An industrial control system configuration generation method, system and device based on a vertical field large model and an agent

Abstract

This invention relates to the field of industrial automation technology, and more particularly to a method, system, and device for generating industrial control system configurations based on a large vertical domain model and intelligent agents. The method includes: S1: Constructing and training a large vertical domain model, using it to analyze and identify intents from multimodal design data, extracting key industrial information, and generating a dual-track output containing structured command sequences; S2: The intelligent agent receives the command sequences, performs task planning and security verification, parses them into sub-tasks, and then calls a professional toolchain to execute them, generating a configuration file; S3: Outputting the execution results, and iteratively optimizing the large vertical domain model and the intelligent agent's planning logic based on user feedback and execution logs. This invention achieves fully automated generation from design data to complete configuration files, possessing security verification, self-learning, and solution reuse capabilities, significantly reducing the technical threshold and labor costs of industrial control system configuration.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, and in particular to a method, system, and device for generating industrial control system configurations based on a large vertical domain model and intelligent agents. Background Technology

[0002] With the deep integration of industrial automation and artificial intelligence technologies, the configuration engineering of industrial control systems (such as DCS, PLC, and SIS) is rapidly developing towards intelligence and automation. Against this backdrop, to improve configuration efficiency and reduce reliance on engineers' professional experience, traditional technologies mainly employ manual configuration, semi-automated code generation tools, and the recently emerging LLM-based assisted generation schemes.

[0003] Manual configuration relies entirely on engineers manually reading design documents such as piping and instrumentation diagrams (P&ID) and input / output (IO) tables, and then writing control logic and drawing human-machine interface (HMI) screens one by one. This method is inefficient, prone to errors, and requires a high level of experience from personnel, making it difficult to meet the rapid delivery needs of modern industrial projects.

[0004] While semi-automated code generation tools (such as hardware configuration wizards and logic template libraries provided by manufacturers) can reduce repetitive work to some extent, their input formats are strictly limited, they cannot directly understand unstructured design documents (such as PDF drawings and handwritten logic descriptions), and knowledge between different manufacturers and industries cannot be reused, resulting in limited intelligence.

[0005] In recent years, with the development of large language modeling technology, research schemes have emerged that utilize LLM to assist in the generation of PLC / DCS control logic. For example, design specifications are converted into natural language for LLM processing through meta-language syntax, or structured text (ST) code snippets conforming to the IEC 61131-3 standard are directly generated from natural language descriptions. However, these schemes have significant drawbacks: firstly, they can only generate fragmented code snippets and cannot achieve fully automated configuration generation from requirements understanding and task planning to calling professional toolchains (such as hardware configuration, I / O parsing, HMI drawing, and specification checking); secondly, these schemes generally lack security verification mechanisms for industrial control scenarios, failing to identify or intercept dangerous operations involving critical safety interlocks, making it difficult to meet the reliability requirements of industrial applications; furthermore, existing schemes lack self-learning capabilities and cannot continuously optimize their performance based on user feedback and execution logs, resulting in difficulty in improving generation quality as the application scenarios expand.

[0006] In actual configuration engineering, different design documents (such as P&ID drawings, IO tables, and logic descriptions) vary significantly in format, complexity, and criticality. While critical control loops (such as emergency stop interlocks) are fewer in number, they require extremely high logical accuracy and safety; whereas most conventional equipment (such as ordinary motor start / stop and indicator lights) have relatively lower requirements for generation accuracy. However, traditional manual configuration and existing AI-assisted solutions, once the processing flow is determined, are difficult to adjust flexibly. All documents use a uniform parsing and generation strategy, leading to over-processing of simple tasks, under-processing of complex tasks, or requiring extensive engineer intervention for corrections, resulting in significant shortcomings in efficiency, economy, and practicality.

[0007] In summary, there is an urgent need for an intelligent method and system that can deeply understand the professional knowledge in the field of industrial control, autonomously plan and call professional toolchains to complete the entire process configuration generation, have built-in security verification mechanisms, and have continuous learning capabilities. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for generating industrial control system configurations based on a large vertical domain model and intelligent agents, comprising the following steps: S1: Construct and train a large vertical domain model, use the large vertical domain model to perform multimodal analysis and intent recognition on multimodal design data, extract key industrial information, and the large vertical domain model generates dual-track output results based on the key industrial information. The dual-track output results include structured configuration operation suggestions and command sequences. The command sequences include tool parameters, execution order and logical dependencies for calling subsequent tools. S2: The intelligent agent receives the command sequence, performs task planning and security verification, parses the command sequence into sub-tasks, and, according to the task planning, calls and orchestrates a professional toolchain to execute the sub-tasks, generating a configuration file for the industrial control system. S3: Output the execution result of the configuration file, and iteratively optimize the planning logic of the vertical domain large model and intelligent agent based on user feedback and execution logs.

[0009] Preferably, in step S1, constructing and training a large model for a vertical domain includes: Based on a general large model, we use professional corpus in the field of industrial control for pre-training, retrieval enhancement to generate RAG and fine-tuning of instructions; A hybrid expert model architecture is adopted to build segmented knowledge bases or sub-models for different design institutes, manufacturers or industries. After receiving the multimodal design data, the corresponding segmented knowledge bases or sub-models are activated through a routing mechanism to perform intent recognition.

[0010] Preferably, in step S1, the vertical domain large model is used to perform multimodal analysis and intent recognition on the multimodal design data to extract key industrial information, including: The sub-model performs multimodal fusion analysis on the multimodal design data to identify the key industrial information, including equipment tag number, control loop type, interlocking logic, and process flow description. Based on the aforementioned key industrial information, a readable text explaining the configuration scheme logic is generated, along with a structured command sequence containing tool parameters for calling subsequent tools, execution order, and logical dependencies.

[0011] Preferably, the structured command sequence, which includes tool parameters for calling subsequent tools, execution order, and logical dependencies, includes: The structured command sequence is a structured intermediate instruction set oriented towards machine parsing, using a preset markup language format; The tool parameters include file path, device tag number, and control logic type; The execution order and the logical dependencies are used to define the serial or concurrent scheduling relationships of tool calls in the subsequent professional toolchain.

[0012] Preferably, in step S2, the agent receives the command sequence and performs task planning and security verification, including: Perform a syntactic and logical integrity scan on the structured command sequence to confirm the consistency between instructions and the ability to execute in a closed loop. Based on the current user's role authorization scope, a security review is conducted to identify and block dangerous operation instructions that involve modifying critical safety interlocks. If the blocking rule is hit, the execution process is frozen and a secondary confirmation request to the user is triggered. After the verification is passed, the data flow and prerequisite dependencies between the instructions are analyzed, and the structured command sequence is decomposed and reassembled into a subtask queue containing concurrent and serial scheduling subtasks.

[0013] Preferably, in step S2, the agent invokes a specialized toolchain according to the plan to execute the sub-tasks to generate a configuration file, including: The intelligent agent strictly follows the dependency graph of the sub-task queue to schedule data parsing tools to extract standardized device tag numbers and lists; The extracted data is passed to the hardware configuration tool to complete the automatic selection and allocation of racks and modules, and drives the logic generation tool to match the control strategy template to generate logic code that conforms to industry standards; The HMI drawing tool is scheduled to generate the interface flowchart synchronously or immediately afterward, automatically establishing dynamic data links between screen elements and underlying tag numbers. Finally, the specification checking tool is scheduled to cross-compile and verify the consistency of the entire set of configuration files generated.

[0014] Preferably, step S3, which iteratively optimizes the planning logic of the vertical domain large model and the agent based on execution logs and user feedback, includes: The execution logs are linked with user evaluations and modification actions. Processes that execute smoothly and receive excellent user evaluations are encapsulated as positive samples, while processes that are interrupted, fail verification, or are heavily modified by users are encapsulated as negative samples. After parsing the positive samples, they are stored in the case library as corpus for subsequent retrieval enhancement or instruction fine-tuning to update the vertical domain model. At the same time, the execution logs of the negative samples are deeply analyzed to accurately locate failure nodes, and the scheduling rules and verification thresholds inside the intelligent body are modified accordingly.

[0015] Preferably, iterative optimization further includes a scheme reuse step: The structured command sequence and execution parameters corresponding to the configuration generation process that serves as a positive sample and meets the preset standardization conditions are saved as an automated configuration scheme template. When new multimodal design data is received and the scene similarity with the automated configuration scheme template is higher than a preset threshold, the template is directly invoked to execute configuration generation, skipping the structured command sequence generation step and some verification steps.

[0016] Based on the same concept, the present invention also provides an industrial control system configuration generation system based on a large vertical domain model and intelligent agents, including: The Vertical Domain Large Model Module is used to build and train a vertical domain large model. The vertical domain large model is used to perform multimodal analysis and intent recognition on multimodal design data and extract key industrial information. The vertical domain large model generates dual-track output results based on the key industrial information. The dual-track output results include structured configuration operation suggestions and command sequences. The command sequences include tool parameters for calling subsequent tools, execution order and logical dependencies. The agent planning and scheduling module is used for the agent to receive the command sequence, perform task planning and security verification, and parse the command sequence into subtasks; The execution toolchain module, scheduled by the agent planning and scheduling module, is used by the agent to call and orchestrate the professional toolchain to execute the sub-tasks according to the task plan, and generate the configuration file of the industrial control system. The self-learning iteration module is used to output the execution results of the configuration file and iteratively optimize the planning logic of the vertical domain large model and intelligent agent based on user feedback and execution logs.

[0017] Based on the same concept, the present invention also provides a computer device, including a memory and one or more processors, wherein the memory stores computer code, and when the computer code is executed by the one or more processors, the one or more processors cause the one or more processors to perform the steps of an industrial control system configuration generation method based on a vertical domain large model and intelligent agent as described in any one of the embodiments.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention constructs and trains a large vertical domain model, and uses the large vertical domain model to perform multimodal analysis and intent recognition on multimodal design data, generating a dual-track output containing structured configuration operation suggestions and command sequences. This realizes the automated understanding and structured transformation of unstructured or semi-structured industrial design data (such as P&ID drawings, IO tables, and logical descriptions), solves the problems of difficult understanding of design data and high dependence on manual labor in the traditional configuration process, and significantly improves the intelligence level and efficiency of configuration design.

[0019] (2) The present invention receives the command sequence through an intelligent agent, performs task planning and security verification, parses the command sequence into sub-tasks, and calls and arranges professional toolchains to execute sub-tasks to generate configuration files for industrial control systems according to the task planning. This realizes the full-chain automated configuration generation from requirement understanding to hardware configuration, logic generation, screen drawing and standard checking. At the same time, the built-in security verification mechanism can identify and intercept dangerous operations involving key safety interlocks, effectively ensuring the safety and controllability of the industrial control system. Compared with existing solutions that can only generate code fragments or require a lot of manual intervention, the present invention truly realizes end-to-end unmanned configuration generation.

[0020] (3) By outputting the execution results of the configuration file, and iteratively optimizing the planning logic of the vertical domain large model and intelligent agent based on user feedback and execution log, the present invention realizes the self-learning and continuous evolution capabilities of the system. It can use successful generated cases as positive samples and failed cases as negative samples. Combined with the latest data crawled from external knowledge sources on a regular basis, it continuously optimizes the model accuracy, enriches the knowledge base and design templates, so that it becomes more and more intelligent with use, gradually improves the adaptability to different industries and different manufacturers' scenarios, reduces the error rate and increases the first pass rate of configuration generation. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0022] Figure 1 This is a flowchart of an industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to the present invention; Figure 2 This is another flowchart of the industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to the present invention; Figure 3 This is a functional architecture diagram of an industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. Obviously, the described embodiments are only some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.

[0024] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a” and “an” used herein, and “the”, may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0025] The technical terms involved in the embodiments of this invention are defined as follows: Vertical domain large model: Unlike general large model, it mainly uses pre-training to enable it to have a relatively complete understanding of a certain professional field. In particular, it has a relatively complete knowledge base of information such as terminology, name abbreviations, professional background knowledge, official or formal information acquisition channels in the field, which is comparable to the level of professional and technical personnel who have worked in this field for at least 5 years.

[0026] Intelligent agent technology: A recent hot topic in AI research, this technology enables agents to act as intelligent dispatchers. It autonomously plans operations based on suggestions from large AI models, and through interfaces, calls existing software or tools to convert textual suggestions into concrete software outputs. This allows it to influence and achieve results in the physical world, simulating human thought and actions. Furthermore, through feedback or knowledge base learning, it can optimize the large model or iterate operations until the goal is achieved or termination conditions are met, demonstrating self-learning and self-optimization capabilities. For example, intelligent agent technology can utilize pre-defined interfaces and authorized permissions to simulate the actions of procurement personnel by operating chat software, online shopping malls, and online payment software. It can autonomously compare prices, negotiate contracts, submit applications for approval, place orders, and make payments, theoretically completely replacing all the thinking and actions of procurement personnel.

[0027] Industrial control systems refer to systems such as DCS, SIS, PLC, and dedicated controllers that are widely used in the field of industrial control. They generally have signal I / O channels, logic controllers, host computers, configuration software, and HMI human-machine interface screens.

[0028] Interface flowchart: refers to the HMI human-machine interface screen in an industrial control system. It is the intuitive interface for operators to interact with the control system. It generally displays the current status of the controlled object, the target setpoint of the control point or control logic loop, parameter settings, activation and deactivation, status settings, etc., and provides data interfaces and feedback interfaces.

[0029] First Embodiment Please see Figure 1 and Figure 2 As shown, this embodiment provides a method for generating industrial control system configurations based on a large vertical domain model and intelligent agents. It aims to improve the automation and intelligence of industrial control system configuration engineering, reduce technical barriers and labor costs, and possesses access control, self-learning, and continuous iteration capabilities. The method includes the following steps: This embodiment first collects multimodal design data in the field of industrial control from multiple sources, including but not limited to: international standards (IEC 61131-3, ISA-88, IEC 61511); hardware manuals, configuration software guides, and function block library descriptions from mainstream manufacturers (Siemens, Rockwell, ABB, Honeywell, Supcon Technology); professional textbooks and academic papers (process control, PLC programming, DCS configuration); industry terminology dictionaries and Chinese-English glossaries; design specifications, I / O tables, logic description documents, HMI design specifications, etc. from actual engineering projects.

[0030] The collected multimodal design data underwent data cleaning and preprocessing, specifically including: removing duplicate, invalid, or low-quality content; converting PDF files and scanned images into parsable plain text using OCR and layout analysis technologies; standardizing and unifying the annotation of technical terms and abbreviations (such as PID, DCS, SIS, ESD, MCC, OMC); segmenting long documents into segments suitable for model training while retaining necessary contextual relationships; and categorizing and storing the data according to technical fields (logic programming, hardware configuration, flowchart design, safety interlocks), vendor types (Siemens, Rockwell, etc., Supcon Technology), and industry (chemical, power, metallurgy, etc.) to provide a data foundation for the subsequent construction of hybrid expert models. The data related to Supcon Technology further includes the OMC system (full-process intelligent operation management and control system) operation manual and the Nyx system (next-generation intelligent control system) configuration development guide.

[0031] S1: Build and train a large vertical domain model. Use the large vertical domain model to perform multimodal analysis and intent recognition on multimodal design data, extract key industrial information, and generate dual-track output results based on the key industrial information. The dual-track output results include structured configuration operation suggestions and command sequences. The command sequences include tool parameters, execution order and logical dependencies for calling subsequent tools.

[0032] Preferably, in step S1, constructing and training a large model for a vertical domain includes: Based on a general large model, this embodiment utilizes a professional corpus in the industrial control field for pre-training, retrieval enhancement to generate RAGs, and instruction fine-tuning. Specifically, in this embodiment, the constructed industrial corpus is used to continuously pre-train the base general large model, employing an autoregressive language modeling task (predicting the next term based on the preceding context) to learn the syntax, semantics, and unique expression patterns of industrial texts. During training, a domain-specific word segmenter is used, treating terms such as "S7-1200," "MODBUS TCP," "OMC," and "Nyx" as independent terms. Perplexity is periodically calculated on the validation set, and pre-training stops once convergence is achieved and the domain term prediction accuracy reaches a preset threshold.

[0033] Retrieval Enhancement Generates (RAG): The cleaned corpus documents are sliced ​​(approximately 256 to 512 tokens per slice). An embedding model is used to convert each slice into a high-dimensional vector, which is then stored in a vector database (such as Milvus). Metadata (including source, manufacturer, industry, version, etc.) is also associated. When a user inputs a query or uploads design materials, the system converts the input into a query vector and performs a similarity search (such as cosine similarity) in the vector database. The top-K most relevant document slices are retrieved as additional context. The retrieved document slices are concatenated with the original query and fed into the larger model. The model is prompted to prioritize the retrieved materials for inference and generation, supplementing the data based on its own parametric knowledge only when the available data is insufficient.

[0034] Instruction Fine-Tuning: Based on real-world industrial application scenarios, a large number of instruction-response pairs are manually constructed or semi-automatically generated. Highly efficient fine-tuning techniques, such as low-rank adaptation, are employed to update only a subset of model parameters, reducing computational resource consumption and preventing overfitting. Multiple rounds of supervised fine-tuning are performed using the instruction dataset to minimize the difference between the predicted output and the standard response using the cross-entropy loss function. Task accuracy is evaluated on a validation set, and fine-tuning stops once the target is met. A hybrid expert model architecture is adopted to construct segmented knowledge bases or sub-models for different design institutes, manufacturers, or industries. After receiving multimodal design data, the corresponding segmented knowledge base or sub-model is activated through a routing mechanism to perform intent recognition. Specifically, in this embodiment, multiple expert sub-models or knowledge bases are designed according to the diversity of industrial scenarios, such as: chemical industry experts (focusing on continuous process control and reactor interlocking), power industry experts (focusing on steam turbine control and AGC / AVC), Siemens product line experts (TIA Portal, S7-1500), safety instrumented systems experts (IEC 61511, SIL calculation), and Supcon Technology product line expert sub-models: familiar with OMC system architecture, Nyx system configuration environment, and functional block characteristics.

[0035] A lightweight gating network is added, which takes user query vectors or extracted task feature vectors as input and dynamically calculates and outputs the weight coefficients of each expert sub-model. The gating network and expert sub-models are trained together, enabling problems involving specific industries or manufacturers to be automatically routed to the most suitable expert sub-model for processing. The knowledge base corresponding to each expert sub-model is stored and maintained independently, and new experts can be dynamically added without affecting the performance of existing experts (such as adding photovoltaic industry experts later). When manufacturers release a new version of their technical manuals, only the vector database of the corresponding expert sub-model needs to be updated to complete the knowledge update.

[0036] Preferably, in step S1, a large vertical domain model is used to perform multimodal analysis and intent recognition on the multimodal design data to extract key industrial information, including: The activated subdivided knowledge base or sub-model performs multimodal fusion analysis on multimodal design data to identify key industrial information, including equipment tag number, control loop type, interlocking logic and process flow description. Based on key industrial information, readable text explaining the configuration scheme logic is generated, along with a structured command sequence containing tool parameters for calling subsequent tools, execution order, and logical dependencies, such as: [{"action": "parse_p&id", "params": {"file": "pid.pdf"}}, {"action": "assign_io", "params":{...}}, {"action": "generate_logic", "params": {"type": "PID_LOOP", "tag": "LIC101"}}, ...].

[0037] Preferably, the structured command sequence includes tool parameters for calling subsequent tools, execution order, and logical dependencies, including: Structured command sequences are structured intermediate instruction sets designed for machine parsing, using a preset markup language format (such as JSON). The tool parameters include file path, device tag number, and control logic type; Execution order and logical dependencies are used to define the serial or concurrent scheduling relationships of tool calls in subsequent professional toolchains.

[0038] S2: The agent receives the command sequence, performs task planning and security verification, parses the command sequence into sub-tasks, and, according to the task plan, calls and orchestrates a professional toolchain to execute the sub-tasks, generating the configuration file of the industrial control system.

[0039] Preferably, in step S2, the agent receives a command sequence and performs task planning and security verification, including: A syntactic and logical integrity scan is performed on the structured command sequence to confirm the consistency between instructions and the closed-loop execution capability. Specifically, it checks whether the command format conforms to the predefined schema, verifies whether the dependencies between instructions form a directed acyclic graph, and confirms whether the necessary parameters are complete. In this embodiment, a step-by-step reasoning chain is used: for the PID loop, the tag number LIC101 (PV is the AI ​​signal and OP is the AO signal) is found in the IO table, and then "LIC101 is PID control, and the setpoint SP is given by the operator" is found in the logical description, thus confirming that it is a single-loop PID control. Based on the current user's role authorization scope, a security review is performed to identify and block dangerous operation instructions that involve modifying critical security interlocks. If the blocking rule is hit, the execution process is frozen and a secondary confirmation request is triggered to the user. The process can only continue after the user has given explicit authorization. After the verification is passed, the data flow and prerequisite dependencies between instructions are analyzed, a directed acyclic graph is constructed based on the dependency relationship, the level of each command is calculated, and the structured command sequence is decomposed and reorganized into a subtask queue containing concurrent and serial scheduling subtasks. Specifically, in this embodiment, commands at the same level and without mutual dependencies are marked as parallel execution groups to improve execution efficiency.

[0040] Preferably, in step S2, the agent invokes a specialized toolchain to execute subtasks to generate a configuration file according to the plan, including: The intelligent agent strictly follows the dependency graph of the subtask queue, scheduling data parsing tools to extract standardized equipment tag numbers and lists. Specifically, in this embodiment, the data parsing tool is invoked to accurately extract IO tables and tag number information from the original document; the hardware configuration tool is invoked to automatically complete the selection and allocation of racks and modules based on the number of IO points. This hardware configuration tool supports hardware configuration rules of Supcon OMC system, Nyx system, and mainstream DCS / PLC manufacturers such as Siemens and Rockwell; the logic generation engine is invoked to generate logic code (such as ladder diagrams, structured text, and sequential function charts) conforming to the IEC 61131-3 standard based on the control strategy template. For the graphical programming environment of Supcon OMC system and the modular configuration characteristics of Nyx system, the logic generation tool has built-in corresponding function block mapping library and control strategy template; the interface flowchart drawing tool is invoked to automatically generate HMI static screens containing dynamic links based on the process flow description and tag number association; and the specification checking tool is invoked to check the consistency, completeness, and specification compliance of the generated configuration. The extracted data is passed to the hardware configuration tool to complete the automatic selection and allocation of racks and modules, and drives the logic generation tool to match the control strategy template to generate logic code that conforms to industry standards; The HMI drawing tool is scheduled to generate the interface flowchart synchronously or immediately afterward, automatically establishing dynamic data links between screen elements and underlying tag numbers. Finally, the specification checking tool is scheduled to cross-compile and verify the consistency of the entire set of configuration files generated.

[0041] S3: Output the execution results of the configuration file. Based on user feedback and execution logs, iteratively optimize the planning logic of the vertical domain's large model and the agent. Specifically, in this embodiment, the agent collects all generated files (including hardware configuration, logic code, HMI screen, and specification check report), packages them into a configuration file compressed package, and generates a download link to provide feedback to the user. The execution report records in detail the execution time, success or failure status, and warning messages generated for each subtask. The user interface also provides feedback components such as satisfaction rating, problem category selection, and text opinion input, and supports users uploading their own modified configuration files for reference.

[0042] Preferably, step S3 involves iteratively optimizing the planning logic of the large vertical domain model and agent based on execution logs and user feedback, including: The execution log is linked to user evaluations and modification actions. Processes that execute smoothly and receive good user evaluations (e.g., rating ≥ 4 stars, modification rate < 5%) are encapsulated as positive samples, while processes that are interrupted, fail verification, or are heavily modified by users (e.g., modification rate > 30%) are encapsulated as negative samples. Positive samples are parsed and stored in a case library as corpus for subsequent retrieval enhancement or instruction fine-tuning to update the large model in the vertical domain. At the same time, the execution logs of negative samples are deeply analyzed to accurately locate failure nodes. Based on this, the scheduling rules and verification thresholds inside the intelligent body are modified accordingly. Specifically, in this embodiment, it also has the ability to continuously update external knowledge: regularly crawl the latest technical information from public or authorized knowledge sources (such as manufacturer websites and standard release platforms), and analyze excellent historical configuration cases within the enterprise to continuously enrich the knowledge base and design templates, so as to realize the continuous evolution of system capabilities.

[0043] Preferably, iterative optimization further includes a scheme reuse step: Save the structured command sequence and execution parameters corresponding to the configuration generation process that serves as a positive sample and meets the preset standardization conditions as an automated configuration scheme template. When new multimodal design data is received and the similarity of the scenario to the automated configuration scheme template is higher than a preset threshold, the template is directly invoked to execute configuration generation, skipping the structured command sequence generation step and some verification steps.

[0044] Second Embodiment Please see Figure 3 As shown, based on the same concept, this invention also provides an industrial control system configuration generation system based on a large vertical domain model and intelligent agents, comprising: User interaction layer: includes user input interface 101 and result feedback interface 102. Through this layer, users can input project information, upload design materials (such as PDF drawings and IO tables), and receive the final generated configuration files and execution reports.

[0045] The vertical domain large model module is bidirectionally connected to the knowledge base and case library module 111. It is used to build and train the vertical domain large model. The vertical domain large model is used to perform multimodal analysis and intent recognition on multimodal design data, extract key industrial information, and generate dual-track output results based on the key industrial information. The dual-track output results include structured configuration operation suggestions and command sequences. The command sequences include tool parameters for calling subsequent tools, execution order, and logical dependencies. Specifically, in this embodiment, the knowledge base and case library module 111 stores vertical domain knowledge, manufacturer information, standard templates, historical success cases, etc., for the large model to call and self-learn and update. The intelligent agent planning and scheduling module 105 is used for the intelligent agent to receive command sequences, perform task planning and security verification, and parse the command sequences into sub-tasks. Specifically, in this embodiment, it is connected to the security management module 106 for permission verification. The execution toolchain module (execution tool layer) is scheduled by the agent planning and scheduling module. It is used by the agent to call and arrange professional toolchains to execute sub-tasks according to the task plan and generate configuration files for the industrial control system. Specifically, in this embodiment, the data parsing tool 106 (extracts IO tables and tag numbers), the hardware configuration tool 107 (generates hardware configuration), the logic generation tool 108 (generates control logic code), the HMI drawing tool 109 (generates flowchart screens), and the specification checking tool 110 (verifies configuration integrity and standardization) are all used. The execution status and results of all tools are returned to the agent module. The self-learning iteration module is used to output the execution results of the configuration file. Based on user feedback and execution logs, iteratively optimizes the planning logic of the large model and agent in the vertical domain. Specifically, in this embodiment, the log and feedback library 112 records the process logs and user feedback for each execution, which are used for subsequent analysis and large model iterative training.

[0046] Third Embodiment In this embodiment, a computer device is provided, including a memory and one or more processors. The memory stores computer code. When the computer code is executed by one or more processors, the one or more processors cause the one or more processors to perform the steps of the industrial control system configuration generation method based on a vertical domain large model and intelligent agent in the first embodiment.

[0047] In some embodiments of this application, a computer-readable storage medium is also provided, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors perform the steps of an industrial control system configuration generation method based on a vertical domain large model and intelligent agent as described in any one of the first embodiments.

[0048] It is understood that, regarding the aforementioned method for generating industrial control system configurations based on a large vertical domain model and intelligent agents, if all components are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer server or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0049] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0050] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating configurations for industrial control systems based on a large vertical domain model and intelligent agents, characterized in that, Includes the following steps: S1: Construct and train a large vertical domain model, use the large vertical domain model to perform multimodal analysis and intent recognition on multimodal design data, extract key industrial information, and the large vertical domain model generates dual-track output results based on the key industrial information. The dual-track output results include structured configuration operation suggestions and command sequences. The command sequences include tool parameters, execution order and logical dependencies for calling subsequent tools. S2: The intelligent agent receives the command sequence, performs task planning and security verification, parses the command sequence into sub-tasks, and, according to the task planning, calls and orchestrates a professional toolchain to execute the sub-tasks, generating a configuration file for the industrial control system. S3: Output the execution result of the configuration file, and iteratively optimize the planning logic of the vertical domain large model and intelligent agent based on user feedback and execution logs.

2. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 1, characterized in that, In step S1, a large vertical domain model is constructed and trained, including: Based on a general large model, we use professional corpus in the field of industrial control for pre-training, retrieval enhancement to generate RAG and fine-tuning of instructions; A hybrid expert model architecture is adopted to build segmented knowledge bases or sub-models for different design institutes, manufacturers or industries. After receiving the multimodal design data, the corresponding segmented knowledge bases or sub-models are activated through a routing mechanism to perform intent recognition.

3. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 1, characterized in that, In step S1, the vertical domain large model is used to perform multimodal analysis and intent recognition on the multimodal design data to extract key industrial information, including: The activated sub-knowledge base or the sub-model performs multimodal fusion analysis on the multimodal design data to identify the key industrial information, including equipment tag number, control loop type, interlocking logic and process flow description. Based on the aforementioned key industrial information, a readable text explaining the configuration scheme logic is generated, along with a structured command sequence containing tool parameters for calling subsequent tools, execution order, and logical dependencies.

4. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 3, characterized in that, The structured command sequence, which includes tool parameters for calling subsequent tools, execution order, and logical dependencies, includes: The structured command sequence is a structured intermediate instruction set oriented towards machine parsing, using a preset markup language format; The tool parameters include file path, device tag number, and control logic type; The execution order and the logical dependencies are used to define the serial or concurrent scheduling relationships of tool calls in the subsequent professional toolchain.

5. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 1, characterized in that, In step S2, the agent receives the command sequence and performs task planning and security verification, including: Perform a syntactic and logical integrity scan on the structured command sequence to confirm the consistency between instructions and the ability to execute in a closed loop. Based on the current user's role authorization scope, a security review is conducted to identify and block dangerous operation instructions that involve modifying critical safety interlocks. If the blocking rule is hit, the execution process is frozen and a secondary confirmation request to the user is triggered. After the verification is passed, the data flow and prerequisite dependencies between the instructions are analyzed, and the structured command sequence is decomposed and reassembled into a subtask queue containing concurrent and serial scheduling subtasks.

6. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 1, characterized in that, In step S2, the agent invokes a specialized toolchain according to the plan to execute the sub-tasks to generate a configuration file, including: The intelligent agent strictly follows the dependency graph of the sub-task queue to schedule data parsing tools to extract standardized device tag numbers and lists; The extracted data is passed to the hardware configuration tool to complete the automatic selection and allocation of racks and modules, and drives the logic generation tool to match the control strategy template to generate logic code that conforms to industry standards; The HMI drawing tool is scheduled to generate the interface flowchart synchronously or immediately afterward, automatically establishing dynamic data links between screen elements and underlying tag numbers. Finally, the specification checking tool is scheduled to cross-compile and verify the consistency of the entire set of configuration files generated.

7. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 1, characterized in that, Step S3 involves iteratively optimizing the planning logic of the vertical domain large model and the intelligent agent based on execution logs and user feedback, including: The execution logs are linked with user evaluations and modification actions. Processes that execute smoothly and receive excellent user evaluations are encapsulated as positive samples, while processes that are interrupted, fail verification, or are heavily modified by users are encapsulated as negative samples. After parsing the positive samples, they are stored in the case library as corpus for subsequent retrieval enhancement or instruction fine-tuning to update the vertical domain model. At the same time, the execution logs of the negative samples are deeply analyzed to accurately locate failure nodes, and the scheduling rules and verification thresholds inside the intelligent body are modified accordingly.

8. The industrial control system configuration generation method based on a large vertical domain model and intelligent agents according to claim 7, characterized in that, The iterative optimization also includes a scheme reuse step: The structured command sequence and execution parameters corresponding to the configuration generation process that serves as a positive sample and meets the preset standardization conditions are saved as an automated configuration scheme template. When new multimodal design data is received and the scene similarity with the automated configuration scheme template is higher than a preset threshold, the template is directly invoked to execute configuration generation, skipping the structured command sequence generation step and some verification steps.

9. An industrial control system configuration generation system based on a large vertical domain model and intelligent agents, characterized in that, include: The Vertical Domain Large Model Module is used to build and train a vertical domain large model. The vertical domain large model is used to perform multimodal analysis and intent recognition on multimodal design data and extract key industrial information. The vertical domain large model generates dual-track output results based on the key industrial information. The dual-track output results include structured configuration operation suggestions and command sequences. The command sequences include tool parameters for calling subsequent tools, execution order and logical dependencies. The agent planning and scheduling module is used for the agent to receive the command sequence, perform task planning and security verification, and parse the command sequence into subtasks; The execution toolchain module, scheduled by the agent planning and scheduling module, is used by the agent to call and orchestrate the professional toolchain to execute the sub-tasks according to the task plan, and generate the configuration file of the industrial control system. The self-learning iteration module is used to output the execution results of the configuration file and iteratively optimize the planning logic of the vertical domain large model and intelligent agent based on user feedback and execution logs.

10. A computer device comprising a memory and one or more processors, the memory storing computer code that, when executed by the one or more processors, causes the one or more processors to perform the steps of the industrial control system configuration generation method based on a vertical domain large model and intelligent agents as described in any one of claims 1-8.

Images