Intelligent scheduling management system for process parameters of garment production line

Abstract

The invention discloses a clothing production line process parameter intelligent scheduling management system, and relates to the technical field of industrial automation, and the system comprises a parameter mapping module which is used for converting clothing design parameters into production working condition demands through a mapping matrix; the intelligent scheduling module is used for generating candidate scheduling schemes based on an association rule mining algorithm and a multi-objective optimization algorithm according to production condition requirements, and selecting an optimal scheduling scheme based on a digital twinborn simulation result; the adaptive learning module is used for optimizing the optimal scheduling scheme and the digital twinborn model in real time through an inner ring and outer ring learning mechanism to obtain an optimal scheduling strategy; the optimal scheduling strategy is used for guiding production of the garment production line. On the basis, the technical problems that design data and production execution are disjointed and production resource scheduling depends on artificial experience in an existing garment production system are solved.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation, specifically an intelligent scheduling and management system for process parameters of a garment production line. Background Technology

[0002] With the deepening of the personalization and fast fashion trends in the consumer market, the garment manufacturing industry faces the challenge of small-batch, multi-variety, and short-delivery production. Traditional garment production lines rely heavily on the personal experience of process engineers and schedulers in the transformation from design to production, resulting in significant technical bottlenecks.

[0003] First, there are data and knowledge barriers between the dimensions, processes, and material parameters generated in the design phase and the specific equipment parameters, working hours, and resource allocation required in the production phase. This disconnect is usually resolved through manual interpretation and experience-based calculations, which is not only inefficient and inconsistent but also difficult to pass on and optimize. This results in long new product introduction cycles and a high dependence on skilled workers, becoming a key obstacle to flexible manufacturing.

[0004] Secondly, at the production scheduling level, existing methods are mostly based on fixed rules or the experience of schedulers, lacking the ability to simultaneously optimize multiple objectives such as production efficiency, equipment utilization, production costs, and product quality. Manual scheduling struggles to cope with complex resource constraints and dynamic disturbances, often resulting in plans with poor feasibility and weak resistance to interference, leading to chaotic production rhythms, resource waste, and delivery delays. Furthermore, existing production management systems are typically static or fixed after configuration, lacking self-learning and adaptability. When equipment performance ages, material characteristics fluctuate, or production modes change, the system cannot automatically adjust process parameters and scheduling strategies, resulting in unstable production quality and a gradual decline in efficiency. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an intelligent scheduling and management system for process parameters of garment production lines, which is used to solve the technical problems of existing garment production systems, such as the disconnect between design data and production execution, reliance on manual experience for production resource scheduling, and difficulty in achieving multi-objective optimization.

[0006] To achieve the above objectives, a first aspect of the present invention provides an intelligent scheduling and management system for process parameters of a garment production line, comprising: The parameter mapping module is used to convert clothing design parameters into production condition requirements through a mapping matrix. The clothing design parameters include size parameters, process requirement parameters, and material characteristic parameters. The production condition requirements include optimal process parameters, equipment configuration, and time allocation. The intelligent scheduling module is used to generate candidate scheduling schemes based on the production conditions and the association rule mining algorithm and the multi-objective optimization algorithm, and to select the optimal scheduling scheme based on the digital twin simulation results. An adaptive learning module is used to optimize the optimal scheduling scheme and digital twin model in real time through inner-loop and outer-loop learning mechanisms to obtain the optimal scheduling strategy; the optimal scheduling strategy is used to guide the production of the garment production line.

[0007] Based on the above technical solutions, this application provides an intelligent scheduling and management system for garment production line process parameters, constructing an intelligent closed loop from design to production scheduling, thereby improving the automation level, resource utilization, and production efficiency of the garment production line. First, this application achieves automatic conversion from garment design parameters to production condition requirements through a parameter mapping module, effectively solving the data disconnect between design and production, and reducing errors and time delays caused by manual intervention. Second, the intelligent scheduling module, combined with association rule mining and multi-objective optimization algorithms, can generate efficient candidate scheduling schemes, which are then verified and optimized using digital twin simulation, improving the accuracy and reliability of scheduling decisions while reducing trial-and-error risks. Furthermore, the adaptive learning module, through inner-loop and outer-loop learning mechanisms, enables the system to continuously learn from production feedback and optimize scheduling strategies and models, adapting to dynamic changes in the production line and achieving system self-improvement and long-term performance enhancement.

[0008] Furthermore, the system also includes a parameter input module, a working condition sensing module, a visualization interaction module, a parameter mapping module, and a data bus module; wherein, The parameter input module is used to receive clothing design parameters from the clothing design system; The working condition sensing module is distributed at each node of the production line and is used to collect equipment status, garment fabric characteristics and environmental data; the equipment status includes equipment speed, temperature and pressure operating parameters, which are used to reflect the real-time working performance and health status of the equipment; the garment fabric characteristics represent physical parameters such as fabric thickness, elasticity and breathability. The visualization interaction module is connected to each module and is used to provide a user interface and real-time monitoring display; The parameter mapping module is connected to the execution control module and the adaptive learning module, and is used to collect production data and update system parameters. The data bus module adopts the industrial Ethernet protocol to realize data exchange and synchronization between modules. The modules are interconnected through the data bus module. The design parameters received by the parameter input module are processed by the parameter mapping module, and then the intelligent scheduling module generates a scheduling scheme. After being verified by the digital twin model, the scheme is sent to the execution control module, and then a closed-loop optimization is formed through the adaptive learning module and the parameter mapping module.

[0009] Furthermore, the workflow of the parameter mapping module includes: A dynamic mapping matrix is ​​constructed. The mapping matrix is ​​trained using historical production data and machine learning methods to map clothing design parameters to production working condition requirements. The mapping matrix includes a basic mapping layer and a dynamic adjustment layer. The basic mapping layer is trained based on historical data to obtain a static mapping relationship, and the dynamic adjustment layer adjusts the mapping weights according to real-time working condition data. Receive clothing design parameters from the parameter input module and standardize the clothing design parameters into vector form to obtain the parameter vector; The parameter vector is mapped to a production condition demand vector using the dynamic mapping matrix. The dynamic adjustment layer is activated based on real-time data collected by the working condition sensing module to adapt to changes in equipment status and fabric characteristics.

[0010] Furthermore, the basic mapping layer obtains static mapping relationships based on historical data, including: Build a dataset by acquiring historical successful order data ;in, Let i be the design parameter vector for the i-th order. Let i be the actual production demand vector for the i-th order. For the production result indicators of the i-th order, the production result indicators include product qualification rate and production efficiency; The design parameter vector is extended with features to calculate the process complexity features; the process complexity features include cutting complexity, sewing complexity, and ironing complexity. The relationship between design parameters and process parameters is modeled as a multivariate nonlinear mapping, and trained using a combination of rule-constrained multivariate nonlinear regression and ensemble learning; the objective function of the multivariate nonlinear regression is: , Let W be the mapping function, and W be the weights of the mapping matrix to be trained. For regularization terms, The regularization coefficient is used. During training, a process knowledge graph is introduced for constraints; where nodes in the process knowledge graph are process parameters, and edges represent physical or logical constraints between process parameters. The dataset is divided into training and validation sets in chronological order, and the validated mapping relationships are solidified into a basic mapping matrix to obtain static mapping relationships.

[0011] Furthermore, the dynamic adjustment layer adjusts the mapping weights based on real-time operating condition data, including: Data streams are obtained from the working condition sensing module, and the data streams include equipment state vectors, fabric characteristic vectors, and environment vectors. The operating condition indicators are calculated based on the data stream, and when the value of the operating condition indicator is greater than a preset threshold, it is determined that the operating condition is abnormal; the operating condition indicators include the equipment health index and the fabric batch offset. The weight adjustment of the basic mapping matrix is ​​treated as a state estimation problem, thereby establishing a state-space model; the state-space model includes state equations. and observation equations , Let be the vectorized form of the fundamental mapping matrix at time t. The observation matrix is ​​composed of the outer product of the current operating condition characteristics and the design parameters. Here is the state transition matrix. For process noise, To observe noise; Based on the state-space model, the Kalman filter recursive algorithm is applied to update the estimated value of the mapping matrix online, resulting in a dynamically adjusted mapping matrix.

[0012] Furthermore, the step of generating candidate scheduling schemes based on the production condition requirements, using an association rule mining algorithm and a multi-objective optimization algorithm, includes: The association rule mining algorithm is used to extract process-equipment-personnel combination patterns from the historical scheduling database to form a rule set; When a new production condition demand vector is received When dealing with the device state vector R, As a rule antecedent, rules that meet the conditions are matched from the rule set to obtain the rule consequent; wherein, the rule consequent is a scheduling fragment in the rule set, and the scheduling fragment is... This indicates that the task will be completed. Assigned to device and personnel From the start time Start, duration is ; By combining the consequents of multiple matching rules, one or more initial scheduling schemes are constructed as the initial population for the multi-objective optimization algorithm. Define a multi-objective function vector for the scheduling problem, including the maximum completion time objective function, the total production cost function, the equipment load balance standard deviation function, and the estimated quality risk function; An improved NSGA-III algorithm is used to optimize the multi-objective function. In each iteration of the algorithm, the generated candidate solutions are input into a digital twin model for simulation verification. The digital twin model outputs the estimated distribution of each objective function in the multi-objective function vector based on equipment failure and material supply fluctuation factors. Calculate the expected performance of each distribution based on the estimated distribution. and robustness For non-dominated sorting and selection; wherein, the expected performance Let i be the expected value of the objective function. Let i be the standard deviation of the objective function. After the algorithm is completed, a set of Pareto optimal solutions is obtained, and the comprehensive proximity of each solution in the solution set is calculated by using the decision analysis method based on the superior-inferior solution distance method TOPSIS. The solution with the highest overall approximation is selected as the final production scheduling solution.

[0013] Furthermore, the process logic of the association rule mining algorithm includes: Each historical scheduling record is transformed into a transaction, which includes a set of processes, equipment allocation, personnel allocation, and production results. Each transaction is assigned a comprehensive weight to obtain a weighted transaction database; the formula for calculating the comprehensive weight is: ;in, It is a time decay function. The time when the transaction occurred. The attenuation coefficient is... For the current time, Here, j is the quality weighting function based on the production result indicators. This represents the j-th transaction; An improved association rule mining algorithm, Apriori, is used to scan a weighted transaction dataset, calculate the weighted support of all 1-itemsets, and filter out frequent 1-itemsets based on a preset support threshold. The formula for calculating weighted support is: , Representation Itemset The weighted support; Candidate k-itemsets are generated iteratively through join and pruning steps, and the weighted support of the candidate k-itemsets is calculated to generate frequent k-itemsets. until no new frequent itemsets can be generated; For each frequent itemset Generate all non-empty proper subsets of the frequent itemset. and calculate the rules Weighted confidence level Only rules with a weighted confidence level exceeding a preset confidence threshold are retained to obtain the first rule set; Based on the weighted confidence and weighted support, the lift and confidence are calculated. Rules whose lift and confidence exceed the corresponding preset thresholds are retained from the first rule set to obtain the final rule set.

[0014] Furthermore, the inner and outer loop learning mechanisms represent a hierarchical collaborative continuous optimization framework, wherein, The inner-loop learning mechanism operates on a minute / hour timescale, and performs parameter optimization and strategy fine-tuning based on real-time production feedback to cope with instantaneous fluctuations in the production line. The outer loop learning mechanism operates on a daily / weekly timescale, and performs model reconstruction and knowledge extraction based on accumulated historical data batches to optimize the long-term performance of the system. Furthermore, the inner loop and the outer loop interact bidirectionally with the model parameter interface through a shared experience buffer pool. That is, the state transition experience of the inner loop provides incremental data to the outer loop, and the outer loop periodically samples data from the experience buffer pool, trains a new global model, and synchronizes the updated model parameters to the inner loop, providing the inner loop with a better initial strategy.

[0015] Furthermore, the workflow of the inner-loop learning mechanism includes: For the dynamic adjustment layer in the parameter mapping module, the mapping weights are updated online using the recursive least squares method; The discrete choices in the scheduling strategy are modeled as a contextual multi-armed slot machine problem; wherein, the discrete choices represent task allocation, equipment selection, and process parameter adjustment; each "arm" of the contextual multi-armed slot machine represents a decision action, including allocating a task to a specific piece of equipment, selecting a specific combination of process parameters, and adjusting the production sequence; the context represents the current production line status, including equipment status, work-in-process quantity, order priority, and current process parameter settings; Based on reinforcement learning theory, according to the current production line state and the expected reward of each decision action, a Thompson sampling strategy is used to sample from the posterior reward distribution of each decision action, select the action with the largest sample value to execute, and update the posterior reward distribution of the decision action according to the actual report, thereby learning the action selection strategy that maximizes the cumulative reward under different states; wherein, the expected reward represents the estimated value of the expected cumulative discount reward calculated based on the current value function or historical average reward after selecting the decision action, the actual reward represents the estimated value of the immediate reward observed after executing the decision action plus the future discount reward, and the posterior reward distribution represents the probability distribution of the expected reward of each decision action; Establish an experience buffer pool to store the state transition experience of the inner loop during the reinforcement learning process.

[0016] Furthermore, the workflow of the outer-loop learning mechanism includes: Collect production data over a complete cycle to form a batch dataset; the production data includes state transition experience in the experience buffer pool. The batch dataset is used to retrain or fine-tune the basic mapping matrix, rule set, and digital twin model parameters in the system; For the scheduling strategy generated by inner-loop learning, outer-loop learning transfers knowledge to a simplified model through knowledge distillation, and trains the simplified model to minimize the KL divergence between the output distribution and the output distribution of the inner-loop action selection strategy. New rules extracted from the trained simplified model are added to the rule set; the knowledge represents the decision patterns and experiences in the inner-loop learning mechanism, and the simplified model is an interpretable mathematical model, including decision trees, linear models, or rule sets.

[0017] Compared with the prior art, the beneficial effects of the present invention are: By constructing a system architecture that includes modules for parameter input, working condition perception, visual interaction, feedback optimization, and data bus, a digital closed loop for the entire process of garment production line from design to execution has been realized. The collaborative work of each module ensures the real-time, accurate, and synchronized data flow, providing a reliable data foundation and interaction channel for intelligent decision-making, and significantly improving the system's integration and ease of operation. First, this invention, through the collaborative design of the basic mapping layer and the dynamic adjustment layer in the parameter mapping module, not only learns process knowledge from historical data but also adaptively fine-tunes based on real-time production status. This effectively overcomes the problem of inaccurate process parameters caused by factors such as changes in equipment performance and batch differences in materials, improving the accuracy and robustness of parameter conversion. In terms of intelligent scheduling, it combines prior knowledge guidance from association rule mining with the global search capability of multi-objective optimization algorithms, and pre-verifies through digital twin simulation. This enables the efficient generation of Pareto-optimal scheduling schemes that balance efficiency, cost, balance, and quality risk, significantly improving the scientific nature and feasibility of production planning.

[0018] Furthermore, the adaptive learning mechanism that combines inner and outer loops is the core advantage of this system. Inner loop learning enables rapid online optimization within minutes, allowing for immediate responses to production fluctuations; outer loop learning, through periodic batch learning and knowledge distillation, deeply optimizes and refines the core model and rule base, ensuring the continuous evolution of the system's strategies. This hierarchical learning framework gives the system both short-term adaptability and long-term optimization capabilities, enabling it to continuously accumulate experience from production practice, transforming complex scheduling strategies into interpretable and maintainable rules, ultimately forming a fully intelligent garment production scheduling system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A system framework diagram of an intelligent scheduling and management system for process parameters of a garment production line provided by the present invention; Figure 2 A flowchart illustrating the parameter mapping module in an intelligent scheduling and management system for process parameters of a garment production line provided by the present invention. Figure 3 A flowchart illustrating the intelligent scheduling module in an intelligent scheduling management system for process parameters of a garment production line provided by the present invention. Figure 4 This is a flowchart illustrating the adaptive learning module in an intelligent scheduling and management system for process parameters of a garment production line. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-4 To address the technical problems in existing technologies, such as the disconnect between design data and production execution, low efficiency due to reliance on manual experience in production resource scheduling, difficulty in achieving multi-objective optimization in scheduling schemes, and the lack of adaptive adjustment capabilities to cope with dynamic production changes, this application provides an intelligent scheduling and management system for process parameters of a garment production line. The system includes: The parameter input module is used to receive clothing design parameters from the clothing design system; The parameter mapping module is used to convert clothing design parameters into production condition requirements through a mapping matrix. The clothing design parameters include size parameters, process requirement parameters, and material characteristic parameters, while the production condition requirements include optimal process parameters, equipment configuration, and time allocation. The intelligent scheduling module is used to generate candidate scheduling schemes based on association rule mining algorithms and multi-objective optimization algorithms according to production conditions, and select the optimal scheduling scheme based on digital twin simulation results. The adaptive learning module is used to optimize the optimal scheduling scheme and digital twin model in real time through inner and outer loop learning mechanisms to obtain the optimal scheduling strategy to guide the production of the garment production line. The working condition sensing module is distributed at various nodes of the production line to collect equipment status, garment fabric characteristics, and environmental data. Among them, the equipment status includes equipment speed, temperature, and pressure operating parameters, which are used to reflect the real-time working performance and health status of the equipment. The garment fabric characteristics represent physical parameters such as fabric thickness, elasticity, and breathability. The visual interaction module connects to each module and provides a user interface and real-time monitoring display; The parameter mapping module is connected to the execution control module and the adaptive learning module, and is used to collect production data and update system parameters; The data bus module adopts the industrial Ethernet protocol to realize data exchange and synchronization between modules. The modules are interconnected through the data bus module. The design parameters received by the parameter input module are processed by the parameter mapping module, and then the intelligent scheduling module generates a scheduling scheme. After being verified by the digital twin model, the scheme is sent to the execution control module, and then a closed-loop optimization is formed through the adaptive learning module and the parameter mapping module.

[0023] Based on this, this application constructs an intelligent closed loop covering the entire process from design to production execution, effectively breaking down the data barriers between design and production. At the same time, through adaptive learning, the system continuously adapts to the dynamic production environment, improving the automation level, resource utilization, and overall production efficiency of the garment production line.

[0024] like Figure 1 As shown in the embodiment of this application, an intelligent scheduling and management system for process parameters of a garment production line includes the following core modules: The parameter mapping module is used to convert clothing design parameters into production condition requirements through a mapping matrix; Among them, clothing design parameters include size parameters, process requirement parameters, and material characteristic parameters; production condition requirements include optimal process parameters, equipment configuration, and time allocation; In some implementations, parameter mapping can be based on statistical analysis of historical production data to construct a mapping matrix, or it can use classic linear mapping combined with industry process experience rules to form a transformation relationship. Alternatively, basic machine learning models such as decision trees and support vector machines can be used to train the mapping relationship. The core logic is to establish a correspondence between design parameters and production parameters, achieving accurate conversion between the two types of parameters through data-driven or experience-guided methods. This breaks down the data barriers between apparel design and production execution, avoiding the inefficiencies and inconsistencies caused by manual experience-based conversions, and ensuring that design intent can be accurately translated into executable production requirements.

[0025] It should be noted that the construction of the mapping matrix can be combined with the industry standard process parameter library to enhance its versatility. Regardless of whether linear mapping, nonlinear mapping or model-based mapping is used, the core goal is to achieve accurate and efficient conversion of design parameters to production condition requirements. Moreover, the conversion process must take into account the completeness and practicality of the parameters to ensure that the production condition requirements can fully cover the key information required for production execution.

[0026] For example, by collecting a large amount of data on size parameters, process requirements, material properties, and corresponding optimal process parameters, equipment configurations, and time allocation for similar types of clothing, statistical methods can be used to calculate the correlation coefficients between each design parameter and production parameter, constructing a basic linear mapping relationship. Then, combined with the process specifications of the clothing manufacturing industry, the mapping results can be corrected in a regularized manner to form the final mapping matrix. When new clothing design parameters are input, the corresponding production requirements can be directly output through this mapping matrix without the need for manual conversion.

[0027] The intelligent scheduling module is used to generate candidate scheduling schemes based on association rule mining algorithms and multi-objective optimization algorithms according to production conditions, and select the optimal scheduling scheme based on digital twin simulation results. Among them, the Association Rule Mining Algorithm is an algorithm that mines potential relationships between variables from large datasets. It is used to discover effective combinations of processes, equipment, and personnel from historical scheduling data, extract reusable scheduling experience, provide prior knowledge for generating reasonable scheduling schemes, reduce the generation of ineffective scheduling schemes, and improve scheduling efficiency. The Multi-Objective Optimization Algorithm is an algorithm that finds the optimal set of solutions under multiple conflicting objective function constraints. It is used to balance multiple objectives such as production efficiency, production cost, equipment utilization, and product quality risk, ensuring the comprehensiveness and practicality of the scheduling scheme.

[0028] In some implementations, the method for generating candidate scheduling schemes is as follows: First, historical scheduling records are analyzed using association rule mining algorithms to extract high-confidence scheduling association rules, forming a scheduling rule base. Then, using production demand as input conditions, and combining the scheduling rule base, preliminary selection of process-equipment-personnel combination schemes that meet basic requirements are made. Subsequently, a multi-objective optimization algorithm is used to set objective functions, such as minimizing the maximum completion time, minimizing the total cost, balancing equipment load, and reducing quality risks. The preliminary selection of combination schemes is used as the initial population, and multiple candidate scheduling schemes that meet multi-objective constraints are generated through iterative search. Finally, all candidate schemes are input into a digital twin model for simulation, simulating the equipment operating status, material flow, and possible disturbances during the production process, evaluating the performance of each candidate scheme in the actual production scenario, and finally selecting the scheduling scheme with the best overall performance.

[0029] It should be noted that during the association rule mining process, effective rules can be filtered and invalid or low-value associations can be eliminated by setting support and confidence thresholds; existing algorithms such as Non-dominated Sorting Genetic Algorithm II (NSGA-II), ideal point method, and weighted sum method can be selected for multi-objective optimization algorithms; digital twin simulation can build a simplified simulation model based on the actual layout of the production line, equipment parameters, and process flow, focusing on simulating the key links that affect the scheduling effect, thereby reducing the complexity of model construction and simulation operation.

[0030] For example, association rules such as "prioritizing the sewing process of knitted fabric garments to high-speed sewing machines" and "assigning the cutting process of high-priority orders to experienced operators" are mined from historical scheduling data to form a scheduling rule base. For a specific batch of small-volume, multi-variety garments, the allocation direction of equipment and personnel for each process is initially determined based on the rule base. Then, using minimizing the production cycle, minimizing production costs, and balancing equipment load as objective functions, the NSGA-II algorithm iteratively generates 20 candidate scheduling schemes. Each scheme is input into a digital twin simulation model to simulate scenarios such as temporary equipment failures and material supply delays that may occur during production, recording data such as completion time, cost consumption, equipment load fluctuations, and estimated quality pass rate for each scheme. Based on the simulation results, the scheme with the best overall performance is selected as the final scheduling scheme.

[0031] The adaptive learning module is used to optimize the optimal scheduling scheme and digital twin model in real time through inner and outer loop learning mechanisms to obtain the optimal scheduling strategy to guide the production of the garment production line.

[0032] The optimal scheduling scheme is an initial scheduling plan generated based on current production demand. While highly targeted, it lacks flexibility and struggles to cope with dynamic changes during production. The optimal scheduling strategy, on the other hand, is a sustainable production guidance strategy formed by dynamically adjusting the optimal scheduling scheme based on real-time production feedback data through inner and outer loop learning mechanisms. It not only includes the core content of the initial scheduling plan but also covers adjustment rules for dynamic situations such as changes in equipment status, fluctuations in material characteristics, order adjustments, and environmental changes. It represents real-time optimization and long-term improvement of the optimal scheduling scheme, ensuring the continuity and stability of the production process.

[0033] In some implementations, the roles of the inner and outer learning mechanisms are as follows: The inner learning mechanism typically employs online learning algorithms, operating on shorter timescales, such as minutes or hours. Common forms include gradient descent, incremental learning, and online gradient boosting. It is used to quickly adjust scheduling parameters and strategies based on real-time production data collected by the condition sensing module, such as abnormal equipment operating parameters, deviations between fabric characteristics and estimates, and adjustments to order urgency. This allows for immediate responses to instantaneous fluctuations in production, preventing small disturbances from escalating into production failures. The outer learning mechanism typically employs offline batch learning, operating on longer timescales, such as days or weeks. Common forms include batch gradient descent, model retraining, and knowledge distillation. Its core function is to update and optimize parameters such as digital twin model parameters, association rule bases, and multi-objective optimization algorithms based on accumulated production data, including adjustment experience recorded during the inner learning process and long-term production effect data. This improves the long-term performance and stability of the system, transforming short-term adjustment experience into long-term reusable knowledge.

[0034] It should be noted that the inner and outer learning mechanisms do not operate in isolation, but rather collaborate through a shared data storage module: the real-time adjustment experience of the inner loop provides rich dynamic data support for the batch learning of the outer loop, ensuring that the outer loop optimization is more in line with actual production scenarios; the optimization results of the outer loop, such as updated model parameters and newly added association rules, provide better initial basis and rule guidance for the immediate adjustment of the inner loop, forming a continuous optimization closed loop of "real-time adjustment - batch optimization - guidance for real-time adjustment".

[0035] For example, the inner-loop learning mechanism uses gradient descent to monitor operating parameters such as equipment speed and temperature in real time. When it detects that a decrease in the speed of a sewing machine is causing a reduction in production efficiency, it immediately adjusts the scheduling strategy, rationally allocating unfinished tasks on that machine to other idle machines of the same type to ensure uninterrupted production. The outer-loop learning mechanism collects production data from the entire production line weekly, including the actual execution effect of each scheduling scheme, inner-loop adjustment records, product qualification rate, equipment failure rate, etc., to retrain the equipment performance parameters and material flow time parameters of the digital twin model, improving the accuracy of the model simulation. At the same time, it updates the association rule base, transforming newly added effective experiences such as "task transfer rules when equipment speed decreases" into association rules and incorporating them into the rule base to support the generation and real-time adjustment of subsequent scheduling schemes.

[0036] Based on the above technical solutions, the intelligent scheduling and management system for garment production line process parameters provided in this application achieves efficient collaboration among its various modules through a data bus module. The parameter mapping module solves the data disconnect between design and production. The intelligent scheduling module enables scientific scheduling under multi-objective constraints. The adaptive learning module empowers the system to cope with dynamic changes. The working condition perception module provides real-time data support for each stage. The visual interaction module enhances the ease of system operation. The parameter mapping module and data bus module ensure closed-loop operation and data synchronization of the system. The entire system can adapt to the needs of garment production lines of different scales and categories, improving the level of intelligent production management and production efficiency.

[0037] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above parameter mapping module can be implemented through the following steps S101, S102, and S103, which are explained in detail below: S101. Construct a dynamic mapping matrix containing a basic mapping layer and a dynamic adjustment layer to map clothing design parameters to production condition requirements.

[0038] The basic mapping layer is trained based on historical production data to obtain static mapping relationships, providing a stable basis for parameter conversion; the dynamic adjustment layer is used to adjust the mapping weights according to real-time operating data, so that the mapping results adapt to the dynamic changes in the production process. The two work together to ensure the accuracy and robustness of the mapping.

[0039] In some implementations, the construction of the base mapping layer includes the following steps: The first step is to obtain historical successful order data and construct a dataset, which is represented as follows: Where N is the total number of historical successful orders, This is the design parameter vector for the i-th order, encompassing core information such as dimensional parameters, process requirements, and material properties. This is the actual production demand vector for the i-th order. It is the production result indicator for the i-th order, specifically including two key evaluation dimensions: product qualification rate and production efficiency.

[0040] The second step involves feature expansion of the design parameter vector and calculation of process complexity features, specifically including: Cutting complexity Where L is the total length of the clipping path. This is the maximum cutting length of the equipment. This represents the actual cutting area. Let the area of ​​the rectangle representing the fabric boundary be . The number of inflection points of the contour curve. These are the weighting coefficients, and their sum is 1; Sewing complexity ;in The weighting coefficient for the k-th suture segment is... For the length of the stitches, The total length of the stitches. For the number of joints, These are the weighting coefficients, and their sum is 1; Ironing complexity ;in This refers to the ironing temperature range. This is the pressure range. The normalization coefficient is... These are the weighting coefficients, and their sum is 1; Introducing these features can more comprehensively characterize the relationship between design parameters and production processes, thereby improving the fitting ability of the mapping model.

[0041] The third step involves modeling the relationship between design parameters and process parameters as a multivariate nonlinear mapping. This is trained using a combination of rule-constrained multivariate nonlinear regression and ensemble learning, where the objective function of the multivariate nonlinear regression is... The objective function is used to minimize the deviation between the actual production process parameters and the model's predicted parameters, while avoiding model overfitting. In the formula... Ω(W) is the mapping function used to characterize the transformation relationship from the design parameter vector to the production demand vector. W is the weight of the mapping matrix to be trained. Ω(W) is the regularization term used to constrain the range of values ​​of the weight parameters. λ is the regularization coefficient used to balance the importance of the prediction bias term and the regularization term. It can be adjusted according to the distribution characteristics of historical data and the model training effect, and its value is usually between 0.001 and 0.1.

[0042] The fourth step is to introduce a process knowledge graph for constraints during the training process. The nodes of the process knowledge graph are various process parameters, and the edges represent physical or logical constraints between process parameters. For example, "high-temperature ironing process requires specific fabric thickness parameters" and "sewing density parameters and fabric elasticity parameters have a positive correlation constraint". Through this constraint, it can be ensured that the mapping relationship obtained by training conforms to the actual production process logic and avoids unreasonable parameter mapping results.

[0043] The fifth step is to divide the dataset into training and validation sets in chronological order. Typically, the training set accounts for 70% to 80% and the validation set accounts for 20% to 30%. Then, the validated mapping relationships will be solidified into a basic mapping matrix, ultimately forming a static mapping relationship.

[0044] It should be noted that the core of the dynamic adjustment layer is to dynamically correct the weights of the base mapping matrix based on real-time operating data. Its construction does not require separate training of the initial model, but rather relies on the weight structure of the base mapping matrix to establish a dynamic response mechanism for weight adjustment.

[0045] For example, the basic mapping matrix is ​​a transformation matrix from the space of clothing design parameters to the space of production process parameters, defining the mathematical mapping relationship between design parameters and process parameters. Its form can be expressed as: Through matrix multiplication The system will input the design parameter vector Mapped to the initial production condition demand vector Each weight value in the matrix has a clear physical meaning: a positive value indicates a positive correlation between the design parameter and the target process parameter, such as increased fabric thickness leading to increased sewing tension; a negative value indicates a negative correlation, such as increased fabric elasticity allowing for lower sewing tension. The absolute value of the weight reflects the strength of the influence. Furthermore, the weight relationships in the matrix are derived through machine learning methods trained on massive amounts of historical successful production data, and incorporate process knowledge constraints such as "thick fabrics require high tension" and "high-elasticity fabrics require low tension," ensuring the rationality and feasibility of the recommended process parameters. This basic mapping matrix is ​​the core knowledge base for the system's intelligent scheduling and can be dynamically fine-tuned based on real-time production feedback through subsequent adaptive learning modules.

[0046] S102. Receive clothing design parameters from the parameter input module and standardize the clothing design parameters into vector form to obtain the parameter vector.

[0047] Among them, clothing design parameters include size parameters, process requirement parameters, and material characteristic parameters. The dimensions and value ranges of different types of parameters are different. The influence of dimensions can be eliminated through standardization.

[0048] In some implementations, standardization can employ methods such as Min-Max standardization or Z-score standardization, with the specific choice depending on the distribution characteristics of the parameters. Furthermore, the standardization process requires first cleaning the received clothing design parameters to remove outliers, then performing the corresponding standardization operations according to parameter type, and finally arranging all standardized parameters in a preset order to form a parameter vector with uniform dimensions.

[0049] S103. The parameter vector is mapped to the production condition demand vector using a dynamic mapping matrix, and the dynamic adjustment layer is enabled based on the real-time data collected by the condition perception module to adapt to changes in equipment status and fabric characteristics.

[0050] The parameter vector is transformed into a production condition demand vector containing optimal process parameters, equipment configuration, and time allocation through linear or nonlinear transformation of the dynamic mapping matrix. The activation of the dynamic adjustment layer is based on real-time condition data. When fluctuations occur in the production environment, equipment status, or material characteristics, the accuracy of the mapping results is ensured by adjusting the weights of the mapping matrix.

[0051] In some implementations, the workflow of the dynamic adjustment layer includes: First, the standardized parameter vector is input into the basic mapping layer, and then the basic mapping matrix is ​​used for preliminary transformation to obtain the initial production condition demand vector. Subsequently, a real-time data stream is acquired from the operating condition sensing module. This data stream includes equipment status vectors, fabric characteristic vectors, and environmental vectors: the equipment status vectors cover operating parameters such as equipment speed, temperature, and pressure; the fabric characteristic vectors include actual measured physical parameters such as fabric thickness, elasticity, and breathability; and the environmental vectors include environmental parameters such as temperature and humidity in the production workshop. Operating condition indicators are calculated based on this data stream, specifically including the equipment health index and fabric batch offset. The equipment health index is calculated by comparing the actual values ​​of various operating parameters of the equipment with the standard values. The formula is as follows: Where HI is the equipment health index, with a value ranging from [0,1]. The closer to 1, the better the equipment health status. M is the number of equipment operating parameters. The weight of the k-th running parameter, This is the real-time value of the k-th running parameter. The standard value of the k-th running parameter. and These are the maximum and minimum allowed values ​​for the k-th operating parameter, respectively. Fabric batch offset is calculated by the deviation between the actual fabric characteristic parameters and the preset fabric characteristic parameters in the design stage. The formula is as follows: Where BO represents the fabric batch offset, and K represents the number of fabric characteristic parameters. The weight of the l-th fabric characteristic parameter is... This represents the real-time detection value of the l-th fabric characteristic parameter. This is the preset design value for the l-th fabric characteristic parameter.

[0052] The preset threshold for the equipment health index is set to 0.8, and the preset threshold for the fabric batch offset is set to 0.1. When the equipment health index is less than 0.8 or the fabric batch offset is greater than 0.1, it is determined to be an abnormal working condition, and the dynamic adjustment layer is activated at this time.

[0053] Next, the dynamic adjustment layer treats the weight adjustment of the basic mapping matrix as a state estimation problem and establishes a state-space model, which includes state equations. and observation equations The state equation describes the dynamic changes in the weights of the mapping matrix. Let be the vectorized form of the fundamental mapping matrix at time t. This is the state transition matrix, reflecting the weight transition relationship from time t to time t+1. It is usually set based on the equipment state change pattern and the fabric property fluctuation characteristics. The noise is the process noise, which follows a mean of 0 and a variance of . The normal distribution The process noise covariance matrix is ​​used; the observation equation is used to establish the correlation between the weights and the observed data, where... Let be the observation vector at time t, composed of the correlation characteristics between real-time operating data and design parameters. The observation matrix is ​​composed of the outer product of the current operating condition characteristics and the design parameters. To observe the noise, it follows a mean of 0 and a variance of . The normal distribution To observe the noise covariance matrix.

[0054] Based on this state-space model, the estimated value of the mapping matrix is ​​updated online using the Kalman filter recursive algorithm. The specific steps are as follows: The first step is to perform state prediction, and obtain , ;in, This is the weight vector at time t predicted based on the data at time t-1. For the prediction error covariance matrix, Let be the filter error covariance matrix at time t-1. for The transpose of the matrix; The second step is to calculate the Kalman gain. In the formula The Kalman gain at time t; The third step is to update the state and obtain... , In the formula Let be the weight vector updated at time t. Let I be the filter error covariance matrix at time t, and let I be the identity matrix. The fourth step is to reconstruct the updated weight vector into a dynamically adjusted mapping matrix, and then input the parameter vector into this matrix to obtain the final production condition demand vector.

[0055] It should be noted that the preset thresholds for equipment health index and fabric batch offset can be flexibly adjusted according to the equipment performance and product quality requirements of the production line; state transition matrix and observation matrix The initial value can be set based on historical operating condition change data and parameter mapping relationships, and can be continuously optimized through an outer loop learning mechanism during system operation; the process noise covariance matrix of the Kalman filter algorithm and observation noise covariance matrix It can dynamically adjust according to the noise level of real-time data to ensure the stability and speed of weight updates.

[0056] For example, suppose the design parameters of a new garment are as follows: Size parameters: shoulder width 42cm, length 75cm, bust 100cm, waist 82cm; Process requirements parameters: topstitching stitch density 12 stitches / cm, overlock width 0.5cm, special seam type: flat seam, medium process complexity; Material characteristics parameters: fabric thickness 0.8mm, elastic recovery rate 85%, good fabric breathability, weight 220g / m². After normalization, the parameter vector is [0.5, 1.0, 1.2, 0.8, 0.9, 0.5, -0.3, 0.7, 1.1, 0.6]. For shoulder width, garment length, chest circumference, waist circumference, stitch density, overlock width, fabric thickness, elastic recovery rate, process complexity, and quality grade, input the basic mapping matrix and obtain the initial production condition demand vector through matrix multiplication. The optimal process parameters include sewing machine speed of 1850 rpm, ironing temperature of 165℃, sewing tension of 4.2N, and stitch length of 2.8mm. The equipment configuration consists of 3 high-speed sewing machines, 2 steam ironing machines, and 1 automatic cutting bed. The time allocation is as follows: cutting process 0.8 hours, sewing process 3.2 hours, ironing process 0.5 hours, and quality inspection and packaging process 0.3 hours.

[0057] Real-time data streams are obtained from the working condition perception module: In the equipment state vector, the real-time value of the rotational speed of a high-speed sewing machine is 2300 rpm (standard value 2500 rpm, maximum allowable value 2800 rpm, minimum allowable value 2200 rpm), and the real-time temperature is 65℃ (standard value 60℃, maximum allowable value 80℃, minimum allowable value 40℃), with weights w1=0.7 and w2=0.3; In the fabric characteristic vector, the actual fabric elasticity recovery rate is 80% (design preset value 85%) and the thickness is 0.9mm (design preset value 0.8mm), with weights v1=0.6 and v2=0.4; In the environment vector, the workshop temperature is 25℃ and the humidity is 50%. The calculated equipment health index HI = 1 - [0.7 × |2300-2500| / (2800-2200) + 0.3 × |65-60| / (80-40)] = 1 - [0.7 × 200 / 600 + 0.3 × 5 / 40] ≈ 1 - [0.233 + 0.0375] = 0.729, which is less than the preset threshold of 0.8; the calculated fabric batch offset BO = 0.6 × |80%-85%| / 85% + 0.4 × |0.9-0.8| / 0.8 ≈ 0.6 × 0.0588 + 0.4 × 0.125 = 0.035 + 0.05 = 0.085, which is less than the preset threshold of 0.1. Based on these calculations, the overall condition is deemed abnormal, and the dynamic adjustment layer is activated. Then set the state transition matrix. process noise covariance matrix Observation matrix The observation noise covariance matrix is ​​composed of the outer product of current operating condition characteristics (equipment health index, fabric batch offset) and design parameter vectors. =0.005I. The Kalman filter algorithm is applied to update the weights of the mapping matrix, and the parameter vector is input back into the matrix to obtain the final production condition demand vector: sewing machine speed is adjusted to 2150 rpm (to compensate for overall efficiency), ironing temperature is adjusted to 160℃, and sewing tension is simultaneously adjusted to 4.8N; regarding equipment configuration, the scheduling system temporarily adds one backup machine to the existing three high-speed sewing machines to cope with potential efficiency fluctuations caused by main equipment in poor condition, while the number of steam ironing machines remains at two; regarding time allocation, due to process adjustments and equipment status, the time is recalculated as follows: cutting process 0.9 hours, sewing process 3.5 hours, ironing process 0.7 hours, and quality inspection and transfer process 0.4 hours, ensuring that production quality and efficiency requirements can still be met even under fluctuating equipment conditions.

[0058] Based on the above technical solution, the parameter mapping module ensures the accuracy of the mapping by relying on historical data and process knowledge through the basic mapping layer. The dynamic adjustment layer realizes the dynamic optimization of weights based on real-time working condition data and Kalman filtering algorithm. The standardized clothing design parameters are accurately converted into production working condition demand vectors containing optimal process parameters, equipment configuration and working time allocation through the dynamic mapping matrix. This effectively overcomes the problem of inaccurate process parameters caused by factors such as changes in equipment performance, differences in material batches, and environmental fluctuations. It provides reliable working condition demand data that fits the actual production scenario for the subsequent intelligent scheduling module.

[0059] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above-mentioned intelligent scheduling module can be implemented through the following steps S201, S202, and S203, which are explained in detail below: S201. Based on the association rule mining algorithm, process-equipment-personnel combination patterns are mined from historical scheduling data to construct a scheduling rule set, providing prior knowledge support for the generation of candidate scheduling schemes.

[0060] Among them, the association rule mining algorithm is used to extract scheduling relationships from massive historical scheduling records, transforming scattered scheduling experience into reusable rules, reducing the blindness in the generation of candidate scheduling schemes, and improving scheduling efficiency and rationality.

[0061] In some implementations, the construction of the scheduling rule set includes the following steps: The first step is transaction transformation. Each historical scheduling record is transformed into a transaction. Each transaction contains a set of work processes, equipment allocation information, personnel allocation information, and production result indicators (such as product qualification rate, production efficiency, etc.). For example, a transaction can be represented as "Work processes: cutting - sewing - ironing; Equipment allocation: cutting machine A - high-speed sewing machine B - steam ironing machine C; Personnel allocation: worker A - worker B - worker C; Production result: qualification rate 98%, production efficiency 85%"; The second step is to construct a weighted transaction database. This involves assigning a comprehensive weight to each transaction. The calculation formula is This is used to balance the time efficiency and quality reliability of transactions; in, This represents the j-th transaction, where j is the transaction index; β is the weight balancing coefficient, which ranges from [0,1] and can be adjusted according to production needs, typically set to 0.4~0.6. Let be the time decay function, and the calculation formula is: This is used to reduce the weight of historical events. For the current time, λ is the time of the transaction, and λ is the decay coefficient, which ranges from 0.001 to 0.01. The larger the λ is, the faster the time decays. The quality weighting function is based on production result indicators. ,in Let the product qualification rate be the product qualification rate of the j-th transaction. Let α1 and α2 be the production efficiency of the j-th transaction, and both are normalized to the interval [0,1]. α1 and α2 are quality weight coefficients, and their sum is 1. They can be adjusted according to production priority. For example, when quality is the priority, α1=0.6 and α2=0.4. The third step is to use the improved Apriori algorithm to mine frequent itemsets.

[0062] First, scan the weighted transaction dataset and calculate the weighted support for all 1-itemsets. The formula for calculating the weighted support is: Where X represents an itemset, and WS(X) represents the weighted support of itemset X; this formula is used to measure the frequency of occurrence of itemset X in a weighted transaction database. Frequent 1-itemsets L1 are selected based on a preset support threshold (usually set to 0.1~0.3); Subsequently, candidate k-itemsets are generated iteratively through join and pruning steps. The weighted support of each candidate k-itemset is calculated, and frequent k-itemsets are selected. This continues until no new frequent itemsets can be generated; the join step refers to resolving the issue in the Apriori algorithm by connecting frequent itemsets. The first step is to connect the itemset to itself to generate a candidate k-item set Ck, ensuring that the first k-2 items are the same and connected in lexicographical order. The pruning step ensures that all non-empty subsets of any frequent itemset are also frequent. If a candidate's non-empty subset is not frequent, then the candidate is definitely not frequent and can be removed from the candidate set.

[0063] The fourth step is rule generation and filtering.

[0064] For each frequent itemset Generate all its non-empty proper subsets s⊂l, and compute the rule. Weighted confidence level The calculation formula is This is used to measure the reliability of the rules; Rules with a weighted confidence level exceeding a preset confidence threshold (usually set to 0.7~0.9) are retained to form the first rule set; Then, lift and confidence are calculated based on weighted confidence and weighted support; where the lift is calculated using the formula: Used to measure rules The correlation strength is calculated as follows: Lift > 1 indicates a positive correlation between s and (ls), and the larger the Lift, the stronger the correlation. The formula for calculating confidence level is... This is used to further verify the validity of the rules; Rules with lift exceeding a preset lift threshold (usually set to 1.2~1.5) and confidence exceeding a preset confidence threshold (usually set to 0.6~0.8) are retained to form a scheduling rule set.

[0065] For example, the scheduling rule set may be as follows: Rule ID Rule antecedent (condition / IF) Rule successor (scheduled action / THEN) Weighted confidence level Weighted support R001 Pp [Sewing Tension] > 5.0 Np [Fabric Thickness] > 1.0 mm R [Sewing Machine Spindle Speed] < 90% of Rated R [Sewing Machine Body Temperature] > 70℃ Task: Main body reinforcement sewing equipment: Heavy-duty sewing machine-01 Personnel: Operator with skill level ≥5 Start time: First sewing process 0.94 0.12 R002 Pp[Sewing Speed]>3000 rpm Pp[Order Priority]="Urgent" R[Equipment Load Rate]<65% R[High-Speed ​​Sewing Machine - Current Pressure]Normal Duration: Standard working hours × 1.3 Task: Regular sewing Equipment: High-speed sewing machine - 02 / 03 Personnel: Operators with skill level ≥ 3 Start Time: Immediately insert order when available Duration: Estimated working hours × 0.85 0.87 0.09 R003 Fabric elasticity > 75%; Stitch pitch < 2.5mm; Presser foot pressure on coverstitch machine adjustable; Workshop humidity between 40% and 60%. Task: Hem Side Sewing Equipment: Automatic Side Sewing Machine-01 Personnel: Specialized staff with skill level ≥ 4 Start Time: After flat sewing and before ironing Duration: Standard working hours × 1.15 0.96 0.10 R004 Pp[Ironing Temperature]>180℃ Pp[Fabric Composition]Contains "Wool" R[Steam Iron - Steam Pressure]<Standard Value R[Ironing Machine - Table Temperature]Unstable Task: Finished product shaping and ironing. Equipment: Allocation of spare steam irons. Personnel: Experienced ironing workers. Start time: After all sewing is completed. Duration: Standard working hours × 1.4 (including equipment preheating). 0.98 0.07 R005 Pp[Cutting Speed]="High Speed" Pp[Process Complexity]High R[Automatic Cutting Machine - Tool Pressure]Abnormal Fluctuation R[Cutting Machine - Vibration Sensor]>Threshold Task: Precision cutting equipment: Switch to standby cutting bed or reduce speed to "precision cutting mode"; Personnel: Cutters with skill level ≥ 5; Start time: Prioritize scheduling after materials are ready; Duration: Standard working hours × 1.5 (including equipment adjustment and quality inspection). 0.93 0.05 When the intelligent scheduling module receives the production condition demand vector and equipment status vector from the parameter mapping module, it matches them with the antecedent of each rule in the rule base. The system then generates a specific scheduling instruction based on the matched consequent. The scheduling fragments of multiple triggered rules are intelligently combined to form a coherent initial process flow, serving as a high-quality initial solution for the multi-objective optimization algorithm, thereby greatly improving the efficiency and quality of scheduling scheme generation.

[0066] S202. Taking production demand as input and combining it with the scheduling rule set, candidate scheduling schemes are generated through multi-objective optimization algorithms, and then simulated and verified using a digital twin model.

[0067] Among them, the production condition requirements come from the optimal process parameters, equipment configuration and time allocation information output by the parameter mapping module; the scheduling rule set provides constraints and guidance for the construction of the initial scheduling scheme; the multi-objective optimization algorithm is used to balance multiple conflicting objectives such as production efficiency, cost, equipment load and quality risk, and generate candidate schemes that take into account multi-dimensional requirements; the digital twin model is used to simulate the actual operation effect of the candidate schemes, provide feedback to the optimization algorithm, and improve the feasibility and robustness of the candidate schemes.

[0068] In some implementations, generating candidate scheduling schemes and performing simulations using digital twin models includes the following steps: Step 1: Construct the initial scheduling scheme.

[0069] Receive the production condition demand vector output by the parameter mapping module and the device state vector SR, As a rule antecedent, it matches rules that meet the conditions from the scheduling rule set; the rule consequent is a scheduling fragment. This indicates that a task is assigned to a device. and personnel ,from Start execution, duration is The equipment state vector is a vector composed of the equipment's rotational speed, temperature, and pressure operating parameters. By combining the consequents of multiple matching rules and incorporating the time allocation constraints in production conditions, one or more initial scheduling schemes are constructed as the initial population for a multi-objective optimization algorithm.

[0070] Step 2: Define the multi-objective function.

[0071] Define the multi-objective function vector of the scheduling problem The four sub-functions each correspond to a different optimization objective: To maximize the completion time, the objective is to minimize it, and the calculation formula is: Tasks is the set of all production tasks. Let i be the completion time of task i; The total production cost is [value], and the objective is to minimize it. The calculation formula is: Devices is a collection of devices. Let j be the unit time usage cost of device j. Let j be the actual running time of device j, and Workers be the set of personnel. Let k be the unit time labor cost. For the actual working time of person k, E represents the unit energy cost, and E represents the total energy consumption of the production process. The standard deviation of equipment load balance is defined as follows, and the objective is to minimize it (i.e., to achieve a more balanced equipment load). The calculation formula is: Where |Devices| represents the total number of devices. The average uptime of all devices; To predict quality risk, the objective is to minimize it, and the calculation formula is: ;in The critical weight of task i, Let be the expected quality pass rate for task i.

[0072] The third step involves optimizing the algorithm using the improved Non-dominated Sorting Genetic Algorithm III (NSGA-III) and verifying it through simulation with digital twins.

[0073] First, set the algorithm parameters: population size. =50~200, maximum number of iterations =50~100; In the g-th iteration of the algorithm (g ranges from 1 to...), ), and the candidate solutions generated for the current population. Then input it into the digital twin model for simulation verification; Digital twin models need to be built based on the actual layout of the production line, equipment parameters, and process flow, focusing on simulating random factors such as equipment failure and material supply fluctuations. =100~500 Monte Carlo simulations; The simulation output shows the estimated distribution of each objective function in the multi-objective function vector, i.e. (i=1,2,3,4); where For the objective function The expected value, i.e., the expected performance. For the objective function The standard deviation of the value is used to reflect the degree of performance fluctuation. The expected performance of each candidate solution is calculated based on the estimated distribution. and robustness ,Will As the core target value, As an additional selection criterion, it is used for non-dominated sorting and selection operations, that is, to give priority to candidate schemes with better expected performance and stronger robustness to enter the next generation of population. Genetic operations such as selection, crossover, and mutation are used to generate a new generation of population. This process is repeated iteratively until the maximum number of iterations is reached, ultimately yielding a set of Pareto optimal solutions. M is the number of Pareto optimal solutions.

[0074] It should be noted that if no matching rule can be found during the initial scheduling scheme construction process, a random allocation strategy can be used to generate an initial scheme and include it in the initial population; the weights of the multi-objective function can be dynamically adjusted according to production needs.

[0075] S203. Perform high-fidelity simulation and decision analysis on the Pareto optimal solution set to select the optimal scheduling scheme with the highest comprehensive approximation.

[0076] The Pareto optimal solution set contains multiple optimization schemes that cannot dominate each other (i.e., no single scheme is superior to all other schemes in all objectives). High-fidelity simulation is needed to improve the accuracy of the objective values, and then decision analysis methods are used to quantify the comprehensive performance of each scheme to finally determine the optimal scheme.

[0077] In some implementations, the steps for generating the optimal scheduling scheme include: First, the Pareto optimal solution set For each solution in the simulation, input it into the digital twin model for high-fidelity simulation. Set the number of simulations to 1000~5000. Compared with the simulation in S202, the high-fidelity simulation needs to refine parameters such as equipment operation model, material flow details, and probability of disturbance occurrence to improve the accuracy of target value calculation. The precise objective value of each Pareto solution is obtained through high-fidelity simulation, and a decision matrix is ​​constructed. In the formula Let be the exact objective function vector of the i-th Pareto solution.

[0078] Then, the decision matrix X is normalized using the Euclidean distance standardization method, and a weight vector is introduced. Normalize each element in the decision matrix X. The weight vector is determined by production demand, and Σwj=1 (j=1 to 4). For example, in the quality-first scenario, set w1=0.2, w2=0.2, w3=0.1, w4=0.5, and in the efficiency-first scenario, set w1=0.5, w2=0.2, w3=0.1, w4=0.2.

[0079] Next, the ideal solution and the negative ideal solution are determined. Among them, the ideal solution A... + It is the set of optimal values ​​for each objective, calculated as follows: ( (The element in the i-th row and j-th column after weighted normalization), that is, the ideal solution is the minimum value of each column, and the negative ideal solution A. - It is the set of worst values ​​for each objective, calculated as follows: In the formula, i = 1 to M; the ideal solution represents the theoretically optimal performance, and the negative ideal solution represents the theoretically worst performance.

[0080] Then, calculate the distance from the solution to the ideal solution. and distance to the negative ideal solution The calculation formula is , ;in To find the ideal solution for the j-th element, The j-th element is the negative ideal solution; Calculate the overall closeness The calculation formula is In the formula The value of is in the range [0,1]. The closer the value is to 1, the closer the solution is to the ideal solution and the better the overall performance.

[0081] Finally, the scheme with the highest overall proximity C_i is selected as the optimal scheduling scheme for the final output. The instructions are then sent to the execution control module to guide production.

[0082] Based on the above technical solutions, the intelligent scheduling module extracts historical scheduling experience through association rule mining and constructs a set of scheduling rules with prior guidance significance; through multi-objective optimization algorithms and digital twin simulation, it generates a Pareto optimal solution set that takes into account efficiency, cost, load balance, and quality risk; and through the TOPSIS decision analysis method, it quantifies the comprehensive performance of the solution and obtains the optimal scheduling solution that fits the actual production needs. This achieves the output of an intelligent scheduling solution that is balanced by multiple objectives, highly feasible, and robust, solving the technical problems of traditional scheduling relying on human experience, difficulty in reconciling multi-objective conflicts, and weak resistance to production disturbances.

[0083] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 4 As shown, the adaptive learning module described above can be implemented through the following steps S301, S302, and S303, which are explained in detail below: S301. Activate the inner-loop learning mechanism to perform rapid online optimization based on real-time production feedback, cope with instantaneous fluctuations in the production line, and simultaneously store state transfer experience to the experience buffer pool.

[0084] The inner-loop learning mechanism operates on a short timescale of minutes or hours to respond instantly to production disturbances such as equipment status fluctuations, changes in material characteristics, and adjustments to order priorities. Through online parameter updates and fine-tuning of decision-making strategies, it maintains the stability and efficiency of the production process.

[0085] In some implementation methods, the specific operation process is as follows: First, for the dynamic adjustment layer in the parameter mapping module, recursive least squares (RLS) is used to update the mapping weights in real time, with the following formula: ,in Let k be the updated mapping weight vector. Let P(k) be the weight vector at time k-1, P(k) be the covariance matrix at time k, x(k) be the input parameter vector at time k (i.e., the standardized apparel design parameter vector), and y(k) be the actual production feedback parameter vector at time k (i.e., the actual process parameters collected by the working condition sensing module). Let x(k) be the transpose of x(k); The formula for updating the covariance matrix is: ,in This is the forgetting factor, with a value ranging from 0.95 to 0.99.

[0086] Next, the discrete choices in the scheduling strategy, such as task allocation, equipment selection, and process parameter adjustment, are modeled as a Contextual Multi-Armed Bandit (CMAB) problem. Here, "arm" corresponds to a specific decision action, such as "assigning the sewing task to equipment A," "adjusting the ironing temperature to 115℃," or "prioritizing order B." "Context" corresponds to the current production line state, encompassing comprehensive information such as equipment state vectors (speed, temperature, health index), work-in-process quantity, order priority, and current process parameter settings. Based on reinforcement learning theory, we first define expected reward and actual reward: Expected reward This represents the estimated expected cumulative discounted reward calculated based on the current value function V(s) or historical average return after choosing action a in state s. The formula is as follows: Where r(s,a) is the immediate reward after performing action a, such as the value of increased production efficiency or reduced cost; is the discount factor, ranging from 0.9 to 0.99; s' is the next state after performing action a; V(s') is the value function of the next state; the actual reward R(a|s) represents the sum of the immediate reward actually observed after performing the action and the estimated future discounted reward, calculated as follows: ,That An immediate reward for actual observation; The estimated value function for the next state; The Thompson sampling strategy is used for action selection. At each decision, random samples are taken from the posterior reward distribution of each action, and the action with the largest sample value is selected for execution. After execution, the cumulative value of the success reward and the cumulative value of the failure penalty are updated according to the actual reward, and the action selection strategy is continuously optimized.

[0087] Simultaneously, an experience buffer pool is established to store the state transition experience tuples (s, a, r, s') during the inner loop learning process, providing incremental data support for the outer loop learning; where s is the current state, a is the selected action, r is the actual reward, and s' is the next state.

[0088] S302. Initiate the outer loop learning mechanism to reconstruct the model and extract knowledge based on batch historical data, thereby optimizing the long-term performance of the system.

[0089] In some implementations, the specific operation flow of step S302 is as follows: First, a dataset is constructed. Production data over a complete cycle is collected to form a batch dataset. The data includes state transition experience tuples (s, a, r, s') from the experience buffer pool, production execution data (completion time of each process, cost consumption, and quality pass rate), equipment full-cycle operation data (cumulative runtime, fault records, and performance degradation data), and material characteristic data (actual parameters of each batch of fabric), ensuring that the dataset covers key information throughout the entire production process.

[0090] Then, the core model of the system is retrained or fine-tuned using a batch dataset: for the basic mapping matrix of the parameter mapping module, the static mapping relationship is updated by retraining using a combination of multivariate nonlinear regression and ensemble learning, consistent with S101. For the rule set of the intelligent scheduling module, new association rules are extracted by statistically analyzing high-return action combinations in batch data, such as "when the equipment health index is below 0.8 and the fabric elasticity offset is greater than 0.1, it is preferentially assigned to equipment C". For digital twin models, adjusting equipment performance parameters (such as equipment failure rate and operating efficiency decay coefficient) and material flow parameters (such as fabric processing time fluctuation range) can improve the prediction accuracy of the simulation model.

[0091] Next, for the action selection strategy based on reinforcement learning generated by inner-loop learning, it is transferred to a simplified model through knowledge distillation. Knowledge distillation uses the inner-loop reinforcement learning strategy model as the teacher model and the simplified interpretable model as the student model. By minimizing the difference in output distribution between the teacher model and the student model, an interpretable student model that closely approximates the decision performance of the teacher model is obtained. The student model is selected from interpretable mathematical models, such as decision trees, linear regression models, or rule sets. During student model training, KL divergence (KLD) is used as the loss function, and the formula is as follows: ,in Let be the probability distribution of the teacher model choosing action a in state s. Let A be the probability distribution of the student model choosing action a in state s, and let A be the action space. It is a very small positive number, usually taken as 1e-8, to avoid the denominator being 0. During training, the state s in the batch data is used as input, and the action probability distribution of the teacher model is used as the supervision signal. The KL divergence loss function is minimized so that the student model can approximate the decision performance of the teacher model.

[0092] Next, new scheduling rules are extracted from the trained student model. For example, the rule "If the device health index is <0.8 and the order priority is high, then the action is assigned to device C" is extracted from the decision tree model and added to the rule set of the intelligent scheduling module to achieve knowledge reuse.

[0093] S303. Achieve collaborative interaction between the inner and outer loops, synchronize optimized model parameters and rules, and output the optimal scheduling strategy for continuous optimization.

[0094] Among them, collaborative interaction is achieved through a shared experience buffer pool and model parameter interface, enabling bidirectional flow of short-term adjustment experience in the inner loop and long-term optimization knowledge in the outer loop, ensuring that the immediate response capability of the inner loop and the long-term evolution capability of the outer loop complement each other; the optimal scheduling strategy is the final execution strategy after integrating the initial scheduling scheme, real-time adjustment actions of the inner loop, and optimization rules of the outer loop, which can simultaneously cope with instantaneous fluctuations and long-term trend changes, and directly guide the operation of the production line execution control module.

[0095] In some implementations, the specific operation flow of step S303 is as follows: Step 1: Two-way data synchronization. Every hour, the inner loop marks newly added state transition experiences in the experience buffer pool as "pending synchronization," while the outer loop periodically samples data from the experience buffer pool for batch training. After each core model update and rule refinement, the outer loop synchronizes the updated basic mapping matrix parameters, digital twin model parameters, and newly added rules to the inner loop through the model parameter interface, overwriting the original initial parameters and rules of the inner loop and providing a better initial basis for the real-time adjustment of the inner loop.

[0096] Step 2: Scheduling Strategy Integration. The optimal scheduling scheme output by the intelligent scheduling module is used as the basic strategy. This is combined with real-time adjustments from the inner loop (such as fine-tuning equipment parameters and temporary adjustments to task allocation) and newly added optimization rules from the outer loop to form the final optimal scheduling strategy. The integration logic is as follows: the basic strategy clarifies the framework for equipment, personnel, and time allocation for each process; the newly added rules from the outer loop act as constraints, limiting the feasible scope of the strategy, such as "high-priority tasks should not be assigned when the equipment health index is <0.8"; the real-time adjustments from the inner loop act as dynamic correction items, fine-tuning the basic strategy based on the current production status, such as "if the speed of equipment A decreases by 5%, extend its task duration by 10%".

[0097] Step 3: Strategy Output and Iteration. The integrated optimal scheduling strategy is distributed to the execution control module to guide actual production. At the same time, production feedback data from the execution control module is continuously input into the inner loop, triggering online optimization of the inner loop. New state transition experiences are stored in the experience buffer pool, providing data for the next batch optimization of the outer loop, forming a continuous iterative closed loop of "inner loop optimization → experience storage → outer loop optimization → parameter synchronization → inner loop re-optimization".

[0098] It should be noted that rule conflicts need to be handled when merging scheduling strategies. For example, when the basic strategy conflicts with newly added rules in the outer loop, the outer loop rules take precedence; when the real-time adjustment actions of the inner loop conflict with the rules of the outer loop, the inner loop actions take precedence. The optimal scheduling strategy needs to be displayed to users through a visual interactive module, including the basic scheme, adjustment items, rule constraints, etc., and supports manual intervention by users.

[0099] Based on the above technical solutions, the adaptive learning module constructs a hierarchical and collaborative continuous optimization framework through a process of rapid online tuning in the inner loop, deep batch optimization in the outer loop, and dual-loop collaborative iteration. Inner-loop learning, based on recursive least squares and reinforcement learning, achieves minute-level instantaneous disturbance response, ensuring the stability of the production process. Outer-loop learning, through batch data training and knowledge distillation, completes the long-term evolution of the core model and rule base, improving the system's versatility and maintainability. The dual-loop collaborative mechanism organically combines short-term adaptability with long-term optimization capabilities, enabling the system to continuously accumulate experience from production practice and continuously optimize scheduling strategies. This module effectively solves the problems of static rigidity and lack of self-learning ability in traditional production management systems, giving the entire garment production line process parameter intelligent scheduling management system the core capabilities of self-improvement and continuous evolution, enhancing the system's adaptability to dynamic production environments and long-term operational performance.

[0100] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0101] Working principle of the invention: First, the parameter mapping module intelligently converts garment parameters from the design system into specific production condition requirements. This module is based on a fundamental mapping relationship trained on historical data and dynamically adjusted in conjunction with real-time collected equipment status and fabric characteristic data to ensure that the conversion results are both accurate and adaptable to changes in the production environment. Subsequently, the intelligent scheduling module receives these production condition requirements and, based on scheduling rules mined from historical data... The system generates an initial scheduling scheme. It employs a multi-objective optimization algorithm to search for optimal solutions under multiple objectives such as efficiency, cost, and quality. A digital twin model is used to perform high-fidelity simulations of candidate schemes, previewing their performance in real production. This allows for the selection of the final scheduling scheme with the best overall performance and robustness, which is then deployed to the production line for execution. Throughout the production process, the adaptive learning module continuously operates. Its inner-loop learning mechanism is based on real-time production feedback at the minute / hour level, rapidly fine-tuning process parameters and scheduling strategies. Its outer-loop learning mechanism is based on historical data accumulated at the day / week level, retraining the core model and refining complex online decision-making knowledge into interpretable new rules, enabling iterative refinement of the system strategy. All modules are interconnected through a unified data bus, forming an intelligent production management system capable of continuous perception, dynamic optimization, and self-evolution. Ultimately, this achieves end-to-end intelligence and flexibility in the garment production line from design to execution, effectively solving problems such as the disconnect between design and production, reliance on manual experience, and weak anti-interference capabilities in traditional production, thereby improving production efficiency and quality stability.

[0102] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A garment production line process parameter intelligent scheduling management system, characterized in that, include: The parameter mapping module is used to convert clothing design parameters into production condition requirements through a mapping matrix. The clothing design parameters include size parameters, process requirement parameters, and material characteristic parameters. The production condition requirements include optimal process parameters, equipment configuration, and time allocation. The intelligent scheduling module is used to generate candidate scheduling schemes based on the production conditions and the association rule mining algorithm and the multi-objective optimization algorithm, and to select the optimal scheduling scheme based on the digital twin simulation results. An adaptive learning module is used to optimize the optimal scheduling scheme and digital twin model in real time through inner-loop and outer-loop learning mechanisms to obtain the optimal scheduling strategy; the optimal scheduling strategy is used to guide the production of the garment production line.

2. The intelligent scheduling management system for process parameters of a garment production line according to claim 1, characterized in that, The system also includes a parameter input module, a working condition sensing module, a visualization interaction module, a parameter mapping module, and a data bus module; among which, The parameter input module is used to receive clothing design parameters from the clothing design system; The working condition sensing module is distributed at each node of the production line and is used to collect equipment status, garment fabric characteristics and environmental data; the equipment status includes equipment speed, temperature and pressure operating parameters, which are used to reflect the real-time working performance and health status of the equipment; the garment fabric characteristics represent physical parameters such as fabric thickness, elasticity and breathability. The visualization interaction module is connected to each module and is used to provide a user interface and real-time monitoring display; The parameter mapping module is connected to the execution control module and the adaptive learning module, and is used to collect production data and update system parameters; The data bus module adopts the industrial Ethernet protocol to realize data exchange and synchronization between modules. The modules are interconnected through the data bus module. The design parameters received by the parameter input module are processed by the parameter mapping module, and then the intelligent scheduling module generates a scheduling scheme. After being verified by the digital twin model, the scheme is sent to the execution control module, and then a closed-loop optimization is formed through the adaptive learning module and the parameter mapping module.

3. The intelligent scheduling management system for process parameters of a garment production line according to claim 2, characterized in that, The workflow of the parameter mapping module includes: A dynamic mapping matrix is ​​constructed. The mapping matrix is ​​trained using historical production data and machine learning methods to map clothing design parameters to production working condition requirements. The mapping matrix includes a basic mapping layer and a dynamic adjustment layer. The basic mapping layer is trained based on historical data to obtain a static mapping relationship, and the dynamic adjustment layer adjusts the mapping weights according to real-time working condition data. Receive clothing design parameters from the parameter input module and standardize the clothing design parameters into vector form to obtain the parameter vector; The parameter vector is mapped to a production condition demand vector using the dynamic mapping matrix. The dynamic adjustment layer is activated based on real-time data collected by the working condition sensing module to adapt to changes in equipment status and fabric characteristics.

4. The intelligent scheduling and management system for process parameters of a garment production line according to claim 3, characterized in that, The basic mapping layer obtains static mapping relationships based on historical data, including: Build a dataset by acquiring historical successful order data ;in, Let i be the design parameter vector for the i-th order. Let i be the actual production demand vector for the i-th order. For the production result indicators of the i-th order, the production result indicators include product qualification rate and production efficiency; The design parameter vector is extended with features to calculate the process complexity features; the process complexity features include cutting complexity, sewing complexity, and ironing complexity. The relationship between design parameters and process parameters is modeled as a multivariate nonlinear mapping, and trained using a combination of rule-constrained multivariate nonlinear regression and ensemble learning; the objective function of the multivariate nonlinear regression is: , Let W be the mapping function, and W be the weights of the mapping matrix to be trained. For regularization terms, The regularization coefficient is used. During training, a process knowledge graph is introduced for constraints; where nodes in the process knowledge graph are process parameters, and edges represent physical or logical constraints between process parameters. The dataset is divided into training and validation sets in chronological order, and the validated mapping relationships are solidified into a basic mapping matrix to obtain static mapping relationships.

5. The intelligent scheduling and management system for process parameters of a garment production line according to claim 4, characterized in that, The dynamic adjustment layer adjusts the mapping weights based on real-time operating data, including: Data streams are obtained from the working condition sensing module, and the data streams include equipment state vectors, fabric characteristic vectors, and environment vectors. The operating condition indicators are calculated based on the data stream, and when the value of the operating condition indicator is greater than a preset threshold, it is determined that the operating condition is abnormal; the operating condition indicators include the equipment health index and the fabric batch offset. The weight adjustment of the basic mapping matrix is ​​treated as a state estimation problem, thereby establishing a state-space model; the state-space model includes state equations. and observation equations , Let be the vectorized form of the fundamental mapping matrix at time t. The observation matrix is ​​composed of the outer product of the current operating condition characteristics and the design parameters. Here is the state transition matrix. For process noise, To observe noise; Based on the state-space model, the Kalman filter recursive algorithm is applied to update the estimated value of the mapping matrix online, resulting in a dynamically adjusted mapping matrix.

6. The intelligent scheduling and management system for process parameters of a garment production line according to claim 1, characterized in that, The step of generating candidate scheduling schemes based on the production condition requirements, using association rule mining algorithms and multi-objective optimization algorithms, includes: The association rule mining algorithm is used to extract process-equipment-personnel combination patterns from the historical scheduling database to form a rule set; When a new production condition demand vector is received When using the device state vector SR, As a rule antecedent, rules that meet the conditions are matched from the rule set to obtain the rule consequent; wherein, the rule consequent is a scheduling fragment in the rule set, and the scheduling fragment is... This indicates that the task will be completed. Assigned to device and personnel From the start time Start, duration is ; By combining the consequents of multiple matching rules, one or more initial scheduling schemes are constructed as the initial population for the multi-objective optimization algorithm. Define a multi-objective function vector for the scheduling problem, including the maximum completion time objective function, the total production cost function, the equipment load balance standard deviation function, and the estimated quality risk function; An improved NSGA-III algorithm is used to optimize the multi-objective function. In each iteration of the algorithm, the generated candidate solutions are input into a digital twin model for simulation verification. The digital twin model outputs the estimated distribution of each objective function in the multi-objective function vector based on equipment failure and material supply fluctuation factors. Calculate the expected performance of each distribution based on the estimated distribution. and robustness For non-dominated sorting and selection; wherein, the expected performance Let i be the expected value of the objective function. Let i be the standard deviation of the objective function. After the algorithm is completed, a set of Pareto optimal solutions is obtained, and the comprehensive proximity of each solution in the solution set is calculated using the TOPSIS decision analysis method based on the superior-inferior solution distance method. The solution with the highest overall approximation is selected as the final production scheduling solution.

7. The intelligent scheduling and management system for process parameters of a garment production line according to claim 6, characterized in that, The workflow logic of the association rule mining algorithm includes: Each historical scheduling record is transformed into a transaction, which includes a set of processes, equipment allocation, personnel allocation, and production results. Each transaction is assigned a comprehensive weight to obtain a weighted transaction database; the formula for calculating the comprehensive weight is: ;in, It is a time decay function. The time when the transaction occurred. The attenuation coefficient is... For the current time, Here, j is the quality weighting function based on the production result indicators. This represents the j-th transaction; An improved association rule mining algorithm, Apriori, is used to scan a weighted transaction dataset, calculate the weighted support of all 1-itemsets, and filter out frequent 1-itemsets based on a preset support threshold. ; Candidate k-itemsets are generated iteratively through join and pruning steps, and the weighted support of the candidate k-itemsets is calculated to generate frequent k-itemsets. until no new frequent itemsets can be generated; For each frequent itemset Generate all non-empty proper subsets of the frequent itemset. and calculate the rules The weighted confidence scores are calculated; only rules with a weighted confidence score exceeding a preset confidence threshold are retained to obtain the first rule set. Based on the weighted confidence and weighted support, the lift and confidence are calculated. Rules whose lift and confidence exceed the corresponding preset thresholds are retained from the first rule set to obtain the final rule set.

8. The intelligent scheduling and management system for process parameters of a garment production line according to claim 1, characterized in that, The inner and outer loop learning mechanisms represent a hierarchical and collaborative continuous optimization framework, wherein... The inner-loop learning mechanism operates on a timescale of minutes / hours, and performs parameter optimization and strategy adjustment based on real-time production feedback to cope with instantaneous fluctuations in the production line. The outer loop learning mechanism operates on a daily / weekly timescale, and performs model reconstruction and knowledge extraction based on accumulated historical data batches to optimize the long-term performance of the system. Furthermore, the inner loop and the outer loop interact bidirectionally with the model parameter interface through a shared experience buffer pool. That is, the state transition experience of the inner loop provides incremental data to the outer loop, and the outer loop periodically samples data from the experience buffer pool, trains a new global model, and synchronizes the updated model parameters to the inner loop, providing the inner loop with a better initial strategy.

9. The intelligent scheduling and management system for process parameters of a garment production line according to claim 8, characterized in that, The workflow of the inner-loop learning mechanism includes: For the dynamic adjustment layer in the parameter mapping module, the mapping weights are updated online using the recursive least squares method; The discrete choices in the scheduling strategy are modeled as a contextual multi-armed slot machine problem; wherein, the discrete choices represent task allocation, equipment selection, and process parameter adjustment; each "arm" of the contextual multi-armed slot machine represents a decision action, including allocating a task to a specific piece of equipment, selecting a specific combination of process parameters, and adjusting the production sequence; the context represents the current production line status, including equipment status, work-in-process quantity, order priority, and current process parameter settings; Based on reinforcement learning theory, according to the current production line state and the expected reward of each decision action, a Thompson sampling strategy is used to sample from the posterior reward distribution of each decision action, select the action with the largest sample value to execute, and update the posterior reward distribution of the decision action according to the actual report, thereby learning the action selection strategy that maximizes the cumulative reward under different states; wherein, the expected reward represents the estimated value of the expected cumulative discount reward calculated based on the current value function or historical average reward after selecting the decision action, the actual reward represents the estimated value of the immediate reward observed after executing the decision action plus the future discount reward, and the posterior reward distribution represents the probability distribution of the expected reward of each decision action; Establish an experience buffer pool to store the state transition experience of the inner loop during the reinforcement learning process.

10. The intelligent scheduling and management system for process parameters of a garment production line according to claim 9, characterized in that, The workflow of the outer loop learning mechanism includes: Collect production data over a complete cycle to form a batch dataset; the production data includes state transition experience in the experience buffer pool. The batch dataset is used to retrain or fine-tune the basic mapping matrix, rule set, and digital twin model parameters in the system; For the scheduling strategy generated by inner-loop learning, outer-loop learning transfers knowledge to a simplified model through knowledge distillation, and trains the simplified model to minimize the KL divergence between the output distribution and the output distribution of the inner-loop action selection strategy. New rules extracted from the trained simplified model are added to the rule set; the knowledge represents the decision patterns and experiences in the inner-loop learning mechanism, and the simplified model is an interpretable mathematical model, including decision trees, linear models, or rule sets.

Images