Decision method for order-based production management index

Abstract

The present application belongs to the technical field of production decision-making, and particularly relates to a decision-making method for production management indexes based on orders. Initial orders are captured from an enterprise system, structured analysis is completed through an improved BERT model to form standardized order data; based on the data, multi-attribute weighted clustering and quantification algorithms are fused to divide order types, a scene-adaptive priority model is constructed to clarify the production priority index; with the index as a constraint, a NSGA-II multi-objective optimization algorithm is used to generate a production scheduling scheme, and a decision-making scheme set is selected through the Pareto optimality principle; the scheme is verified through a digital twin model, production line load, equipment health and material completeness rate are evaluated, and the scheme with the optimal order acceptance capacity index is selected. The present application solves the problems of traditional methods, such as dependence on manual work and lack of quantitative basis for priority, and improves the intelligence, scientificity and executability of production decision-making.

Description

Technical Field

[0001] This invention belongs to the field of production decision-making technology, and in particular relates to a decision-making method based on production management indicators for orders. Background Technology

[0002] The manufacturing industry is currently facing multiple challenges, including diversified orders and compressed delivery cycles, with multi-batch, small-volume orders becoming the mainstream. In traditional production management, order parsing relies on manual processing, leading to problems such as redundant information interference, inaccurate extraction of core fields, and inconsistent industry terminology, resulting in low efficiency in generating structured data. Order classification and prioritization largely depend on experience, lacking quantitative basis and making it difficult to balance urgency and profit demands. Production scheduling does not fully consider the actual load on the production line and equipment status, easily leading to resource misallocation and delayed order delivery. Summary of the Invention

[0003] In view of the technical problems existing in the above-mentioned background art, the present invention proposes a decision-making method based on production management indicators of orders.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Retrieve initial orders from the enterprise system, and use an improved BERT to perform structured parsing of the original orders to form standardized order data;

[0006] S2. Based on standardized order data, multi-attribute weighted clustering and quantification algorithms are integrated to classify order types, construct a priority model for scoring and ranking, and clarify the production priority index;

[0007] S3. Using the production priority index as a constraint, apply a multi-objective optimization algorithm to generate various production scheduling schemes, and use the Pareto optimality principle to select a set of production decision schemes.

[0008] S4. Verify the set of production decision schemes through digital twins, evaluate the current production line load status, equipment health, and material availability rate, calculate the order fulfillment capacity index, and select the scheme with the best order fulfillment capacity index as the optimal production decision scheme.

[0009] Preferably, the specific implementation of step S1, which involves retrieving the initial order from the enterprise system and using an improved BERT to perform structured parsing of the original order, is as follows:

[0010] S11. Perform standardized preprocessing on the initial order text retrieved: remove redundant symbols, standardize industry terminology, and mark missing values ​​for core fields; and enhance the BERT basic word embedding layer for domain-specific terminology representation requirements in the production management field. ,in, For the first Domain-enhanced word vectors for each order term. For the first The generic word vectors output by the original BERT model for order vocabulary. This is a professional word vector pre-trained based on a dictionary in the production management field. For adaptive fusion weights;

[0011] S12. Introduce order location-aware masking and length-adaptive attention mechanisms into the Transformer attention layer of BERT. Divide the order text into three functional regions: header, middle, and footer. Construct a location-aware mask matrix and perform length-adaptive attention calculation. ,in, These are the query vector, key vector, and value vector of the Transformer attention layer, respectively. For vector dimensions, For location-aware masking matrix, Multiply the corresponding elements; the final BERT output is: ,in, For length adaptive coefficients, This is the output feature vector of the previous Transformer module;

[0012] S13. Directly outputting BERT at its output layer Mapping to structured field annotation probabilities, defining the core business rule set R for production orders, and calculating rule-based annotation probabilities: ,in, For the word at position t Label as field category The probability, For the output layer weights and bias terms, Scoring based on matching business rules. These are the rule constraint coefficients;

[0013] S14. Select the label category with the highest probability for each position to form an order field label sequence. Extract the corresponding text content according to the label category to obtain standardized order data.

[0014] Preferably, step S2, based on standardized order data, integrates multi-attribute weighted clustering and quantization algorithms to classify order types, construct a priority model scoring and ranking, and clarifies the specific implementation of the production priority index as follows:

[0015] S21. Extract the two core factors of urgency and profit from the standardized order data and perform normalization preprocessing.

[0016] S22. Analyze the normalized dual-attribute data and determine the number of clusters to 4, corresponding to four core order types: urgent high-profit, urgent low-profit, non-urgent high-profit, and non-urgent low-profit.

[0017] S23. Based on the attribute variation coefficient to quantify the urgency and the dispersion of the profit attribute, determine the influence weight on the clustering result. The higher the dispersion, the greater the influence weight of the attribute on the clustering result. This can be calculated directly without initial clustering. The calculation method for the influence weight is as follows: , Where a and b are the normalized urgency and profit, respectively, and CV is the coefficient of variation, which is obtained by the ratio of the standard deviation of the attribute to the mean.

[0018] S24. Incorporate dynamic weights into the clustering objective function, optimize the cluster center update logic, and output four types of order sets;

[0019] S25. For each order set, calculate the mean of urgency and normalized profit within the category. Based on the mean of attributes within the category, assign unique urgency and profit score weights to each order category, aligning with the core characteristics of the category: , ,in , respectively, represent the urgency level within the class and the mean of normalized profit, where c is the index of the class;

[0020] S26. Combining the order's category weight with its own attribute values, calculate the urgency quantification score and profit quantification score for a single order: , ,in, These are the urgency quantification score and profit quantification score for the j-th order, respectively. These represent the urgency level and the normalized profit value of the j-th order, respectively.

[0021] S27. Calculate the basic weights of the categories using the entropy weight method;

[0022] S28. Construct a scenario-adaptive priority index model, introducing a business scenario coefficient to dynamically adapt to different business objectives of the enterprise. The formula is: ,in, As the basic weight of the category, Let j be the production priority index of the j-th order. As the business scenario coefficient, all pending production orders are sorted in descending order of production priority index, and the final output is a clear production priority index and sorting result.

[0023] Preferably, step S24 incorporates dynamic weights into the clustering objective function, optimizes the cluster center update logic, and outputs the four types of order sets as follows:

[0024] S241. Initialize 4 cluster centers;

[0025] S242. Constructing a weighted clustering objective function: Where t is the number of iterations and n is the total number of orders. These are the cluster centers for the urgency and profit of the c-th order category in the t-th iteration, respectively. These are the influence weights for the t-th iteration. As an indicator variable, a value of 1 indicates that the j-th order in the t-th iteration belongs to class c, otherwise it is 0;

[0026] S243. After each round of clustering, the coefficient of variation of the urgency normalization value and the coefficient of variation of the profit normalization value need to be recalculated. Based on these two new coefficients of variation, the urgency clustering weight and profit clustering weight corresponding to the next iteration are determined, and the cluster centers are updated.

[0027] S244. When the weight change in the two rounds is less than the set threshold or the maximum number of iterations is reached, a set of 4 types of orders is obtained.

[0028] Preferably, the specific implementation of step S27, which uses the entropy weight method to calculate the basic weights of the categories, is as follows:

[0029] S271. For each type of order, first calculate the sum of the urgency quantification score and the profit quantification score for each order in that type. Then, sum these sums for all orders in that type. Finally, divide the sum by the total number of orders in that type to obtain the average comprehensive score for that type of order. ;

[0030] S272, Calculate the entropy value ;

[0031] S273. Obtain the basic category weights using entropy values: ,in, .

[0032] Preferably, step S3 applies a multi-objective optimization algorithm with a production priority index as a constraint to generate multiple production scheduling schemes, and uses the Pareto optimality principle to select the set of production decision schemes. The specific implementation of this step is as follows:

[0033] S31. To minimize order delay time, calculate the portion of actual delivery time exceeding planned delivery time for all orders, excluding orders without delay. Finally, take the sum of all delay portions as the objective function, with the production priority index as the constraint.

[0034] S32. Design a time window-based coding scheme to encode the production scheduling scheme into a vector of a specific length. The length of the vector is consistent with the total number of orders to be scheduled. Each element in the vector corresponds to the start time of a single order. At the same time, define the time window constraint: the start time of each order must fall between the earliest start time and the latest start time of that order, and the earliest start time must not be later than the latest start time.

[0035] S33. The NSGA-II multi-objective optimization algorithm is adopted, the algorithm initialization parameters are set, the production priority index constraint is incorporated into the fitness function of the algorithm, and the initial population is generated.

[0036] S34. Run the multi-objective algorithm to generate production scheduling schemes that meet the constraints. Apply the Pareto optimality principle to sort all generated schemes in a non-dominated order and divide them into multiple non-dominated layers: the first non-dominated layer contains all schemes that have no other schemes that are better than it in terms of objectives; the second non-dominated layer contains schemes that have no other schemes that are better than it in terms of objectives except for the first non-dominated layer, and so on for subsequent non-dominated layers.

[0037] S35. Calculate the congestion distance for each scheme within the non-dominated layer, and prioritize the scheme with the largest congestion distance;

[0038] S36. Starting from the first non-dominated layer, select schemes sequentially until the number of selections reaches the preset upper limit, thus obtaining a set of production decision schemes.

[0039] Preferably, step S4 verifies the production decision scheme set using digital twins, evaluates the current production line's load status, equipment health, and material availability rate, calculates the order fulfillment capacity index, and selects the scheme with the optimal order fulfillment capacity index as the optimal production decision scheme. Specifically, this involves: constructing a digital twin model; performing digital twin simulation on each scheme; using real-time data to drive the evaluation of the production line's load status, equipment health, and material availability rate; normalizing the production line's load status, equipment health, and material availability rate; and obtaining the order fulfillment capacity index for each scheme through weighted summation; selecting the scheme with the largest order fulfillment capacity index as the optimal production decision scheme.

[0040] Compared with existing technologies, the advantages and positive effects of this invention are as follows: By improving the BERT model to achieve structured parsing of orders, it avoids the inefficiency and bias of manual processing, and improves the accuracy of core field extraction and data standardization efficiency; It uses dynamic weighted clustering based on the coefficient of variation to classify order types, and combines entropy weight method and business scenario coefficients to build a priority model, providing a quantitative basis for order ranking and effectively balancing urgency and profit demands; It generates high-quality scheduling solutions based on the NSGA-II algorithm and Pareto optimality principle, taking into account both diversity and optimality; It uses digital twin verification to comprehensively evaluate production line load, equipment health and material availability rate, avoid resource mismatch, reduce order delay risk, and improve the intelligence, scientificity and feasibility of production decisions overall. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating a decision-making method based on order-based production management indicators. Detailed Implementation

[0043] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0044] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0045] In the context of the rapid development of the manufacturing industry towards multi-variety, small-batch, and short-delivery-cycle production, enterprise production organization models are facing significant challenges. As customer orders become increasingly personalized, they exhibit high variability in delivery deadlines and profit levels. Traditional experience-based production management and scheduling methods struggle to respond promptly and accurately to complex and ever-changing order demands, leading to resource misallocation and delayed order delivery. This invention proposes an intelligent production decision-making method driven by orders and based on production management indicators. This method constructs a complete closed loop covering order understanding, decision generation, and executability verification through a technical path encompassing order structured analysis, order classification and priority quantification, multi-objective production scheduling, and digital twin verification of decisions. Its specific implementation is as follows... Figure 1 As shown.

[0046] Firstly, in order to efficiently obtain real order data from the company's existing information system and solve the problems of scattered order sources, inconsistent formats, and excessive text noise, we first extracted initial orders from the company's system and used an improved BERT to perform structured parsing of the original orders to form standardized order data.

[0047] Specifically, the initial order text retrieved from the enterprise order management system undergoes standardized preprocessing to eliminate differences in expression between orders from different sources. Preprocessing steps include: cleaning irrelevant symbols, redundant punctuation, and non-business characters from the order text; uniformly replacing synonymous but inconsistent industry terms based on a pre-defined production management terminology mapping table, such as converting delivery dates, batch sizes, and process names with different descriptions into standardized expressions; and marking missing key fields not explicitly given in the order to provide a basis for subsequent model recognition and rule correction. After text standardization, to improve the model's ability to express production management semantics, a domain adaptation enhancement mechanism is introduced into the BERT model's basic word embedding layer. For each order term, both its general semantic vector in the original BERT model and its specialized semantic vector pre-trained based on a production management domain dictionary are obtained. These two types of vectors are then weighted and fused using adaptive fusion weights to form domain-enhanced word vectors that balance general language understanding and production management knowledge expression capabilities. Domain adaptation enhancement is then applied to the BERT basic word embedding layer. ,in, For the first Domain-enhanced word vectors for each order term. For the first The generic word vectors output by the original BERT model for order vocabulary. This is a professional word vector pre-trained based on a dictionary in the production management field. For adaptive fusion weights.

[0048] After obtaining the domain-enhanced word embedding vector sequence, it is input into the Transformer attention layer of the BERT model for context modeling. To fully utilize the inherent structural regularities of order text, a position-aware mask and length-adaptive attention mechanism are introduced during the attention calculation process. The order text is divided into three functional regions: header, middle, and tail. A position-aware mask matrix is ​​constructed, and length-adaptive attention is then calculated. ,in, These are the query vector, key vector, and value vector of the Transformer attention layer, respectively. For vector dimensions, For location-aware masking matrix, Multiply the corresponding elements; the final BERT output is: ,in, For length adaptive coefficients, This is the output feature vector of the previous Transformer module.

[0049] After completing the semantic modeling of the multi-layer Transformer module, the final output feature vector is directly input into the output layer of BERT to generate the initial annotation probabilities of words at each position in the order text corresponding to different field categories. To avoid ambiguity or misjudgment in key business fields due to pure model prediction, a set of core business rules for production orders is introduced to constrain and correct the annotation results. In the specific implementation, the feature vector is directly input into the output layer of BERT. Mapping to structured field annotation probabilities, defining the core business rule set R for production orders, and calculating rule-based annotation probabilities: ,in, For the word at position t Label as field category The probability, For the output layer weights and bias terms, Scoring based on matching business rules. This represents the rule constraint coefficient.

[0050] After semantic modeling of the multi-layer Transformer module, the final output feature vector is directly input into the BERT output layer to generate initial label probabilities for different field categories corresponding to words at various positions in the order text. To avoid ambiguity or misjudgment in key business fields due to pure model prediction, a core business rule set for production orders is introduced to constrain and correct the labeling results. Subsequently, based on the field labeling sequence, consecutive words with the same field category are aggregated, their corresponding original text fragments are extracted, and mapped to structured fields such as order number, product model, order quantity, delivery deadline, and process requirements according to a predefined data structure format. For core fields marked as missing in previous steps, the corresponding missing identifier is retained in the structured output for subsequent decision-making modules to complete or assess risks. Through the above methods, semantically accurate, field-clear, and formatted structured order data is finally obtained.

[0051] Based on standardized order data, multi-attribute weighted clustering and quantification algorithms are integrated to classify order types, construct a priority model for scoring and ranking, and clarify the production priority index.

[0052] Specifically, to reflect the production value of orders with the fewest but most critical indicators and address the issues of complexity and poor interpretability in multi-attribute evaluation models, a scheme using urgency and profit as core attributes for analysis was adopted. Urgency and profit, the two core factors, were extracted from standardized order data and preprocessed using normalization. Then, to identify the inherent structural differences in orders without relying on manual rules, the normalized dual-attribute data was analyzed, resulting in a cluster size of 4, corresponding to four core order types: urgent high-profit, urgent low-profit, non-urgent high-profit, and non-urgent low-profit.

[0053] Then, based on the attribute variation coefficient, the urgency and profit attribute dispersion are quantified to determine their influence weights on the clustering results. The higher the dispersion, the greater the influence weight of the attribute on the clustering results. This can be calculated directly without initial clustering. The calculation method for the influence weights is as follows: , , where a and b are the normalized urgency and profit, respectively, and CV is the coefficient of variation, which is obtained by the ratio of the standard deviation to the mean of the attribute.

[0054] Dynamic weights are incorporated into the clustering objective function to optimize the cluster center update logic, outputting four types of order sets. Specifically, four cluster centers are initialized; a weighted clustering objective function is constructed: Where t is the number of iterations and n is the total number of orders. These are the cluster centers for the urgency and profit of the c-th order category in the t-th iteration, respectively. These are the influence weights for the t-th iteration. As an indicator variable, a value of 1 indicates that the j-th order in the t-th iteration belongs to class c, otherwise it is 0. After each round of clustering, the coefficients of variation of the urgency normalized value and the profit normalized value need to be recalculated. Based on these two new coefficients of variation, the urgency cluster weights and profit cluster weights for the next iteration are determined, and the cluster centers are updated. The coefficients of variation of urgency and profit are recalculated after each iteration, and the cluster centers are updated. , Then, the cluster centers are updated accordingly: , , where t is the number of iterations. When the weight change in two rounds is less than a set threshold or the maximum number of iterations is reached, a set of four types of orders is obtained.

[0055] Next, for each order category, calculate the mean of urgency and normalized profit within the category. Based on the mean of attributes within the category, assign specific urgency and profit score weights to each order category, aligning with the core characteristics of the category: , ,in These represent the urgency level within the class and the mean of normalized profit, respectively. c is the index of the class. Combining the order's category weight with its own attribute values, the urgency quantification score and profit quantification score for a single order are calculated. , ,in, These are the urgency quantification score and profit quantification score for the j-th order, respectively. These represent the urgency level of the j-th order and the normalized profit value, respectively.

[0056] The entropy weight method is used to calculate the basic weights of each category. Specifically, for each category of orders, the sum of the urgency quantification score and the profit quantification score for each order in that category is first calculated. Then, this sum is added together for all orders in that category. Finally, the sum is divided by the total number of orders in that category to obtain the average comprehensive score for that category of orders. Calculate the entropy value. ; and obtain the basic category weights through entropy: ,in, Specifically, by summarizing the urgency and profit quantification scores of each order in each order category, and further calculating the average comprehensive score of that category, the average value level and delivery pressure characteristics of that category of orders within the current order set can be reflected from an overall perspective. Based on this, the entropy weight method is introduced to analyze the distribution of comprehensive scores across different categories, using entropy values ​​to characterize the dispersion of information distribution within each order category. A larger entropy value indicates smaller differences between orders in that category, providing relatively limited effective distinguishing information; a smaller entropy value indicates significant differences in order characteristics, providing higher reference value for production decisions. Furthermore, the entropy value is used to inversely calculate the category-based weights, allowing order categories with greater information content and stronger distinguishing ability to receive higher weights in overall production decisions. The category-based weights formed in this way are entirely derived from the distribution characteristics of the order data itself, adaptively reflecting changes in order structure under different production stages and market environments. Compared to fixed weights or manually assigned weights, this step has advantages in objectivity, adaptability, and interpretability, which is conducive to building a more robust and reasonable order priority evaluation system.

[0057] Finally, a scenario-adaptive priority index model is constructed, incorporating a business scenario coefficient to dynamically adapt to different business objectives of the enterprise. The formula is as follows: ,in, As the basic weight of the category, Let j be the production priority index of the j-th order. To determine the operational scenario coefficient, all pending production orders are sorted in descending order of production priority index, ultimately outputting a clear production priority index and ranking result. Among these, The production priority index for the j-th order ranges from [0,1]. The operational scenario coefficient is determined based on delivery-oriented scenarios (peak season / high default risk). =0.3, balanced guidance scenario takes =0.5, taking the profit-oriented scenario (off-season / surplus capacity) as 0.5. =0.7.

[0058] Then, using a production priority index as a constraint, a multi-objective optimization algorithm is applied to generate various production scheduling schemes. The Pareto optimality principle is then used to select the set of production decision schemes. Specifically, the objective function is to minimize order delay time, i.e., to calculate the portion of all orders whose actual delivery time exceeds the planned delivery time (orders without delay are excluded). The sum of all delayed portions is taken as the objective function, constrained by the production priority index. A time window-based encoding scheme is designed, encoding the production scheduling scheme into a vector of a specific length, consistent with the total number of orders to be scheduled. Each element in the vector corresponds to the start time of a single order. Simultaneously, a time window constraint is explicitly defined: the start time of each order must fall between its earliest and latest start times, and the earliest start time cannot be later than the latest start time. A complete production scheduling scheme is encoded into a fixed-length real-number vector, consistent with the total number of orders to be scheduled. Each element in the vector uniquely corresponds to the production start time of a single order. This one-to-one encoding method allows each individual to intuitively reflect the production arrangement of all orders on the timeline. The NSGA-II multi-objective optimization algorithm is adopted, and the algorithm initialization parameters are set, including population size, maximum number of iterations, crossover probability, and mutation probability. The population size controls the number of candidate scheduling schemes in each generation, preferably set to a moderate number matching the order size to balance search efficiency and solution diversity. The maximum number of iterations limits the algorithm's running time, ensuring a stable solution set is obtained within an acceptable computational cost. The crossover probability and mutation probability are used to adjust the degree of recombination and perturbation of the original scheduling scheme when new individuals are generated, respectively, to avoid premature convergence. After parameter initialization, an initial population is randomly generated according to the aforementioned time window-based encoding rules, and a multi-objective fitness value is calculated for each individual, where the fitness function is the objective function. Subsequently, the algorithm enters the iterative optimization phase, sequentially performing non-dominated sorting, crowding distance calculation, selection, crossover, and mutation operations in each generation. The algorithm employs several steps: non-dominated sorting to categorize individuals into different quality levels, crowding distance to maintain the uniform distribution of the solution set in the target space, selection prioritizing individuals with high non-dominated levels and high crowding, and crossover and mutation operations to generate a new generation of scheduling schemes while satisfying time window constraints. The algorithm terminates when the number of iterations reaches a set upper limit or the solution set convergence condition is met, outputting the final Pareto-optimal production scheduling scheme set. The Pareto optimality principle is then applied to all generated schemes for non-dominated sorting, dividing the algorithm into multiple non-dominated layers: the first non-dominated layer contains all schemes for which no other scheme is superior in all aspects of the target; the second non-dominated layer contains schemes for which no other scheme is superior in all aspects of the target, except for the first non-dominated layer; and so on for subsequent non-dominated layers. For each non-dominated layer, the crowding distance is calculated, prioritizing the scheme with the largest crowding distance.A larger crowding distance indicates a sparser region in the target space, representing a differentiated scheduling strategy. Therefore, during the solution selection process, solutions with larger crowding distances are prioritized to ensure that the final set of solutions maintains good diversity while maintaining similar performance. Starting from the first non-dominated layer, solutions are selected sequentially until the preset upper limit is reached, resulting in the production decision solution set.

[0059] Finally, the production decision set is validated through digital twins to assess the current production line's load status, equipment health, and material availability rate. Production line load status is assessed by real-time data collection of equipment operation and work-in-process, quantified by equipment utilization and process queue length to reflect resource saturation. Equipment health is assessed by monitoring runtime, fault records, and key parameters, with weighted scoring reflecting equipment integrity and operational reliability. Material availability rate is calculated by comparing scheduled material requirements with inventory and delivery plans to determine the proportion of materials met, reflecting the availability of production materials. An order fulfillment capacity index is calculated, and the optimal production decision is selected based on this index. Specifically, first, a digital twin model corresponding one-to-one with the actual production line is constructed. Virtual production units are established, and data such as equipment operating status, production cycle time, work-in-process quantity, and material inventory are accessed in real-time through interfaces with the on-site data acquisition system, achieving virtual-real synchronization. In this way, the digital twin model can realistically reflect the current production line's operating characteristics and resource constraints. Second, for each candidate production scheduling scheme, it is used as input to drive the digital twin model for simulation. During the simulation, based on the order start time and process sequence in the plan, the flow of each order in the production line is simulated, and the production line load status, equipment health, and material availability rate are statistically analyzed in real time. The production line load status is evaluated through equipment utilization and process queue length; equipment health is comprehensively determined through equipment runtime, fault records, and status parameters; and the material availability rate is calculated by comparing the materials required at the simulation time with actual inventory and delivery plans. Subsequently, the above three evaluation results are normalized to ensure a consistent numerical range and eliminate the influence of differences in units. Then, the production line load status, equipment health, and material availability rate are weighted and summed according to preset weights to calculate the order fulfillment capacity index corresponding to each production scheduling plan. Finally, the order fulfillment capacity indices of all candidate plans are compared, and the plan with the highest index value is selected as the optimal production decision plan.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A decision-making method based on production management indicators for orders, characterized in that, Includes the following steps: S1. Retrieve initial orders from the enterprise system, and use an improved BERT to perform structured parsing of the original orders to form standardized order data; S2. Based on standardized order data, multi-attribute weighted clustering and quantification algorithms are integrated to classify order types, construct a priority model for scoring and ranking, and clarify the production priority index; S3. Using the production priority index as a constraint, apply a multi-objective optimization algorithm to generate various production scheduling schemes, and use the Pareto optimality principle to select a set of production decision schemes. S4. Verify the set of production decision schemes through digital twins, evaluate the current production line load status, equipment health, and material availability rate, calculate the order fulfillment capacity index, and select the scheme with the best order fulfillment capacity index as the optimal production decision scheme.

2. The decision-making method based on production management indicators according to claim 1, characterized in that, The specific implementation of retrieving initial orders from the enterprise system and using an improved BERT to perform structured parsing of the original orders in step S1 is as follows: S11. Perform standardized preprocessing on the initial order text retrieved: remove redundant symbols, standardize industry terminology, and mark missing values ​​for core fields; and enhance the BERT basic word embedding layer for domain-specific terminology representation requirements in the production management field. ,in, For the first Domain-enhanced word vectors for each order term. For the first The generic word vectors output by the original BERT model for order vocabulary. This is a professional word vector pre-trained based on a dictionary in the production management field. For adaptive fusion weights; S12. Introduce order location-aware masking and length-adaptive attention mechanisms into the Transformer attention layer of BERT. Divide the order text into three functional regions: header, middle, and footer. Construct a location-aware mask matrix and perform length-adaptive attention calculation. ,in, These are the query vector, key vector, and value vector of the Transformer attention layer, respectively. For vector dimensions, For location-aware masking matrix, Multiply the corresponding elements; the final BERT output is: ,in, For length adaptive coefficients, This is the output feature vector of the previous Transformer module; S13. Directly outputting BERT at its output layer Mapping to structured field annotation probabilities, defining the core business rule set R for production orders, and calculating rule-based annotation probabilities: ,in, For the word at position t Label as field category The probability, For the output layer weights and biases, Scoring based on matching business rules. These are the rule constraint coefficients; S14. Select the label category with the highest probability for each position to form an order field label sequence. Extract the corresponding text content according to the label category to obtain standardized order data.

3. The decision-making method based on order-based production management indicators according to claim 1, characterized in that, Step S2, based on standardized order data, integrates multi-attribute weighted clustering and quantization algorithms to classify order types, construct a priority model for scoring and ranking, and clarifies the specific implementation of the production priority index as follows: S21. Extract the two core factors of urgency and profit from the standardized order data and perform normalization preprocessing. S22. Analyze the normalized dual-attribute data and determine the number of clusters to 4, corresponding to four core order types: urgent high-profit, urgent low-profit, non-urgent high-profit, and non-urgent low-profit. S23. Based on the attribute variation coefficient to quantify the urgency and the dispersion of the profit attribute, determine the influence weight on the clustering result. The higher the dispersion, the greater the influence weight of the attribute on the clustering result. This can be calculated directly without initial clustering. The calculation method for the influence weight is as follows: , Where a and b are the normalized urgency and profit, respectively, and CV is the coefficient of variation, which is obtained by the ratio of the standard deviation of the attribute to the mean. S24. Incorporate dynamic weights into the clustering objective function, optimize the cluster center update logic, and output four types of order sets; S25. For each order set, calculate the mean of urgency and normalized profit within the category. Based on the mean of attributes within the category, assign unique urgency and profit score weights to each order category, aligning with the core characteristics of the category: , ,in , respectively, represent the urgency level within the class and the mean of normalized profit, where c is the index of the class; S26. Combining the order's category weight with its own attribute values, calculate the urgency quantification score and profit quantification score for a single order: , ,in, These are the urgency quantification score and profit quantification score for the j-th order, respectively. These represent the urgency level and normalized profit value of the j-th order, respectively. S27. Calculate the basic weights of the categories using the entropy weight method; S28. Construct a scenario-adaptive priority index model, introducing a business scenario coefficient to dynamically adapt to different business objectives of the enterprise. The formula is: ,in, As the basic weight of the category, Let j be the production priority index of the j-th order. As the business scenario coefficient, all pending production orders are sorted in descending order of production priority index, and the final output is a clear production priority index and sorting result.

4. The decision-making method based on production management indicators according to claim 3, characterized in that, The specific implementation of step S24, which incorporates dynamic weights into the clustering objective function, optimizes the cluster center update logic, and outputs four types of order sets, is as follows: S241. Initialize 4 cluster centers; S242. Constructing a weighted clustering objective function: Where t is the number of iterations and n is the total number of orders. These are the cluster centers for the urgency and profit of the c-th order category in the t-th iteration, respectively. These are the influence weights for the t-th iteration. As an indicator variable, a value of 1 indicates that the j-th order in the t-th iteration belongs to class c, otherwise it is 0; S243. After each round of clustering, the coefficient of variation of the urgency normalization value and the coefficient of variation of the profit normalization value need to be recalculated. Based on these two new coefficients of variation, the urgency clustering weight and profit clustering weight corresponding to the next iteration are determined, and the cluster centers are updated. S244. When the weight change in the two rounds is less than the set threshold or the maximum number of iterations is reached, a set of 4 types of orders is obtained.

5. The decision-making method based on production management indicators according to claim 3, characterized in that, The specific implementation of step S27, which uses the entropy weight method to calculate the basic weights of the categories, is as follows: S271. For each type of order, first calculate the sum of the urgency quantification score and the profit quantification score for each order in that type. Then, sum these sums for all orders in that type. Finally, divide the sum by the total number of orders in that type to obtain the average comprehensive score for that type of order. ; S272, Calculate the entropy value ; S273. Obtain the basic category weights using entropy values: ,in, .

6. The decision-making method for production management indicators based on orders according to claim 1, characterized in that, Step S3, which uses a multi-objective optimization algorithm with a production priority index as a constraint to generate multiple production scheduling schemes and selects the production decision scheme set using the Pareto optimality principle, is specifically implemented as follows: S31. To minimize order delay time, calculate the portion of actual delivery time exceeding planned delivery time for all orders, excluding orders without delay. Finally, take the sum of all delay portions as the objective function, with the production priority index as the constraint. S32. Design a time window-based coding scheme to encode the production scheduling scheme into a vector of a specific length. The length of the vector is consistent with the total number of orders to be scheduled. Each element in the vector corresponds to the start time of a single order. At the same time, define the time window constraint: the start time of each order must fall between the earliest start time and the latest start time of that order, and the earliest start time must not be later than the latest start time. S33. The NSGA-II multi-objective optimization algorithm is adopted, the algorithm initialization parameters are set, the production priority index constraint is incorporated into the fitness function of the algorithm, and the initial population is generated. S34. Run the multi-objective algorithm to generate production scheduling schemes that meet the constraints. Apply the Pareto optimality principle to sort all generated schemes in a non-dominated order and divide them into multiple non-dominated layers: the first non-dominated layer contains all schemes that have no other schemes that are better than it in terms of objectives; the second non-dominated layer contains schemes that have no other schemes that are better than it in terms of objectives except for the first non-dominated layer, and so on for subsequent non-dominated layers. S35. Calculate the congestion distance for each scheme within each non-dominated layer, and prioritize the scheme with the largest congestion distance; S36. Starting from the first non-dominated layer, select schemes sequentially until the number of selections reaches the preset upper limit, thus obtaining a set of production decision schemes.

7. The decision-making method based on production management indicators according to claim 1, characterized in that, Step S4 verifies the production decision set using digital twins, assesses the current production line's load status, equipment health, and material availability rate, calculates the order fulfillment capacity index, and selects the optimal production decision plan based on the plan with the optimal order fulfillment capacity index. Specifically, this involves: constructing a digital twin model; performing digital twin simulation on each plan; using real-time data to drive the assessment of production line load status, equipment health, and material availability rate; normalizing the production line load status, equipment health, and material availability rate; and obtaining the order fulfillment capacity index for each plan through weighted summation; selecting the plan with the largest order fulfillment capacity index as the optimal production decision plan.

Images