A method and system for automatically generating a context operation task based on a Wenxun intelligent agent

Abstract

The application relates to the field of artificial intelligence and education technology, and particularly discloses a context operation task automatic generation method and system based on a Wenxun intelligent agent, which comprises an input layer, a Wenxun intelligent agent engine layer, a knowledge layer, a simulation layer and an output layer; wherein the Wenxun intelligent agent engine layer comprises multiple intelligent agents, is used for analyzing natural language instructions and automatically generating complete experiment packages; a first intelligent agent receives the natural language instructions and generates a structured requirement description; a second intelligent agent generates a matched experiment framework according to the analysis result; a third intelligent agent generates detailed operation steps of each experiment stage based on the experiment framework; a fourth intelligent agent maps the operation steps to a three-dimensional equipment model and a mathematical simulation model of a simulation platform and completes parameter binding to form a complete experiment package; and a fifth intelligent agent verifies the experiment package and deploys qualified experiment packages to a virtual simulation training platform. The application is suitable for a vocational skill training scene requiring a large number of virtual simulation experiment resources.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and education technology, and in particular to a method and system for automatically generating contextual operation tasks based on a textual and training intelligent agent. Background Technology

[0002] Virtual simulation training experiments are an important vehicle for vocational skills training. They can effectively avoid the problems of high safety risks, large equipment wear and tear, and difficulty in reproducing scenarios in on-site training. They have been widely used in professional training scenarios such as campus integrated energy dispatching, meter installation and connection, metering device inspection, and source-grid-load-storage coordination operation. They are an important support for realizing large-scale, standardized, and personalized training teaching.

[0003] In virtual simulation training platforms, scenario-based operational tasks are built upon real-world work scenarios, including complete operational procedures and executable interactive logic, providing instructors with an immersive learning experience. However, current technology requires professional personnel to manually design and build scenario-based operational tasks, involving multiple stages such as requirements analysis, content writing, parameter configuration, platform integration, and review testing. This results in a long development cycle and significant manpower investment. Once the training content is solidified, it is difficult to update, hindering rapid responses to business and technological iterations. Quality depends on the professional level and experience of the writers, lacking a systematic compliance review mechanism. The rationality of experimental parameters is difficult to guarantee, posing a risk of deviating from real equipment behavior.

[0004] The aforementioned problems have resulted in a long-term state of insufficient supply, lagging updates, and uneven quality in the construction of practical training resources, which seriously restricts the large-scale promotion and application of virtual simulation training. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for automatically generating contextual operation tasks based on textual training agents. Using natural language as the input interface, and through the collaborative execution and iterative verification of multiple specialized agents, it automatically outputs a complete training experiment package that can be directly run on a simulation platform. This method is suitable for vocational skills training scenarios that require a large number of virtual simulation experiment resources.

[0006] According to one aspect of the present invention, a context-based automatic task generation system for training agents is provided, comprising: The input layer is used to receive natural language instructions; The Wenxun Intelligent Agent Engine Layer comprises multiple intelligent agents used to parse natural language instructions and automatically generate complete experimental packages. Specifically: a first intelligent agent receives natural language instructions and performs intent parsing and information feature extraction to obtain a structured requirement description; based on the structured requirement description, a second intelligent agent automatically generates an experimental framework matching the natural language instructions; based on the experimental framework, a third intelligent agent generates detailed operation steps for each experimental stage within the framework; a fourth intelligent agent maps the detailed operation steps to the 3D device model and mathematical simulation model of the simulation platform, performs parameter binding, and obtains a complete experimental package containing parameter binding; a fifth intelligent agent performs verification operations on the experimental package, and deploys the verified experimental package to the virtual simulation training platform; the content of the experimental package includes one or more scenario-based operation tasks. The knowledge layer is used to enhance the knowledge base through RAG retrieval and provide professional support. The simulation layer is used to interface with the campus integrated energy system to bind parameters to the model; The output layer is used to output the complete experimental package that has passed verification.

[0007] Optionally, the intelligent agent encapsulates the large language model, role-specific prompts, callable tool interfaces, and loop control logic.

[0008] Optionally, the structured requirements description includes the experiment type, target learner level, difficulty level, and the knowledge domains and skills involved.

[0009] Optionally, the fourth intelligent agent maps the detailed operation steps to the three-dimensional equipment model and mathematical simulation model of the simulation platform, performs parameter binding, and obtains a complete experimental package containing parameter binding. Specifically, it maps upward to the three-dimensional equipment model and downward to the mathematical simulation model, and injects the equipment's rated parameters, operating conditions, and boundary conditions. The rated parameters of the equipment include the equipment nameplate database and the historical operating condition sample database.

[0010] Optionally, each bound parameter can be subjected to both hard and soft constraint verification, which includes hard constraint verification and soft constraint verification.

[0011] Optionally, hard constraint verification is based on the physical characteristics and technical specifications of the equipment, setting value boundaries for various parameters. If any parameter exceeds the hard boundary, the fifth intelligent system will directly determine that it fails, mark the violation field and the corresponding specification item in the audit report, and force a backtracking and repair.

[0012] Optionally, soft constraint verification is based on the statistical patterns of historical operating data, calculating the mean and standard deviation of each parameter under typical operating conditions, and constructing a reasonable reference interval.

[0013] Optionally, the verification operation involves reviewing the complete experimental package from five aspects: process integrity, operational compliance, safety specification coverage, simulation parameter rationality, and assessment standard feasibility.

[0014] Optionally, the collaboration of multiple agents in the Wenxun agent engine layer is driven by a two-layer mechanism: the orchestration layer and the prompt word engineering layer. The orchestration layer adopts a directed acyclic graph. The prompt word engineering layer equips each agent with an independent system prompt word and embeds four functional modules: role definition, output format template, prohibited behavior constraints, and domain knowledge background injection.

[0015] According to another aspect of the present invention, a method for automatically generating contextual operation tasks based on a textual training agent is provided, comprising: S1: Receives natural language commands from the user through the input layer; S2: The Wenxun intelligent agent parses natural language instructions and automatically generates a complete experimental package; wherein, the Wenxun intelligent agent includes multiple intelligent agents, the first intelligent agent receives natural language instructions, and performs intent parsing and information feature extraction on the instructions to obtain a structured requirement description; S3: Based on intent parsing and information features, an experimental framework that automatically generates natural language instructions using a second intelligent agent; S4: Based on the experimental framework, generate detailed operation steps for each experimental stage under the experimental framework according to the third agent; S5: Based on the fourth intelligent agent, the detailed operation steps are mapped to the 3D device model and mathematical simulation model of the simulation platform, and parameter binding is performed to obtain a complete experimental package containing parameter binding. S6: The fifth intelligent agent performs a verification operation on the experimental package and deploys the verified experimental package to the virtual simulation training platform; the content of the experimental package includes one or more scenario operation tasks; S7: Output the validated experimental package through the output layer to provide feedback to the user on the final scenario operation task.

[0016] According to another aspect of the invention, a storage medium is provided that non-transitory stores computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in any of the above aspects is performed.

[0017] According to another aspect of the present invention, an electronic device is provided, comprising: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any of the above aspects.

[0018] The above technical solution has the following advantages or beneficial effects: This invention utilizes a large-model-driven multi-agent architecture, allowing instructors to simply describe their needs in natural language. The system then automatically generates a complete, directly runnable virtual simulation training experiment—an operable scenario task. Instructors can adjust the experiment content, difficulty, and assessment methods at any time, generating tasks according to the target learners' levels and knowledge domains, thus achieving personalized instruction.

[0019] Through multi-agent collaboration, each agent performs its own function and completes the entire process from requirement analysis to experimental delivery through pipeline collaboration. Finally, the process is reviewed in five aspects: the integrity of the execution process of the verification agent, the compliance of operation, the coverage of safety specifications, the rationality of simulation parameters, and the feasibility of assessment standards. If the review fails, a targeted backtracking and repair mechanism is triggered, rather than generating an uncontrollable result all at once. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 Here is the overall architecture diagram of the contextual operation task automatic generation system based on textual training agent in Example 1; Figure 2 This is a flowchart of the Wenxun intelligent agent engine layer in Example 1. Detailed Implementation

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0025] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 Terminology Explanation: The Wenxun Intelligent Agent is a multi-functional composite intelligent agent based on large-scale model technology, specifically designed for practical training tasks to intelligently generate scenario-based operational tasks. "Wen" refers to the large-scale model's ability to understand and reason about text, while "Xun" refers to the scenario-based tasks closely aligned with skills training. It deeply integrates job standards and knowledge graphs, extracts keywords and understands semantics from input statements, and generates complete practical training solutions in real time through natural language interaction, including work context, task requirements, implementation steps, and assessment standards. This completely changes the traditional situation of long development cycles and high costs for scenario-based operational tasks, making it possible to produce personalized, on-the-spot-driven training resources. It enables skills training to highly focus on real job requirements, truly achieving "training generated from text, integration of training and application," and propelling technical skills into a new stage of on-demand training and personalized solutions for each individual.

[0028] Scenario-based operational tasks: These are comprehensive training solutions generated in real-world job scenarios by the Wenxun intelligent agent. Leveraging a large-scale model to semantically understand and extract keywords from user input, the solution instantly integrates the work context, task requirements, implementation steps, and assessment standards. It deeply integrates job standards, knowledge graphs, and simulation models, reproducing frontline operational scenarios in highly structured task units, ensuring skills training precisely focuses on core job requirements. This task format can generate personalized content on demand, supporting a new stage of precision and differentiation in technical skills training.

[0029] Figure 1 This is a schematic diagram of the structure of an automatic generation system for contextual operation tasks based on a textual training agent, provided in an exemplary embodiment of the present invention. For example... Figure 1 The context-based task generation system based on textual training agents, as shown, includes: The input layer is used to receive natural language instructions; The Wenxun Intelligent Agent Engine Layer comprises multiple intelligent agents used to parse natural language instructions and automatically generate complete experimental packages. Specifically: a first intelligent agent receives natural language instructions and performs intent parsing and information feature extraction to obtain a structured requirement description; based on the structured requirement description, a second intelligent agent automatically generates an experimental framework matching the natural language instructions; based on the experimental framework, a third intelligent agent generates detailed operation steps for each experimental stage within the framework; a fourth intelligent agent maps the detailed operation steps to the 3D device model and mathematical simulation model of the simulation platform, performs parameter binding, and obtains a complete experimental package containing parameter binding; a fifth intelligent agent performs verification operations on the experimental package, and deploys the verified experimental package to the virtual simulation training platform; the content of the experimental package includes one or more scenario-based operation tasks. The knowledge layer is used to enhance the knowledge base through RAG retrieval and provide professional support. The simulation layer is used to interface with the campus integrated energy system to bind parameters to the model; The output layer is used to output the complete experimental package that has passed verification.

[0030] Specifically, large models or agents are used to generate general content, such as text, poems, essays, and general web page code. These capabilities are inherent to the large models themselves and do not involve simulation calculations specific to any particular field.

[0031] Furthermore, the experimental content of existing virtual simulation training platforms is fixed; whatever is purchased is what is available, limiting the scope and content that can be trained. Experimental resources are manually designed by professionals, resulting in long development cycles, significant manpower investment, and difficulty in updating content once it is solidified. If an instructor wants to customize an experiment to address a student's weak knowledge points or add a new training scenario, they lack software development capabilities and cannot modify or expand the experimental content on their own.

[0032] This system utilizes a large-model-driven multi-agent architecture, allowing instructors to simply describe their needs in natural language, and the system will automatically generate a complete, ready-to-run virtual simulation training experiment. Instructors can adjust the experiment content, difficulty, and assessment methods at any time, generating experiments according to the target learners' levels and knowledge domains, thus achieving personalized instruction.

[0033] like Figure 1 As shown, the present invention proposes an automatic generation system for contextual operation tasks based on a textual training agent, comprising: 1. Input layer, used to receive natural language instructions, including text and speech.

[0034] For example, if an instructor inputs a description of an experimental requirement in natural language, such as generating a source-network-load-storage scheduling experiment for students at a certain level.

[0035] 2. The Wenxun Intelligent Agent Engine Layer includes multiple intelligent agents for parsing natural language instructions and automatically generating complete experimental packages. Specifically: the first intelligent agent receives natural language instructions and performs intent parsing and information feature extraction to obtain a structured requirement description; based on the structured requirement description, the second intelligent agent automatically generates an experimental framework that matches the natural language instructions; based on the experimental framework, the third intelligent agent generates detailed operation steps for each experimental stage under the experimental framework; the fourth intelligent agent maps the detailed operation steps to the 3D device model and mathematical simulation model of the simulation platform, performs parameter binding, and obtains a complete experimental package containing parameter binding; the fifth intelligent agent performs verification operations on the experimental package, and deploys the verified experimental package to the virtual simulation training platform; the content of the experimental package includes one or more scenario operation tasks.

[0036] The Wenxun Intelligent Agent Engine Layer, such as Figure 2 As shown, it includes 5 intelligent agents with different functions. These 5 intelligent agents with different functions are executed in series in a fixed order to form a closed-loop mechanism of generation, verification and repair.

[0037] The intelligent agent is driven by a large language model, which serves as the inference engine. It encapsulates the large language model, role-specific prompts, callable tool interfaces, and loop control logic. The large language model, as the underlying inference engine, is responsible for understanding natural language and generating text and code.

[0038] The prompts for specific roles are compiled by combining knowledge base content and industry experts to create a prompt template library for roles.

[0039] Callable tool interfaces are requests that an agent outputs during inference. An external program executes the tool and returns the result to the agent as a message. These include various capabilities such as knowledge retrieval tools, precise calculations, operating business systems, and processing multimedia information.

[0040] Loop control logic is an iterative execution mechanism within an agent that drives a large model to perform multi-step inference, tool calls, and result verification. It is a fixed program coded and implemented by developers at the agent framework level.

[0041] The coordination of multiple agents is pipeline-like: the output of one agent becomes the input of the next. Simultaneously, a central control loop monitors whether the overall goal has been achieved; if not, it drives the corresponding agent to continue working or to correct the issue. This central control loop validates the agents; the generator framework verifies the agents' outputs, and the requirements parsing needs to output structured results: whether they are empty, and whether they meet the requirements.

[0042] The first intelligent agent is the requirement parsing agent, responsible for requirement parsing. It receives natural language input from the user, performs intent recognition and information extraction, and outputs a structured requirement description JSON. The structured requirement description JSON includes: experiment type, target learner level, difficulty level, and relevant knowledge domains and skills.

[0043] Using a language model pre-trained on a general large-scale corpus as a base, further pre-training (i.e., domain-adaptive pre-training) is performed on the base model using unlabeled text from a specific domain. This allows the model to fully learn the contextual collocations and semantic relationships of domain-specific terminology, thereby acquiring solid domain-specific language knowledge representations. Subsequently, a downstream task dataset containing terminology annotation information is constructed, and the domain-adapted model is subjected to supervised fine-tuning. This enables the model to acquire specific natural language understanding capabilities such as intent recognition and entity extraction. The final model, when understanding domain-specific text, can achieve high-accuracy terminology recognition and extraction based on the learned prior knowledge of terminology, rather than relying solely on guesswork based on general language understanding capabilities.

[0044] The input to the requirement parsing agent is the experiment requirements described by the instructor in natural language. It generates a source-network-load-storage scheduling experiment for a specific learner level. Through multiple rounds of dialogue with the instructor, the requirement parsing agent repeatedly confirms details such as the experiment type, target learner level, difficulty level, and covered knowledge points. The output is a structured requirement description.

[0045] The second agent is the experiment planning agent, used for experiment planning. It takes a structured requirement description JSON as input, retrieves the most matching template type from the SOP template library, and automatically plans the number of experiment stages, learning objectives and assessment dimensions for each stage based on the template framework. It outputs an experiment framework JSON, which is then used by subsequent agents to fill in the detailed operation steps and content corresponding to the work process stage by stage.

[0046] The SOP template library is a standardized operating procedure library, which is created based on the standardized operating instructions for the corresponding positions in the scenario-based operation tasks.

[0047] The input to the experiment planning agent is the structured requirement description from the requirement parsing agent. Based on the structured requirement description, the experiment planning agent retrieves the most matching template type from the SOP template library, combines the professional knowledge in the knowledge base, plans the number of experiment stages, the learning objectives and assessment dimensions for each stage, and finally outputs the experiment framework plan.

[0048] The third agent is a step-generating agent, used to generate the various operational steps within the experimental framework of the second agent.

[0049] Using a structured experimental framework JSON as input, which explicitly defines the overall experimental process, objectives for each stage, required equipment, and key operational objects, the system automatically triggers a Retrieval Enhancement Generation (RAG) pipeline for each stage of the experimental framework. First, key information is extracted from the stage description to form a semantic query vector for that stage. Then, a pre-built domain knowledge base is used to accurately retrieve multi-source documents related to the semantic query vector for that stage, including operating procedures, safety specifications, standardized operating instructions, and historical experimental cases. These retrieved multi-source documents are then sorted by relevance, redundancy removed, and fragment fused to form a high-quality reference context for that stage. Subsequently, the generation module of the step-generating agent feeds the structured experimental framework JSON and the high-quality reference context into a large language model, forcing the model to follow a four-tuple structure of "operation object—operation action—expected phenomenon—safety prompt," progressively generating detailed operational steps for that stage.

[0050] The system defines the operation object as the equipment, material, or system interface to be operated on; the operation action describes the specific behaviors and standard procedures that must be performed; the expected phenomena provide the results, data, or status changes that should be observed after normal operation, serving as a basis for verification; and the safety prompts highlight potential risks, prohibited operations, and protective measures in this step, referencing relevant safety regulations. During the generation process, historical cases are prioritized for supplementing common errors and handling experience, making the steps more practical. Once all stages have completed the generation of RAG-based four-tuple steps, the system assembles these steps according to the stage sequence and hierarchical relationship of the experimental framework, automatically outputting a structured, executable, and fully safe experimental operation task.

[0051] Simultaneously, based on the RAG search operation procedures and historical failure cases retrieved from a pre-built domain knowledge base, normal workflow paths and failure handling branch paths are generated and integrated into the experimental operation tasks in parallel. The normal path is used for routine experimental execution, while the failure path handles abnormal operating conditions. Setting up dual paths allows for the pre-setting of risk handling plans, improves emergency response logic, and ensures that the entire experimental process is safe, controllable, and feasible.

[0052] The Retrieval Enhancement Generation (RAG) mechanism is based on a pre-stored multi-domain knowledge base. All documents are segmented and converted into high-dimensional vectors through an embedding model. Before the third agent generates detailed operation steps each time, the target description of the current experimental stage is used as the query vector. The most relevant Top-K segments of professional materials are selected based on semantic similarity as the generation context. The professional materials, experimental framework constraints, and SOP template format are concatenated into prompt words and input into the large language model to generate detailed operation steps.

[0053] The experimental framework constraint is the outline (framework) that the experimental planning agent outputs based on the structured requirements description, by retrieving the most matching template type from the SOP template library, combining the professional knowledge in the knowledge base, planning the number of experimental stages, the learning objectives and assessment dimensions for each stage, and generating the experimental task.

[0054] The database includes multiple vertical domain knowledge bases: a domain training outline library, storing training objectives and course standards for different learner levels; an equipment operation procedure library, containing standard operating procedures and precautions for various types of equipment; a safety operation specification library, covering safety briefing templates, personal protective equipment requirements, and emergency response plans; an SOP template library, containing standard framework structures for different types of experiments; a technical standard library, storing relevant industry and national standard documents; and a historical experiment case library, storing validated experiment packages for reference in similar scenarios.

[0055] The retrieved professional materials are filtered using a relevance threshold. The system presets a semantic similarity threshold, retaining only professional materials that are highly relevant to the operational needs of the current experimental stage, and removing low-relevance and irrelevant content. This ensures the quality of the context injected into the large model, so that the generated content can be traced back to specific procedure items, and the final generated detailed operation steps are verifiable and interpretable.

[0056] The experimental framework is input into the step-generating agent. For each experimental stage, relevant operating procedures, safety specifications, and historical cases are retrieved from the knowledge base through semantic retrieval. Detailed operating steps are generated, including a four-tuple structure of "operation object, operation action, expected phenomenon, and safety prompt." Parallel generation of normal processes and fault handling branch paths is also supported. Finally, the complete operating steps for each stage are output.

[0057] The fourth intelligent agent is a parameter-binding agent, used to bind the mathematical model parameters of the equipment to the corresponding 3D model, and to bidirectionally map detailed operation steps with a four-tuple structure to the simulation platform. Specifically, it maps upward to the 3D equipment model, binding interactive events and visual feedback; and maps downward to the mathematical simulation model, injecting the equipment's rated parameters, operating conditions, and boundary conditions. The equipment's rated parameters include the equipment nameplate database and the historical operating condition sample library. To ensure the authenticity and rationality of the injected parameters, both soft and hard constraints are used to verify their authenticity.

[0058] For example, the switching status of a disconnector: when the 3D device model is modeled, a control interface is included. The parameters are bound to the intelligent system for querying and configuring parameters in the project to link with the 3D device model, and mapped to the control status variable in a specific disconnector-related module of the mathematical model.

[0059] The system achieves precise correlation between operational steps and the simulation model through a two-way data interface between the fourth intelligent agent and the simulation platform. To ensure the authenticity and rationality of the injected parameters, the system performs both soft and hard constraint verification on each bound parameter.

[0060] The dual constraint verification includes both hard constraint verification and soft constraint verification. Hard constraint verification sets boundary values ​​for various parameters based on the equipment's physical characteristics and technical specifications. If any parameter exceeds the hard boundary, the fifth-level intelligent system will directly determine it as a failure, marking the violation field and corresponding specification item in the audit report, and forcibly triggering backtracking and repair. Soft constraint verification calculates the mean and standard deviation of each parameter under typical operating conditions based on the statistical patterns of historical operating data, constructing a reasonable reference range.

[0061] Specifically, the hard constraint verification is based on the physical characteristics and technical specifications of the equipment, setting insurmountable value boundaries for various parameters. Equipment physical characteristics include: rated capacity upper limit, safe operating voltage range, temperature protection threshold, etc. Technical specifications include: national standards, industry standards, technical manuals, design specifications, safety regulations, etc. The mandatory technical indicators, parameter limits, operating boundaries, and design principles explicitly stipulated in the technical specifications for the equipment are legally mandated and industry-mandated requirements. If any parameter exceeds the hard boundaries, the fifth-party intelligent system will directly determine it as a failure, marking the violation field and corresponding specification item in the audit report, and forcibly triggering a backtracking and repair process.

[0062] Soft constraint verification is based on the statistical patterns of historical operating data. The mean and standard deviation of each parameter under typical operating conditions are calculated to construct a reasonable reference interval. Specifically: first, historical operating data for each parameter under typical operating conditions are extracted, and abnormal interference values ​​are removed; the mean μ and standard deviation σ of the data are calculated, and a reference interval [μ] is constructed according to the 3σ criterion. [3σ, μ+3σ]. Real-time parameters are compared with the interval; values ​​exceeding this range are deemed unreasonable, triggering a soft constraint warning. Parameter values ​​falling outside the 3σ range trigger a soft warning, and the fifth agent marks "deviation from typical operating range" in the review report, prompting manual review but not forcibly rejecting the request, thus maintaining tolerance for special operating conditions.

[0063] Historical operating data was obtained through background records of typical operating conditions and on-site surveys.

[0064] The dual constraint mechanism of soft and hard ensures that every parameter in the simulation experiment is supported by data, and the system response obtained by the trainees when operating on the simulation platform is consistent with the behavior of the real equipment, avoiding the formation of incorrect operating perceptions by trainees due to parameter distortion.

[0065] The complete operational steps for each stage are input parameters and bound to an intelligent agent. This agent bidirectionally maps each operational step to the simulation platform: upwards to the 3D equipment model, binding interactive events and visual feedback; downwards to the mathematical simulation model, injecting the equipment's rated parameters, operating conditions, and boundary conditions. Parameter sources include the equipment nameplate database and a historical operating condition sample library, and their authenticity is ensured through both hardware and software constraints. The final output is a complete experimental package with fully bound parameters.

[0066] The fifth agent is the verification agent, which reviews the generated complete experimental package from five directions to achieve a full-process self-testing and verification process.

[0067] The complete experimental package is generated collaboratively by various agents: from the moment the request is received, the entire process output (including scenarios, steps, standards, etc.) is synchronized to the system configuration system in real time. Configuration files are gradually completed as the process progresses, evolving from initial requirements into structured project definitions, ultimately forming the complete experimental package. The system provides a `build_release` interface tool, which is injected during agent configuration. Once the agent confirms the configuration is complete, it can call this tool at any time to mark the release and package and deliver the package with one click.

[0068] The five audit areas include: Process completeness: verifying the coverage and logical coherence of steps; Operational compliance: verifying the matching degree between the operation content and the pre-set specification library; Safety specification coverage: verifying the completeness of the three elements of safety briefing / protective measures / emergency response; Simulation parameter rationality: verifying the pass rate of parameters through the dual constraints of soft and hard mechanisms; and Assessment standard executability: confirming whether the assessment scoring nodes can be accurately quantified and measured by the simulation platform. If any dimension fails the audit, the verification agent will output a structured discrepancy report, triggering a targeted backtracking and repair mechanism.

[0069] Structured discrepancy reports come in two forms. The first is the content structure that the agent is required to generate, as explicitly required by the system process, such as the data format of the required content and the configuration of the simulation algorithm. The second is to verify the subjective comprehensive judgment conclusion made by the agent based on professional rules and to clarify whether the overall review is passed or not.

[0070] The targeted retrospective repair mechanism combines the incomplete and non-standard items clearly identified in the first category of information in the structured difference report with the supplementary verification of the agent's subjective judgment results. It accurately locates the content to be corrected and the corresponding responsible agent, and directly outputs precise repair instructions to guide the relevant agents to complete targeted rectification.

[0071] A professional knowledge base built on industry standards will be provided to the verification agent. The agent will call the knowledge base query tool to verify and confirm the standardization of experimental content, complete the review of process integrity, operational compliance, safety specification coverage, simulation parameter rationality, and assessment standard feasibility, and support full-process iterative verification.

[0072] The complete experimental package with bound parameters and the requirement description from the requirement parsing agent are input into the verification agent. This agent reviews the experimental package from five aspects: process completeness, operational compliance, safety specification coverage, simulation parameter rationality, and assessment standard feasibility. The rationality of simulation parameters is verified using both hard and soft constraints. Hard constraints set insurmountable value boundaries based on the physical characteristics of the equipment, while soft constraints construct a rationality reference range based on historical operating data. The final output is of two types: one is verification passed, in which case the experimental package is delivered to the instructor for direct use; the other is failure to verify, in which case a structured discrepancy report is output and directed to the corresponding agent for repair and re-verification.

[0073] Furthermore, the collaboration of multiple agents in the text-based intelligent agent engine layer is driven by a two-layer mechanism: an orchestration layer and a prompt word engineering layer. The orchestration layer employs a directed acyclic graph; the prompt word engineering layer equips each agent with an independent system prompt word, embedding four functional modules: role definition, output format template, prohibited behavior constraints, and domain knowledge background injection. Specifically: The collaboration of the five agents in the Wenxun agent engine layer relies on the joint driving mechanism of the orchestration layer and the prompt word engineering layer.

[0074] The orchestration layer uses a directed acyclic graph (DAG) to describe the dependencies and data flow between agents. Each agent's output is encapsulated as a JSON structure, with field types and required fields strongly validated by a predefined schema. This ensures that upstream data is in a valid format before being passed to downstream agents, preventing errors from silently propagating throughout the process. The SOP orchestrator manages the execution status, scheduling order, and triggering and termination conditions for loop iterations of each agent.

[0075] like Figure 2 As shown, natural language input is used as the initial input, which may include instructor needs and training needs. This input then passes through a first agent to parse the needs, identifying the need type and college level. A second agent performs experiment planning, including template matching and experimental framework design. A third agent generates steps within the experimental framework, including procedure writing and RAG retrieval. A fourth agent binds parameters, including model mapping and data injection. A fifth agent performs self-testing and verification, including compliance review and quality checks. If the verification by the fifth agent passes, a complete experimental package is output, including experimental steps, operating procedures, safety precautions, simulation parameters, and assessment criteria. If the verification fails, the system returns to the agent that encountered the problem for targeted repair, eliminating the need to restart the entire process, reducing iteration overhead, and improving the efficiency of automatic generation of scenario-based operational tasks.

[0076] The prompt word engineering layer equips each agent with an independent system prompt word, embedding four functional modules: role definition, output format template, prohibited behavior constraints, and domain knowledge background injection. Taking the third agent as an example: its system prompt word stipulates that the output must include a four-tuple field of "operation object, operation action, expected phenomenon, and safety prompt", and injects the SOP format example corresponding to the current experiment type into the prompt word, so that the model output conforms to the preset specifications in both structure and content.

[0077] When the fifth agent's judgment result fails, a structured difference report containing three elements—problem field location, violation specification entry, and repair suggestion—is generated and fed back to the corresponding agent. That is, the third agent performs targeted repair instead of triggering a full process regeneration, in order to reduce iteration overhead and speed up convergence.

[0078] Furthermore, this invention equips the intelligent agent not with general-purpose development tools, but with a function library and secondary development interfaces for professional simulation algorithms (such as power system power flow calculation and energy storage scheduling algorithms). The intelligent agent can call this dedicated toolchain to generate professional simulation software with underlying simulation calculation capabilities, rather than just implementing simple page displays. This ensures the professionalism and realistic simulation capabilities of the generated experiments from the tool level, solving the pain points of existing technologies that rely on general-purpose tools, lack simulation logic, and can only generate display-type content.

[0079] The intelligent agent generated by this invention for textual training is a directly runnable virtual simulation experiment software. Unlike general large models that only output text or code snippets, this invention outputs a complete, independently runnable virtual simulation training experiment, including a complete chain of simulation logic, parameter binding, interactive operations, and data visualization.

[0080] Through a closed-loop mechanism of multi-agent collaboration and verification and repair, multiple agents perform their respective duties and complete the entire process from requirement analysis to experimental delivery through pipeline collaboration. Finally, the verification agent reviews the process in five aspects: completeness of execution process, compliance of operation, coverage of safety specifications, rationality of simulation parameters, and feasibility of assessment standards. If the process fails, a targeted backtracking and repair mechanism is triggered, rather than generating an uncontrollable system all at once.

[0081] 3. Knowledge layer, used to enhance the knowledge base through RAG retrieval and provide professional support; A RAG-enhanced knowledge base was constructed, comprising: State Grid training outlines, equipment operation procedures, safety operation specifications, and a SOP template library. The State Grid training outlines include professional certifications for new and existing employees; equipment operation procedures cover photovoltaic, energy storage, power distribution, and chiller / heater units; safety operation specifications include safety briefings and emergency response; and the SOP template library covers equipment knowledge, equipment operation, equipment maintenance, and optimized equipment scheduling.

[0082] The knowledge base of this invention is customized for specific professional fields. The system incorporates multiple vertical domain knowledge bases, including domain training outlines, equipment operation procedures, safety operation specifications, SOP templates, technical standards, and historical experimental case studies. During the planning and generation process, the intelligent agent retrieves relevant content from these knowledge bases through semantic retrieval as the basis for generation, ensuring that the generated experimental content is professionally accurate and traceable.

[0083] 4. Simulation Layer: This layer interfaces with the campus integrated energy system, binding parameters to the model through a parameter contract dictionary. The campus integrated energy system includes power sources, distribution networks, electrical loads, energy storage, and corresponding source-grid-load-storage model engines. Sources include photovoltaic, wind power, and combined heat and power (CHP); networks include distribution networks, microgrids / grid interfaces; loads include electrical / cooling / heating loads and building energy consumption; and storage includes energy storage systems and thermal / cold storage. The model engine includes 3D models and multi-energy coupling mathematical models. The parameter contract dictionary is constructed based on general professional references and the specific equipment of the campus integrated energy system.

[0084] 5. Output layer, used to output the complete experimental package that has passed verification.

[0085] The output layer outputs a complete experimental package, which includes experimental steps, operating procedures, safety precautions, simulation parameters, and assessment criteria.

[0086] Scenario 1: An instructor at a vocational skills training institution needs to prepare a virtual simulation training experiment on "source-grid-load-storage coordination and scheduling" for students.

[0087] The instructor inputs: "Help me generate a virtual simulation training experiment on source-grid-load-storage coordination and scheduling, aimed at intermediate students, with a medium level of difficulty." Step 1: Request Analysis and Agent Response The agent followed up with, "What knowledge points should the experiment cover, such as new energy output prediction, load response strategies, and energy storage scheduling algorithms? Which ones should be included?" The instructor replied: "It just needs to include wind and solar power output forecasting and energy storage charging and discharging strategies." The agent followed up with, "Does the experiment require trainees to manually adjust the parameters and observe the changes in the results?" The instructor replied, "Yes, trainees need to be able to adjust the energy storage capacity and charging / discharging power." The structured requirements for intelligent agent output are as follows: The experiment type is source-grid-load-storage scheduling, the target learners are intermediate level, the difficulty is moderate, the knowledge points include wind and solar power output prediction and energy storage charging and discharging strategies, and the interaction method is adjustable parameters and real-time observation of simulation results.

[0088] Step 2: Experimentally plan the work of the intelligent agent. Read the requirement description, retrieve matching experimental templates from the SOP template library, and query professional models and algorithms related to "source-grid-load-storage" in the knowledge base.

[0089] The experiment is planned to include the following stages: wind and solar power output cognition stage, energy storage system operation stage, and coordinated dispatch comprehensive exercise stage, with clear learning objectives and assessment dimensions for each stage.

[0090] Output the experimental framework scheme.

[0091] Step 3: Generating the intelligent agent's work. For each experimental phase, relevant content is retrieved from the equipment operation procedure database and the safety operation specification database.

[0092] Detailed operation steps are generated for each stage, and each step includes four elements: operation object, operation action, expected phenomenon, and safety prompts.

[0093] At the same time, fault handling branch paths are generated, such as emergency operation procedures when energy storage overcharge protection is triggered.

[0094] Step 4: Parameter binding for agent operation Each operation step is mapped to the 3D device model and mathematical simulation model of the simulation platform.

[0095] Input parameters such as the rated capacity, upper and lower limits of charge and discharge power, and temperature protection threshold of the energy storage battery from the equipment nameplate database.

[0096] Hard constraints verify that the parameters do not exceed the physical limits of the equipment, while soft constraints verify that the parameters fall within the typical operating range.

[0097] Output the complete experimental package with all parameters bound.

[0098] Step 5: Verify the agent's operation Perform a five-dimensional audit on the experimental package: whether the operation steps completely cover all experimental stages, whether the operation process conforms to the specifications in the procedure library, whether the safety briefing and protection measures are complete, whether the simulation parameters pass both soft and hard constraints, and whether the assessment and scoring nodes can be quantified and measured by the simulation platform.

[0099] Once all tests are passed, the experimental kit will be delivered to the instructor, who can then deploy and use it directly on the virtual simulation training platform, allowing students to conduct the experiments.

[0100] This invention utilizes a large-model-driven multi-agent architecture, allowing instructors to simply describe their needs in natural language, and the system automatically generates a complete, ready-to-run virtual simulation training experiment. Instructors can adjust the experiment content, difficulty, and assessment methods at any time, generating experiments according to the target learners' levels and knowledge domains, thus achieving personalized instruction.

[0101] Example 2 This embodiment provides a method for automatically generating contextual operation tasks based on textual intelligence agents, including: S1: Receives natural language commands from the user through the input layer; S2: The Wenxun intelligent agent parses natural language instructions and automatically generates a complete experimental package; wherein, the Wenxun intelligent agent includes multiple intelligent agents, the first intelligent agent receives natural language instructions, and performs intent parsing and information feature extraction on the instructions to obtain a structured requirement description; S3: Based on intent parsing and information features, an experimental framework that automatically generates natural language instructions using a second intelligent agent; S4: Based on the experimental framework, generate detailed operation steps for each experimental stage under the experimental framework according to the third agent; S5: Based on the fourth intelligent agent, the detailed operation steps are mapped to the 3D device model and mathematical simulation model of the simulation platform, and parameter binding is performed to obtain a complete experimental package containing parameter binding. S6: The fifth intelligent agent performs a verification operation on the experimental package and deploys the verified experimental package to the virtual simulation training platform; the content of the experimental package includes one or more scenario operation tasks; S7: Output the validated experimental package through the output layer to provide feedback to the user on the final scenario operation task.

[0102] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0103] The proposed method can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A context-based operational task automatic generation system based on a textual training agent, characterized in that, include: The input layer is used to receive natural language instructions; The Wenxun intelligent agent engine layer includes multiple intelligent agents used to parse natural language instructions and automatically generate complete experimental packages; specifically, the first intelligent agent receives natural language instructions and performs intent parsing and information feature extraction on the natural language instructions to obtain a structured requirement description; Based on structured requirements description, an experimental framework matching natural language instructions is automatically generated using a second intelligent agent. Based on this framework, a third intelligent agent generates detailed operational steps for each experimental stage. A fourth intelligent agent maps these detailed operational steps to the 3D device model and mathematical simulation model of the simulation platform, performing parameter binding to obtain a complete experimental package containing parameter bindings. A fifth intelligent agent performs verification operations on the experimental package, and the verified package is deployed to the virtual simulation training platform. The experimental package includes one or more scenario-based operational tasks. The knowledge layer is used to enhance the knowledge base through RAG retrieval and provide professional support. The simulation layer is used to interface with the campus integrated energy system to bind parameters to the model; The output layer is used to output the complete experimental package that has passed verification.

2. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 1, characterized in that, The multiple intelligent agents encapsulate the large language model, role-specific prompts, callable tool interfaces, and loop control logic.

3. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 1, characterized in that, The structured requirements description includes the experiment type, target learner level, difficulty level, and the knowledge domains and skills involved.

4. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 1, characterized in that, The fourth intelligent agent maps detailed operation steps to the 3D equipment model and mathematical simulation model of the simulation platform, performs parameter binding, and obtains a complete experimental package containing parameter binding. Specifically, it maps upward to the 3D equipment model and downward to the mathematical simulation model, and injects the equipment's rated parameters, operating conditions, and boundary conditions. The rated parameters of the equipment include the equipment nameplate database and the historical operating condition sample database.

5. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 1, characterized in that, Each bound parameter is subjected to both hard and soft constraint verification, which includes hard constraint verification and soft constraint verification.

6. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 5, characterized in that, The hard constraint verification is based on the physical characteristics and technical specifications of the equipment, and sets value boundaries for various parameters. Once any parameter exceeds the hard boundary, the fifth intelligent system will directly determine that it fails, mark the violation field and the corresponding specification item in the audit report, and force a backtracking repair.

7. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 5, characterized in that, The soft constraint verification is based on the statistical patterns of historical operating data, calculating the mean and standard deviation of each parameter under typical operating conditions, and constructing a reasonable reference interval.

8. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 1, characterized in that, The verification operation involves reviewing the complete experimental package from five aspects: process completeness, operational compliance, safety specification coverage, simulation parameter rationality, and assessment standard feasibility.

9. The automatic generation system for contextual operation tasks based on a textual training agent according to claim 1, characterized in that, The collaboration of multiple agents in the Wenxun intelligent agent engine layer is driven by a two-layer mechanism: the orchestration layer and the prompt word engineering layer. The orchestration layer uses a directed acyclic graph; the prompt word engineering layer equips each agent with an independent system prompt word, and embeds four functional modules: role definition, output format template, prohibited behavior constraints, and domain knowledge background injection.

10. A method for automatically generating contextual operation tasks based on a textual training agent, characterized in that, include: S1: Receives natural language commands from the user through the input layer; S2: The Wenxun intelligent agent parses natural language instructions and automatically generates a complete experimental package; wherein, the Wenxun intelligent agent includes multiple intelligent agents, the first intelligent agent receives natural language instructions, and performs intent parsing and information feature extraction on the instructions to obtain a structured requirement description; S3: Based on intent parsing and information features, an experimental framework that automatically generates natural language instructions using a second intelligent agent; S4: Based on the experimental framework, generate detailed operation steps for each experimental stage under the experimental framework according to the third agent; S5: Based on the fourth intelligent agent, the detailed operation steps are mapped to the 3D device model and mathematical simulation model of the simulation platform, and parameter binding is performed to obtain a complete experimental package containing parameter binding. S6: The fifth intelligent agent performs a verification operation on the experimental package and deploys the verified experimental package to the virtual simulation training platform; the content of the experimental package includes one or more scenario operation tasks; S7: Output the validated experimental package through the output layer to provide feedback to the user on the final scenario operation task.

Images