Intelligent production scheduling and exception management method based on large language model

Abstract

The invention provides an intelligent production scheduling and exception management method based on a large language model, and relates to the technical field of artificial intelligence, and the method comprises the steps: converting a user request into structured data through a semantic parser, and complementing fuzzy information in combination with context; modeling production data by adopting an attention-based relational graph neural network, and retrieving related context information in combination with a retrieval enhancement generation framework; adjusting a scheduling strategy in real time according to user feedback and production environment changes; and generating a scheduling instruction by using a large language model in combination with the structured data and the scheduling strategy. Through dynamic knowledge base construction, hierarchical memory mechanism design, attention-based relational graph neural network modeling and semantic understanding ability of a large language model, the intelligent level, logicality and adaptability of the production scheduling system are significantly improved.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to an intelligent production scheduling and exception management method based on a large language model. Background Art

[0002] With the rapid development of intelligent manufacturing technology, the complexity of production tasks is increasing. How to achieve efficient production scheduling has become a key issue in the field of intelligent manufacturing. Traditional production scheduling methods usually rely on static rules and predefined processes. Although they perform well in structured and less changing environments, they often show obvious limitations when facing dynamic and changing production scenarios. These methods are difficult to quickly adapt to sudden changes in production tasks, such as equipment failures, process adjustments, or temporarily inserted emergency tasks, resulting in reduced production efficiency and waste of resources.

[0003] In recent years, artificial intelligence technology, especially natural language processing (NLP) and deep learning methods based on large models, has been increasingly used in production scheduling. These methods can automatically generate scheduling plans and improve scheduling efficiency by analyzing massive amounts of data. However, the current intelligent scheduling system still faces the following major problems: In a complex production environment, there are high dependencies and sequential logic between different tasks, but existing scheduling systems often find it difficult to capture these relationships, resulting in task conflicts or logical confusion in the generated scheduling plans. Production tasks usually require reference to multiple rounds of instructions and historical data, but existing methods are limited in their processing of context, especially in capturing key information hidden in the context. For example, production instructions may have clearly stated key conditions or restrictions in early communication, but because the expression becomes increasingly vague or colloquial in subsequent communication, the system ignores these key information, thereby misleading the generated scheduling plan. Existing scheduling systems are difficult to respond to changes in the production environment in real time, such as task priority adjustments, equipment status changes, or sudden anomalies, and lack flexibility and dynamic adjustment capabilities. Production data is usually scattered in multiple heterogeneous systems, with complex data structures and insufficient semantic relevance. Existing methods find it difficult to effectively integrate and manage this data, resulting in limited scheduling accuracy. Summary of the invention

[0004] The present application provides an intelligent production scheduling and exception management method based on a large language model to solve one of the above technical problems.

[0005] The technical solution adopted in this application is: The embodiment of the present application provides an intelligent production scheduling and exception management method based on a large language model, including: The semantic parser converts user requests into structured data and completes ambiguous information based on the context. An attention-based graph neural network is used to model production data, and a retrieval-enhanced generation framework is used to retrieve relevant contextual information. Adjust scheduling strategies in real time based on user feedback and changes in the production environment; A large language model is used in combination with the structured data and the scheduling policy to generate scheduling instructions.

[0006] According to an embodiment of the present application, the context information is stored and managed through a three-layer structure of a global planning layer, a process logic layer, and an operation step layer.

[0007] According to an embodiment of the present application, the user request is converted into structured data by a semantic parser, and the ambiguous information is completed in combination with the context, specifically: Extract key parameters through a semantic parser, and combine historical context information in a hierarchical memory mechanism to complete the missing conditions in the user request; If the user request is insufficient, guide the user to add details through a prompt mechanism; The supplemented key parameters are output as the structured data.

[0008] According to an embodiment of the present application, the attention-based relational graph neural network is used to model the production data, and the relevant context information is retrieved in combination with the retrieval enhancement generation framework, specifically: Integrate the production data from different sources into a knowledge graph, where nodes represent production entities and edges represent relationships between entities; Matching the user request with the node in the knowledge graph using a vectorized storage model; Attention-based graph neural network captures implicit associations and key dependencies between tasks.

[0009] According to an embodiment of the present application, the scheduling strategy is adjusted in real time according to user feedback and changes in the production environment, specifically: Constructing a reward signal through the user feedback and the production environment changes to guide the optimization of the scheduling strategy; Using a policy gradient-based algorithm, scheduling policy parameters are dynamically adjusted according to real-time data to maximize the cumulative reward.

[0010] According to an embodiment of the present application, the structured data and the scheduling strategy are combined to generate a scheduling instruction using a large language model, specifically: Generate the scheduling instruction by combining the structured data and the scheduling strategy using a large language model; Automatically adjust instruction content based on changes in user needs or environmental emergencies.

[0011] According to one embodiment of the present application, it also includes: Monitor production status in real time, quickly identify and classify anomalies, and dynamically adjust scheduling plans.

[0012] According to an embodiment of the present application, the real-time monitoring of production status, rapid identification and classification of anomalies, and dynamic adjustment of the scheduling plan are specifically as follows: Track the status of production equipment, task execution, and environmental parameters; Identify and classify abnormal situations in the production process and take appropriate measures based on the severity; Automatically generate response plans and dynamically adjust task scheduling to minimize the impact of exceptions on production.

[0013] A computer-readable storage medium stores a program, which implements the steps in the method when executed by a processor.

[0014] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein the processor implements the steps in the method when executing the program.

[0015] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows: This application significantly improves the intelligence, logic and adaptability of the production scheduling system through dynamic knowledge base construction, hierarchical memory mechanism design, attention-based relational graph neural network (R-GAT) modeling and semantic understanding capabilities of the large language model (LLM).

[0016] Through the hierarchical memory mechanism of layered storage, production tasks are divided into three levels: global planning, process logic and operation steps, ensuring that the sequential relationship, dependency structure and execution conditions of scheduling tasks are clear and controllable, avoiding task conflicts or logical confusion.

[0017] Combined with the contextual understanding capabilities of large language models, multiple rounds of interactions in production tasks are dynamically managed, and contextual information is stored in layers through a hierarchical memory mechanism to capture logical conditions and background information implicit in early instructions, thus avoiding scheduling errors caused by loss of contextual information.

[0018] Through the reinforcement learning optimization module, user feedback and production environment change information are collected in real time, and task priorities and scheduling strategies are dynamically adjusted, so that the system can flexibly respond to emergencies such as equipment failures and emergency task insertions, thereby improving the adaptability and efficiency of the production system.

[0019] The dynamic knowledge base is combined with the attention-based relational graph neural network (R-GAT) to deeply model the knowledge graph and capture the implicit associations and key dependencies between tasks. Through the retrieval-augmented generation (RAG) framework, the dynamic knowledge base supports real-time updates and efficient retrieval, providing accurate and reliable context data for complex scheduling tasks.

[0020] By utilizing the semantic generation capability of the large language model, combined with the dynamic knowledge base and reinforcement learning optimization results, we can dynamically generate production instructions with rigorous logic and detailed content. We support semantic completion of fuzzy input and priority adjustment of sudden tasks, ensuring that the generated instructions can be directly applied to actual production scenarios.

[0021] By real-time monitoring of equipment status, task progress and environmental parameters, combined with anomaly detection algorithms, problems can be quickly identified and classified, and scheduling plans can be adjusted through hierarchical memory mechanisms and dynamic knowledge bases to ensure the continuity of production tasks and minimize the impact of anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A flow chart of an intelligent production scheduling and exception management method based on a large language model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to more clearly illustrate the overall concept of the present application, a detailed description is given below in an illustrative manner in conjunction with the accompanying drawings.

[0024] In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present application is not limited by the specific embodiments disclosed below. It should be noted that the embodiments of the present application and the features in each embodiment may be combined with each other without conflict.

[0025] In the present application, unless otherwise clearly specified and limited, a first feature "above" or "below" a second feature may be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples.

[0026] Example 1 like Figure 1 As shown, an intelligent production scheduling and exception management method based on a large language model includes: The semantic parser converts user requests into structured data and combines the context to complete ambiguous information.

[0027] Specifically, first, multiple forms of input from the user are received, including but not limited to text, voice commands, and direct scheduling instructions. These inputs may include content about scheduling tasks (for example, "arrange maintenance of equipment A"), status queries (for example, "the current status of equipment B"), or process adjustments (such as "optimize production line processes"). Through semantic parsing technology, key elements are extracted from the user input, including task type (such as maintenance, inspection, production, etc.), parameters (such as equipment name, time requirements, priority, etc.), and the specific intention of the request is determined. Considering the completeness of complex instructions, a hierarchical memory mechanism is used to extract background information from previous rounds of interaction to supplement the key conditions or dependencies that may be omitted in the user's current input. For example, in a multi-round dialogue, the early communication may have clarified the prerequisites or restrictions of certain operations, but these details are not explicitly mentioned in the subsequent instructions. At this time, the context completion mechanism will come into play to ensure that all necessary background information is taken into account to avoid inaccurate instructions or logical confusion caused by missing information. After the above steps, the user's natural language request is converted into a structured instruction format, such as "Task type: equipment inspection; Equipment name: Equipment A; Priority: High; Time: Immediate". This structured representation facilitates further processing and use by subsequent modules.

[0028] When a user submits a request, the semantic parser first performs a preliminary analysis of the natural language request, attempting to extract key information such as the task type, parameter conditions, and priority. If some necessary information is missing or not clearly stated (for example, the device name is missing, the time requirement, etc.), the next step is entered. The information that is required but not provided is determined based on preset rules or a knowledge base based on historical data. For example, when scheduling a maintenance task, if the specific device name is not specified, this is considered an information gap. For the identified information gaps, the system automatically generates targeted prompts or questions to guide the user to provide the required specific information. These prompts can be in the form of direct questions, such as "Please specify the specific device name that needs maintenance", or more detailed instructions. Prompts can also be based on background information provided by the context completion mechanism. For example, if a device has been mentioned before as having a possible problem, and the current request seems to be related to this but is not clearly stated, the prompt may contain content such as "Did you mean device A mentioned last time?" After receiving the prompt, the user can fill in these information gaps through further input. This may include providing missing specific parameter values, clarifying intent, or confirming previous assumptions. The system will parse the user's additional input again and repeat the above process until all necessary information is accurately obtained. Finally, the system will conduct a comprehensive check of all collected information to ensure that nothing is missing and that all provided information is clear and consistent. If there are still questions or conflicts, the system will continue to request further clarification from the user through the feedback mechanism. Once it is confirmed that all relevant information is complete and accurate, the system will convert this information into a structured instruction format, ready for subsequent task execution or other operations.

[0029] The process of converting user requests into structured data through a semantic parser and completing fuzzy information in context is mainly to ensure that the system can accurately understand the user's intention and generate precise and executable instructions based on complete context information. This process not only improves the accuracy of instructions, but also enhances the system's ability to cope with changes in complex production environments.

[0030] For example, suppose a dispatcher at a manufacturing company wants to schedule a maintenance task for an equipment, but due to being busy or negligent, does not provide all the necessary details in the initial request. Here is how the process unfolds: Initial user request: "Schedule equipment maintenance." The system first tries to extract the key elements from the sentence. In this example, "Task Type: Equipment Maintenance" is identified, but important information such as the name of the specific equipment to be maintained, the priority, and the expected completion time are missing.

[0031] Given the incomplete information, the system uses a hierarchical memory mechanism to find relevant contextual information. For example, if a potential problem with a specific device (e.g., device A) was discussed in a previous interaction, the system might infer that the current request might be related to this device.

[0032] At the same time, the system may also refer to recent task records or known production plans to infer which equipment is most likely to require maintenance.

[0033] The system will automatically generate targeted questions or suggestions to fill in the missing information. "Do you mean to perform maintenance on device A? If so, please confirm the maintenance priority and expected completion time." Suppose the user replies, "Yes, it is maintenance on device A, with a high priority, and it would be best if it could be completed today." Based on the user's supplementary information, the system converts the original request into a detailed structured instruction: "Task type: equipment inspection; Equipment name: Equipment A; Priority: High; Time requirement: Complete within today." The system will conduct a comprehensive check on this structured instruction to ensure that all necessary information is included and consistent with each other. If everything is normal, the instruction will be passed to the dynamic knowledge base module for further retrieval of relevant context information to support subsequent task execution.

[0034] Through the above steps, even if the user's initial request is very brief or even ambiguous, the intelligent production scheduling system can still accurately capture the user's true intention through effective semantic parsing and context completion mechanisms, and generate precise and executable instructions. This approach not only improves the flexibility and adaptability of the system, but also greatly enhances the user experience.

[0035] Furthermore, in addition to the basic semantic parser, more advanced natural language processing (NLP) technologies, such as the large language model (LLM) of the Transformer architecture, can be introduced to improve the ability to understand complex sentence structures and contexts. External knowledge bases or industry-specific knowledge graphs can be used to enrich the system's contextual understanding capabilities, so that the system can not only understand information in the current field, but also understand and apply related knowledge across fields.

[0036] By analyzing the user's historical behavior and preferences, the system automatically adjusts the instruction generation strategy to provide more personalized services. For example, for users who often pay attention to efficiency, the system can prioritize the recommendation of quick-execution task arrangements. Based on the current production environment status, historical data and predictive analysis, the system provides users with the best operation suggestions or alternatives to help users make better decisions.

[0037] An attention-based graph neural network is used to model production data, and a retrieval-enhanced generation framework is combined to retrieve relevant contextual information.

[0038] Specifically, an embedding model (such as Transformer) is used to convert the nodes and relationships of the knowledge graph into vector representations, so that user requests can quickly locate relevant nodes and contextual information through vector matching.

[0039] The formula used in this process includes query embedding model , and nodes or documents in the knowledge base Embedding model converted to vector . Then cosine similarity is used to measure the relevance of user queries to knowledge base content.

[0040] In the retrieval phase, the RAG framework selects the k nodes most relevant to the user request to form a contextual supplementary dataset.

[0041] In the generation phase, context information and user input are combined to generate production scheduling instructions that better meet the needs of the actual scene. This approach effectively avoids the hallucination problem in traditional generation methods and improves the accuracy and relevance of the generated content.

[0042] The dynamic knowledge base works closely with R-GAT to form a complete closed loop from data organization to semantic modeling. The dynamic knowledge base ensures that the knowledge graph can accurately reflect the latest status in the production scenario by updating node attributes and relationship weights in real time; while R-GAT performs deep modeling based on the updated graph to capture implicit associations and task dependencies, thereby improving the semantic expression ability of the data.

[0043] The output retrieval results and node embeddings are directly passed to the hierarchical memory mechanism as the basic input for context management, which enhances the coherence of task logic and improves the system's adaptability to complex production tasks.

[0044] The combination of the attention-based relational graph neural network (R-GAT) and the retrieval-augmented generation (RAG) framework enables the system to not only capture deep semantic associations when processing complex production data, but also efficiently retrieve and generate relevant contextual information according to the user's specific needs, greatly improving the intelligence and flexibility of production scheduling.

[0045] For example, suppose in a manufacturing plant, a scheduler needs to schedule a new production line task, and this task depends on the status of other equipment and some previous process steps. Due to the complex production environment involving multiple equipment, processes and their interrelationships, an intelligent system is needed to help the scheduler make the best decision.

[0046] First, the dynamic knowledge base will integrate heterogeneous data from different sources into a production-based knowledge graph. For example: Nodes: represent different production entities, such as equipment A, equipment B, process step X, etc. Edges: represent the relationship between these entities, such as equipment A must start working after equipment B completes a certain operation; process step X is a key link in the process flow, etc.

[0047] Next, use R-GAT to perform deep modeling on this knowledge graph: The attention weights between each node and its neighbor nodes are calculated to highlight the connections that are critical to the current task. For example, when scheduling a new task, if the status of device A directly affects whether the new task can be carried out smoothly, a higher attention weight is given to device A. Based on the results of the attention weights, R-GAT will perform multiple embedding updates on each node, not only considering the information of local neighbors, but also reflecting the correlation between global tasks. For example, considering that the status of device A may affect the efficiency of the entire production chain, the system will comprehensively analyze the status of all relevant nodes to ensure the optimization of the overall scheduling plan.

[0048] When a user submits a new request, such as "arrange the best task sequence for a new production line", the system will perform the following steps: Use an embedding model (such as Transformer) to convert the user's request and the nodes in the knowledge graph into vector representations. Then use cosine similarity matching to find the nodes and context information most relevant to the user's query. For example, identify which equipment is currently idle, which process steps have not been completed but are critical to the new task, etc. The k most relevant nodes selected from the retrieval stage form a context supplementary data set. Then, the system combines this context information with user input to generate a detailed production scheduling instruction. For example, "Start equipment A first, and start the operation of equipment B immediately after it is completed; the total time is expected to be no more than 4 hours." In this process, R-GAT captures the implicit task dependency chain, such as the completion of equipment A is a prerequisite for the operation of equipment B; and RAG uses this information to efficiently locate and generate scheduling instructions that meet actual needs.

[0049] From the above examples, we can see that the combination of attention-based relational graph neural network (R-GAT) and retrieval-augmented generation (RAG) framework enables the intelligent production scheduling system to not only deeply understand the complex production environment and its internal relationships, but also quickly respond to the specific needs of users and provide accurate and efficient scheduling suggestions. This approach greatly improves the intelligence level and adaptability of production scheduling, and helps to achieve more flexible and efficient production management.

[0050] Furthermore, in addition to the existing structured modeling and semantic association analysis, time series analysis, machine learning models and other technologies can be introduced to predict possible future events, such as equipment failure warnings, raw material demand forecasts, etc., so as to prepare in advance and reduce downtime and costs. Integrate external knowledge bases from different fields (such as supply chain management, market demand analysis, etc.) into the current system, and capture the implicit relationships between these fields through R-GAT to provide more comprehensive support for production scheduling. In the instruction generation module, the format of the generated instructions (concise or detailed) can be customized according to the preferences of different users, or special instruction templates can be designed for the needs of specific industries to improve the applicability and user experience of the instructions.

[0051] Adjust scheduling strategies in real time based on user feedback and changes in the production environment.

[0052] Specifically, the system will collect direct feedback from users, including subjective opinions such as evaluation of the logic of instructions and the rationality of priority settings. In addition to user feedback, the system will also automatically collect actual execution data of production tasks, such as objective indicators such as task completion time and equipment utilization. Based on the above information, a reward function is designed to quantitatively evaluate the effectiveness of the current scheduling strategy. For example:

[0053] in, Indicates the task completion efficiency, Indicates resource utilization, Indicates user satisfaction, , , It is the weight coefficient, which is used to balance the importance of various indicators.

[0054] In the reinforcement learning framework, the production environment and task state are defined as states , and the dispatch instruction acts as an action .

[0055] A policy gradient-based method (such as the REINFORCE algorithm) is used to optimize the scheduling policy. Specifically, the policy Choose actions based on the current state to maximize the cumulative reward The policy update follows the following formula:

[0056] in, is the objective function of the strategy, is the discount factor, It's time t Instant rewards received. Use the latest feedback data to continuously adjust scheduling parameters, such as task priority and resource allocation schemes, to adapt to changes in the current production environment. When faced with different types of production environments, policy migration technology can be used to quickly adapt to the needs of the new environment while keeping the core scheduling principles unchanged. During the multi-round task generation process, if the existing scheduling scheme is found to be inefficient or wrong, the system can correct it through a dynamic optimization mechanism to ensure the consistency of the instruction logic with the actual production needs.

[0057] Continuously monitor the status of production equipment, task progress, and environmental parameters to identify any factors that may affect the production process. Once an abnormal situation (such as equipment failure, task delay) is found, the predefined response plan is immediately initiated and the task scheduling order is dynamically adjusted. Based on the new data collected and user feedback, the parameters in the reinforcement learning model are regularly updated to make the scheduling strategy more in line with the actual situation and improve the adaptability and efficiency of the overall system.

[0058] Through this mechanism, the intelligent production scheduling system can not only formulate a reasonable scheduling plan in the initial stage, but also make timely adjustments according to changes in external conditions during operation, ensuring the flexibility and stability of production. This method greatly improves the intelligence level of the system, enabling it to maintain efficient operation in a complex and changing industrial environment.

[0059] For example, suppose that in a manufacturing plant, the task scheduling of the production line depends on the status of multiple devices, the requirements of the process flow, and external factors (such as raw material supply). The system needs to be able to dynamically adjust the scheduling strategy based on real-time user feedback and changes in the production environment to ensure efficient production operation.

[0060] Current task arrangement: The production line plans to complete three main processes in sequence: A (preparing materials), B (assembling parts), and C (quality inspection).

[0061] Resource allocation: Equipment X is responsible for process A, equipment Y is responsible for process B, and equipment Z is responsible for process C.

[0062] Expected goal: All processes must be completed within the next 8 hours, and seamless connection between each link must be ensured.

[0063] The dispatcher provided feedback through the system interface: "The actual execution time of process B is 30% longer than expected, which may affect the on-time completion of subsequent processes." Changes in the production environment: The system monitoring module detects a minor fault in equipment Y, which reduces its efficiency by 25%. At the same time, the warehouse management system reports that the supply of raw materials is sufficient, but the transportation vehicle is delayed due to traffic jams, which may cause the specific material required for process A to arrive 1 hour late.

[0064] After receiving user feedback and information about changes in the production environment, the reinforcement learning optimization module begins to evaluate the impact of these changes on the overall production plan. For example, calculate the time delay of the entire production chain due to the reduced efficiency of equipment Y and the delayed arrival of raw materials. Adjust the parameters in the reward function according to the new situation. For example, increase the weight of completing process B on time, and consider that process A may be delayed, and appropriately relax the strict requirements in the initial stage.

[0065] The system re-evaluates the priority of each process and decides to allocate more resources to process B to speed up the progress, while considering whether some preparation work for process C can be started in advance.

[0066] If possible, call in spare equipment or personnel from other production lines to support the work of process B, or adjust the workload of equipment Y to reduce unnecessary downtime.

[0067] Based on the latest forecast data, re-plan the specific execution time of each process. For example, postpone the start time of process A to match the actual arrival time of raw materials, and adjust the plans of subsequent processes accordingly.

[0068] The system generates detailed scheduling instructions through a large language model, clearly indicating the new schedule, required resources and precautions for each step. For example, "Please start process A as soon as the raw materials arrive, and the expected delay is 1 hour; increase the manpower input of process B to ensure that it is completed within the original time; process C can pre-check some non-critical items in advance." During the implementation of the new scheduling strategy, the system continues to monitor changes in the production environment. If any new problems or opportunities are found (such as the repair speed of equipment Y exceeding expectations), the strategy will be adjusted again immediately to ensure that it is always moving in the optimal direction.

[0069] Furthermore, historical data and machine learning models are used to predict and analyze equipment status, identify potential failure risks in advance, and schedule maintenance work before problems occur. Based on the needs of current production tasks and the actual operating conditions of equipment, maintenance plans are intelligently adjusted to avoid production interruptions caused by excessive or neglected maintenance.

[0070] By real-time monitoring of raw material inventory, combined with market demand forecasts, and dynamically adjusting procurement plans, we ensure that the production line always has sufficient raw material supply while reducing inventory backlogs. Based on external factors such as traffic information and weather forecasts, we optimize the raw material transportation routes and schedules to ensure that materials arrive at the factory on time.

[0071] Create a digital twin model of the physical production line, test different scheduling schemes and their effects in a virtual environment, select the optimal solution and then apply it to actual production. The digital twin model is synchronized with the actual production line in real time, and any changes in the production environment can be immediately reflected in the digital model, facilitating timely response.

[0072] A large language model is used in combination with the structured data and the scheduling policy to generate scheduling instructions.

[0073] Specifically, through user request input and processing, the user's natural language request is converted into structured data containing key information such as task type, target entity, priority, etc. The dynamic knowledge base is combined with the attention-based relational graph neural network (R-GAT) to model multi-source heterogeneous production data, and obtain relevant contextual background information through the retrieval-augmented generation (RAG) framework.

[0074] Using reinforcement learning optimization, the scheduling strategy is adjusted in real time according to user feedback and changes in the production environment to ensure that task priorities, resource allocation, etc. are optimally set. The scheduling strategy not only considers the current task requirements, but also integrates historical data, equipment status, process flow and other factors to ensure the efficiency and continuity of production.

[0075] As a core technology, the large language model runs through the entire process, especially in parsing user requests, managing context, reasoning and generating instructions. Instruction generation utilizes the powerful semantic understanding and generation capabilities of the large language model to integrate information obtained from the dynamic knowledge base, R-GAT and hierarchical memory mechanism to generate specific executable scheduling instructions. These instructions usually include but are not limited to detailed information such as task type, target entity, execution conditions, priority, time requirements and subsequent steps. For example, "Check the operating status of equipment C, give priority to the power module, and notify the maintenance team after completion. If any abnormality is found, log it and send a report." When generating instructions, the large language model refers to the hierarchical storage information (global planning, process logic, and operation steps) provided by the hierarchical memory mechanism to ensure the consistency of task sequence, dependencies, and completion conditions.

[0076] At the same time, combined with the deep reasoning results provided by R-GAT, the large language model can understand the task context, thereby improving the applicability and accuracy of generated instructions. According to the results of the reinforcement learning optimization module, the large language model can also adjust the instruction format and priority. For example, in an emergency task scenario, high-priority tasks will be generated first.

[0077] When user input is ambiguous or certain details are omitted, the large language model can automatically complete these missing parts based on the context, ensuring that the generated instructions are complete and logical. In the face of emergencies or changes in user needs, the system can respond quickly, adjust the priority of existing instructions, insert new urgent tasks, and regenerate related instructions.

[0078] Suppose the user enters a relatively vague request: "Adjust production line 1." The system first converts it into structured data, then combines it with the current state of the production environment (for example, which processes are currently being carried out on production line 1, which equipment is idle, etc.), and determines the optimal scheduling strategy through the reinforcement learning optimization module. Then, the large language model is used to generate specific scheduling instructions, such as: "Adjust the priority of the task 'assembly' on production line 1 to high, and it must be completed within 2 hours. Immediately start the 'test' task after completion." For example, suppose in a manufacturing plant, the task scheduling of the production line depends on the status of multiple devices, the requirements of the process flow, and external factors (such as raw material supply). The system needs to be able to dynamically adjust the scheduling strategy based on real-time user requests and changes in the production environment, and generate detailed scheduling instructions through a large language model.

[0079] Current task arrangement: The production line plans to complete three main processes in sequence: A (preparing materials), B (assembling parts), and C (quality inspection).

[0080] Resource allocation: Equipment X is responsible for process A, equipment Y is responsible for process B, and equipment Z is responsible for process C.

[0081] Expected goal: All processes must be completed within the next 8 hours, and seamless connection between each link must be ensured.

[0082] User request: The dispatcher enters a natural language request: "Please optimize the task sequence of production line 2 and give priority to completing the 'packaging' task." Semantic parsing and structured conversion: The system converts the above request into structured data through the semantic parser: "Task type: optimization; production line number: 2; priority task: packaging." Retrieve contextual information: The dynamic knowledge base is combined with the attention-based relational graph neural network (R-GAT) to model production data and retrieve relevant contextual information. For example, it is known that the current task of production line 2 is "assembly" and it is expected to take another 2 hours to complete; at the same time, the "packaging" task depends on the completion of the "assembly" task.

[0083] Formulate scheduling strategies: The reinforcement learning optimization module adjusts the scheduling strategies in real time according to the latest changes in the production environment (such as equipment status, task progress, etc.). Considering that the "packaging" task is given a high priority, the system decides to start the "packaging" task immediately after the "assembly" task is completed, and postpone the scheduling of other non-critical tasks.

[0084] Comprehensive analysis and decision support: The large language model receives the search results from the dynamic knowledge base, the deep reasoning output provided by R-GAT, and the context information in the hierarchical memory mechanism. It will integrate this information, considering the overall goals and priorities of the global planning layer, the sequential relationship and dependency structure of the process logic layer, and the specific status and execution conditions of the operation step layer.

[0085] Generate specific executable scheduling instructions: Based on the above analysis, the large language model generates specific scheduling instructions as follows: Equipment maintenance task instructions: Check the operating status of equipment C, give priority to the power module, and notify the maintenance team after completion. If any abnormality is found, record the log and send a report.

[0086] Production flow instructions: Adjust the priority of the task "Assembly" of production line 2 to the normal level, and it is expected to take another 2 hours to complete. After completion, immediately start the "Packaging" task, set the priority to high, and it is expected to take 1 hour. Then continue with the remaining "Testing" tasks.

[0087] Task assignment instructions: Assign the "Material Transportation" task to operator Li Si, using forklift No. 2. The estimated completion time is 1 hour. The status needs to be updated after the task is completed.

[0088] Instruction content and logic consistency: In the process of generating instructions, the large language model ensures the logical clarity and adaptability of the instruction content. For example, in the production process instructions, it is clearly stated that the "packaging" task will start after the "assembly" task is completed, avoiding potential task conflicts or logical confusion.

[0089] At the same time, for ambiguous or omitted information, the large language model can automatically complete it based on the context. For example, if the user only mentions "optimize production line 2" but does not specify which tasks need to be adjusted, the system can infer the most likely part to be adjusted based on historical data and current production conditions, and generate detailed instructions accordingly.

[0090] Real-time monitoring and feedback: Once a scheduling instruction is generated and issued, the system will continue to monitor changes in the production environment and be ready to adjust the scheduling strategy and instruction content based on new user feedback or environmental changes to ensure production continuity and efficiency.

[0091] Furthermore, machine learning models are used to conduct predictive analysis of production processes, identify risk factors that may affect production progress (such as equipment failure, raw material shortages, etc.) in advance, and automatically generate preventive instructions through large language models. Combining real-time monitoring data and prediction results, the system can dynamically adjust the scheduling strategy and generate corresponding adjustment instructions through large language models to ensure the continuity and efficiency of production.

[0092] By analyzing the feedback and behavior patterns of different users, the system can learn the user's preferences and customize more personalized dispatch instructions accordingly. For example, some users may prefer a concise instruction format, while others require detailed step-by-step instructions. It supports the generation of dispatch instructions in multiple formats (such as concise version and detailed version) to meet different user needs and application scenarios.

[0093] According to an embodiment of the present application, the context information is stored and managed through a three-layer structure of a global planning layer, a process logic layer, and an operation step layer.

[0094] Specifically, the global planning layer,functions: record the overall goals and priorities of production tasks, and provide global directional,guidance for the scheduling instructions generated by the system.

[0095] Specific content: Overall scheduling of tasks: define the approximate arrangement of each task on the timeline, such as which tasks need to be completed first and which can be handled later. Resource allocation strategy: determine how to allocate resources (such as manpower, equipment, raw materials, etc.) required for different tasks to maximize resource utilization and avoid conflicts. Priority setting: prioritize tasks based on business needs or urgency to ensure that high-priority tasks can be handled in a timely manner.

[0096] Process logic layer, function: store the sequential relationship and dependency structure of tasks to ensure the logical consistency of tasks in the process flow.

[0097] Specific content: Task dependency: Define the order and dependency between tasks. For example, Task A can only start after Task B is completed, or Task C can only start after both Task D and Task E are completed. Process flow rules: Specify the standard operating steps in a specific production process and their logical connections to ensure that each process is executed in the correct order. Exception handling rules: Predetermine the response measures when a task cannot be completed on time, such as skipping the task or adjusting the schedule of subsequent tasks.

[0098] Operation step layer, function: track the specific status and execution conditions of each task, and reflect the completion progress and waiting conditions (such as equipment status or resource availability) of the task in real time.

[0099] Specific content: Task status update: Continuously monitor and update the status of each task (such as not started, in progress, completed, etc.) to keep abreast of the current production progress. Execution condition check: Check whether all necessary prerequisites are met before each task starts (such as whether the required equipment is idle, whether the raw materials are in place, etc.), and trigger the task start when the conditions are met. Dynamic adjustment: According to the changes in task status monitored in real time (such as equipment failure causing a task delay), automatically adjust the associated task plan to reduce the impact on the overall production process.

[0100] In order to ensure that the system can respond to changes in the production environment in a timely manner, the hierarchical memory mechanism adopts a dynamic management and update mechanism: Incremental update rules: Incremental update is used to adjust the memory content by real-time monitoring of task status changes (such as equipment failure, task delay, etc.). For example, when the task progress is updated, the update formula of the memory status is:

[0101] in, represents the memory state at time t, is the updated weight, is the task status feature at time t, It is the mapping function of the task state characteristics.

[0102] During the multi-round task generation process, the module will dynamically integrate user input, dynamic knowledge base search results and R-GAT reasoning output, adjust the content of each level of memory, and ensure that the generated instructions are consistent with the context logic. For example, when the user modifies the task priority or adds a new task, the module can readjust the scheduling order and logical dependencies of each task to avoid instruction conflicts or logical errors.

[0103] Through this hierarchical storage and dynamic management design, the intelligent production scheduling system can maintain logical coherence and real-time adaptability in complex task chains, providing contextual support and decision-making basis for the efficient operation of intelligent production scheduling. This approach not only improves the flexibility and adaptability of the system, but also enhances the accuracy and efficiency of production scheduling.

[0104] According to an embodiment of the present application, the user request is converted into structured data by a semantic parser, and the ambiguous information is completed in combination with the context, specifically: Extract key parameters through a semantic parser, and combine historical context information in a hierarchical memory mechanism to complete the missing conditions in the user request; If the user request is insufficient, guide the user to add details through a prompt mechanism; The supplemented key parameters are output as the structured data.

[0105] Specifically, receiving user requests: the system first receives the user's natural language requests or instructions, which can be text input, voice commands, or direct scheduling instructions, etc.

[0106] For example, a user might enter: "Schedule maintenance for device A."

[0107] Semantic parsing: Use a semantic parser to analyze user input, identify and extract key elements. These elements usually include task type (such as maintenance, inspection, production), target entity (such as equipment name), priority, time requirement, etc.

[0108] In the above example, the semantic parser will identify "Task Type: Equipment Maintenance" and "Target Entity: Equipment A". However, if the user does not explicitly provide a priority or time requirement, this information is considered missing.

[0109] Incorporating historical contextual information in a hierarchical memory mechanism Contextual completion: The hierarchical memory mechanism includes the global planning layer, process logic layer, and operation step layer, which stores the historical context information of past interactions. The system will use this information to complete the missing conditions in the current request.

[0110] For example, if the previous conversation has made it clear that all urgent maintenance tasks have a high priority and need to be completed as soon as possible, then the system can infer that the priority in the current request is high based on this background information, and infer that the time requirement is "as soon as possible" or a specific time limit.

[0111] Dynamically integrated context: The system not only relies on static historical records, but also updates and integrates the latest production environment status in real time. For example, if the current production line is in a high-load operation stage, the system may automatically adjust the priority or suggest users to choose other time periods for maintenance to reduce interference with production.

[0112] If the user request is insufficient, guide the user to add details through a prompt mechanism Identify information gaps: If after initial analysis it is found that some necessary information is still missing (such as specific time requirements), the system will proceed to the next step.

[0113] Generate prompt feedback: The system automatically generates targeted questions or prompts to guide users to provide more details. For example, "By what time do you want the maintenance work on device A to be completed?" or "Please confirm whether the maintenance of device A has the highest priority." User response and supplement: The user provides additional information based on the prompt, such as "I want to complete the maintenance of device A within the next two hours and set it as the highest priority." Output the supplemented key parameters as the structured data Integrate information: The system combines the original request information provided by the user with the supplementary information obtained through context completion and prompt mechanisms to form a complete task description.

[0114] Structured conversion: Finally, the integrated information is converted into a structured data format to facilitate further processing by subsequent modules. For example, "Task type: equipment maintenance; Target entity: Equipment A; Priority: High; Time requirement: Complete within two hours." Validation and optimization: The generated structured data is fully checked to ensure that all necessary information is included and consistent. If there are any questions or conflicts, the system will continue to request further clarification from the user through the feedback mechanism.

[0115] Suppose the user enters a relatively brief request: "Adjust the order of tasks on production line 1." The system first tries to extract key parameters from it, but finds that the specific task name to be adjusted and the new priority setting are missing. Therefore, the system accesses the hierarchical memory mechanism to find the early interaction records related to production line 1 and finds that there was a recent discussion about the need to increase the priority of the "packaging" task. Based on this background, the system generates a prompt: "Do you want to adjust the priority of the 'packaging' task on production line 1 to high? If so, please confirm the estimated completion time." The user responded: "Yes, set the priority of the 'packaging' task to high and hope to complete it within the next 3 hours." The system then converted this request into structured data: "Task type: adjust order; production line number: 1; priority task: packaging; priority: high; time requirement: complete within 3 hours." In this way, the system can not only accurately understand the user's intentions, but also ensure that the generated instructions are both complete and logical, greatly improving the efficiency and accuracy of production scheduling.

[0116] According to an embodiment of the present application, the attention-based relational graph neural network is used to model the production data, and the relevant context information is retrieved in combination with the retrieval enhancement generation framework, specifically: Integrate the production data from different sources into a knowledge graph, where nodes represent production entities and edges represent relationships between entities; Matching the user request with the node in the knowledge graph using a vectorized storage model; Attention-based graph neural network captures implicit associations and key dependencies between tasks.

[0117] Specifically, multi-source heterogeneous data integration: Dynamic knowledge base: First, the system integrates heterogeneous production data from different sources into a production-based knowledge graph. These data may include equipment information, process flow, task records, and environmental data.

[0118] Nodes and edges: In this knowledge graph, nodes represent production entities (such as equipment A, task B, process step C, etc.), while edges represent the relationships between these entities (such as task dependency, equipment linkage, process sequence, etc.). For example, a node can be a specific production equipment or process step, and an edge indicates whether there is a certain operational dependency between these equipment.

[0119] Standardization and real-time updates: Semantic labeling and data standardization technology: To ensure the consistency and comparability of data, semantic labeling and data standardization technology are used to convert raw data into a unified structured form.

[0120] Dynamic update: The dynamic knowledge base supports real-time updates and can adjust node attributes (such as task priority) and relationship weights (such as task dependency strength) according to the latest production status to reflect changes in the current production environment.

[0121] Matching using vectorized storage models Vectorized storage and retrieval: Embedding model: Use an embedding model (such as Transformer) to convert user requests and nodes and relationships in the knowledge graph into high-dimensional vector representations. This allows user requests to quickly locate relevant nodes and contextual information through vector matching.

[0122] Cosine similarity matching: The relevance between the user request vector and the knowledge graph node embedding is measured by calculating the cosine similarity between them. The formula is as follows:

[0123] in, It is a query embedding model that converts user request q into a vector. The nodes in the knowledge base Embedding model converted to vector.

[0124] In the retrieval phase, the RAG framework selects the k nodes most relevant to the user's request to form a contextual supplementary dataset. This approach effectively avoids the hallucination problem in traditional generation methods and improves the accuracy and relevance of generated content.

[0125] Generation phase: Combine context information and user input to generate production scheduling instructions that better meet the needs of the actual scenario. This context-based generation method helps improve the quality and applicability of instructions.

[0126] Capturing implicit relationships using attention-based graph neural networks Relational Attention Mechanism: Dynamically assign weights: Dynamically assign weights to different relations in the knowledge graph through the relational attention mechanism to highlight the semantic associations between key tasks. Specifically, the attention weights of nodes u and 𝑣 under the relation 𝑟 are It can be calculated by the following formula:

[0127] in, and are the embedding representations of nodes u and v respectively, is the feature weight matrix of relation 𝑟.

[0128] Combined with the results of the attention mechanism, R-GAT performs multiple embedding updates on each node to capture its semantic association with neighboring nodes. The update formula is as follows:

[0129] in, is the embedding representation of node 𝑣 at layer l+1, is the activation function (such as ReLU).

[0130] Deep modeling capability: R-GAT's multi-level modeling capability ensures that the system can capture implicit task dependency chains and task sequence optimization in process flows in complex task scenarios. For example, in a multi-task dependency chain, the delay of a task may affect the scheduling of a series of subsequent tasks. R-GAT can identify and handle such complex dependencies.

[0131] Assume that a user submits a natural language request: "Schedule maintenance for device A". The system will perform the following steps: Data integration: Obtain all information related to device A from the knowledge graph, including the device status, last maintenance time, current task dependencies, etc.

[0132] Vectorized storage and retrieval: Convert the user request "arrange maintenance of device A" into a vector and match it with the nodes in the knowledge graph to find all nodes related to device A.

[0133] Application of attention mechanism: Through the relational attention mechanism, the system identifies the key dependencies between device A and other devices or tasks, such as which tasks need to wait until device A completes maintenance before they can start.

[0134] Generate scheduling instructions: Based on the above analysis results, the system generates detailed scheduling instructions: "Please schedule maintenance work for device A immediately, which is expected to take 2 hours. During this period, please suspend all tasks that depend on device A." In this way, the system can not only deeply understand the complex production environment and the various relationships within it, but also efficiently respond to the specific needs of users and provide accurate and efficient scheduling suggestions, thereby greatly improving the intelligence level and adaptability of production scheduling.

[0135] According to an embodiment of the present application, the scheduling strategy is adjusted in real time according to user feedback and changes in the production environment, specifically: Constructing a reward signal through the user feedback and the production environment changes to guide the optimization of the scheduling strategy; Using a policy gradient-based algorithm, scheduling policy parameters are dynamically adjusted according to real-time data to maximize the cumulative reward.

[0136] Specifically, subjective evaluation: the system collects direct feedback from users, including subjective opinions such as evaluation of the logic of instructions and the rationality of priority settings.

[0137] Objective data: In addition to user feedback, the system will automatically collect actual execution data of production tasks, such as task completion time, equipment utilization and other objective indicators.

[0138] Constructing the reward function: Quantitative evaluation: Based on the above information, a reward function is designed to quantitatively evaluate the effect of the current scheduling strategy. For example, the following formula can be defined to measure the performance of the scheduling strategy:

[0139] in, Indicates the task completion efficiency, Indicates resource utilization, Indicates user satisfaction, , , It is the weight coefficient, which is used to balance the importance of various indicators.

[0140] According to different production goals or business needs, these weight coefficients can be dynamically adjusted to adapt to different optimization focuses (such as improving efficiency or increasing user satisfaction).

[0141] State representation: Define the production environment and task status as states For example, the status can include the state of the device (running, idle, faulty), the progress of the task (not started, in progress, completed), the availability of resources, etc.

[0142] Action Representation: Scheduling Instructions as Actions For example, an action can be to start a device, pause a task, adjust task priority, etc.

[0143] Policy gradient method: Use a policy gradient-based method (such as the REINFORCE algorithm) to optimize the scheduling policy. Specifically, the policy , select actions based on the current state to maximize the cumulative reward , the policy update follows the following formula:

[0144] in, is the objective function of the strategy, is a discount factor to account for the time value of future rewards, is the immediate reward obtained at time 𝑡.

[0145] Real-time data-driven: The system uses real-time monitoring data (such as device status changes, task completion, user feedback, etc.) to continuously adjust policy parameters to make them more in line with the current production environment. For example, if a minor failure is detected on a device, the system can dynamically adjust the priority of related tasks to reduce reliance on the device or schedule maintenance work in advance.

[0146] Strategy Iteration Optimization: Through continuous trials and feedback loops, the system gradually optimizes its scheduling strategy. After each adjustment, the system will evaluate the effect of the new strategy and further fine-tune the parameters based on the results to achieve the best production effect.

[0147] Specific implementation steps for real-time adjustment of scheduling strategies Monitoring and data analysis: Continuously monitor the status of production equipment, task progress and environmental parameters to identify any factors that may affect the production process. For example, monitor the operating status of equipment, raw material inventory levels, task execution progress, etc.

[0148] Abnormal detection and response: Once an abnormal situation (such as equipment failure, task delay) is detected, the predefined response plan is immediately launched and the task scheduling order is dynamically adjusted. For example, when a failure of device Y is detected, the system can reallocate tasks to backup devices and adjust the schedule of subsequent tasks to minimize the impact on overall production.

[0149] Generate new scheduling instructions: Based on the latest forecast data and optimized strategies, the system generates detailed scheduling instructions through a large language model. For example, "Please start process B immediately after the repair of device Y is completed. It is estimated to take 2 hours. During this period, please suspend all tasks that depend on device Y and reschedule other executable tasks." Continuous monitoring and further adjustment: During the implementation of the new scheduling strategy, the system continues to monitor changes in the production environment. If any new problems or opportunities are found (such as the repair speed of equipment Y exceeds expectations), the strategy is immediately adjusted again to ensure that it is always moving in the optimal direction.

[0150] Assume that in a manufacturing plant, the task scheduling of the production line depends on the status of multiple devices, the requirements of the process flow, and external factors (such as raw material supply). The system needs to be able to dynamically adjust the scheduling strategy based on real-time user feedback and changes in the production environment.

[0151] Initial state Current task arrangement: The production line plans to complete three main processes in sequence: A (preparing materials), B (assembling parts), and C (quality inspection).

[0152] Resource allocation: Equipment X is responsible for process A, equipment Y is responsible for process B, and equipment Z is responsible for process C.

[0153] Expected goal: All processes must be completed within the next 8 hours, and seamless connection between each link must be ensured.

[0154] User feedback and production environment changes User feedback: The dispatcher provided feedback through the system interface: "The actual execution time of process B is 30% longer than expected, which may affect the timely completion of subsequent processes." Changes in the production environment: The system monitoring module detected a minor fault in device Y, causing its work efficiency to decrease by 25%.

[0155] At the same time, the warehouse management system reports that the raw materials are in sufficient supply, but the transport vehicles are delayed due to traffic jams, which may cause the specific materials required for process A to arrive 1 hour late.

[0156] The process of adjusting the scheduling strategy in real time Build reward signals: The system builds reward signals based on user feedback and changes in the production environment, and adjusts the weights in the reward function. For example, it increases the weight for completing process B on time, while taking into account that process A may be delayed, and appropriately relaxes the strict requirements in the initial stage.

[0157] Policy update: Use policy gradient-based methods (such as the REINFORCE algorithm) to optimize scheduling strategies. Based on the latest changes in the production environment (such as reduced efficiency of equipment Y and delayed arrival of raw materials), the system re-evaluates the priority of each process and decides to allocate more resources to process B to speed up the progress, while considering whether some preparation work for process C can be started in advance.

[0158] Generate new scheduling instructions: The system generates detailed scheduling instructions through a large language model, clearly indicating the new schedule, required resources and precautions for each step. For example, "Please start process A as soon as the raw materials arrive, and the expected delay is 1 hour; increase the manpower input of process B to ensure that it is completed within the original time; process C can pre-check some non-critical items in advance." In this way, "adjusting scheduling strategies in real time based on user feedback and changes in the production environment" not only helps solve unexpected problems, but also improves the flexibility and adaptability of the entire production process, ensuring efficient operation even in the face of uncertainty. This approach greatly improves the intelligence level of the system, enabling it to maintain efficient operation in a complex and changing industrial environment.

[0159] According to an embodiment of the present application, the structured data and the scheduling strategy are combined to generate a scheduling instruction using a large language model, specifically: Generate the scheduling instruction by combining the structured data and the scheduling strategy using a large language model; Automatically adjust instruction content based on changes in user needs or environmental emergencies.

[0160] Specifically, integrating structured data with scheduling strategies: Structured data preparation: Through the user request input and processing module, the user's natural language request is converted into structured data containing key information such as task type, target entity, priority, etc.

[0161] Scheduling strategy formulation: Utilize the reinforcement learning optimization module to adjust the scheduling strategy in real time based on user feedback and changes in the production environment to ensure that task priority, resource allocation, etc. are optimally set.

[0162] Application of large language model: Instruction content generation: The large language model is used as a core technology throughout the entire process, especially in parsing user requests, managing context, reasoning, and instruction generation. The instruction generation module uses the powerful semantic understanding and generation capabilities of the large language model to integrate the information obtained from the dynamic knowledge base, R-GAT, and hierarchical memory mechanism to generate specific executable scheduling instructions.

[0163] For example, suppose the system receives the following structured data: "Task type: equipment maintenance; Equipment name: Equipment A; Priority: High; Time requirement: Complete today." Combined with the current scheduling strategy (for example, Equipment B depends on the normal operation of Equipment A), the large language model can generate specific scheduling instructions: "Immediately schedule maintenance work for Equipment A, with a high priority and an estimated time of 4 hours. During this period, please suspend all tasks that depend on Equipment A and reschedule other executable tasks." Logical consistency and adaptability: Context fusion: When generating instructions, the large language model refers to the hierarchical storage information (global planning, process logic, and operation steps) provided by the hierarchical memory mechanism to ensure the consistency of task sequence, dependencies, and completion conditions.

[0164] Dynamic adjustment: In the face of emergencies or changes in user needs, the system can respond quickly, adjust the priority of existing instructions, insert new emergency tasks, and regenerate related instructions. For example, if a failure of device A is suddenly detected, the system can instantly update the dispatch instruction: "Due to the failure of device A, please immediately start the emergency maintenance procedure and notify relevant personnel to come to the scene to handle it." Automatically adjust instruction content based on changes in user needs or environmental emergencies Real-time monitoring and feedback loop: Continuous Monitoring: The system continuously monitors the status of production equipment, task progress and environmental parameters to identify any factors that may affect the production process.

[0165] Feedback mechanism: Once an abnormal situation is detected (such as equipment failure, task delay) or a new user demand is received (such as the temporary addition of an emergency task), the system will trigger the feedback mechanism, re-evaluate the current scheduling strategy, and adjust the scheduling instructions accordingly.

[0166] Automation adjustments and instruction updates: Automatically generate new instructions: Based on the latest scheduling strategies and user needs, the large language model automatically generates new scheduling instructions. For example, when the supply of raw materials is delayed, the system can generate the following instructions: "Please postpone the start time of process A until the raw materials arrive, which is expected to be delayed by 1 hour; at the same time, start some preparation work for process C in advance to reduce the impact on the overall production plan." Instruction optimization: In addition to directly generating new instructions, the system can also optimize existing instructions. For example, when it detects that the working efficiency of a certain device is lower than expected, the system can adjust the schedule of all tasks on the device and generate corresponding adjustment instructions.

[0167] Initial state: The production line is planned to complete three main processes in sequence: A (prepare materials), B (assemble parts), and C (quality inspection).

[0168] User demand changes: The user submitted a new request: "Please complete the 'packaging' task as soon as possible." The system first converts this request into structured data: "Task type: optimization; production line number: 2; priority task: packaging." Combined with the status of the current production environment (such as which processes are being carried out on production line 2, which equipment is idle, etc.), and the optimal scheduling strategy is determined through the reinforcement learning optimization module - it is decided to start the "packaging" task immediately after the "assembly" task is completed, and postpone the scheduling of other non-critical tasks.

[0169] The large language model generates specific scheduling instructions: "Adjust the priority of the 'assembly' task on production line 2 to the normal level, and it is estimated to take another 2 hours to complete. Immediately after completion, start the 'packaging' task, set the priority to high, and it is estimated to take 1 hour. Then continue with the remaining 'testing' tasks." Responding to emergencies Emergency: Assume that during the execution of the above instructions, device Y suddenly fails.

[0170] The system discovered this problem through real-time monitoring and immediately adjusted the scheduling strategy.

[0171] The large language model generates new scheduling instructions: "Due to the failure of device Y, please arrange for a maintenance team to inspect and repair it immediately. During this period, please suspend all tasks that depend on device Y and reschedule other available resources to perform non-critical tasks." In this way, the process of generating scheduling instructions using a large language model by combining structured data and scheduling strategies not only improves the intelligence level and scenario adaptability of instruction generation, but also ensures the orderliness and practicality of task execution in each link, greatly improving the efficiency and accuracy of production scheduling. This method enables the system to maintain efficient operation in a complex and changing industrial environment and respond to various changes and challenges in a timely manner.

[0172] According to one embodiment of the present application, it also includes: Monitor production status in real time, quickly identify and classify anomalies, and dynamically adjust scheduling plans.

[0173] According to an embodiment of the present application, the real-time monitoring of production status, rapid identification and classification of anomalies, and dynamic adjustment of the scheduling plan are specifically as follows: Track the status of production equipment, task execution, and environmental parameters; Identify and classify abnormal situations in the production process and take appropriate measures based on the severity; Automatically generate response plans and dynamically adjust task scheduling to minimize the impact of exceptions on production.

[0174] Specifically, real-time monitoring of production status Track the status of production equipment: Equipment health monitoring: The system continuously monitors the operating status of production equipment, including but not limited to key parameters such as temperature, pressure, vibration frequency, etc., to evaluate the working condition of the equipment.

[0175] Failure early warning: By setting thresholds or using machine learning models to predict potential equipment failures, warning signals can be issued in advance so that preventive measures can be taken in time.

[0176] Task execution progress monitoring: Progress tracking: Track the actual execution progress of each production task in real time, compare and analyze it with the planned schedule, and ensure that all tasks are carried out according to the established schedule.

[0177] Bottleneck Detection: Identifies task bottlenecks that may reduce production line efficiency and provides optimization suggestions.

[0178] Environmental parameter monitoring: Environmental condition monitoring: Monitor environmental parameters such as temperature, humidity, and air quality to ensure they are within a range suitable for production, thus avoiding product quality problems or equipment damage caused by environmental factors.

[0179] Quickly identify and classify anomalies Anomaly Detection: Multi-source data analysis: Using sensor data, historical records, and other relevant information, algorithms automatically detect deviations from normal operating ranges. For example, when the temperature of a device exceeds the safe range, the system will mark it as abnormal.

[0180] Pattern recognition: Use advanced pattern recognition techniques (such as deep learning) to identify complex abnormal patterns, not only limited to single-point anomalies, but also include correlated anomalies between multiple variables.

[0181] Abnormal classification: Severity classification: Classify anomalies according to their impact on production, such as minor, moderate, and severe. This helps prioritize those issues that have the greatest impact on production.

[0182] Type classification: Classify anomalies into different types, such as equipment failure, raw material shortage, human error, etc., to facilitate the development of targeted solutions.

[0183] Dynamically adjust the scheduling plan Response Plan Generation: Automated decision support: Once an anomaly is identified, the system automatically generates a response plan, including specific repair steps, required resources, and estimated recovery time. For example, for an equipment failure, the system may recommend immediately starting a backup device and arranging for a maintenance team to inspect it.

[0184] Notification mechanism: timely notify relevant personnel or departments about abnormal situations and their handling plans to ensure the timeliness and accuracy of information transmission.

[0185] Dynamically adjust task scheduling: Re-planning of task order: Based on the current abnormal situation and available resources, the system will dynamically adjust the order of task scheduling to minimize the impact of the abnormality on overall production. For example, during a device failure, tasks that depend on the device can be suspended and other unrelated tasks can be executed instead.

[0186] Resource reallocation: If a bottleneck occurs in a certain link, the system can reallocate human and material resources according to the actual situation to alleviate the bottleneck effect.

[0187] Example Initial state: Assume that a production line plans to complete three processes in sequence: A (material preparation), B (parts assembly), and C (quality inspection). Equipment X is responsible for process A, equipment Y is responsible for process B, and equipment Z is responsible for process C.

[0188] Abnormality occurs: During the production process, the system detects that equipment Y has a minor fault, causing its work efficiency to decrease by 25%.

[0189] Identification and classification: The system recognizes this anomaly and classifies it as "moderate" based on the impact on production, because although efficiency is reduced, work has not stopped completely.

[0190] Response plan: The system automatically generates a response plan, recommending to temporarily reduce the task load on equipment Y and arrange for the maintenance team to conduct an inspection; at the same time, the dispatcher is notified to re-plan the execution order of subsequent tasks.

[0191] Dynamic adjustment: The scheduling plan is dynamically adjusted to postpone some tasks that depend on equipment Y and start some preparation work of process C in advance to minimize the impact on the overall production plan.

[0192] Through the above mechanism, the system can monitor the production status in real time, quickly identify and classify abnormal situations, and dynamically adjust the scheduling plan accordingly to ensure that production continuity and efficiency can be maintained even in the face of unexpected problems. This approach greatly improves the flexibility and adaptability of the system, enabling it to operate stably in complex and changing industrial environments.

[0193] A computer-readable storage medium stores a program, which implements the steps in the method when executed by a processor.

[0194] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein the processor implements the steps in the method when executing the program.

[0195] Anything not described in this application can be achieved by adopting or drawing on existing technologies.

[0196] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0197] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. An intelligent production scheduling and exception management method based on a large language model, characterized in that: include: The semantic parser converts user requests into structured data and completes ambiguous information based on the context. An attention-based graph neural network is used to model production data, and a retrieval-enhanced generation framework is used to retrieve relevant contextual information. Adjust scheduling strategies in real time based on user feedback and changes in the production environment; A large language model is used in combination with the structured data and the scheduling policy to generate scheduling instructions.

2. The method according to claim 1, characterized in that The context information is stored and managed through a three-layer structure of a global planning layer, a process logic layer, and an operation step layer.

3. The method according to claim 1, characterized in that The semantic parser converts the user request into structured data and completes the ambiguous information in combination with the context, specifically: Extract key parameters through a semantic parser, and combine historical context information in a hierarchical memory mechanism to complete the missing conditions in the user request; If the user request is insufficient, guide the user to add details through a prompt mechanism; The supplemented key parameters are output as the structured data.

4. The method according to claim 1, characterized in that: The attention-based graph neural network is used to model production data, and the retrieval enhancement generation framework is combined to retrieve relevant context information, specifically: Integrate the production data from different sources into a knowledge graph, where nodes represent production entities and edges represent relationships between entities; Matching the user request with the node in the knowledge graph using a vectorized storage model; Attention-based graph neural network captures implicit associations and key dependencies between tasks.

5. The method according to claim 1, characterized in that The scheduling strategy is adjusted in real time according to user feedback and changes in the production environment, specifically: Constructing a reward signal through the user feedback and the production environment changes to guide the optimization of the scheduling strategy; Using a policy gradient-based algorithm, scheduling policy parameters are dynamically adjusted according to real-time data to maximize the cumulative reward.

6. The method according to claim 1, characterized in that The combining of the structured data and the scheduling strategy and using a large language model to generate a scheduling instruction is specifically: Generate the scheduling instruction by combining the structured data and the scheduling strategy using a large language model; Automatically adjust instruction content based on changes in user needs or environmental emergencies.

7. The method according to claim 1, characterized in that Also includes: Monitor production status in real time, quickly identify and classify anomalies, and dynamically adjust scheduling plans.

8. The method according to claim 7, characterized in that The real-time monitoring of production status, rapid identification and classification of anomalies, and dynamic adjustment of scheduling plans are specifically as follows: Track the status of production equipment, task execution, and environmental parameters; Identify and classify abnormal situations in the production process and take appropriate measures based on the severity; Automatically generate response plans and dynamically adjust task scheduling to minimize the impact of exceptions on production.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method according to any one of claims 1 to 8 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.

Images