Manufacturing enterprise intelligent production scheduling resource optimization method based on artificial intelligence

Abstract

The invention discloses a manufacturing enterprise intelligent production scheduling resource optimization method based on artificial intelligence, and the method comprises the following steps: S1, collecting and preprocessing multi-source heterogeneous data, and constructing a scheduling input data set; s2, an improved MLP-Mixer network model is constructed, and structure initialization and predictive capability modeling are completed; s3, introducing an improved spider monkey optimization algorithm, and jointly optimizing the structure parameters and the training parameters; s4, outputting a scheduling priority score vector based on the optimization model, generating an initial scheduling sequence and performing conflict correction; s5, issuing the scheduling plan to a manufacturing execution system, and collecting actual execution and resource state data; and S6, an error evaluation index is constructed, and model adaptive updating is carried out. According to the method, the improved MLP-Mixer network and the spider monkey optimization algorithm are fused, so that high-precision prediction and resource optimization allocation of manufacturing task scheduling are realized, and the scheduling efficiency and the self-adaptive capability of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent production scheduling resource optimization method for manufacturing enterprises based on artificial intelligence. Background Technology

[0002] With the continuous advancement of Industry 4.0 and intelligent manufacturing systems, manufacturing enterprises are facing increasingly complex market environments and diversified production demands. Achieving efficient and flexible production scheduling has become a core issue of concern in the industry. As a crucial bridge connecting enterprise production plans and workshop execution, the scheduling system's operational efficiency and level of intelligence directly determine the resource allocation efficiency, task responsiveness, and product delivery cycle of the manufacturing system. In traditional manufacturing systems, production scheduling largely relies on fixed rules or expert experience for task sequencing and resource allocation. This approach often suffers from problems such as delayed response, resource waste, and unstable scheduling results when facing complex manufacturing environments with diverse equipment, frequent order changes, and dynamic resource allocation. It fails to meet the current manufacturing enterprises' demands for flexible, efficient, and dynamically collaborative intelligent scheduling.

[0003] In recent years, the application of artificial intelligence (AI) technology in industrial scenarios has deepened, injecting new vitality into manufacturing scheduling systems. Especially driven by the development of deep learning, reinforcement learning, and swarm intelligence optimization algorithms, more and more research institutions and enterprises are attempting to introduce neural networks and intelligent optimization models into the manufacturing scheduling process to solve problems such as task sequencing, resource allocation, and improving equipment utilization. AI-based production scheduling methods typically possess advantages such as strong modeling capabilities for complex patterns, good generalization ability, and the ability to achieve end-to-end prediction, gradually becoming the mainstream technical route to replace traditional methods.

[0004] At the technical implementation level, structures such as Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Graph Neural Network (GNN) are widely used for feature modeling and scheduling priority prediction in manufacturing tasks. However, these models generally suffer from weak temporal modeling capabilities, simple feature interaction structures, and insufficient modeling of high-dimensional scheduling factors, making it difficult to achieve accurate scheduling decisions in resource-intensive and heterogeneous concurrent manufacturing scenarios. In recent years, the MLP-Mixer network, with its simple structure, all-MLP architecture, and strong scalability, has become an emerging form of neural network, demonstrating good performance in tasks such as image modeling and sequence prediction. However, the original MLP-Mixer model does not consider key issues in manufacturing scheduling, such as task feature selection, differences in task importance, and adaptive modeling of task complexity. It lacks targeted optimization for manufacturing scheduling scenarios, and its direct application still faces performance bottlenecks.

[0005] On the other hand, swarm intelligence optimization algorithms, due to their strong global search capabilities and ease of parallel computation, are widely used in manufacturing scheduling for parameter optimization and scheduling scheme solving. For example, Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Whale Optimization (WOA) have been validated in typical scheduling problems. However, these algorithms often suffer from slow convergence speed, susceptibility to local optima, and unstable search direction, making it difficult to achieve fine-grained search and dynamic tuning in high-dimensional and complex scheduling parameter spaces. Spider Monkey Optimization (SMO), a novel swarm intelligence optimization algorithm proposed in recent years, is inspired by the distributed leadership and subgroup cooperative foraging behavior of spider monkeys. It features dynamic leadership switching, adaptive subgroup splitting and reconstruction, and global and local cooperative search, demonstrating superior global search capabilities and convergence stability compared to traditional optimization algorithms. However, in current technology, the SMO algorithm has not yet been deeply integrated with artificial intelligence models for manufacturing scheduling optimization scenarios, and its potential in structural parameter tuning and scheduling model feedback learning has not been fully explored.

[0006] Meanwhile, in current AI scheduling systems, there is a disconnect between the model optimization process and the scheduling execution feedback. Feedback information collected during actual execution, such as task completion status, equipment utilization efficiency, and scheduling latency, is not incorporated into the model training and structure update process. This lack of a closed-loop feedback mechanism limits the scheduling model's ability to adapt to changing environments and execution deviations in actual production, hindering the continuous improvement of model performance and scheduling strategy accuracy.

[0007] Therefore, how to provide intelligent production scheduling and resource optimization methods for manufacturing enterprises based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose an intelligent production scheduling resource optimization method for manufacturing enterprises based on artificial intelligence. This invention integrates an improved MLP-Mixer network model and a spider monkey optimization algorithm, designs a dynamic feature selection layer, a task attention fusion layer, and a channel compression mechanism, and constructs a closed-loop optimization mechanism by combining task execution feedback. It describes in detail the entire process of intelligently predicting scheduling priority scores, generating legal scheduling schemes, and dynamically updating the scheduling model, and has the advantages of high scheduling accuracy, high resource utilization, and strong model adaptability.

[0009] The intelligent production scheduling resource optimization method for manufacturing enterprises based on artificial intelligence according to embodiments of the present invention includes the following steps: S1. Collect multi-source heterogeneous data from manufacturing enterprises, preprocess the multi-source heterogeneous data, and construct a scheduling input dataset; S2. Construct the MLP-Mixer network model. The MLP-Mixer network model receives the scheduled input dataset and completes the initialization of the MLP-Mixer network model structure and modeling of its predictive capabilities. S3. An improved spider monkey optimization algorithm is introduced to jointly optimize the structural parameters and training parameters of the MLP-Mixer network model; S4. Based on the optimized MLP-Mixer network model, output the manufacturing task scheduling priority score vector, generate the initial scheduling sequence of manufacturing tasks based on the manufacturing task scheduling priority score vector, and perform scheduling legality verification and conflict correction according to resource availability and process constraints to form a scheduling plan. S5. Send the scheduling plan to the manufacturing execution system, construct the scheduling execution results based on the task completion status, equipment utilization rate and scheduling time, and generate error evaluation indicators based on the comparison between the scheduling plan and the scheduling execution results. S6. Construct a feedback error dataset based on the error evaluation index, and use the spider monkey optimization algorithm to adaptively update the MLP-Mixer network structure parameters to complete the dynamic iterative optimization of the scheduling strategy.

[0010] Optionally, the multi-source heterogeneous data specifically includes order information, equipment status, personnel scheduling, and material inventory, which are used to support the construction of the manufacturing task scheduling input dataset and scheduling decision modeling.

[0011] Optionally, the preprocessing of multi-source heterogeneous data specifically includes data cleaning, standardization, and feature encoding, which are used to improve the consistency of multi-source heterogeneous data and its effectiveness in manufacturing task scheduling modeling.

[0012] Optionally, S2 specifically includes: S21. Construct the input matrix from the scheduling input dataset. ,in To determine the number of manufacturing tasks. For each task's feature dimensions, It is the set of real numbers; S22, Input matrix The input is fed into the dynamic feature selection layer to obtain the feature selection weight vector. The feature selection weight vector is composed of the feature selection weight matrix. and bias terms The input matrix is ​​controlled and adjusted by weighting column by column to obtain the feature-selected matrix. ; S23. The matrix after feature selection The input is fed into the Patch partitioning module and divided into several task vector blocks, each with a dimension of [dimension value missing]. ; S24. The partitioned task vectors are input into the Token-Mixing module for feature interaction mapping between tasks, and then input into the task attention fusion layer. The task attention fusion layer generates a task importance weight vector based on the feature vector of each task. The task importance weight vector is derived from the task attention mapping matrix. and bias terms control; S25. The fused matrix is ​​input to the Channel-Mixing module, which includes an adjustable channel compression mechanism and a channel compression ratio. Or the dimensions of the compressed hidden layer As a structural parameter, it controls the adjustment of the number of neurons in the hidden layer; S26. Add residual connections and layer normalization operations between the Token-Mixing module, the task attention fusion layer, and the Channel-Mixing module respectively to construct an improved MLP-Mixer network structure that includes a dynamic feature selection layer, a task attention fusion layer, and a channel compression mechanism. Complete the initialization of the MLP-Mixer network model structure and the modeling of its predictive capabilities for use in generating task scheduling priority score prediction.

[0013] Optionally, S3 specifically includes: S31. Combine the three parameters to be optimized in the MLP-Mixer network model to form an optimization variable vector: ; in, Choose a weight matrix for the features. For task attention mapping matrix, This refers to the channel compression ratio. S32. Based on the functional modules to which the parameters belong, the optimization variable vector is divided into three subgroups according to its structural function, corresponding to the feature selection subgroup, the task attention subgroup, and the channel compression subgroup, respectively, to construct a task feature-guided grouping mechanism. S33. Initialize the spider monkey optimization algorithm population, where each spider monkey represents a set of parameter combinations. The total population size is Each subgroup is divided into an average of 1000. Spider monkey individuals, and initialize local leader spider monkey individuals within each subgroup; S34. Define the fitness function Used for comprehensive evaluation of parameter combinations Accuracy and ranking rationality of manufacturing task scheduling priority score prediction: ; in, Indicates combining parameters After being applied to the MLP-Mixer network model, the predicted scheduling priority score for the i-th manufacturing task is... For true priority labels, For the set of true precedence pairs, These are weighting coefficients. The number of samples for the manufacturing task. To find the maximum value, Indicates combining parameters The predicted scheduling priority score of the j-th manufacturing task after being applied to the MLP-Mixer network model; S35. In each iteration, a local search phase is performed, where each spider monkey individual within a subgroup updates its position based on its local leader and its group neighbors: ; in, This represents the intermediate parameter vector updated through local search in the t-th iteration. This represents the parameter vector of the i-th individual in the t-th iteration. This is a local leader guiding factor, used to control the degree of influence of local leaders on the search direction. The neighbor-individual guiding factor is used to control the degree to which interactions between individual spider monkeys affect the search direction. This represents the parameter vector of the local leader individual in the subgroup in the current t-th iteration. Let be the parameter vector of a neighboring individual in the same subgroup as the i-th spider monkey individual; S36. Execute the global boot phase and establish the elite memory bank. It is used to store the best historical individuals and sets the maximum lifespan for each elite individual. Each iteration updates and replaces based on the current globally optimal result; S37. Select the spider monkey individual with the highest fitness from the elite memory bank. It is used to guide the current population to perform a global position update, based on intermediate update values: ; in, Let represent the final updated parameter vector position of individual i in round t+1. These are random numbers used to adjust the influence strength of the global guiding factor; S38. After each iteration, calculate the fitness value of all spider monkey individuals, and update the lifecycle of the global leader, the local leaders of each subgroup, and the spider monkey individuals in the elite memory bank based on the fitness values. ,like If so, the individual spider monkey will be removed from the memory bank; S39. If a subgroup is in continuous If there is no significant improvement in fitness within a round, the structural restructuring mechanism of the subgroup is triggered, the subgroup individuals are randomly re-initialized or migration is guided by other subgroups, and a local leader is re-selected; S310. Once the maximum number of iterations or fitness convergence condition is met, select the spider monkey individual with the highest fitness from the elite memory. As the optimal configuration of structural parameters, the updated structural parameters and training parameters are assigned to the MLP-Mixer network model respectively, thus completing the joint optimization of structural parameters and training parameters.

[0014] Optionally, S4 specifically includes: S41. Input the scheduling input dataset into the parameter combination... The configured MLP-Mixer network model yields a manufacturing task scheduling priority score vector. : ; in, Indicates the first The scheduling priority prediction score for each task. , Total number of manufacturing tasks; S42. Construct a task scheduling priority index sequence based on the scores of each task in the manufacturing task scheduling priority score vector; S43. Generate an initial scheduling sequence based on the task scheduling priority index sequence. ,in Indicates the first Manufacturing tasks at each scheduling location ; S44. For each task in the initial scheduling sequence Verify resource allocation conditions sequentially. Equipment availability status Process constraints If there are tasks that do not meet the conditions Perform conflict correction operations, where... Indicates the first Manufacturing tasks at each scheduling location Indicates the first Manufacturing tasks at each scheduling location; S45. Perform a final legality check on the conflict-corrected scheduling sequence and renumber it into a sequential index format. A manufacturing task scheduling plan is constructed based on the executable time and resource allocation of each manufacturing task. ,in, Indicates the start time of task scheduling. Indicates the assigned device number, This represents the i-th manufacturing task after conflict correction; S46. Manufacturing task scheduling plan The output is sent to the scheduling and execution module for scheduling, distribution, and execution.

[0015] Optionally, S5 specifically includes: S51. The scheduling plan table generated and verified by the manufacturing task scheduling priority score vector is sent to the manufacturing execution system, including the execution time, task number and resource allocation information of each task. S52. The manufacturing execution system receives the scheduling plan and controls each task to be executed in a predetermined order, time and equipment, and records the actual start time, end time and equipment number of the task. S53. After the task is completed, collect the task's execution status information; S54. Collect usage data of manufacturing resources during the scheduling cycle, including the utilization rate, idle rate, task switching frequency and resource conflict records of each device, construct a set of device usage status, and generate resource status data. S55. By comparing the scheduling plan with the actual task execution results, construct a task-level difference dataset to identify scheduling deviations, including task delays, differences in equipment allocation, and changes in execution order. S56. Based on the task-level differential dataset and resource status data, generate error evaluation indicators to evaluate the accuracy, stability and resource utilization efficiency of the scheduling results.

[0016] Optionally, the task execution status information specifically includes whether the task was completed on time, the task execution time, and the device status and switching status during task execution. This information is used to evaluate the execution deviation of the scheduling plan and the actual completion status of the task, and to provide basic data for generating error evaluation indicators.

[0017] Optionally, the feedback error dataset includes task execution deviation features and resource usage deviation information to reflect the difference between the predictions of the MLP-Mixer network model and the actual scheduling results, supporting the adaptive optimization and updating of the MLP-Mixer network model.

[0018] The beneficial effects of this invention are: This invention significantly enhances the ability to identify key features, model differences in task importance, and compress high-dimensional scheduling factors during manufacturing task scheduling priority prediction by constructing an improved MLP-Mixer network model that integrates a dynamic feature selection layer, a task attention fusion layer, and an adjustable channel compression mechanism. This solves the problems of feature redundancy, simplistic interactions, and poor model generalization inherent in traditional neural networks in manufacturing scheduling. Furthermore, this invention introduces the spider-monkey optimization algorithm to jointly optimize model structure parameters and training parameters, and proposes a task feature-guided grouping mechanism and an elite memory mechanism. While ensuring global search capabilities, this improves the structural integrity and stability of the optimization process, effectively enhancing the convergence speed and accuracy of the scheduling prediction model in complex scenarios. During the scheduling execution phase, this invention outputs a scheduling priority score vector based on the optimized model, generates an initial scheduling sequence for manufacturing tasks, and performs validity checks and conflict corrections based on resource availability, equipment status, and process constraints, thereby obtaining a scheduling plan with high feasibility and execution efficiency. Furthermore, this invention establishes a feedback mechanism for extracting key data such as task completion status and resource utilization efficiency from scheduling execution results. It constructs error evaluation indicators and a feedback error dataset, and introduces them into the dynamic optimization process of the scheduling model. This enables the scheduling model to continuously adapt to changes in the production environment, achieving iterative evolution of scheduling strategies and long-term performance maintenance. Compared to existing technologies, this invention achieves substantial improvements in scheduling accuracy, response flexibility, resource matching, and model self-evolution capabilities, possessing significant engineering application value and promising prospects for industrial promotion. Attached Figure Description

[0019] 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:

[0020] Figure 1 This is a flowchart of the intelligent production scheduling and resource optimization method for manufacturing enterprises based on artificial intelligence proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved MLP-Mixer network model of the intelligent production scheduling resource optimization method for manufacturing enterprises based on artificial intelligence proposed in this invention. Detailed Implementation

[0021] 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.

[0022] refer to Figure 1 and Figure 2The intelligent production scheduling resource optimization method for manufacturing enterprises based on artificial intelligence includes the following steps: S1. Collect multi-source heterogeneous data from manufacturing enterprises, preprocess the multi-source heterogeneous data, and construct a scheduling input dataset; S2. Construct the MLP-Mixer network model. The MLP-Mixer network model receives the scheduled input dataset and completes the initialization of the MLP-Mixer network model structure and modeling of its predictive capabilities. S3. An improved spider monkey optimization algorithm is introduced to jointly optimize the structural parameters and training parameters of the MLP-Mixer network model; S4. Based on the optimized MLP-Mixer network model, output the manufacturing task scheduling priority score vector, generate the initial scheduling sequence of manufacturing tasks based on the manufacturing task scheduling priority score vector, and perform scheduling legality verification and conflict correction according to resource availability and process constraints to form a scheduling plan. S5. Send the scheduling plan to the manufacturing execution system, construct the scheduling execution results based on the task completion status, equipment utilization rate and scheduling time, and generate error evaluation indicators based on the comparison between the scheduling plan and the scheduling execution results. S6. Construct a feedback error dataset based on the error evaluation index, and use the spider monkey optimization algorithm to adaptively update the MLP-Mixer network structure parameters to complete the dynamic iterative optimization of the scheduling strategy.

[0023] This invention constructs an MLP-Mixer network model that integrates a dynamic feature selection layer, a task attention fusion layer, and a channel compression mechanism. This model achieves efficient modeling and priority score prediction of multi-dimensional heterogeneous features of manufacturing tasks, improving the accuracy of scheduling and ranking, and enhancing the model's ability to express differences in task importance. For model optimization, an improved spider-monkey optimization algorithm is introduced, combined with task feature-guided grouping and an elite memory mechanism, effectively improving the efficiency and stability of joint search of structural and training parameters, and enhancing the model's scheduling generalization ability in complex task scenarios. By constructing a scheduling score vector and generating an initial scheduling sequence based on resource constraints, this invention can achieve reasonable ranking and conflict correction of manufacturing tasks, ensuring that the scheduling results are executable and resource allocation is reasonable. During the scheduling execution phase, a feedback mechanism is constructed to evaluate the error in task completion and resource usage status, forming a closed-loop optimization process. Based on the error feedback, the model is iteratively updated to continuously improve the performance of the scheduling strategy. This method has advantages such as high scheduling accuracy, strong execution efficiency, and adaptive model optimization, making it suitable for dynamic, complex, and multi-task parallel intelligent manufacturing environments.

[0024] In this embodiment, the multi-source heterogeneous data specifically includes order information, equipment status, personnel scheduling, and material inventory, which are used to support the construction of the manufacturing task scheduling input dataset and scheduling decision modeling.

[0025] In this embodiment, the preprocessing of multi-source heterogeneous data specifically includes data cleaning, standardization, and feature encoding, which are used to improve the consistency of multi-source heterogeneous data and its effectiveness in manufacturing task scheduling modeling.

[0026] In this embodiment, S2 specifically includes: S21. Construct the input matrix from the scheduling input dataset. ,in To determine the number of manufacturing tasks. For each task's feature dimensions, It is the set of real numbers; S22, Input matrix The input is fed into the dynamic feature selection layer to obtain the feature selection weight vector. The feature selection weight vector is composed of the feature selection weight matrix. and bias terms The input matrix is ​​controlled and adjusted by weighting column by column to obtain the feature-selected matrix. ; S23. The matrix after feature selection The input is fed into the Patch partitioning module and divided into several task vector blocks, each with a dimension of [dimension value missing]. ; S24. The partitioned task vectors are input into the Token-Mixing module for feature interaction mapping between tasks, and then input into the task attention fusion layer. The task attention fusion layer generates a task importance weight vector based on the feature vector of each task. The task importance weight vector is derived from the task attention mapping matrix. and bias terms control; S25. The fused matrix is ​​input to the Channel-Mixing module, which includes an adjustable channel compression mechanism and a channel compression ratio. Or the dimensions of the compressed hidden layer As a structural parameter, it controls the adjustment of the number of neurons in the hidden layer; S26. Add residual connections and layer normalization operations between the Token-Mixing module, the task attention fusion layer, and the Channel-Mixing module respectively to construct an improved MLP-Mixer network structure that includes a dynamic feature selection layer, a task attention fusion layer, and a channel compression mechanism. Complete the initialization of the MLP-Mixer network model structure and the modeling of its predictive capabilities for use in generating task scheduling priority score prediction.

[0027] To improve the modeling accuracy and structural flexibility of MLP-Mixer networks in manufacturing task scheduling scenarios, this invention adds three structural parameters: a feature selection weight matrix, a task attention mapping matrix, and a channel compression ratio. The feature selection weight matrix is ​​introduced to enable the model to dynamically select features. By explicitly weighting each dimension of the input features, redundant information is automatically suppressed, and key scheduling factors are highlighted, effectively addressing the problem of insufficient feature importance recognition in the original MLP-Mixer model. The task attention mapping matrix is ​​introduced to calculate the relative importance scores between tasks, enabling the model to identify the dependencies between high-priority and key tasks in multi-task scheduling scenarios, improving the consistency and interpretability of scheduling predictions. The channel compression ratio, as a structural parameter controlling the changes in hidden layer dimensions, introduces an adjustable channel compression mechanism, which can flexibly adjust the channel width according to data complexity, reducing network redundancy and balancing modeling capability and computational efficiency.

[0028] Based on the aforementioned structural parameters, the improved MLP-Mixer network model comprises three key modules: a dynamic feature selection layer, a task attention fusion layer, and a channel compression module. Residual connections and layer normalization are added between each sub-module to ensure complete information transfer and training stability. The model uses a scheduled input dataset as its foundation. First, the feature selection layer performs weighted filtering of the original features. Then, the features sequentially enter the Patch partitioning, Token-Mixing, task attention fusion, and Channel-Mixing modules, ultimately completing the predictive modeling of task priority scores. By introducing adjustable structural parameters, this invention significantly enhances the model's adaptability and structural flexibility, providing a more efficient and accurate prediction basis for complex manufacturing task scheduling. It possesses advantages such as lightweight design, strong scheduling adaptability, and a large space for parameter optimization.

[0029] In this embodiment, S3 specifically includes: S31. Combine the three parameters to be optimized in the MLP-Mixer network model to form an optimization variable vector: ; in, Choose a weight matrix for the features. For task attention mapping matrix, This refers to the channel compression ratio. S32. Based on the functional modules to which the parameters belong, the optimization variable vector is divided into three subgroups according to its structural function, corresponding to the feature selection subgroup, the task attention subgroup, and the channel compression subgroup, respectively, to construct a task feature-guided grouping mechanism. S33. Initialize the spider monkey optimization algorithm population, where each spider monkey represents a set of parameter combinations. The total population size is Each subgroup is divided into an average of 1000. Spider monkey individuals, and initialize local leader spider monkey individuals within each subgroup; S34. Define the fitness function Used for comprehensive evaluation of parameter combinations Accuracy and ranking rationality of manufacturing task scheduling priority score prediction: ; in, Indicates combining parameters After being applied to the MLP-Mixer network model, the predicted scheduling priority score for the i-th manufacturing task is... For true priority labels, For the set of true precedence pairs, These are weighting coefficients. The number of samples for the manufacturing task. To find the maximum value, Indicates combining parameters The predicted scheduling priority score of the j-th manufacturing task after being applied to the MLP-Mixer network model; S35. In each iteration, a local search phase is performed, where each spider monkey individual within a subgroup updates its position based on its local leader and its group neighbors: ; in, This represents the intermediate parameter vector updated through local search in the t-th iteration. This represents the parameter vector of the i-th individual in the t-th iteration. This is a local leader guiding factor, used to control the degree of influence of local leaders on the search direction. The neighbor-individual guiding factor is used to control the degree to which interactions between individual spider monkeys affect the search direction. This represents the parameter vector of the local leader individual in the subgroup in the current t-th iteration. Let be the parameter vector of a neighboring individual in the same subgroup as the i-th spider monkey individual; S36. Execute the global boot phase and establish the elite memory bank. It is used to store the best historical individuals and sets the maximum lifespan for each elite individual. Each iteration updates and replaces based on the current globally optimal result; S37. Select the spider monkey individual with the highest fitness from the elite memory bank. It is used to guide the current population to perform a global position update, based on intermediate update values: ; in, Let represent the final updated parameter vector position of individual i in round t+1. These are random numbers used to adjust the influence strength of the global guiding factor; S38. After each iteration, calculate the fitness value of all spider monkey individuals, and update the lifecycle of the global leader, the local leaders of each subgroup, and the spider monkey individuals in the elite memory bank based on the fitness values. ,like If so, the individual spider monkey will be removed from the memory bank; S39. If a subgroup is in continuous If there is no significant improvement in fitness within a round, the structural restructuring mechanism of the subgroup is triggered, the subgroup individuals are randomly re-initialized or migration is guided by other subgroups, and a local leader is re-selected; S310. Once the maximum number of iterations or fitness convergence condition is met, select the spider monkey individual with the highest fitness from the elite memory. As the optimal configuration of structural parameters, the updated structural parameters and training parameters are assigned to the MLP-Mixer network model respectively, thus completing the joint optimization of structural parameters and training parameters.

[0030] This invention structurally improves upon the original spider-monkey optimization algorithm by introducing a task feature-guided grouping mechanism and an elite memory management mechanism. This enables efficient joint optimization of the structural and training parameters of the MLP-Mixer network model, enhancing the search capability and convergence quality of the manufacturing task scheduling priority prediction model. By constructing an optimization variable vector containing three key parameters—feature selection weight matrix, task attention mapping matrix, and channel compression ratio—and dividing them into three functional subgroups based on their respective functional modules, a structure-oriented search architecture is built. This allows the optimization process to focus more on the synergistic improvement within each functional module of the model. Each subgroup independently performs local search, achieving fine-grained updates through information interaction between local leaders and neighbors, thus improving the local accuracy of structural parameter search. The introduction of the elite memory mechanism preserves the parameter combinations with the best fitness performance in the global search history and avoids premature convergence and over-reliance on historical optimal solutions through a lifecycle management mechanism, improving the diversity and robustness of the overall search. Furthermore, by setting intermediate position update values ​​and a global guiding factor to perform global position updates in stages, the algorithm's convergence stability and global guidance capability in high-dimensional parameter spaces are enhanced. Ultimately, based on the dual evaluation mechanism of integrating prediction error and ranking consistency in the fitness function, this invention can dynamically select the optimal parameter combination to optimize the scheduling prediction model, significantly improving the accuracy of scheduling priority score prediction and the rationality of ranking. This improved spider-monkey optimization algorithm possesses advantages such as a clear search structure, efficient parameter evolution, excellent convergence quality, and strong model adaptability, providing a reliable algorithmic foundation for model adaptive evolution and dynamic strategy optimization in complex scheduling environments of manufacturing enterprises.

[0031] In this embodiment, S4 specifically includes: S41. Input the scheduling input dataset into the parameter combination... The configured MLP-Mixer network model yields a manufacturing task scheduling priority score vector. : ; in, Indicates the first The scheduling priority prediction score for each task. , Total number of manufacturing tasks; S42. Construct a task scheduling priority index sequence based on the scores of each task in the manufacturing task scheduling priority score vector; S43. Generate an initial scheduling sequence based on the task scheduling priority index sequence. ,in Indicates the first Manufacturing tasks at each scheduling location ; S44. For each task in the initial scheduling sequence Verify resource allocation conditions sequentially. Equipment availability status Process constraints If there are tasks that do not meet the conditions Perform conflict correction operations, where... Indicates the first Manufacturing tasks at each scheduling location Indicates the first Manufacturing tasks at each scheduling location; S45. Perform a final legality check on the conflict-corrected scheduling sequence and renumber it into a sequential index format. A manufacturing task scheduling plan is constructed based on the executable time and resource allocation of each manufacturing task. ,in, Indicates the start time of task scheduling. Indicates the assigned device number, This represents the i-th manufacturing task after conflict correction; S46. Manufacturing task scheduling plan The output is sent to the scheduling and execution module for scheduling, distribution, and execution.

[0032] This invention, based on an optimized MLP-Mixer network model, efficiently outputs manufacturing task scheduling priority score vectors and constructs a task scheduling ranking index accordingly. This enables precise task-level ranking decisions and significantly improves the scheduling model's ability to predict task order. By constructing an index sequence and an initial scheduling sequence, the model's prediction results are directly mapped to an executable task scheduling plan structure, establishing a seamless "prediction-ranking-scheduling" link and effectively enhancing the engineering usability of the model's output. During the scheduling generation process, this invention incorporates multiple constraints such as resource availability, equipment operating status, and process flow to detect and correct conflicts in unreasonable task pairs within the scheduling sequence, constructing a scheduling plan with executability and resource matching. This scheduling plan, through sequential processing and number conversion, can be directly used in the scheduling execution system, ensuring the operability and stable execution of tasks in actual production environments. The final generated scheduling plan not only considers the priority relationships between tasks but also integrates resource allocation efficiency and equipment scheduling logic, possessing advantages such as high scheduling rationality, few execution conflicts, and strong plan adaptability. This provides a scheduling generation method with deep predictive support and engineering feasibility for multi-task concurrent scheduling in intelligent manufacturing environments.

[0033] In this embodiment, S5 specifically includes: S51. The scheduling plan table generated and verified by the manufacturing task scheduling priority score vector is sent to the manufacturing execution system, including the execution time, task number and resource allocation information of each task. S52. The manufacturing execution system receives the scheduling plan and controls each task to be executed in a predetermined order, time and equipment, and records the actual start time, end time and equipment number of the task. S53. After the task is completed, collect the task's execution status information; S54. Collect usage data of manufacturing resources during the scheduling cycle, including the utilization rate, idle rate, task switching frequency and resource conflict records of each device, construct a set of device usage status, and generate resource status data. S55. By comparing the scheduling plan with the actual task execution results, construct a task-level difference dataset to identify scheduling deviations, including task delays, differences in equipment allocation, and changes in execution order. S56. Based on the task-level differential dataset and resource status data, generate error evaluation indicators to evaluate the accuracy, stability and resource utilization efficiency of the scheduling results.

[0034] This invention establishes a complete application path from the output of the predictive model to the actual production scheduling control by distributing the generated and verified scheduling plan to the Manufacturing Execution System (MES), thus realizing the task-level scheduling implementation of the model prediction results. The MES organizes each task to execute sequentially according to the planned instructions and records the actual start and end times of the tasks, the status information of the executing equipment, etc., in real time, providing a detailed raw data foundation for subsequent scheduling evaluation. After the task execution is completed, the system collects task status data and resource operation data, systematically constructing a task execution status set and an equipment usage status set to comprehensively reflect the execution performance and resource scheduling efficiency during the production process. By comparing the scheduling plan with the actual execution, a task-level difference dataset is constructed, which can accurately identify execution deviations such as task delays, equipment allocation offsets, and task sorting deviations, enabling quantitative analysis of scheduling execution deviations. The final generated error evaluation index system integrates three dimensions: scheduling accuracy, scheduling stability, and resource utilization efficiency, providing multi-dimensional feedback for model optimization. This mechanism not only enhances the observability and controllability of the scheduling system, but also provides key support for subsequent adaptive model updates and dynamic iterative optimization of strategies. It has significant advantages such as strong operational closed-loop performance, high feedback accuracy, and comprehensive scheduling quality assessment.

[0035] In this embodiment, the task execution status information specifically includes whether the task is completed on time, the task execution time, and the device status and switching status during task execution. This information is used to evaluate the execution deviation of the scheduling plan and the actual completion status of the task, and to provide basic data for generating error evaluation indicators.

[0036] In this embodiment, the feedback error dataset includes task execution deviation characteristics and resource usage deviation information, which are used to reflect the difference between the prediction of the MLP-Mixer network model and the actual scheduling results, and support the adaptive optimization and updating of the MLP-Mixer network model.

[0037] Example 1: To verify the feasibility of this invention in practice, it was applied to the flexible assembly production line of electrical control cabinets in a large intelligent manufacturing enterprise. The performance of the AI-based intelligent production scheduling resource optimization method proposed in this invention was compared with that of the traditional manual rule-based scheduling method during the same time period from 8:00 AM to 11:00 AM. This production line handles more than 150 customized tasks daily, and the equipment includes multi-station automated assembly robotic arms, intelligent handling trolleys, and vision inspection units. Previously, production scheduling was manually configured by the scheduling supervisor every two hours. Under fixed rules, the task sequence and resource allocation could not quickly respond to order fluctuations and equipment status changes, leading to equipment overload, production rhythm imbalance, and frequent conflicts.

[0038] This invention's system first automatically collects heterogeneous data from multiple sources, including order process requirements, equipment operating status, personnel scheduling, and material inventory, from MES and ERP systems. This data is then preprocessed to construct a scheduling input dataset. This data is input into an improved MLPMXer network model to complete structural initialization and priority score prediction. Subsequently, a spider-monkey optimization algorithm jointly optimizes three key parameters to generate a scheduling priority score vector. Based on the score ranking and combined with resource availability and process constraints, the system automatically generates and verifies an initial scheduling sequence, and after conflict correction, forms an executable plan. After the scheduling plan is issued, the manufacturing execution system executes tasks according to the plan, recording the actual start and completion times of tasks, actual equipment allocation, and execution status in real time; simultaneously, it collects resource status data such as equipment utilization rate, idle rate, switching frequency, and conflict status. The system compares the plan with the execution results, constructs a scheduling deviation and resource difference dataset, and generates error evaluation indicators to drive the next round of adaptive iterative optimization of the model.

[0039] Table 1 Comparison of Scheduling Methods ; As shown in Table 1, the AI-based intelligent production scheduling resource optimization method proposed in this invention outperforms traditional scheduling methods in several key indicators. Regarding task scheduling time, the scheduling plan generated by the method of this invention significantly reduces the average task delay time from 7.4 minutes in the traditional method to 2.6 minutes, a delay reduction of over 60%, indicating that the prediction priority model and conflict correction mechanism effectively improve the timeliness and accuracy of scheduling.

[0040] Regarding the matching degree of task execution order, traditional methods mainly rely on human experience and static rules, resulting in an overall task order matching degree of only 75% with the actual production line rhythm. However, this invention, through a deep prediction model and structural optimization mechanism, makes the scheduling order more in line with actual production, achieving a matching degree of 93.3%, significantly improving the executability of scheduling results and the stability of production rhythm.

[0041] In terms of resource utilization efficiency, the optimized scheduling model of this invention significantly improves equipment utilization, increasing the average resource utilization rate from 68.0% in the traditional method to 80.5%. This result reflects that the method has stronger optimization capabilities in terms of task and resource matching and equipment allocation rationality, effectively reducing resource idleness and inefficient use.

[0042] In terms of equipment conflict control, traditional scheduling methods lack conflict prediction and validity verification mechanisms, resulting in approximately eight equipment scheduling conflicts every three hours on average, affecting task switching efficiency and system stability. This invention, however, significantly reduces the number of equipment conflicts to less than one by verifying the validity of scheduling sequences and correcting conflicts, ensuring scheduling continuity and execution reliability.

[0043] Furthermore, this invention demonstrates a significant efficiency advantage in scheduling response speed. Traditional methods typically require manual judgment and production scheduling, with an average time of 24 minutes to generate an effective scheduling plan. However, by adopting the method of this invention, relying on model prediction and intelligent optimization algorithms, a high-quality scheduling plan can be automatically generated in less than 3 minutes, meeting the real-time requirements of high-frequency and highly variable production environments.

[0044] In summary, this invention demonstrates comprehensive advantages over traditional methods in terms of scheduling accuracy, resource utilization, execution stability, and scheduling response efficiency, validating its significant beneficial effects and application value in practical manufacturing systems. This system is suitable for intelligent manufacturing scenarios involving large-volume, multi-variety production, strict production cycle requirements, or highly coupled resources, and possesses promising prospects for widespread adoption.

[0045] 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. An artificial intelligence-based intelligent production scheduling resource optimization method for manufacturing enterprises, characterized in that, The method comprises the following steps: S1, collecting multi-source heterogeneous data in a manufacturing enterprise, preprocessing the multi-source heterogeneous data, and constructing a scheduling input dataset; S2, constructing an MLP-Mixer network model, the MLP-Mixer network model receiving the scheduling input dataset, completing MLP-Mixer network model structure initialization and prediction capability modeling; S3, introducing an improved spider monkey optimization algorithm to jointly optimize the structure parameters and training parameters of the MLP-Mixer network model; S4, based on the optimized MLP-Mixer network model, outputting a manufacturing task scheduling priority score vector, generating an initial scheduling sequence of the manufacturing task based on the manufacturing task scheduling priority score vector, and performing scheduling legality verification and conflict correction according to resource availability and process constraint conditions to form a scheduling plan; S5, issuing the scheduling plan to a manufacturing execution system, constructing a scheduling execution result according to the task completion state, device utilization rate and scheduling time, and generating an error evaluation index based on the comparison between the scheduling plan and the scheduling execution result; S6, constructing a feedback error dataset based on the error evaluation index, and using the spider monkey optimization algorithm to adaptively update the MLP-Mixer network structure parameters, and completing dynamic iterative optimization of the scheduling strategy. 2.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, The multi-source heterogeneous data specifically includes order information, device status, personnel scheduling and material inventory, and is used to support the construction of a manufacturing task scheduling input dataset and scheduling decision modeling. 3.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, The preprocessing of the multi-source heterogeneous data specifically includes data cleaning, standardization and feature encoding, and is used to improve the consistency of the multi-source heterogeneous data and the effectiveness of the manufacturing task scheduling modeling. 4.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, S2 specifically includes: S21, constructing the dispatch input dataset into an input matrix wherein is the number of manufacturing tasks, is the feature dimension for each task, is a real number set; S22, input matrix to the dynamic feature selection layer to obtain a feature selection weight vector , wherein the feature selection weight vector is controlled by a feature selection weight matrix and a bias term , and the input matrix is adjusted by a column-by-column weighting manner to obtain a feature-selected matrix ; S23, the matrix after feature selection Input to the Patch division module, and divided into several task vector blocks, each vector block dimension is ; S24, input the divided task vector to the Token-Mixing module, perform feature interaction mapping between tasks, and then input to the task attention fusion layer, the task attention fusion layer generates a task importance weight vector according to the feature vector of each task , the task importance weight vector is controlled by a task attention mapping matrix and a bias term ​ S25, input the fused matrix to a Channel-Mixing module, the Channel-Mixing module includes an adjustable channel compressor mechanism, and the channel compression ratio or the compressed hidden layer dimension As a structure parameter, adjust the number of hidden layer neurons; S26, adding residual connections and layer normalization operations between the Token-Mixing module, the task attention fusion layer and the Channel-Mixing module, constructing an improved MLP-Mixer network structure containing a dynamic feature selection layer, a task attention fusion layer and a channel compression mechanism, and completing MLP-Mixer network model structure initialization and prediction capability modeling for manufacturing task scheduling priority score prediction. 5.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, S3 specifically includes: S31, combining the three parameters to be optimized in the MLP-Mixer network model to form an optimization variable vector: ； wherein, is a feature selection weight matrix, is a task attention mapping matrix, is a channel compression ratio; S32, according to the function modules of the parameters, dividing the optimization variable vector into three subgroups according to the structure function, respectively corresponding to the feature selection subgroup, the task attention subgroup and the channel compression subgroup, and constructing a task feature guided grouping mechanism; S33, initializing a population of spider monkey optimization algorithm, each spider monkey individual representing a set of parameter combinations , the total population size is , each sub-population is divided into spider monkey individuals on average, and a local leader spider monkey individual is initialized in each sub-population; S34, defining a fitness function , for comprehensive evaluation of parameter combinations Accuracy and reasonability of ranking for manufacturing task scheduling priority score prediction: ； wherein, combining parameters the predicted scheduling priority score of the i-th manufacturing task after being applied to the MLP-Mixer network model, is a real priority label, is a set of real priority pairs, is a weighting coefficient, is a number of samples of manufacturing tasks, is a max operation, combining parameters the predicted scheduling priority score of the j-th manufacturing task after being applied to the MLP-Mixer network model; S35, in each iteration, perform a local search phase, and each subgroup spider monkey individual updates its position according to the local leader and neighbors in the same subgroup: ； wherein, denotes the intermediate parameter vector updated by local search in the tth iteration, denotes the parameter vector of the ith individual in the tth iteration, is a local leader guiding factor for controlling the degree of influence of the local leader on the search direction, is a neighbor individual guiding factor for controlling the degree of influence of the interaction between the spider monkey individuals on the search direction, denotes the parameter vector of the local leader individual of the subgroup in the tth iteration, is the parameter vector of one neighbor individual in the same subgroup as the ith spider monkey individual. S36, performing a global guidance stage, establishing an elite memory pool for storing historical optimal individuals, setting a maximum life cycle for each elite individual updating and replacing according to a current global optimal result in each round of iteration S37. Selecting the spider monkey individual with the highest fitness from the elite memory bank for guiding the current population for global position update, based on the intermediate update value ； wherein, denotes the final updated parameter vector position of individual i in the t+1th round, is a random number used to adjust the influence strength of the global guide factor; S38. After each iteration, calculate the fitness value of all spider monkey individuals, and update the lifecycle of the global leader, the local leaders of each subgroup, and the spider monkey individuals in the elite memory bank based on the fitness values. ,like If so, the individual spider monkey will be removed from the memory bank; S39, if a sub-population has no significant fitness improvement in continuous S39, if a sub-population has no significant fitness improvement in continuous S39, if a sub-population has no significant fitness improvement in continuous S310, when the maximum number of iterations or fitness convergence condition is met, the spider monkey individual with the highest fitness in the elite memory bank is selected As the optimal structure parameter configuration, the updated structure parameters and the training parameters are respectively assigned to the MLP-Mixer network model, and the joint optimization of the structure parameters and the training parameters is completed. 6.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, S4 specifically includes: S41, inputting the scheduling input data set into the parameter combination The configured MLP-Mixer network model obtains a manufacturing task scheduling priority score vector : ； wherein, represents a scheduling priority prediction score for the , is the total number of manufacturing tasks;​ S42, constructing a task scheduling priority index sequence according to the score of each task in the manufacturing task scheduling priority score vector; S43, generating an initial scheduling sequence based on the task scheduling priority index sequence wherein denotes a manufacturing task at the th scheduling position, ; S44. For each task in the initial scheduling sequence Verify the resource allocation conditions in sequence. Equipment availability status Process constraints If there are tasks that do not meet the conditions Perform conflict correction operations, where... Indicates the first Manufacturing tasks at each scheduling location Indicates the first Manufacturing tasks at each scheduling location; S45, performing final legality check on the conflict-corrected scheduling sequence, and renumbering into sequential index form and constructing a manufacturing task scheduling table based on the executable time and resource allocation of each manufacturing task wherein, denotes the scheduling start time of the task, denotes the allocated equipment number, denotes the conflict-corrected i-th manufacturing task; S46, scheduling the manufacturing task plan table output to the scheduling execution module for scheduling and execution. 7.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, S5 specifically includes: S51, issuing the scheduling plan table generated and verified by the manufacturing task scheduling priority score vector to the manufacturing execution system, including the execution time, task number and allocated resource information of each task; S52, receiving the scheduling plan by the manufacturing execution system and controlling actual execution of each task according to the predetermined order, time and device, recording actual start time, completion time and used device number of the task; S53, collecting execution state information of the task after completion of the task execution; S54, collecting usage data of the manufacturing resource in the scheduling period, including the utilization rate, idle rate, task switching frequency and resource conflict record of each device, constructing a device usage state set and generating resource state data; S55, constructing a task level difference data set by comparing the scheduling plan and the actual task execution result, identifying scheduling deviations, including task delay, device allocation difference and execution order change; S56, generating an error evaluation index based on the task level difference data set and the resource state data, evaluating the accuracy, stability and resource utilization efficiency of the scheduling result. 8.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 7, wherein, The execution state information of the task specifically includes whether the task is completed on time, task execution time consumption, device state and switching condition in the task execution process, which is used to evaluate the execution deviation of the scheduling plan and the actual completion condition of the task, and provides basic data for generating the error evaluation index. 9.The AI-based manufacturing enterprise intelligent production scheduling resource optimization method of claim 1, wherein, The feedback error data set includes task execution deviation characteristics and resource usage deviation information, which is used to reflect the difference between the MLP-Mixer network model prediction and the actual scheduling result, and supports adaptive optimization and update of the MLP-Mixer network model.

Images