A natural language scheduling intention recognition method for discrete manufacturing workshops

Abstract

This invention discloses a natural language scheduling intent recognition method for discrete manufacturing workshops, comprising: constructing a multi-turn dialogue thought chain dataset for discrete manufacturing workshop scheduling; constructing an intent classification system for discrete manufacturing workshop scheduling, defining work order-type operation intents and equipment-type operation intents; fine-tuning a large language model using the multi-turn dialogue thought chain dataset to determine the intent type to which the user request belongs; extracting product name, order quantity, process flow, start time, priority, and operation target from natural language and generating structured parameters based on the identified intent type; performing fuzzy matching and knowledge base completion on the extracted structured parameters, and retrieving and completing missing process information from the knowledge base. This invention, by constructing a multi-turn dialogue thought chain dataset covering six scenario types and various dialogue turn distributions, explicitly annotates the reasoning process, enabling the model to internalize the intent recognition standards of the scheduling domain, thus lowering the operational threshold of the production scheduling system.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and natural language processing technology, specifically to a natural language scheduling intent recognition method for discrete manufacturing workshops. Background Technology

[0002] Production scheduling in discrete manufacturing workshops is characterized by diverse product types, small batches, and complex processes, requiring comprehensive consideration of multiple factors such as equipment capacity, order priority, and process constraints. Currently, workshop scheduling mainly relies on professional schedulers using form interfaces and optimization algorithms, resulting in long response cycles and high operational barriers, making it difficult to meet the rapid response needs in dynamic scenarios such as temporary order insertions and equipment malfunctions.

[0003] With the development of large language model technology, the application of natural language interaction in intelligent manufacturing has attracted widespread attention. Intent recognition and slot filling technologies are relatively mature in general dialogue systems such as intelligent customer service and voice assistants. Existing research attempts to apply large language models to manufacturing execution systems to issue scheduling instructions in natural language, or to combine them with reinforcement learning to optimize job shop scheduling problems. However, the application of existing large language models in production scheduling still has obvious limitations: on the one hand, large language models are essentially language models rather than professional optimization algorithms, and are prone to "illusion" phenomena, potentially generating scheduling schemes that do not conform to actual constraints; on the other hand, the model output results lack consistency and predictability, and are highly dependent on high-quality historical production data for training.

[0004] Applying general intent recognition technology directly to discrete manufacturing workshop scheduling scenarios still presents the following specific problems: First, the existing general intent classification system cannot cover the professional operation types of workshop scheduling, such as emergency order insertion, sequence adjustment, order merging and splitting, resulting in insufficient intent recognition accuracy; Second, there is a lack of dedicated training data for the workshop scheduling domain, making it difficult for general large language models to understand scheduling terminology and business logic, resulting in poor performance in actual deployment; Third, existing technologies lack end-to-end intent recognition and parameter extraction mechanisms for workshop scheduling scenarios, making it difficult to achieve a complete conversion from natural language input to structured scheduling instructions.

[0005] Therefore, developing a natural language scheduling intent recognition method that is compatible with discrete manufacturing workshop scenarios, including dedicated intent classification, domain data fine-tuning, fuzzy matching completion, and multi-turn dialogue compatibility, has become an urgent technical problem to be solved in the field of intelligent manufacturing. Summary of the Invention

[0006] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a natural language scheduling intent recognition method for discrete manufacturing workshops. This method enables end-to-end conversion from operator natural language input to executable structured scheduling instructions, thereby reducing the threshold for scheduling operations and improving the dynamic scheduling response speed and stability.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a natural language scheduling intent recognition method for discrete manufacturing workshops, comprising the following steps:

[0009] S1. Construct a multi-turn dialogue thinking chain dataset for discrete manufacturing workshop scheduling. Extract real work order data from the workshop's historical production scheduling database, design multiple scenario types and various dialogue round distributions, and explicitly label the thinking chain reasoning process in the dialogue samples.

[0010] S2. Construct an intent classification system for discrete manufacturing workshop scheduling, and define work order-type operation intents and equipment-type operation intents;

[0011] S3. Fine-tune the large language model using a multi-turn dialogue thought chain dataset, receive natural language input from operators and perform intent recognition to determine the intent type to which the user's request belongs;

[0012] S4. Based on the identified intent type, extract product name, order quantity, process flow, start time, priority, and operation target from natural language and generate structured parameters;

[0013] S5. Perform fuzzy matching and knowledge base completion on the extracted structured parameters, match the product abbreviation to the standard complete product name, and retrieve and complete the missing process information from the knowledge base.

[0014] As a preferred technical solution, the method for constructing the multi-turn dialogue thought chain dataset in step S1 is as follows:

[0015] S101. Data preparation: Extract historical work order data from the workshop scheduling database. Each work order includes product name, product number, order quantity, start time, and key fields of manufacturing process.

[0016] S102. Scenario Division: Divide the dataset into six categories: basic order placement scenario, parameter change scenario, production scheduling intervention scenario, information query scenario, process control scenario, and exception handling scenario, and allocate the sample proportion according to the preset ratio.

[0017] S103, Round Setting: Divide the dialogue into single-round dialogue, 2-3 round short dialogue, and 4 or more round long dialogue, and distribute them according to a preset ratio;

[0018] S104. Sample generation: Using the template filling method, real work order parameters are filled into a preset dialogue template to generate dialogue samples.

[0019] S105, Mind Chain Annotation: In each sample, the complete reasoning chain is annotated sequentially, including the task intent type, key information extraction, knowledge base matching results, missing information completion content, and finally, the JSON structured output.

[0020] S106. Quality Verification: Ensure that the dialogue logic, parameter format, and reasoning steps are compliant and effective through manual spot checks.

[0021] As a preferred technical solution, in step S102, the prime number-based order placement scenario is used to train the creation and confirmation of a single order;

[0022] The parameter change scenario is used to train operations that modify the quantity and time parameters;

[0023] The production scheduling intervention scenario is used to train operators on emergency order insertion and specified order operations;

[0024] The information query scenario is used to train order and device status queries;

[0025] The process control scenario is used to train and confirm execution and undo operations;

[0026] The aforementioned anomaly handling scenario is used to train fuzzy matching and slot completion for anomalies.

[0027] As a preferred technical solution, in step S2, the work order operation intentions include: adding an order, inserting an emergency order, adjusting the order, deleting an order, modifying the quantity, modifying the time, merging orders, and splitting orders;

[0028] The device-related operation intentions are: status query, device query, confirmation of execution, and cancellation of operation.

[0029] As a preferred technical solution, in step S3, the fine-tuning of the large language model using a multi-turn dialogue thought chain dataset specifically involves:

[0030] First, all the original pre-trained weights of the large language model are frozen. Only in the attention layer of the Transformer structure of the model, the multi-turn dialogue thought chain dataset is organized into standardized Prompt training samples of dialogue history, user natural language requests, intent labels, and thought chain reasoning process. A domain-adaptive fine-tuning training set is constructed. Then, with the intent classification cross-entropy loss and the thought chain reasoning generation loss as joint optimization objectives, the low-rank trainable matrix is ​​iteratively trained to complete the exclusive adaptation of the large language model in the discrete manufacturing workshop scheduling scenario.

[0031] As a preferred technical solution, in step S3, the semantic equivalence principle is adopted to determine the intent type to which the user request belongs. When the intent output by the large language model is consistent with the expected intent semantics, the intent recognition result is determined to be correct.

[0032] As a preferred technical solution, in step S4, the structured parameter extraction adopts the slot filling method, defining the following slot fields: action field represents the operation type, product_name field represents the product name, quantity field represents the order quantity, process_flow field represents the process flow, start_date field represents the start time, priority field represents the priority, target field represents the operation target, and optimization_goal field represents the scheduling optimization target.

[0033] As a preferred technical solution, in step S5, the fuzzy matching adopts a two-level matching strategy, specifically:

[0034] First, a string similarity algorithm is used to calculate the similarity between the input and all product names in the knowledge base. The product with the highest similarity that exceeds the preset threshold is selected as the matching result.

[0035] If no match is found by string similarity, semantic similarity is further calculated to find the semantically most similar product by utilizing the semantic representation capability of the pre-trained language model.

[0036] As a preferred technical solution, step S5, knowledge base completion, specifically involves:

[0037] When the user does not provide complete process information, the system will retrieve the default process flow for the product from the standard process library based on the product name and automatically complete the missing process information.

[0038] As a preferred technical solution, step S5 also includes a multi-turn dialogue context understanding mechanism:

[0039] When the user enters "same process", the process flow from the previous order is retrieved from the recent conversation history and applied to the current order; when the user enters "add N more" without specifying a product name, the product name from the previous order is retrieved from the conversation history and applied to the current order; when the user enters "delete it" without explicitly specifying a target, the operation target is inferred from the conversation context.

[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0041] This invention provides a natural language scheduling intent recognition method for discrete manufacturing workshops. First, it proposes a dedicated intent classification system for workshop scheduling, which, compared to the intent classification of general dialogue systems, can accurately identify workshop-specific scheduling operations such as urgent order insertion, order adjustment, and order merging / splitting. Second, through the construction and model fine-tuning of a multi-turn dialogue thought chain dataset, the large language model learns professional knowledge and reasoning patterns in the workshop scheduling domain, effectively avoiding the "illusion" problem of general large language models and improving the consistency and reliability of output results. Third, the end-to-end intent recognition and parameter extraction mechanism, combined with fuzzy matching, knowledge base completion, and contextual understanding, effectively solves the problem of non-standard operator expressions in the workshop environment, achieving a complete conversion from natural language input to structured scheduling instructions, and significantly improving the response efficiency of workshop scheduling.

[0042] In summary, the intent recognition method provided by this invention enables workshop operators to interact directly with the scheduling system through natural language without having to learn a complex operating interface. This lowers the threshold for production scheduling operations, improves the workshop's response speed to unexpected production events, and has significant practical application value for promoting the intelligent upgrading of discrete manufacturing workshops. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is an overall flowchart of the intent recognition method in this embodiment, showing the complete processing flow from natural language input to structured output, including steps such as intent classification, parameter extraction, fuzzy matching, knowledge base completion, and context association.

[0045] Figure 2 This is a flowchart of the multi-turn dialogue thought chain dataset construction process in this embodiment, showing the complete construction process of six types of scenario design, dialogue turn distribution, template filling generation, and thought chain annotation.

[0046] Figure 3 This is a classification system diagram of 12 intent categories in this embodiment, showing two major categories: work order operations (8 types) and equipment operations (4 types), as well as the specific intent types and typical trigger expressions contained therein.

[0047] Figure 4 This is a schematic diagram of the parameter structure extraction and knowledge base completion in this embodiment, showing the definition of the 8 slot fields and the implementation of fuzzy matching and knowledge base completion. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0049] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0050] like Figure 1 As shown in the figure, this embodiment provides a natural language scheduling intent recognition method for discrete manufacturing workshops, which includes the following steps:

[0051] Step S1: Construct a multi-turn dialogue thinking chain dataset for discrete manufacturing workshop scheduling. Extract real work order data from the workshop's historical production scheduling database, design six types of scenarios and multiple dialogue turn distributions, and explicitly label the thinking chain reasoning process in the dialogue samples for domain fine-tuning of the large language model.

[0052] Furthermore, such as Figure 2 As shown, the method for constructing the multi-turn dialogue thought chain dataset in step S1 specifically includes:

[0053] Step S101: Prepare real work order data.

[0054] Specifically, historical work order data is extracted from the scheduling database of the discrete manufacturing workshop. Each work order contains key fields such as product name, product number, order quantity, order start time, and product manufacturing process, covering a variety of discrete processes such as cutting, engraving, printing, UV treatment, and packaging.

[0055] Step S102: Design of six types of scenarios.

[0056] Specifically, six scenario types are designed to cover various scheduling operations in actual production. The first type is the basic order placement scenario, accounting for 25%, used to train the creation and confirmation of a single order, covering the new_order intent; the second type is the parameter change scenario, accounting for 20%, used to train operations such as modifying parameters such as quantity and time, covering the modify, merge, and split intents; the third type is the scheduling intervention scenario, accounting for 20%, used to train operations such as urgent order insertion and specifying order, covering the urgent, reorder, and delete intents; the fourth type is the information query scenario, accounting for 15%, used to train order and equipment status queries, covering the status and machine intents; the fifth type is the process control scenario, accounting for 10%, used to train confirmation and cancellation operations, covering the confirm and cancel intents; and the sixth type is the exception handling scenario, accounting for 10%, used to train fuzzy matching and slot completion for exceptions such as product not found and incomplete information.

[0057] Step S103: Dialogue turn distribution design.

[0058] Specifically, the design incorporates dialogue rounds of varying lengths to simulate real-world human-computer interaction. Single-turn dialogues account for approximately 15%, used for simple queries and one-time order placement scenarios; short-range dialogues (2-3 rounds) account for approximately 55%, used for typical processes of order placement, modification, and confirmation; and long-range dialogues (4 rounds or more) account for approximately 30%, used for scenarios involving multiple modifications and complex adjustments.

[0059] Step S104: Template filling generation.

[0060] Specifically, multiple dialogue templates are pre-set for each scenario type. Work orders are randomly selected from a real work order database, and parameters such as product name, quantity, time, and process are filled into the corresponding positions in the template. To ensure data diversity, each work order is used only once to avoid overfitting caused by data duplication.

[0061] Step S105: Mind chain annotation.

[0062] Specifically, the reasoning process of the thought chain is explicitly labeled in the dialogue samples, with the label format using specific tags to wrap the reasoning content. The reasoning steps include: first, identifying the type of intent of the user input; second, extracting key information from the input; third, matching and completing the missing information using a knowledge base; and finally, generating a compliant JSON output. By explicitly labeling the reasoning chain, the model learns the complete reasoning path from user input to structured output during fine-tuning.

[0063] Step S106: Quality control.

[0064] Specifically, manual sampling is used to verify the correctness of the dialogue logic, JSON format, and thought chain reasoning steps to ensure the quality of the dataset.

[0065] Step S2: Construct an intent classification system for discrete manufacturing workshop scheduling, and define work order-type operation intents and equipment-type operation intents;

[0066] Furthermore, such as Figure 3 As shown, in this embodiment, the work order type operation intents include eight types: adding an order, urgent order insertion, order adjustment, order deletion, quantity modification, time modification, order merging, and order splitting; the equipment type operation intents include four types: status query, equipment query, confirmation of execution, and cancellation. The specific definitions are shown in Table 1 below:

[0067] Table 1

[0068] category Intent Name Intent identifier Typical expression Work order operation New orders new_order Add one more order for 1000 rabbit puzzles, to be made tomorrow. Work order operation Emergency order urgent_order Urgently add 500 pandas, rush order, express delivery Work order operation Order adjustment reorder Put the panda first and move the rabbit to the back. Work order operation Cancel order delete Delete the tiger; I don't want this anymore. Work order operation Number of modifications modify_quantity Change the number of rabbits to 1500, then add 2000 more. Work order operation Modified time modify_time Postponed to 2 PM, changed to tomorrow. Work order operation Order consolidation merge Combine two batches of rabbits into one order. Work order operation Order splitting split Divided into two batches of 500 each Equipment operation Status query status_query Where is Panda now? What orders are still pending? Equipment operation Equipment Inquiry machine_query What is Cutting 01 doing? Which machine is idle? Equipment operation Confirm Execution confirm Okay, let's start. That's it. Equipment operation Cancel operation cancel No, withdraw; the previous one doesn't count.

[0069] Specifically, there are eight types of work order operation intentions:

[0070] (1) Add a new order (new_order), create a new production order. Typical expressions include "Add an order for 1,000 rabbit puzzles to be made tomorrow" and "Make 3,000 pandas";

[0071] (2) Urgent order insertion (urgent_order): High-priority order insertion, typically triggered by keywords such as "urgently add 500 pandas", "insertion order", and "urgent".

[0072] (3) Reordering: Adjusting the order of order execution. Typical expressions include "put the panda first" and "move the rabbit to the back."

[0073] (4) Delete order (delete) Remove the specified order. Typical expressions include "Delete the tiger" and "I don't want this anymore";

[0074] (5) Modify quantity: Change the order quantity. Typical expressions include "Change the number of rabbits to 1,500" and "Add another 2,000".

[0075] (6) Modify time (modify_time), which changes the start time. Typical expressions include "postponed to 2 p.m." and "changed to tomorrow";

[0076] (7) Order merging: Merging similar orders. Typical expressions include "merging two batches of rabbits into one order";

[0077] (8) Order splitting: splitting large orders, typically expressed as "split into two batches of 500 each".

[0078] Device operation intentions include four types:

[0079] (9) Status query (status_query): Query the order status. Typical expressions include "Where is Panda now?" and "What orders are still pending?"

[0080] (10) Machine query (machine_query) queries the status of the equipment. Typical expressions include "What is Cutting 01 doing" and "Which machine is idle";

[0081] (11) Confirm execution: Confirm the current plan. Typical expressions include "Okay, let's start" and "That's it."

[0082] (12) Cancel operation: Undo the previous operation. Typical expressions include "No, undo" and "That one doesn't count".

[0083] Step S3: Use the multi-turn dialogue thought chain dataset to fine-tune the large language model using LoRA. The specific implementation process is as follows: First, freeze all the original pre-trained weights of the large language model. Only inject a low-rank trainable matrix into the attention layer of the Transformer structure of the model to significantly reduce the number of fine-tuning parameters and computational overhead. Organize the multi-turn dialogue thought chain dataset into standardized Prompt training samples of dialogue history, user natural language requests, intent labels, and thought chain reasoning process to construct a domain-adapted fine-tuning training set. Then, use the intent classification cross-entropy loss and the thought chain reasoning generation loss as joint optimization objectives to iteratively train the low-rank trainable matrix to complete the exclusive adaptation of the large language model in the discrete manufacturing workshop scheduling scenario. Finally, receive the operator's natural language input and perform intent recognition to determine the intent type to which the user's request belongs.

[0084] Furthermore, the intent recognition adopts the semantic equivalence principle: when the intent identifier output by the model is semantically consistent with the predefined identifier, it is determined that the recognition is correct, and the existence of synonyms is allowed.

[0085] S4. Based on the identified intent type, extract the product name, order quantity, process flow, start time, priority, and operation objective from natural language and generate structured parameters, as specifically defined in Table 2 below:

[0086] like Figure 4As shown, the structured parameter extraction in step S4 uses a slot-filling method, defining 8 slot fields: the action field is the operation type, a required field that must match one of the 12 predefined intent type identifiers; the product_name field is the product name, supporting fuzzy matching and knowledge base completion; the quantity field is the order quantity, which must be a positive integer; the process_flow field is the process flow, in the format "process 1-process 2-process 3"; the start_date field is the start time, in the format "YYYY-MM-DD HH:MM:SS"; the priority field is the priority, 1 indicates urgent, 2 indicates normal, and the default value is 2; the target field is the operation target, used to specify the object for operations such as modify and delete; the optimization_goal field is the scheduling optimization target, with optional values ​​including makespan, utilization, flow_time, and balanced.

[0087] Table 2

[0088] field name Data types illustrate action string The operation type must exactly match the predefined API name. product_name string Product name, supports fuzzy matching and knowledge base completion. quantity integer The order quantity must be a positive integer. process_flow string The process flow is formatted as "Process 1 - Process 2 - Process 3". start_date datetime Start time, in the format "YYYY-MM-DD HH:MM:SS" priority integer Priority: 1 = Urgent, 2 = Normal target string Operation target, used to specify the object for operations such as modify / delete. optimization_goal string The strategy value is madespan / utilization / flow_time / balanced

[0089] S5. Perform fuzzy matching and knowledge base completion on the extracted parameters, match the product abbreviation to the complete product name, and retrieve and complete the missing process information from the knowledge base;

[0090] Furthermore, in step S5, fuzzy matching and knowledge base completion are specifically as follows: the fuzzy matching mechanism adopts a two-level matching strategy. First, a string similarity algorithm (such as edit distance) is used to calculate the similarity between the input and all product names in the knowledge base, and the product with the highest similarity and exceeding the preset threshold is selected as the matching result. If the string similarity fails to find a match, semantic similarity calculation is further used to find the semantically most similar product by utilizing the semantic representation capability of the pre-trained language model.

[0091] Furthermore, in step S5, the knowledge base completion mechanism retrieves the default process flow for the product from the standard process library based on the product name. The standard process library pre-stores standard manufacturing process chains for various products.

[0092] Furthermore, step S5 also includes a multi-turn dialogue context understanding mechanism, specifically: the system maintains a dialogue history in memory, recording the user's historical inputs and the system's historical responses. Specific processing rules include: when the user inputs "same process," the system retrieves the process flow from the most recent dialogue history and applies it to the current order; when the user inputs "add N more" without specifying a product name, the system retrieves the product name from the previous order and applies it to the current order; when the user inputs "delete it" without explicitly specifying a target, the system infers the operation target based on the dialogue context.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A natural language scheduling intent recognition method for discrete manufacturing workshops, characterized in that, Includes the following steps: S1. Construct a multi-turn dialogue thinking chain dataset for discrete manufacturing workshop scheduling. Extract real work order data from the workshop's historical production scheduling database, design multiple scenario types and various dialogue round distributions, and explicitly label the thinking chain reasoning process in the dialogue samples. S2. Construct an intent classification system for discrete manufacturing workshop scheduling, and define work order-type operation intents and equipment-type operation intents; S3. Fine-tune the large language model using a multi-turn dialogue thought chain dataset, receive natural language input from operators and perform intent recognition to determine the intent type to which the user's request belongs; S4. Based on the identified intent type, extract product name, order quantity, process flow, start time, priority, and operation target from natural language and generate structured parameters; S5. Perform fuzzy matching and knowledge base completion on the extracted structured parameters, match the product abbreviation to the standard complete product name, and retrieve and complete the missing process information from the knowledge base.

2. The natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S1, the method for constructing the multi-turn dialogue thought chain dataset is as follows: S101. Data preparation: Extract historical work order data from the workshop scheduling database. Each work order includes product name, product number, order quantity, start time, and key fields of manufacturing process. S102. Scenario Division: Divide the dataset into six categories: basic order placement scenario, parameter change scenario, production scheduling intervention scenario, information query scenario, process control scenario, and exception handling scenario, and allocate the sample proportion according to the preset ratio. S103, Round Setting: Divide the dialogue into single-round dialogue, 2-3 round short dialogue, and 4 or more round long dialogue, and distribute them according to a preset ratio; S104. Sample generation: Using the template filling method, real work order parameters are filled into a preset dialogue template to generate dialogue samples. S105, Mind Chain Annotation: In each sample, the complete reasoning chain is annotated sequentially, including the task intent type, key information extraction, knowledge base matching results, missing information completion content, and finally, the JSON structured output. S106. Quality Verification: Ensure that the dialogue logic, parameter format, and reasoning steps are compliant and effective through manual spot checks.

3. The natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S102, the prime number-based order placement scenario is used to train the creation and confirmation of a single order; The parameter change scenario is used to train operations that modify the quantity and time parameters; The production scheduling intervention scenario is used to train operators on emergency order insertion and specified order operations; The information query scenario is used to train order and device status queries; The process control scenario is used to train and confirm execution and undo operations; The aforementioned anomaly handling scenario is used to train fuzzy matching and slot completion for anomalies.

4. The natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S2, the work order type operation intentions include: adding an order, inserting an urgent order, adjusting the order, deleting an order, modifying the quantity, modifying the time, merging orders, and splitting orders; The device-related operation intentions are: status query, device query, confirmation of execution, and cancellation of operation.

5. The natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S3, the fine-tuning of the large language model using a multi-turn dialogue thought chain dataset specifically involves: First, all the original pre-trained weights of the large language model are frozen. Only in the attention layer of the Transformer structure of the model, the multi-turn dialogue thought chain dataset is organized into standardized Prompt training samples of dialogue history, user natural language requests, intent labels, and thought chain reasoning process. A domain-adaptive fine-tuning training set is constructed. Then, with the intent classification cross-entropy loss and the thought chain reasoning generation loss as joint optimization objectives, the low-rank trainable matrix is ​​iteratively trained to complete the exclusive adaptation of the large language model in the discrete manufacturing workshop scheduling scenario.

6. A natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S3, the semantic equivalence principle is used to determine the intent type to which the user request belongs. When the intent output by the large language model is consistent with the expected intent semantics, the intent recognition result is determined to be correct.

7. A natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S4, the structured parameter extraction adopts the slot filling method, defining the following slot fields: action field represents the operation type, product_name field represents the product name, quantity field represents the order quantity, process_flow field represents the process flow, start_date field represents the start time, priority field represents the priority, target field represents the operation target, and optimization_goal field represents the scheduling optimization target.

8. A natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S5, fuzzy matching employs a two-level matching strategy, specifically: First, a string similarity algorithm is used to calculate the similarity between the input and all product names in the knowledge base. The product with the highest similarity that exceeds the preset threshold is selected as the matching result. If no match is found by string similarity, semantic similarity is further calculated to find the semantically most similar product by utilizing the semantic representation capability of the pre-trained language model.

9. A natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, In step S5, the knowledge base completion is specifically as follows: When the user does not provide complete process information, the system will retrieve the default process flow for the product from the standard process library based on the product name and automatically complete the missing process information.

10. A natural language scheduling intent recognition method for discrete manufacturing workshops according to claim 1, characterized in that, Step S5 also includes a multi-turn dialogue context understanding mechanism: When the user enters "same process", the process flow of the previous order is retrieved from the recent conversation history and applied to the current order; when the user enters "add N more" without specifying a product name, the product name of the previous order is retrieved from the conversation history and applied to the current order; when the user enters "delete it" without explicitly specifying a target, the operation target is inferred from the conversation context.

Images