dynamic resource allocation; achieving efficient and stable scheduling has become a pressing issue for manufacturing enterprises.

[0003] Existing production line scheduling methods mostly employ heuristic algorithms, rule engines, or static optimization models, such as genetic algorithms and particle swarm optimization. These methods are effective in low-dimensional, relatively stable production scenarios, but when faced with high-dimensional, strongly coupled, and dynamically changing e-cigarette production lines, their scheduling strategies often struggle to adapt to complex operating conditions, exhibiting problems such as slow convergence speed and getting trapped in local optima. Furthermore, traditional algorithms have weak modeling capabilities for time-series data, task dependencies, and resource constraints, making it difficult to deeply mine potential patterns in historical production data and dynamically adjust scheduling strategies based on real-time feedback.

[0004] In recent years, with the development of deep learning technology, more and more research has begun to apply neural networks to manufacturing scheduling problems, especially Transformer networks and graph neural networks, which have shown great potential in modeling multidimensional heterogeneous information. However, most existing research focuses on building static models and optimizing single objectives, lacking real-time perception and feedback mechanisms for the production line's operating status, making it difficult to achieve continuous optimization of the scheduling model. In addition, existing scheduling strategies are usually applied directly to the actual production line after generation, lacking simulation verification and multi-objective adaptive evaluation, which easily leads to insufficient stability and poor performance of the strategies during actual execution.

[0005] Therefore, how to provide a deep learning-based simulation method for scheduling electronic atomization production lines is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based simulation method for scheduling electronic atomization production lines. This invention fully utilizes spatiotemporal collaborative Transformer networks and multi-objective optimization algorithms to describe in detail the technical path for achieving intelligent scheduling and continuous optimization in the electronic atomization production process. It can dynamically adjust the scheduling strategy based on real-time production data and continuously optimize the production line scheduling strategy through multi-objective simulation evaluation and feedback mechanisms. It has the advantages of strong adaptability, high optimization accuracy, good scheduling stability, and the ability to handle complex working conditions and quickly respond to changes in production demand.

[0007] The deep learning-based electronic atomization production line scheduling simulation method according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect historical production data from the electronic atomization production line and preprocess it to form a standardized dataset;

[0009] S2. Based on a standardized dataset, an electronic atomization production line scheduling model is constructed and trained using a multi-dimensional spatiotemporal collaborative Transformer network.

[0010] S3. Input the real-time production line status information into the trained electronic atomization production line scheduling model to generate a candidate scheduling strategy set, and generate an optimized candidate scheduling strategy set by evaluating the production line operation indicators.

[0011] S4. Construct a multi-objective simulation environment for the electronic atomization production line, simulate the execution of the optimized candidate scheduling strategy set, and output the simulation evaluation results.

[0012] S5. Based on the simulation evaluation results, an improved multi-objective ant colony optimization algorithm is used to optimize the candidate scheduling strategies for simulated execution and select the optimal scheduling strategy.

[0013] S6. Apply the optimal scheduling strategy to actual production scheduling and collect feedback information in real time during the operation process;

[0014] S7. Based on feedback information, continuously iterate and optimize the multi-dimensional spatiotemporal collaborative Transformer network.

[0015] Optionally, step S1 specifically includes:

[0016] S11. The production data includes equipment status data, production task information, raw material supply information, production progress information, task dependencies, resource usage information, order delivery constraints, production feedback information, and historical scheduling execution records.

[0017] S12. The preprocessing includes data format unification, missing value imputation, outlier removal, timestamp alignment, and removal of unusable or noisy information.

[0018] S13. Standardize and encode the preprocessed data according to field type and time dimension to form a structured standard dataset. The standardization and encoding includes: normalizing numerical data, one-hot encoding categorical data, and converting timestamp data into numerical features.

[0019] Optionally, step S2 specifically includes:

[0020] S21. Divide the standardized dataset into task data, equipment data, resource data, production progress data, and time series data, and input them into a multidimensional spatiotemporal collaborative Transformer network.

[0021] S22. The multi-dimensional spatiotemporal collaborative Transformer network includes a spatiotemporal feature encoder and a target adaptive decoder. The spatiotemporal feature encoder includes a shared Transformer encoder and a graph neural network encoder. It optimizes task information, equipment status, resource relationships and time dependencies in the electronic atomization production line through multi-task learning. The target adaptive decoder dynamically adjusts the weight priority between scheduling optimization targets based on real-time production feedback information.

[0022] S23. Process task data, equipment data, resource data, production progress data and time series data through a shared Transformer encoder, and weight the input data through a multi-head self-attention mechanism to capture the dependencies between tasks, between tasks and equipment, resources and time series data, extract the common features of the data and map them to the shared latent space to generate shared features.

[0023] S24. The multi-head self-attention mechanism introduces time encoding. When calculating attention weights, it weights the time-series data by the relative positional relationship between time steps, optimizes the time order and dependency of task scheduling, and introduces domain knowledge enhancement vectors to weight the domain features of device type and task priority, and adjusts the similarity between query and key vectors.

[0024] S25. The graph neural network encoder represents the dependencies between equipment and tasks and the constraints of resource allocation in the production line through a graph structure, and updates the node features through graph convolution operations to construct the graph relationship between equipment, tasks and resources, and outputs node embedding features, which include equipment embedding features, task embedding features and resource embedding features.

[0025] S26. The shared features and node embedding features are weighted and averaged to obtain preliminary fused features, and the preliminary fused features are weighted by a self-attention mechanism to generate collaborative features;

[0026] S27. The objective of each task is individually optimized through multi-task loss, which includes task priority, production efficiency, resource allocation efficiency and time cost, and the parameters of the shared Transformer encoder and graph neural network encoder are updated through backpropagation.

[0027] S28. The target adaptive decoder independently decodes each task target based on collaborative features to generate a corresponding task target. The target adaptive decoder introduces dynamic target weight adjustment and dynamically adjusts the weight priority between scheduling optimization targets according to real-time production feedback information.

[0028] S29. Based on changes in the production environment, the target adaptive decoder continuously adjusts task priorities, optimizes the weights between targets, ensures that the scheduling of each target meets actual needs, and outputs the trained electronic atomization production line scheduling model.

[0029] Optionally, step S3 specifically includes:

[0030] S31. Input the real-time production line status information into the trained electronic atomization production line scheduling model to obtain a predicted set of candidate scheduling strategies. The production line status information includes equipment operating status, task progress, resource occupancy, raw material inventory, production feedback data, and scheduling-related data.

[0031] S32. Calculate the production line operation indicators for each candidate scheduling strategy. The production line operation indicators include resource utilization rate, production cycle, bottleneck risk and energy consumption cost. The resource utilization rate is evaluated by the ratio of resource usage to total resources. The production cycle measures the duration of the entire production process. The bottleneck risk is measured by the average task delay. The energy consumption cost is measured by the energy cost consumed by the equipment.

[0032] S33. Evaluate each candidate scheduling strategy by weighted summation of the operating indicators of each production line to obtain a comprehensive score for each candidate scheduling strategy, and select Top-A candidate scheduling strategies as the optimized candidate scheduling strategy set based on the comprehensive score. Top-A means selecting the A candidate scheduling strategies with the highest scores based on the comprehensive score.

[0033] Optionally, step S4 specifically includes:

[0034] S41. A simulation environment for the electronic atomization production line is constructed using simulation software, and the optimized candidate scheduling strategy set is input into the simulation environment for simulation execution. The simulation environment can simulate the execution effect of the candidate scheduling strategy in the actual production line.

[0035] S42. By evaluating the task completion rate, equipment load balancing degree, and scheduling stability, the execution effect of each candidate strategy is obtained, and the simulation evaluation results are output. The task completion rate measures the proportion of tasks completed by each candidate scheduling strategy within a predetermined time. The equipment load balancing degree measures the distribution of equipment load during the production process for each candidate scheduling strategy. The scheduling stability measures the volatility of task execution time for each candidate scheduling strategy during the production process.

[0036] Optionally, step S5 specifically includes:

[0037] S51. Based on the simulation evaluation results, an improved multi-objective ant colony optimization algorithm is used to optimize the candidate scheduling strategy for simulation execution. The improved multi-objective ant colony optimization algorithm introduces a multi-level pheromone optimization mechanism to dynamically update the pheromone at the levels of path selection, task priority, resource allocation, scheduling stability and response speed.

[0038] S52. Through the improved multi-objective ant colony optimization algorithm, individual ants dynamically adjust the weight of each objective level according to changes in the production environment during the search process. The pheromone update mechanism updates the pheromone concentration based on the task execution path during the modeling process, and the pheromone increment is calculated in combination with the scheduling feedback weight.

[0039] S53. During the optimization process, non-dominated sorting is used to evaluate the scheduling results at multiple objective levels. Combining the performance of the scheduling strategy in terms of path selection, task priority, resource allocation, scheduling stability and response speed, the candidate scheduling strategy with the best convergence is selected from the non-dominated front as the optimal scheduling strategy.

[0040] Optionally, step S6 specifically includes:

[0041] S61. Deploy the optimal scheduling strategy to the actual electronic atomization production line to control the task scheduling order, resource allocation path and equipment operating status;

[0042] S62. During the scheduling and execution process, feedback information is collected in real time, including task plan deviation, equipment resource fluctuation, task delay status and node response data;

[0043] S63. Organize and archive the feedback information to form a feedback data record.

[0044] Optionally, step S7 specifically includes:

[0045] S71. Compare the feedback data records with the optimal scheduling strategy, analyze task execution deviations, resource usage anomalies, and scheduling response anomalies, and generate a feedback enhancement dataset.

[0046] S72. Input the feedback enhancement dataset into the multidimensional spatiotemporal collaborative Transformer network for incremental training to continuously optimize the scheduling performance of the electronic atomization production line scheduling model;

[0047] S73. Based on the updated electronic atomization production line scheduling model, a new scheduling strategy is generated to complete the adaptive learning and dynamic scheduling optimization of the electronic atomization production line.

[0048] The beneficial effects of this invention are:

[0049] First, this invention constructs a multi-dimensional spatiotemporal collaborative Transformer network model. By combining a shared Transformer encoder and a graph neural network encoder, it fully integrates multi-dimensional data such as tasks, equipment, resources, and time series. This model can model local constraints while capturing global dependencies, improving the representation ability of complex production data and the accuracy of scheduling modeling. By introducing time encoding and domain knowledge enhancement mechanisms, the electronic atomization production line scheduling model can accurately identify the temporal dependencies of task execution, focusing on critical tasks and core equipment, effectively improving the performance of scheduling strategies in terms of time rationality and domain adaptability.

[0050] Secondly, in the process of generating scheduling strategies, this invention introduces a real-time feedback mechanism and a target adaptive decoder, which can dynamically adjust the weights and priorities among multiple objectives according to the production status. This enhances the adaptive learning and continuous optimization capabilities of the electronic atomization production line scheduling model, and strengthens its ability to cope with unexpected situations such as production line fluctuations and resource bottlenecks. Simultaneously, a comprehensive scoring mechanism is used to evaluate candidate scheduling strategies across multiple indicators, and the strategy execution effect is verified through a simulation environment, ensuring that the final deployed scheduling strategy exhibits good performance in terms of task completion rate, equipment load balancing, and scheduling stability.

[0051] Furthermore, this invention introduces a multi-level pheromone optimization mechanism into the improved multi-objective ant colony optimization algorithm. This mechanism enables fine-grained guidance at multiple scheduling objective levels, such as task path selection, priority ranking, and resource allocation, significantly improving the global optimality and execution robustness of the scheduling strategy. Moreover, the optimal scheduling scheme is extracted from simulation evaluation results using a non-dominated sorting method, further ensuring the quality and reliability of scheduling decisions.

[0052] In summary, this invention significantly improves the intelligent generation capability and execution stability of scheduling strategies for electronic atomization production lines, achieves the organic integration of deep learning models and simulation optimization processes, and possesses the capabilities of adaptive scheduling, dynamic optimization, and efficient execution. It enhances the overall efficiency and intelligence level of production line operation and has promising industrial application prospects and promotional value. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a schematic diagram of the deep learning-based electronic atomization production line scheduling simulation method proposed in this invention;

[0055] Figure 2 This is a schematic diagram of the multidimensional spatiotemporal collaborative Transformer network structure in this invention;

[0056] Figure 3 This is a flowchart of the scheduling strategy selection process based on simulation and multi-objective optimization in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1-3 A deep learning-based simulation method for scheduling electronic atomization production lines includes the following steps:

[0059] S1. Collect historical production data from the electronic atomization production line and preprocess it to form a standardized dataset;

[0060] S2. Based on a standardized dataset, an electronic atomization production line scheduling model is constructed and trained using a multi-dimensional spatiotemporal collaborative Transformer network.

[0061] S3. Input the real-time production line status information into the trained electronic atomization production line scheduling model to generate a candidate scheduling strategy set, and generate an optimized candidate scheduling strategy set by evaluating the production line operation indicators.

[0062] S4. Construct a multi-objective simulation environment for the electronic atomization production line, simulate the execution of the optimized candidate scheduling strategy set, and output the simulation evaluation results.

[0063] S5. Based on the simulation evaluation results, an improved multi-objective ant colony optimization algorithm is used to optimize the candidate scheduling strategies for simulated execution and select the optimal scheduling strategy.

[0064] S6. Apply the optimal scheduling strategy to actual production scheduling and collect feedback information in real time during the operation process;

[0065] S7. Based on feedback information, continuously iterate and optimize the multi-dimensional spatiotemporal collaborative Transformer network.

[0066] In this embodiment, step S1 specifically includes:

[0067] S11. The production data includes equipment status data, production task information, raw material supply information, production progress information, task dependencies, resource usage information, order delivery constraints, production feedback information, and historical scheduling execution records.

[0068] S12. The preprocessing includes data format unification, missing value imputation, outlier removal, timestamp alignment, and removal of unusable or noisy information.

[0069] S13. Standardize and encode the preprocessed data according to field type and time dimension to form a structured standard dataset. The standardization and encoding includes: normalizing numerical data, one-hot encoding categorical data, and converting timestamp data into numerical features.

[0070] In this embodiment, step S2 specifically includes:

[0071] S21. Divide the standardized dataset into task data, equipment data, resource data, production progress data, and time series data, and input them into a multidimensional spatiotemporal collaborative Transformer network.

[0072] S22. The multi-dimensional spatiotemporal collaborative Transformer network includes a spatiotemporal feature encoder and a target adaptive decoder. The spatiotemporal feature encoder includes a shared Transformer encoder and a graph neural network encoder. It optimizes task information, equipment status, resource relationships and time dependencies in the electronic atomization production line through multi-task learning. The target adaptive decoder dynamically adjusts the weight priority between scheduling optimization targets based on real-time production feedback information.

[0073] S23. Process task data, equipment data, resource data, production progress data and time series data through a shared Transformer encoder, and weight the input data through a multi-head self-attention mechanism to capture the dependencies between tasks, between tasks and equipment, resources and time series data, extract the common features of the data and map them to the shared latent space to generate shared features.

[0074] S24. The multi-head self-attention mechanism introduces temporal encoding. When calculating attention weights, it weights the temporal data based on the relative positional relationship between time steps, optimizing the temporal order and dependencies of task scheduling. It also introduces domain knowledge enhancement vectors to weight domain features such as device type and task priority, adjusting the similarity calculation between query and key vectors.

[0075] ;

[0076] in, For a moment Attention weights For query vector, For key vectors, This indicates the similarity between the query vector and the key vector. Augment vectors for domain knowledge As a moderating factor, it controls the influence weight of domain knowledge;

[0077] In this invention, the multi-head self-attention mechanism incorporates temporal encoding and domain knowledge enhancement vectors, which improves the ability of the electronic atomization production line scheduling model to perceive and represent complex production data. Temporal encoding quantifies the relative positional relationships between time steps, enabling the model to capture the execution order and temporal dependencies of tasks, thus enhancing the rationality and foresight of the scheduling strategy in the time dimension. The domain knowledge enhancement vectors, combined with prior information on task priority and equipment type, adjust the similarity weights between query and key vectors in attention calculation, making the model pay more attention to the scheduling relationships of key tasks and core equipment, thereby improving the accuracy of task scheduling and its domain adaptability.

[0078] S25. The graph neural network encoder represents the dependencies between equipment and tasks and the constraints of resource allocation in the production line through a graph structure, and updates node features through graph convolution operations to construct graph relationships between equipment, tasks, and resources, outputting node embedding features, which include equipment embedding features, task embedding features, and resource embedding features:

[0079] ;

[0080] in, Let v be the update feature of node v. Let v be the set of adjacent nodes. The normalization coefficient is... Let be the weight matrix of the k-th layer. For activation functions;

[0081] In this invention, a shared Transformer encoder and a graph neural network encoder work together to achieve deep fusion modeling of task, equipment, resource, and time characteristics in an electronic atomization production line. The shared Transformer encoder utilizes a multi-head self-attention mechanism to extract global dependencies between tasks, equipment, resources, and time-series data, capturing key common patterns in the production process; the graph neural network encoder uses graph structures to characterize the local structural information of task dependencies, resource constraints, and equipment connections, enhancing the ability to model complex production structures.

[0082] S26. The shared features and node embedding features are weighted and averaged to obtain preliminary fused features, and the preliminary fused features are weighted by a self-attention mechanism to generate collaborative features;

[0083] S27. The objective of each task is individually optimized through multi-task loss, which includes task priority, production efficiency, resource allocation efficiency and time cost, and the parameters of the shared Transformer encoder and graph neural network encoder are updated through backpropagation.

[0084] S28. The target adaptive decoder independently decodes each task target based on collaborative features to generate corresponding task targets. The target adaptive decoder introduces dynamic target weight adjustment, dynamically adjusting the weight priority between scheduling optimization targets according to real-time production feedback information. The update formula for target weights is:

[0085] ;

[0086] in, The dynamic weights of the scheduling target t, As an adjustment factor, it controls the sensitivity of the feedback. This is a real-time feedback value, representing feedback information about the current production line status. The set feedback threshold is used to determine the trigger point for adjusting the target priority, ensuring that the model adjusts the target priority in real time based on production feedback;

[0087] In this invention, the target adaptive decoder is based on dynamic target weight adjustment, which can dynamically update the weights of each scheduling target according to real-time production feedback information, effectively improving the model's adaptability to complex and ever-changing production environments. By sensing feedback signals such as task completion rate, resource usage fluctuations, and critical node delays, the relative importance between different optimization targets, such as task priority, resource utilization, production efficiency, and time cost, is adjusted in real time, prompting the decoder to focus on the core needs of the current production stage. Furthermore, even under conditions of sudden changes in task load or resource bottlenecks, the electronic atomization production line scheduling model can still generate scheduling strategies that meet the current operating conditions, thereby improving system stability, response speed, and optimization effectiveness.

[0088] S29. Based on changes in the production environment, the target adaptive decoder continuously adjusts task priorities, optimizes the weights between targets, ensures that the scheduling of each target meets actual needs, and outputs the trained electronic atomization production line scheduling model.

[0089] In this embodiment, step S3 specifically includes:

[0090] S31. Input the real-time production line status information into the trained electronic atomization production line scheduling model to obtain a predicted set of candidate scheduling strategies. The production line status information includes equipment operating status, task progress, resource occupancy, raw material inventory, production feedback data, and scheduling-related data.

[0091] S32. Calculate the production line operation indicators for each candidate scheduling strategy. The production line operation indicators include resource utilization rate, production cycle, bottleneck risk and energy consumption cost. The resource utilization rate is evaluated by the ratio of resource usage to total resources. The production cycle measures the duration of the entire production process. The bottleneck risk is measured by the average task delay. The energy consumption cost is measured by the energy cost consumed by the equipment.

[0092] S33. Each candidate scheduling strategy is evaluated by weighted summation of the operating indicators of each production line to obtain a comprehensive score for each candidate scheduling strategy. Based on the comprehensive score, the Top-A candidate scheduling strategies are selected as the optimized candidate scheduling strategy set. The Top-A represents the A candidate scheduling strategies with the highest comprehensive scores. The comprehensive score is:

[0093] ;

[0094] in, For the first The overall score of each candidate scheduling strategy For the first The weight of each indicator, For the first The first candidate scheduling strategy Individual production line operating indicators.

[0095] In this embodiment, step S4 specifically includes:

[0096] S41. A simulation environment for the electronic atomization production line is constructed using simulation software, and the optimized candidate scheduling strategy set is input into the simulation environment for simulation execution. The simulation environment can simulate the execution effect of the candidate scheduling strategy in the actual production line.

[0097] S42. By evaluating the task completion rate, equipment load balancing degree, and scheduling stability, the execution effect of each candidate strategy is obtained, and the simulation evaluation results are output. The task completion rate measures the proportion of tasks completed by each candidate scheduling strategy within a predetermined time. The equipment load balancing degree measures the distribution of equipment load during the production process for each candidate scheduling strategy. The scheduling stability measures the volatility of task execution time for each candidate scheduling strategy during the production process.

[0098] In this invention, the optimized candidate scheduling strategy set is simulated and executed in a simulation environment. This allows for accurate evaluation of the performance of each candidate scheduling strategy in terms of task completion rate, equipment load balancing, and scheduling stability without affecting the actual production line operation. This enables the selection of scheduling schemes with strong adaptability and excellent execution effect, thereby improving the reliability and practicality of scheduling decisions.

[0099] In this embodiment, step S5 specifically includes:

[0100] S51. Based on the simulation evaluation results, an improved multi-objective ant colony optimization algorithm is used to optimize the candidate scheduling strategy for simulation execution. The improved multi-objective ant colony optimization algorithm introduces a multi-level pheromone optimization mechanism to dynamically update the pheromone at the levels of path selection, task priority, resource allocation, scheduling stability and response speed.

[0101] S52. Through the improved multi-objective ant colony optimization algorithm, individual ants dynamically adjust the weight of each objective level according to changes in the production environment during the search process. The pheromone update mechanism updates the pheromone concentration based on the task execution path during the modeling process.

[0102] ;

[0103] in, Indicates the first Generation Zhong, No. Task execution path at the target level pheromone concentration, For the first The pheromone volatile factor of a target The pheromone increment is calculated by combining the scheduling feedback weights:

[0104] ;

[0105] in, For the first Dynamic weights at the target level. Indicates the task execution path In the Loss value at the target level;

[0106] In this invention, the improved multi-objective ant colony optimization algorithm introduces a multi-level pheromone optimization mechanism, which can establish a more refined guidance strategy among multi-dimensional scheduling objectives, effectively improving the decision-making accuracy of individual ants in complex search spaces. This mechanism dynamically updates pheromones through path selection, task priority, resource allocation, scheduling stability, and response speed, enabling the algorithm to have stronger convergence ability and the ability to escape local optima when facing complex situations such as task conflicts, resource bottlenecks, or scheduling changes. This improves the global optimality, execution stability, and adaptability to the dynamic environment of the production line in the final scheduling scheme.

[0107] S53. During the optimization process, non-dominated sorting is used to evaluate the scheduling results at multiple objective levels. Combining the performance of the scheduling strategy in terms of path selection, task priority, resource allocation, scheduling stability and response speed, the candidate scheduling strategy with the best convergence is selected from the non-dominated front as the optimal scheduling strategy.

[0108] For example, in practical applications, based on real-time status information of an e-cigarette production line, such as equipment operation, resource usage, and task progress, a trained e-cigarette production line scheduling model generates multiple candidate scheduling strategies. Taking mixed orders received within a day as an example, a total of 50 candidate scheduling strategies are generated. Subsequently, resource usage rate, production cycle, bottleneck risk, and energy consumption cost are calculated for each candidate strategy, and they are ranked using a weighted comprehensive scoring method. The top 10 strategies (Top-10) are selected as optimized candidate scheduling strategies. The Top-10 candidate scheduling strategies are input into a multi-objective simulation environment, and simulated execution and evaluation are performed on three dimensions: task completion rate, equipment load balancing, and scheduling stability. Based on the simulation results, an improved multi-objective ant colony optimization algorithm is used for global optimization and non-dominated solution screening. Finally, the optimal scheduling strategy is determined and deployed to the actual production line.

[0109] In this embodiment, step S6 specifically includes:

[0110] S61. Deploy the optimal scheduling strategy to the actual electronic atomization production line to control the task scheduling order, resource allocation path and equipment operating status;

[0111] S62. During the scheduling and execution process, feedback information is collected in real time, including task plan deviation, equipment resource fluctuation, task delay status and node response data;

[0112] S63. Organize and archive the feedback information to form a feedback data record.

[0113] In this embodiment, step S7 specifically includes:

[0114] S71. Compare the feedback data records with the optimal scheduling strategy, analyze task execution deviations, resource usage anomalies, and scheduling response anomalies, and generate a feedback enhancement dataset.

[0115] S72. Input the feedback enhancement dataset into the multidimensional spatiotemporal collaborative Transformer network for incremental training to continuously optimize the scheduling performance of the electronic atomization production line scheduling model;

[0116] S73. Based on the updated electronic atomization production line scheduling model, a new scheduling strategy is generated to complete the adaptive learning and dynamic scheduling optimization of the electronic atomization production line.

[0117] Example 1:

[0118] To verify the feasibility of this invention in practice, it was applied to an intelligent production line scheduling optimization project of an electronic atomization product manufacturing company. This production line exhibits typical flexible manufacturing characteristics of multi-category, small-batch, and rapid changeover, encompassing multiple parallel processes such as liquid injection, assembly, testing, and packaging. It has over 40 production devices, with complex task dependencies and resource allocation constraints between workstations. The production line needs to complete dynamic scheduling of approximately 2000 products daily and is required to complete mixed production of various orders within 24 hours.

[0119] Under traditional scheduling methods, the company mainly relies on manual scheduling plans, combined with an automated scheduling system based on rule engines. However, during peak production periods, it often faces problems such as insufficient capacity, task congestion, and uneven equipment utilization. Especially when the types of tasks increase and the production pace accelerates, traditional scheduling methods are unable to respond to task changes in a timely manner, resulting in increased task delay rates, unbalanced equipment load, and significant fluctuations in the overall production rhythm, which seriously affects the stable operation of the production line.

[0120] During implementation, historical production data for the production line over 30 consecutive days was first collected, covering equipment operating status, task start and completion times, resource usage details, raw material supply cycle time, order delivery time windows, and dependencies between tasks. Through standardized processing, missing value imputation, and standardization, a structured standard dataset containing 78 fields and approximately 2.2 million records was ultimately formed. During model training, the standard dataset was input into a multi-dimensional spatiotemporal collaborative Transformer network to perform multimodal encoding of task data, resource data, equipment data, and time-series information. The encoder of this multi-dimensional spatiotemporal collaborative Transformer network integrates a shared Transformer encoder and a graph neural network encoder, enabling it to capture high-order dependencies between tasks, equipment, and resources, reflecting changes in scheduling rhythm over time. The decoder incorporates a dynamic objective weight adjustment mechanism, continuously optimizing the focus on objectives such as task priority, rational resource allocation, and scheduling response speed based on feedback information. After training, real-time production line status data for one day during actual production is collected and input into the electronic atomization production line scheduling model. Multiple candidate scheduling strategies are output. The scheduling strategies are simulated in a multi-objective simulation environment to assess task completion rate, equipment load balance, and scheduling stability. The improved multi-objective ant colony optimization algorithm is used for simulation optimization, and the optimal scheduling strategy is selected by combining non-dominated sorting.

[0121] To further verify the practical effect of the present invention, two comparative schemes were used as benchmarks for performance evaluation: a heuristic scheduling method and a deep learning scheduling method. The heuristic scheduling method formulates scheduling strategies based on a rule engine and human experience, mainly relying on task priority ranking and resource availability matching for static scheduling. The deep learning scheduling method uses recurrent neural networks to model historical production data and generate scheduling strategies, but it has limited capabilities in handling long-distance dependencies, complex relationships between tasks, and multi-objective feedback optimization, making it difficult to meet the requirements of dynamic production lines for scheduling flexibility and stability. Simulation tests were conducted using production data from five consecutive days to quantitatively compare the different scheduling methods in terms of daily average task completion rate, average equipment load rate, and scheduling response time. Table 1 shows the specific evaluation results of the different scheduling methods:

[0122] Table 1. Performance Comparison of Different Scheduling Methods in Electronic Atomization Production Lines

[0123]

[0124] As shown in Table 1, this invention outperforms both heuristic and deep learning scheduling methods in several key metrics. Regarding the daily average task completion rate, this invention achieves 96.4%, a significant improvement over the heuristic method's 87.2% and the deep learning method's 91.5%, indicating stronger task processing capabilities under high load conditions. In terms of average device load rate, this invention achieves 84.7%, significantly better than the other two methods, demonstrating more effective allocation of task resources and improved device utilization efficiency. Regarding scheduling response time, this invention's response time is only 3.1 seconds, a significant improvement compared to the heuristic method's 12.4 seconds and the deep learning method's 7.6 seconds, effectively enhancing the ability to respond to real-time state changes.

[0125] Furthermore, this invention demonstrates outstanding performance in task delay control, significantly reducing the average daily number of delayed tasks to 12, which is significantly better than the 46 delayed tasks of heuristic methods and the 28 delayed tasks of deep learning methods, effectively improving the on-time delivery capability of orders. In terms of system resource utilization efficiency, this invention achieves 80.5%, an improvement of over 10% compared to traditional methods, fully demonstrating the optimization effect of resource allocation. Regarding the proportion of unstable scheduling, this invention controls it at 6.5%, significantly lower than the 22.3% of heuristic scheduling methods and the 12.6% of deep learning scheduling methods, indicating that this invention has stronger scheduling stability in dynamic environments and can effectively avoid system fluctuations caused by frequent scheduling changes.

[0126] In summary, the electronic atomization production line scheduling method proposed in this invention demonstrates superior performance in multiple key dimensions such as task completion rate, scheduling response speed, resource utilization efficiency, and system stability. It is significantly better than existing scheduling methods and has outstanding practical application value and broad prospects for promotion.

[0127] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based simulation method for scheduling electronic atomization production lines, characterized in that, Includes the following steps: S1. Collect historical production data from the electronic atomization production line and preprocess it to form a standardized dataset; S2. Based on a standardized dataset, a multi-dimensional spatiotemporal collaborative Transformer network is used to construct and train an electronic atomization production line scheduling model, specifically including: S21. Divide the standardized dataset into task data, equipment data, resource data, production progress data, and time series data, and input them into a multidimensional spatiotemporal collaborative Transformer network. S22. The multi-dimensional spatiotemporal collaborative Transformer network includes a spatiotemporal feature encoder and a target adaptive decoder. The spatiotemporal feature encoder includes a shared Transformer encoder and a graph neural network encoder. It optimizes task information, equipment status, resource relationships and time dependencies in the electronic atomization production line through multi-task learning. The target adaptive decoder dynamically adjusts the weight priority between scheduling optimization targets based on real-time production feedback information. S23. Process task data, equipment data, resource data, production progress data and time series data through a shared Transformer encoder, and weight the input data through a multi-head self-attention mechanism to capture the dependencies between tasks, between tasks and equipment, resources and time series data, extract the common features of the data and map them to the shared latent space to generate shared features. S24. The multi-head self-attention mechanism introduces time encoding. When calculating attention weights, it weights the time-series data by the relative positional relationship between time steps, optimizes the time order and dependency of task scheduling, and introduces domain knowledge enhancement vectors to weight the domain features of device type and task priority, and adjusts the similarity between query and key vectors. S25. The graph neural network encoder represents the dependencies between equipment and tasks and the constraints of resource allocation in the production line through a graph structure, and updates the node features through graph convolution operations to construct the graph relationship between equipment, tasks and resources, and outputs node embedding features, which include equipment embedding features, task embedding features and resource embedding features. S26. The shared features and node embedding features are weighted and averaged to obtain preliminary fused features, and the preliminary fused features are weighted by a self-attention mechanism to generate collaborative features; S27. The objective of each task is individually optimized through multi-task loss, which includes task priority, production efficiency, resource allocation efficiency and time cost, and the parameters of the shared Transformer encoder and graph neural network encoder are updated through backpropagation. S28. The target adaptive decoder independently decodes each task target based on collaborative features to generate a corresponding task target. The target adaptive decoder introduces dynamic target weight adjustment and dynamically adjusts the weight priority between scheduling optimization targets according to real-time production feedback information. S29. Based on changes in the production environment, the target adaptive decoder continuously adjusts task priorities, optimizes the weights between targets, ensures that the scheduling of each target meets actual needs, and outputs the trained electronic atomization production line scheduling model. S3. Input the real-time production line status information into the trained electronic atomization production line scheduling model to generate a candidate scheduling strategy set, and generate an optimized candidate scheduling strategy set by evaluating the production line operation indicators. S4. Construct a multi-objective simulation environment for the electronic atomization production line, simulate the execution of the optimized candidate scheduling strategy set, and output the simulation evaluation results. S5. Based on the simulation evaluation results, an improved multi-objective ant colony optimization algorithm is used to optimize the candidate scheduling strategies for simulated execution and select the optimal scheduling strategy. S6. Apply the optimal scheduling strategy to actual production scheduling and collect feedback information in real time during the operation process; S7. Based on feedback information, continuously iterate and optimize the multi-dimensional spatiotemporal collaborative Transformer network.

2. The deep learning-based electronic atomization production line scheduling simulation method according to claim 1, characterized in that, Step S1 specifically includes: S11. The production data includes equipment status data, production task information, raw material supply information, production progress information, task dependencies, resource usage information, order delivery constraints, production feedback information, and historical scheduling execution records. S12. The preprocessing includes data format unification, missing value imputation, outlier removal, timestamp alignment, and removal of unusable or noisy information. S13. Standardize and encode the preprocessed data according to field type and time dimension to form a structured standard dataset. The standardization and encoding includes: normalizing numerical data, one-hot encoding categorical data, and converting timestamp data into numerical features.

3. The deep learning-based electronic atomization production line scheduling simulation method according to claim 1, characterized in that, Step S3 specifically includes: S31. Input the real-time production line status information into the trained electronic atomization production line scheduling model to obtain a predicted set of candidate scheduling strategies. The production line status information includes equipment operating status, task progress, resource occupancy, raw material inventory, production feedback data, and scheduling-related data. S32. Calculate the production line operation indicators for each candidate scheduling strategy. The production line operation indicators include resource utilization rate, production cycle, bottleneck risk and energy consumption cost. The resource utilization rate is evaluated by the ratio of resource usage to total resources. The production cycle measures the duration of the entire production process. The bottleneck risk is measured by the average task delay. The energy consumption cost is measured by the energy cost consumed by the equipment. S33. Evaluate each candidate scheduling strategy by weighted summation of the operating indicators of each production line to obtain a comprehensive score for each candidate scheduling strategy, and select Top-A candidate scheduling strategies as the optimized candidate scheduling strategy set based on the comprehensive score. Top-A means selecting the A candidate scheduling strategies with the highest scores based on the comprehensive score.

4. The deep learning-based electronic atomization production line scheduling simulation method according to claim 1, characterized in that, Step S4 specifically includes: S41. A simulation environment for the electronic atomization production line is constructed using simulation software, and the optimized candidate scheduling strategy set is input into the simulation environment for simulation execution. The simulation environment can simulate the execution effect of the candidate scheduling strategy in the actual production line. S42. By evaluating the task completion rate, equipment load balancing, and scheduling stability, the execution effect of each candidate strategy is obtained, and the simulation evaluation results are output. The task completion rate measures the proportion of tasks completed by each candidate scheduling strategy within a predetermined time. The equipment load balancing measures the distribution of equipment load during the production process for each candidate scheduling strategy. The scheduling stability measures the volatility of task execution time during the production process for each candidate scheduling strategy.

5. The deep learning-based electronic atomization production line scheduling simulation method according to claim 1, characterized in that, Step S5 specifically includes: S51. Based on the simulation evaluation results, an improved multi-objective ant colony optimization algorithm is used to optimize the candidate scheduling strategy for simulation execution. The improved multi-objective ant colony optimization algorithm introduces a multi-level pheromone optimization mechanism to dynamically update the pheromone at the levels of path selection, task priority, resource allocation, scheduling stability and response speed. S52. Through the improved multi-objective ant colony optimization algorithm, individual ants dynamically adjust the weight of each objective level according to the changes in the production environment during the search process. The pheromone update mechanism updates the pheromone concentration based on the task execution path during the modeling process, and the pheromone increment is calculated in combination with the scheduling feedback weight. S53. During the optimization process, non-dominated sorting is used to evaluate the scheduling results at multiple objective levels. Combining the performance of the scheduling strategy in terms of path selection, task priority, resource allocation, scheduling stability and response speed, the candidate scheduling strategy with the best convergence is selected from the non-dominated front as the optimal scheduling strategy.

6. The deep learning-based electronic atomization production line scheduling simulation method according to claim 1, characterized in that, Step S6 specifically includes: S61. Deploy the optimal scheduling strategy to the actual electronic atomization production line to control the task scheduling order, resource allocation path and equipment operating status; S62. During the scheduling and execution process, feedback information is collected in real time, including task plan deviation, equipment resource fluctuation, task delay status and node response data; S63. Organize and archive the feedback information to form a feedback data record.

7. The deep learning-based electronic atomization production line scheduling simulation method according to claim 1, characterized in that, Step S7 specifically includes: S71. Compare the feedback data records with the optimal scheduling strategy, analyze task execution deviations, resource usage anomalies, and scheduling response anomalies, and generate a feedback enhancement dataset. S72. Input the feedback enhancement dataset into the multidimensional spatiotemporal collaborative Transformer network for incremental training to continuously optimize the scheduling performance of the electronic atomization production line scheduling model; S73. Based on the updated electronic atomization production line scheduling model, a new scheduling strategy is generated to complete the adaptive learning and dynamic scheduling optimization of the electronic atomization production line.

Images