Order balancing hybrid production method and device based on strategy generator and LLM, and medium

Abstract

The application provides an order balancing hybrid production method and device based on a strategy generator and an LLM, and a medium, which comprises the following steps: obtaining order information, inventory information and equipment capacity information; based on a preset prompt word, processing the order information, the inventory information and the equipment capacity information through a large language model to obtain a strategy selection and strategy parameters; and processing the strategy selection and the strategy parameters through a strategy generator to return production scheduling information. Based on the LLM technology, the application combines the current order information, the inventory information and the equipment capacity information, dynamically adjusts the number of orders, mixes production of production orders and standby inventory orders, balances the utilization rate of equipment, realizes early standby of products, greatly reduces the delivery time for small orders and multi-variety production orders, and improves the production scheduling efficiency based on the LLM and the strategy generator.

Description

Technical Field

[0001] This invention relates to the field of production scheduling optimization technology based on LLM, specifically to a method, apparatus, and medium for order balancing hybrid production based on a strategy generator and LLM. Background Technology

[0002] In typical smart card order processing, balancing production capacity with production orders is a common challenge. Currently, order inflows and production scheduling are separate processes. Specifically, orders are typically entered into the sales or customer system, then aggregated and directly imported into the ERP (Enterprise Resource Planning) system. The ERP system then generates planned production instructions based on its built-in scheduling rules. However, this process often requires frequent production order changes as order batches decrease and product variety increases. This results in significant downtime during production, severely impacting overall output efficiency.

[0003] In recent years, the industry has begun to explore the introduction of large language models into the field of production scheduling. For example, Chinese patent application CN121212696A discloses an LLM (Large Language Model) driven method for complex manufacturing production scheduling, which optimizes scheduling by constructing a manufacturing-specific LLM; some related studies explore using large models as end-to-end self-decision engines for shop floor scheduling.

[0004] However, the aforementioned existing technologies generally employ an architecture that directly generates production scheduling instructions end-to-end using large language models, which has several inherent drawbacks. The decision-making process of large language models is like a black box, making the resulting production scheduling results difficult to interpret. When there is a significant conflict between the production scheduling plan and human experience, it is impossible to trace the reasoning logic, resulting in insufficient credibility in industrial applications. Furthermore, such solutions are ill-suited to adapting to changing operational strategies. Enterprise operational goals often need to dynamically switch between "meeting delivery deadlines," "reducing inventory," and "improving utilization." Existing technologies require retraining the large language model or repeatedly adjusting prompts, resulting in slow response and poor decision consistency. In addition, existing technologies focus more on macro-level production sequencing and do not truly address the order change bottleneck caused by order fragmentation, neglecting the value of using backup production to smooth equipment fluctuations in small-batch, multi-variety scenarios.

[0005] Furthermore, Chinese invention patent application CN119962938A utilizes a large language model for intelligent production scheduling and anomaly management; Chinese invention patent application CN120975468A discloses a resource-constrained parallel machine rule-based scheduling method based on a large language model. This method obtains an initial heuristic rule set by constructing a heuristic rule generator based on the large language model, and then generates a scheduling scheme for the current shift based on the initial heuristic rule set, monthly order demand, and resource constraints. However, this scheme focuses on generating scheduling rules based on real-time order demand and dynamic characteristics, and does not design a mechanism to proactively produce partial inventory to smooth production fluctuations. Summary of the Invention

[0006] The primary objective of this invention is to provide a hybrid production method for balanced order production based on a strategy generator and LLM, thereby addressing the problem of decreased production scheduling efficiency caused by frequent order changes in existing production processes.

[0007] A second objective of the present invention is to provide a computer device for implementing the above-described order balancing hybrid production method based on a strategy generator and LLM.

[0008] A third objective of this invention is to provide a computer-readable storage medium for implementing the above-described order balancing hybrid production method based on a strategy generator and LLM.

[0009] To achieve the aforementioned first objective, this invention provides a hybrid production method for order balancing based on a strategy generator and LLM, comprising the following steps: acquiring order information, inventory information, and equipment capacity information; processing the order information, inventory information, and equipment capacity information through a large language model based on preset prompt words to obtain strategy selection and strategy parameters; and processing the strategy selection and strategy parameters through a strategy generator to return production scheduling information.

[0010] As can be seen from the above scheme, this invention utilizes LLM to determine strategies based on existing order information, inventory information, and equipment capacity information. A strategy generator then generates production scheduling information based on the production strategy to guide production. Based on LLM technology and combining current order information, inventory information, and equipment capacity information, this invention dynamically adjusts order quantities, mixing production orders and reserve orders for production. This achieves balanced equipment utilization and ultimately enables advance product preparation. For small-batch, multi-variety production orders, it significantly reduces delivery time. This invention improves production scheduling efficiency based on LLM and a strategy generator.

[0011] A further approach is to use a large language model to process order information, inventory information, and equipment capacity information after feature extraction and standardization to obtain strategy selection and strategy parameters.

[0012] Therefore, obtaining a unified data format makes it easier for large language models to process the data directly.

[0013] A further solution involves feature extraction and standardization, including: determining whether the confidence level of key fields in the currently processed order is lower than a set threshold; if so, diverting the currently processed order to a manual review queue.

[0014] This demonstrates that it can improve the accuracy of data processing.

[0015] A further approach is that the strategy generator includes pre-configured production strategy templates of different types; the production strategy template defines the following mathematical mapping relationship: strategy S_i = f_i(order queue Q, current inventory I, equipment utilization rate U, delivery cycle set D), where different strategy functions f_i correspond to different production target preferences.

[0016] Therefore, by setting up a strategy generator, different strategy functions can be executed according to different production goals, such as ensuring backup inventory, eliminating downtime, and prioritizing delivery time, and the corresponding production quantities can be output.

[0017] A further approach is to include production strategy templates of three types: guaranteed inventory, full equipment utilization, and delivery priority. The weight vectors and objective functions of different types of production strategy templates are different. The weight vectors are used to indicate the weights of the order queue, current inventory, equipment utilization, and delivery cycle set. The objective function of the guaranteed inventory type is to maximize the inventory fulfillment rate for a set number of future days. The objective function of the full equipment utilization type is to minimize equipment idle time. The objective function of the delivery priority type is to minimize order delay penalties.

[0018] This demonstrates that different strategies can be dynamically set to achieve different production orientations, allowing for different production strategies to be adopted during idle and busy periods. For example, to ensure the inventory needs for the next 5 days and calculate the order quantity, or to ensure that all workshop processes do not stop and calculate the order quantity.

[0019] A further solution is to return the production scheduling information to the ERP system.

[0020] Therefore, this invention can assist ERP in scheduling and achieve linkage between the front end and the production end.

[0021] A further approach is to include the production scheduling information in terms of the number of production orders and the number of backup orders.

[0022] This demonstrates that by combining orders with backup orders, delivery time and capacity utilization can be balanced.

[0023] A further approach is to pre-set prompts including environment coding and feature abstraction prompts, multi-strategy adaptation scoring prompts, and output format requirements.

[0024] To achieve the second objective mentioned above, the present invention provides a computer device comprising a processor and a memory, wherein: the memory stores a computer program, and when the computer program is executed by the processor, it implements the above-mentioned order balancing hybrid production method based on strategy generator and LLM.

[0025] To achieve the third objective described above, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements the order balancing hybrid production method based on strategy generator and LLM as described in the above claims. Attached Figure Description

[0026] Figure 1 This is a flowchart of an embodiment of the order balancing hybrid production method based on strategy generator and LLM of the present invention.

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0028] The order balancing hybrid production method based on strategy generator and LLM of the present invention realizes that when processing order scheduling, the order quantity is dynamically adjusted according to the development trend of orders, so as to provide a basis for subsequent frequent order changes, thereby reducing the order change time and optimizing scheduling efficiency.

[0029] Example of an order balancing hybrid production method based on strategy generator and LLM: This embodiment is implemented using a computer program, as described below. Figure 1 Detailed explanation.

[0030] First, execute step S11 to obtain order information, inventory information, and equipment capacity information.

[0031] Order information refers to the raw sales order data from customers, generally including order identifier, customer code, product model, order quantity, and delivery time. The order identifier uniquely identifies the order. The customer code identifies the customer who placed the order. The product model indicates the cardioid model involved in the order. The order quantity indicates the required quantity of the cardioid model involved in the order. The delivery time indicates the delivery cycle requirements or delivery deadline for the order, such as a 30-day delivery cycle explicitly stated by the customer, or a specified latest arrival date.

[0032] Inventory information refers to data such as the current inventory quantity, inventory turnover rate, and inventory status (e.g., shortage, normal, backlog) of various product models in the warehouse.

[0033] Equipment capacity information refers to the total available capacity time, currently occupied capacity time, and the resulting equipment utilization rate curve for each key piece of equipment (such as a laminator) in the production line within a specified future time window.

[0034] Next, step S12 is executed to extract features and standardize the order information, inventory information, and equipment capacity information.

[0035] By performing feature extraction and standardization on order information, inventory information, and equipment capacity information, these data can be transformed into a unified parameter format that can be directly processed by large oracle models, enabling them to better understand and handle these data. Specifically, feature extraction and standardization operations include multimodal input parsing, parameter standardization mapping, and multidimensional feature concatenation.

[0036] In this embodiment, the order information can be recorded in the form of email, PDF file, image, paper scan, EDI file or Excel file, etc.

[0037] Modal input parsing enables differentiated parsing of different order information recording formats. For order information in the form of emails or text, a deployed natural language processing model is used for named entity recognition to extract key fields such as order identifier, product model, order quantity, and delivery time. For order information in the form of PDFs, images, or scanned paper documents, a pre-trained large-scale visual model is invoked to directly extract information such as quantity and delivery date through semantic understanding, without relying on fixed table lines or OCR templates, thus solving the problem of different table layouts for customers. For structured data such as EDI or Excel files, a preset field mapping table is used to directly map customer system field names to internal standard field names.

[0038] Parameter standardization mapping resolves the semantic confusion and inconsistent units in order information, inventory information, and equipment capacity information.

[0039] For order information, a product model alias mapping table is established to uniformly map various ways of writing customer descriptions (such as "S50 white card", "S50 chip card K") to internal material codes (such as MTRL-S50-001-V2); by converting all vague terms related to delivery time (such as "urgent" and "ASAP") into absolute time functions, the delivery cycle is uniformly expressed as delta_t = delivery date - production schedule date, in hours; Optionally, logical validation can also be performed. For example, if the same order information has both the "regular product" and "urgent delivery date" tags, it will be automatically marked as abnormal and await manual review. If the confidence level of any key field of the order information (such as product model or order quantity) is lower than the set threshold of 95% (e.g., in cases where the text in the image is extremely blurry), the order information will be automatically diverted to the manual review queue and will not be processed.

[0040] For inventory information, the parameter standardization mapping process discretizes the original inventory turnover rate value into abstract features that can be understood by the LLM. Specifically, based on the current inventory level and historical average daily consumption of the product model, the estimated number of days available is calculated, and the inventory is marked as "high turnover / shortage", "normal" or "low turnover / overstock" according to the set threshold. Finally, the inventory is stored in the inventory_feature object with the material code as the key.

[0041] For equipment capacity information, the parameter standardization mapping process first calculates the equipment utilization time series for a future set time (72 hours in this embodiment) based on the total available capacity and the occupied capacity. Then, by scanning this series, periods with utilization rates continuously below the set threshold are identified as idle windows and named with a unified shift or time identifier, and stored together with the utilization lower limit in a capacity_feature object. This processing time transforms the original capacity figures into an equipment status sequence that can be directly semantically matched by LLM.

[0042] Multidimensional feature stitching aligns the order data, inventory data, and capacity data after feature extraction and standardization in the feature space, generating a standardized JSON structure that can be directly input into an LLM.

[0043] The standardized input format for the final output to the LLM is shown below: JSON { "task_id": "T-20260429-001", “order_batch”: [ { "order_no": "PO-8842", "material": "IC-KB01", “quantity”: 15000 "deadline_timestamp": "2026-05-03T18:00:00Z", "priority_feature": 0.96 / / Derived feature: Delivery urgency score} ], “inventory_feature”: { “IC-KB01”: { "stock_qty": 4000 "pick_rate": "low" / / Abstract characteristic: slow turnover rate } }, “capacity_feature”: { “press_machine”: { "util_rate_lower_bound": 0.55, “slot_window”: “Shifts-4_to_9” / / Idle window period } } } The JSON structure is structured as follows: "task_id" serves as the unique identifier for this production scheduling task; the "order_batch" array contains standardized order batches, with each order specifying the product model (i.e., "material" in the JSON structure), quantity (i.e., "quantity" in the JSON structure), absolute delivery timestamp (i.e., "deadline_timestamp" in the JSON structure), and a derived delivery urgency score; "inventory_feature" uses the material code as the key to record the current inventory quantity and abstract inventory turnover characteristics (e.g., "low" indicates slow turnover); and "capacity_feature" describes the lower limit of key equipment utilization and the predicted idle window period. This JSON structure, as a unified parameter format, allows LLM to directly proceed with strategy matching calculations without needing to understand the heterogeneity of the data source.

[0044] It's important to note that in the JSON data structure described above, `order_batch` is an array that can contain multiple order information entries. For example, it lists the order information for order identifier "PO-8842". This means that the `order_batch` in the JSON data structure does not only refer to newly received, unscheduled orders, but also covers all confirmed but incomplete sales orders, including both orders awaiting production and those already in production but not yet completed. The "inventory_feature" and "capacity_feature" in the JSON are global background data. For example, "util_rate_lower_bound: 0.55" describes the lower limit of the overall utilization rate of the press machine within the current production cycle, and "slot_window: Shifts-4_to_9" describes the idle time slots for the next 72 hours. "inventory_feature" and "capacity_feature" serve as feature data describing inventory level trends and equipment status sequences, respectively. They represent the shared production scheduling context for all order information. LLM calculates the fit score with each strategy template by matching the order information in the order_batch with these shared capacity and inventory features. Correspondingly, the equipment occupancy time reflected in capacity_feature is calculated by combining the remaining working hours of these work-in-process orders.

[0045] Therefore, in conjunction with global inventory levels and trend characteristics, this JSON structure provides LLM with a complete and self-consistent snapshot of the production scenario, enabling strategy matching to be completed in one go based on the real-time status of the entire factory.

[0046] Then, step S13 is executed, based on preset prompt words, and after feature extraction and standardization of order information, inventory information and equipment capacity information through large language model processing, to obtain strategy selection and strategy parameters.

[0047] In this embodiment, the large language model is used for strategy matching. Specifically, based on the input preset prompt words and JSON format data obtained after feature extraction and standardization of order information, inventory information, and equipment capacity information, it calculates the fit score between the current production scenario and each strategy template in the preset strategy library, and outputs the strategy selection with the highest score and its corresponding strategy parameters. The preset prompt words include environment-encoded and feature-abstracted prompt words, multi-strategy fit scoring prompt words, and output format requirements.

[0048] Specifically, after receiving the JSON format data obtained in step S12, the Big Prophecy model processes it based on environment encoding and feature abstraction prompts.

[0049] The environment coding and feature abstraction prompts in this embodiment are: "You are a manufacturing system state analysis model. Please complete the quantitative assessment of the current production environment based on the following input data:" a. Order Status Analysis: Calculate the average delivery urgency of the order queue (based on the ratio of delivery cycle to standard working hours); Identify whether there are high-volume orders (i.e., demand exceeding three times the daily production capacity of the production line); b. Inventory level assessment: Based on finished goods inventory turnover data, each product model is marked as shortage, normal, or overstocked; c. Equipment fluctuation assessment: Analyze the equipment utilization curve for the next 72 hours to identify idle windows with utilization below 60% and overload zones with utilization above 90%. Please output the above analysis results in JSON format, only outputting JSON, without adding any explanatory text. Therefore, based on the aforementioned environment coding and feature abstraction prompts, LLM evaluates the input JSON-formatted data and outputs environmental feature JSON. The environmental feature JSON includes average delivery urgency score, high-volume order identifiers, inventory status markers for each product model, and a list of idle and overloaded window periods.

[0050] Then, the LLM executes multi-strategy fit scoring prompts. In this embodiment, the multi-strategy fit scoring prompts are: […]. The current production environment characteristics are as follows (output from the previous stage): {environmental_features_json} The existing candidate strategy library is defined as follows: S_1 (Backup Stock Type): The optimization goal is to maximize the inventory fulfillment rate for the next X days (X is 3 in this example), and accept that the equipment may be idle for a short period of time.

[0051] S_2 (Full Load Type): The optimization goal is to minimize equipment idle time, while allowing for a moderate increase in the backup database.

[0052] S_3 (Delivery Priority): The optimization goal is to minimize order delays, which can split the entire order into multiple batches.

[0053] Please use this as a strategy evaluation model to strictly calculate the fitness score (0.0-1.0) for each strategy S_i according to the following three dimensions: Target fit (weight 0.4): The degree to which the current production situation matches the optimization target of this strategy.

[0054] Constraint Satisfaction (Weight 0.3): Whether the strategy triggers constraints such as upper / lower inventory limits and equipment operating hours when executed.

[0055] Switching smoothness (weight 0.3): Whether switching from the current strategy to S_i will cause a large fluctuation in the output rate.

[0056] Output format requirement: JSON {"S_1": {"Target Fit": 0.85, "Constraint Satisfaction": 0.90, "Switching Smoothness": 0.60, "Overall Score": 0.78}, “S_2”: {“...”: ..., “Overall Score”: 0.92}, “S_3”: {“...”: ..., “Overall Score”: 0.45} } Output only JSON. In this context, `{environmental_features_json}` is replaced with the environment JSON output by the LLM based on environment encoding and feature abstraction prompts. The LLM then calculates the three-dimensional scores and weighted composite score for each strategy S_i. The strategy with the highest composite score is selected as the current strategy, and the LLM outputs the required strategy parameters based on the current scenario characteristics. For example, assuming S_2 (equipment at full capacity) has the highest score, the LLM outputs the strategy selection "S_2" and strategy parameters, which may include the target minimum equipment utilization rate (0.85) and the maximum reserve quantity limit (e.g., 8000 units). These parameters are not fixed default values ​​but rather boundary conditions dynamically set by the LLM based on the current inventory level, idle window size, and order urgency.

[0057] Then, step S14 is executed, where the strategy generator processes the strategy selection and strategy parameters to obtain production scheduling information.

[0058] The strategy generator is a rule engine and strategy library independent of the LLM. It internally stores a variety of pre-configured production strategy templates, and each strategy template defines the following mathematical mapping relationship: Strategy S_i = f_i(order queue Q, current inventory I, equipment utilization rate U, delivery cycle set D), where different strategy functions f_i correspond to different production objective preferences. The weight vectors for order queue Q, current inventory I, equipment utilization rate U, and delivery cycle set D differ across strategy functions. These weight vectors define the optimization preference of the strategy function among multiple production objectives (such as order fulfillment, inventory health, equipment utilization, and on-time delivery), enabling the quantitative calculations of the strategy generator to automatically approach the core objective of the strategy. For example, the following are weight vectors set for different strategies: Strategy S_1 (Guaranteed Backup Inventory): Weight vector W = (0.3, 0.5, 0.1, 0.1), objective function is to maximize the inventory satisfaction rate in the next N days (N is a set number of days). It can be seen that the current inventory I has the highest weight of 0.5.

[0059] Strategy S_2 (Full Load Type): Weight vector W = (0.1, 0.1, 0.7, 0.1), objective function is to minimize device idle time. It can be seen that the device occupancy rate U has the highest weight of 0.7.

[0060] Strategy S_3 (Delivery Priority): Weight vector W = (0.4, 0.1, 0.1, 0.4), objective function is to minimize order delay penalty. It can be seen that the order queue Q and delivery cycle set D) both have a maximum value of 0.7.

[0061] Based on the strategy generator, the qualitative decisions (strategy selection and strategy parameters) output by the LLM are transformed into precise quantitative production instructions. Taking the LLM selection of strategy S_2 (equipment full load type) with strategy parameters {target_min_utilization: 0.85, max_stockpile_qty: 8000} as an example, the specific working process of the strategy generator is explained.

[0062] Assume the strategy generator receives the following raw data: 5000 orders to be produced, product model M1; production of this order requires 10 hours of capacity; total available capacity within the future production cycle is 20 hours; current equipment utilization is 10h / 20h = 50%; available inventory is product model M1, current inventory is 500 units, historical daily average consumption is 200 units; standard production rate is 1000 units / hour. The calculation steps are as follows: The first step is to calculate the capacity gap. To achieve the target utilization rate of 85% specified by the LLM, the target utilization time is 20 hours × 0.85 = 17 hours. Existing orders occupy 10 hours, therefore the capacity gap that needs to be filled is 17 - 10 = 7 hours. This step quantifies the vague state of "insufficient equipment" into a precise 7 hours of idle time.

[0063] The second step is the calculation of standby orders. The strategy generator calls its internal idle filling function to multiply the 7-hour idle window by the standard production rate of 1000 units / hour, calculating a standby order quantity of 7000 units. Constraint verification is then performed: 7000 units is less than the `max_stockpile_qty` (8000 units) given by the LLM, and does not exceed the standby order limit, so the verification passes. If the calculated standby order quantity exceeds this limit, the limit is used to avoid overproduction.

[0064] It should be noted that during the standby order calculation process, the strategy generator makes the following trade-offs based on the weight vector: When there are multiple optional standby production plans (such as filling idle capacity, partially filling, or retaining idle capacity), the dimension with the highest value in the weight vector (such as the equipment occupancy rate weight of 0.7 in the equipment full-load strategy) will dominate the overall score, so that the plan that best improves the indicator of that dimension will be executed first. For example, in actual production, there are usually multiple standby plans to fill the same idle window. Since the strategy generator adopts the equipment full-load strategy, even if increasing inventory will slightly affect the inventory health under the equipment full-load strategy, the strategy generator will still choose the plan that completely fills the idle capacity to maximize equipment utilization because the inventory weight is only 0.1.

[0065] The third step is to generate production scheduling information. The final production scheduling information output by the strategy generator is as follows: 5,000 sales orders, estimated to take 10 hours; 7,000 backup orders, estimated to take 7 hours; total production of 12,000 orders, total capacity utilization of 17 hours, and equipment utilization rate of 85%.

[0066] Finally, step S15 is executed to return the production scheduling information to the ERP system.

[0067] Production scheduling information includes a clear quantity of production orders and a quantity of standby orders, which is returned to the ERP system in the form of structured data. The ERP system then uses this information to generate detailed workshop operation instructions.

[0068] In summary, this invention enables hybrid production of sales orders and standby orders. By using LLM to identify idle periods where equipment utilization is below a preset threshold, it generates standby orders to fill capacity gaps and prevents inventory buildup through constraints such as standby limit limits in strategy parameters. Simultaneously, the combination of preset strategy templates in the strategy generator and dynamic strategy matching in LLM allows enterprises to flexibly switch between different operational objectives such as ensuring delivery time, reducing inventory, and improving utilization based on real-time order status, inventory levels, and equipment usage. This eliminates the need for retraining the model or significantly adjusting prompts, and the decision-making process is transparent and traceable. Ultimately, delivery efficiency is improved, and the standby strategy shifts from passive prediction to proactive strategy-driven idle filling using LLM. This assists ERP in achieving intelligent linkage from order receipt to production execution, making it particularly suitable for scenarios with highly fragmented orders, such as small-batch, multi-variety smart card manufacturing.

[0069] Computer device embodiment: The computer device in this embodiment includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the above-described embodiment of the order balancing hybrid production method based on strategy generator and LLM.

[0070] A computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that a computer device may include more or fewer components, or a combination of certain components, or different components; for example, a computer device may also include input / output devices, network access devices, buses, etc.

[0071] For example, a processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microcontroller or any conventional processor. The processor is the control center of a computer device, connecting all parts of the computer device through various interfaces and lines.

[0072] The memory can be used to store computer programs and / or modules. The controller implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. For example, the memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (e.g., sound receiving function, sound-to-text function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (e.g., audio data, text data, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0073] Examples of computer-readable storage media: If the modules integrated into the computer device in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes of the embodiment of the order balancing hybrid production method based on a strategy generator and LLM can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a controller, it can implement the steps of the above embodiment of the order balancing hybrid production method based on a strategy generator and LLM. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in computer-readable media may be appropriately added to or subtracted from the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, computer-readable media may not include electrical carrier signals and telecommunication signals, in accordance with legislation and patent practice.

[0074] Finally, it should be emphasized that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hybrid production method for order balancing based on strategy generator and LLM, characterized in that, Includes the following steps: Obtain order information, inventory information, and equipment capacity information; Based on preset prompt words, the order information, inventory information and equipment capacity information are processed through a large language model to obtain strategy selection and strategy parameters; The strategy generator processes the strategy selection and strategy parameters, and returns production scheduling information.

2. The order balancing hybrid production method based on strategy generator and LLM as described in claim 1, characterized in that: The order information, inventory information, and equipment capacity information, after feature extraction and standardization, are processed by the large language model to obtain the strategy selection and the strategy parameters.

3. The order balancing hybrid production method based on strategy generator and LLM as described in claim 1, characterized in that: The feature extraction and standardization operations include: determining whether the confidence level of the key fields of the currently processed order is lower than a set threshold; if so, diverting the currently processed order to a manual review queue.

4. The order balancing hybrid production method based on strategy generator and LLM as described in claim 1, characterized in that: The strategy generator includes pre-configured production strategy templates of different types; The production strategy template defines the following mathematical mapping relationship: strategy S_i = f_i(order queue Q, current inventory I, equipment utilization rate U, delivery cycle set D), where different strategy functions f_i correspond to different production target preferences.

5. The order balancing hybrid production method based on strategy generator and LLM as described in claim 4, characterized in that: The types of production strategy templates include: guaranteed inventory type, equipment full capacity type, and delivery date priority type; The weight vectors and objective functions of the different types of production strategy templates are different. The weight vectors are used to indicate the weights of the order queue, the current inventory, the equipment utilization rate, and the delivery cycle set. The objective function of the backup inventory type includes maximizing the inventory fulfillment rate for a set number of future days. The objective function of the equipment full-load type includes minimizing the equipment idle time. The objective function of the delivery priority type includes minimizing the order delay penalty.

6. The order balancing hybrid production method based on strategy generator and LLM as described in claim 1, characterized in that: The production scheduling information is then returned to the ERP system.

7. The order balancing hybrid production method based on strategy generator and LLM as described in claim 1, characterized in that: The production scheduling information includes the number of production orders and the number of backup orders.

8. The order balancing hybrid production method based on strategy generator and LLM as described in claim 1, characterized in that: The preset prompts include environment coding and feature abstraction prompts, multi-strategy adaptation scoring prompts, and output format requirements.

9. A computer device comprising a processor and a memory, characterized in that: The memory stores a computer program, which, when executed by the processor, implements the order balancing hybrid production method based on strategy generator and LLM as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the order balancing hybrid production method based on strategy generator and LLM as described in any one of claims 1 to 8.

Images