A Closed-Loop Training and Capability Assessment Method Based on Multi-Agent Collaboration
By using a multi-agent collaborative framework, the problems of accuracy and personalization of training content in high-risk industry training systems have been solved, enabling dynamic personalized training and precise retraining. This breaks the situation of isolated storage of implicit experience data and ensures that the knowledge base is closely integrated with actual operating scenarios.
Patent Information
- Application Number
- CN202611343228.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
In high-risk and highly specialized industries, existing training systems lack task division and result verification mechanisms among multiple agents, making it difficult to guarantee the authority and accuracy of training content generation. This makes it impossible to achieve dynamic personalized training and precise retraining, and the implicit experience generated during the training process is difficult to accumulate and provide feedback.
A multi-agent collaborative framework is adopted, including agents for demand identification, capability assessment, course generation, training implementation, training path planning, case extraction, and knowledge review and updating. Through multi-agent collaborative verification, the accuracy of training content is ensured, enabling dynamic personalized training and precise retraining, and tacit experience is automatically deposited into a structured knowledge base.
It ensures the authority and accuracy of training content, supports highly targeted training that responds to immediate needs, forms an objective and calculable competency assessment and a dynamically evolving knowledge base, and improves training efficiency and the matching degree of trainees' abilities.
Smart Images

Figure CN123118318A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and digital education and training technology, specifically to a closed-loop training and competency assessment method based on multi-agent collaboration. Background Technology
[0002] In many high-risk and highly specialized industries (such as civil aviation operations and maintenance, rail transit, and special equipment operations and maintenance), where professional knowledge is highly privatized, regulations are frequently updated, and job safety and compliance requirements are extremely high, routine training and qualification assessments for employees are the core means to ensure operational safety and maintain job proficiency. Such employee training is typically characterized by short individual sessions, high overall frequency, and a heavy reliance on expert instructors' experience.
[0003] Currently, although some artificial intelligence (AI) intelligent teaching assistants or scheduling assistance systems have emerged in the market, existing technologies still have the following significant shortcomings in terms of full-chain training and competency assessment when facing specific industries with high levels of professional privatization, confidentiality, or complex regulatory systems: First, most existing training systems rely on general large models or single-agent architectures, lacking task division and result verification mechanisms among multiple agents. This makes it difficult to guarantee the authority and accuracy of training content generation. Industry-specific knowledge bases are often highly proprietary and logically fragmented. General large models lack a deep understanding of industry-specific regulations, manuals, and cases, easily leading to "illusion" phenomena. Furthermore, a single agent cannot simultaneously handle complex multi-dimensional tasks such as demand identification, professional retrieval, lesson plan generation, behavior assessment, and knowledge verification. Without a central control and orchestration agent to verify the consistency and confidence of the output results, it is difficult to ensure that the generated training content strictly adheres to authoritative regulations, easily leading to training content deviating from actual business standards.
[0004] Secondly, existing training models mostly adopt static or general curriculum systems, lacking targeted dynamic demand triggering and personalized training task package generation mechanisms, resulting in slow improvement of practitioners' capabilities. Traditional training relies heavily on manual periodic scheduling or the use of standardized textbooks, making it difficult to capture immediate training needs arising from changes in regulations and manuals, the commissioning of new equipment, the expiration of employee qualifications, or unexpected operational anomalies. At the same time, existing course recommendation logic cannot combine the dynamic differences between employees' current personalized capability profiles and job target capability profiles, resulting in a lack of job-specific content and failing to meet the requirement of precise matching of personnel capabilities under the rapid pace of knowledge updates.
[0005] Third, existing competency assessment methods are too simplistic and fail to comprehensively integrate multimodal process behavior data to form an objective and calculable competency profile and a precise closed-loop retraining path. Current assessments primarily rely on traditional manual grading or single online test papers, failing to deeply integrate and analyze multimodal learning behavior data during training, including answer time, operational step sequences, simulator operation results, and scenario handling choices. Furthermore, they lack scientific quantitative calculation models that consider time decay and data reliability. This results in overly one-sided assessments. When practitioners' competencies fall short, the system cannot accurately identify the specific competency items and knowledge gaps, thus failing to generate clearly prioritized retraining paths based on knowledge dependencies and risk levels. Consequently, training fails to form an effective "assessment-retraining" closed loop.
[0006] Fourth, the existing training system is severely disconnected from actual operational scenarios, making it difficult to retain and utilize the tacit experience and knowledge generated during training. In traditional and existing online training processes, multimodal data such as trainees' summary materials, answers, explanations, typical handling paths, and instructor corrections are usually stored in isolation or discarded directly after the training ends. The lack of automated entity recognition, conflict detection, desensitization, and regulatory / version consistency verification mechanisms prevents these high-quality cases and tacit knowledge derived from front-line practical operations and training exercises from being automatically transformed and updated back into the structured knowledge base, hindering the continuous dynamic evolution of the knowledge base. Summary of the Invention
[0007] The following provides a brief overview of one or more aspects to offer a basic understanding of them. This overview is not an exhaustive summary of all conceived aspects, nor is it intended to identify key or decisive elements of all aspects, nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form to prepare for the more detailed descriptions that follow.
[0008] The purpose of this invention is to solve the above problems and provide a closed-loop training and capability assessment method based on multi-agent collaboration. This method achieves the following: ensuring the accuracy and authority of training content through multi-agent collaborative verification; realizing a closed loop of dynamic personalized training and precise retraining based on capability profile differences; and realizing the automated feedback and evolution of the knowledge base based on multi-dimensional verification of training data.
[0009] The technical solution of this invention is as follows: This invention discloses a closed-loop training and capability assessment method based on multi-agent collaboration, the method comprising: The demand identification agent identifies the training subjects and target positions from the training needs, and the capability assessment agent generates a target capability profile based on the job requirements of the target positions. The ability assessment agent obtains the current ability profile of the trainee, the course generation agent combines the knowledge content in the structured knowledge base to generate a personalized training task package based on the difference between the target ability profile and the current ability profile, and the training implementation agent conducts training on the trainee and collects multimodal training behavior data during the training process. The capability assessment agent evaluates and updates the current capability profile of the trainee based on the multimodal training behavior data; calculates the capability gap between the updated current capability profile and the target capability profile; if it is determined that the preset training objective has not been achieved, the training path planning agent generates a supplementary training path based on the capability gap, and the overall control and orchestration agent re-triggers the generation and implementation of the training task; If the preset training objective is determined to be achieved, the case extraction agent extracts candidate knowledge from the multimodal training behavior data and the materials submitted by the trainees. The knowledge review and update agent verifies and reviews the candidate knowledge and updates the approved candidate knowledge into the structured knowledge base.
[0010] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the requirement identification agent, the capability assessment agent, the course generation agent, the training implementation agent, the training path planning agent, the case extraction agent, and the knowledge review and update agent constitute a professional task capability agent, which together with the overall control and orchestration agent constitutes a multi-agent collaborative framework. The task dependencies and verification rules between the various professional task capability agents are established based on the overall control and orchestration agent. During the operation of the multi-agent collaborative framework, the overall control and orchestration agent performs consistency and / or confidence verification on the output results of each agent, and triggers a retry, review, or manual intervention process when the verification fails.
[0011] According to an embodiment of the closed-loop training and competency assessment method based on multi-agent collaboration of the present invention, the target competency profile includes multiple job competency requirements and corresponding target competency levels.
[0012] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the construction process of the structured knowledge base includes: Acquire text and image data, including regulations, manuals, job standards, cases, courses, and historical training data, and perform format conversion, paragraph segmentation, duplicate content removal, sensitive information processing, version identification, validity period identification, and source marking on the data to form standard knowledge units; Entities, attributes, and relationships are extracted from the standard knowledge units to establish a hierarchical knowledge structure or knowledge graph that includes nodes for legal provisions, operating procedures, job competencies, equipment models, risk events, training courses, assessment questions, and case studies. Each knowledge unit is assigned a source identifier, version identifier, validity period identifier, credibility identifier, and access permission identifier.
[0013] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the agents communicate with each other through standard message objects, which include at least several of the following: training task identifier, employee anonymity identifier, job position and model identifier, input data address, required knowledge scope, output data type, confidence level, reference basis, prerequisite task identifier, and processing status.
[0014] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the training requirements are generated by the requirement identification agent according to demand-triggered conditions or periodic triggering conditions, wherein: The triggering conditions include at least one of the following: job change, new equipment or system put into operation, change of regulations or operation manual version, qualification about to expire, occurrence of job-related event, a certain ability indicator in the ability profile falling below a preset threshold, abnormal operation, and management personnel initiating a special task. The periodic triggering conditions include: annual, quarterly, monthly, or qualification periodic training plans.
[0015] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the generation process of the personalized training task package includes: The knowledge retrieval agent retrieves knowledge units corresponding to the target capability items in the differences from the structured knowledge base, and returns a set of knowledge evidence with source and version information to the course generation agent; The course-generating agent generates training content based solely on the set of knowledge evidence that has passed validity verification, and establishes a binding relationship between each piece of training content and the corresponding knowledge node, target ability item, and assessment item.
[0016] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the multimodal training behavioral data includes: At least several of the following: course completion status, answer content, answer duration, number of answer revisions, sequence of operation steps, simulator operation results, scenario handling selection, written materials, intelligent question and answer interaction records, and instructor evaluation data; The multimodal training behavior data is converted into unified training behavior features and associated with training subjects, training tasks, knowledge nodes, and competency items.
[0017] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the capability assessment agent calculates the first [agent's value] based on the multimodal training behavior data. Item Ability Value The formula is:
[0018] in, For the first Class 1 evaluation data corresponds to the first The weight of each capability; For the first Standardized scores in class assessment data; For the first The credibility coefficient of the evaluation data; This is the time decay coefficient.
[0019] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the determination of situations in which the preset training objective has not been achieved includes: If the current capability value of any required capability item is lower than the corresponding threshold, the total weighted capability gap exceeds the set threshold, a high-risk capability item is not passed, the completeness of the assessment evidence is lower than the set threshold, or the confidence level of the capability assessment is lower than the set threshold; The supplementary training path is regenerated by the training path planning agent according to the following priorities: knowledge nodes directly associated with the unmet competency items, prerequisite knowledge nodes that lead to errors, case nodes similar to the error scenarios, training methods that have historically proven effective, and mandatory content with high risk levels.
[0020] According to an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention, the step of verifying and reviewing the candidate knowledge and updating the structured knowledge base with the approved candidate knowledge includes: The case extraction agent performs text segmentation, entity recognition, relation extraction, distinction between factual and opinion content, identification of conflicting content, desensitization processing, and generation of candidate knowledge units on the text materials submitted by the training subjects. The knowledge review and update agent performs legal consistency, version consistency, source credibility, and duplication checks on the candidate knowledge units, and writes the candidate knowledge units that pass the checks into the structured knowledge base, while establishing traceable associations with the original training cases, training subjects' competency items, and knowledge sources.
[0021] Compared with the prior art, the present invention has the following advantages: First, this invention ensures the authority, accuracy, and rigor of the generated and output training content. It employs a multi-agent collaborative framework comprised of specialized task capability agents (requiring demand identification, capability assessment, course generation, training implementation, training path planning, case extraction, and knowledge review and updating) and a central control and orchestration agent. During operation, the central control and orchestration agent performs consistency and / or confidence checks on the outputs of each agent, triggering retry, review, or manual intervention processes when checks fail. Compared to traditional systems relying on general large models or single agents, this invention, through clear role division and standard message communication among multiple agents, combined with the constraint verification mechanism of the central control and orchestration agent, strictly binds the content generated by the agents to knowledge evidence in a structured knowledge base. This overcomes the "illusion" of large models and the problem of knowledge version confusion, significantly improving the sufficiency and accuracy of the training content generation and evaluation basis.
[0022] Secondly, it achieves real-time demand response and highly targeted dynamic personalized training covering all scenarios. This invention generates training needs through a demand identification agent that generates needs based on triggering conditions such as job changes, regulatory version changes, impending qualification expiration, and operational events. A capability assessment agent generates a target capability profile based on job requirements, and a course generation agent, combined with a structured knowledge base, generates personalized training task packages based on the differences between the target capability profile and the current capability profile. This feature changes the rigid model of traditional fixed-cycle scheduling and the use of general teaching materials. It can capture real-time needs such as changes in operational scenarios and qualification expiration, and accurately match personalized learning content based on the dynamic changes in practitioners' capability profiles, effectively improving the efficiency of learners' capability improvement and job matching in a high-frequency knowledge update environment.
[0023] Third, it achieves objective and quantifiable personnel competency assessment and a closed-loop control of "assessment-precise retraining". This invention collects multimodal training behavior data, including answer content, answer duration, operation step sequences, and simulator operation results, through a training implementation agent and converts it into unified training behavior characteristics. A competency assessment agent calculates the current competency value by comprehensively considering weights, standardized scores, reliability coefficients, and time decay coefficients. When the preset training objectives are not met, a training path planning agent regenerates supplementary training paths based on priorities such as knowledge nodes directly related to the unmet competency items, prerequisite knowledge nodes, and high-risk mandatory content. The overall control and orchestration agent then re-triggers the implementation. This feature comprehensively utilizes multi-source process behavior data to replace single-session scoring and introduces a time decay coefficient that conforms to the forgetting curve, forming an objectively calculated competency profile. Combined with high-risk item assessment and a prioritized reverse-tracking retraining mechanism, it achieves a precise closed loop of "assessment-retraining".
[0024] Fourth, this invention achieves the automated accumulation of implicit training experience and the dynamic evolution and updating of the knowledge base. When a preset training objective is met, the invention uses a case extraction agent to perform text segmentation, entity recognition, relationship extraction, fact-opinion differentiation, conflict content identification, and anonymization on the text materials submitted by the trainees to generate candidate knowledge units. A knowledge review and update agent then performs checks for regulatory consistency, version consistency, source credibility, and repeatability before writing the data into a structured knowledge base, while simultaneously establishing traceable relationships. This feature breaks the isolated storage of training process data, automatically transforming excellent handling processes, typical errors, and first-hand summaries generated during training into structured knowledge to feed back into the knowledge base. This ensures that the knowledge base is closely integrated with actual operational scenarios, achieving continuous evolution of the knowledge system. Attached Figure Description
[0025] The above-described features and advantages of the present invention will be better understood after reading the following detailed description of embodiments of the present disclosure in conjunction with the accompanying drawings. In the drawings, components are not necessarily drawn to scale, and components having similar related characteristics or features may have the same or similar reference numerals.
[0026] Figure 1 This is a flowchart of an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention.
[0027] Figure 2 A simplified logic diagram for the interaction between the multi-agent collaboration framework and the knowledge base and profile database.
[0028] Figure 3 This is a logic diagram of the interaction between a multi-agent framework and the training process in the civil aviation industry.
[0029] Figure 4A flowchart for retraining individual trainees or groups of trainees when they do not meet the required competency criteria. Detailed Implementation
[0030] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that the aspects described below with reference to the accompanying drawings and specific embodiments are merely exemplary and should not be construed as limiting the scope of protection of the present invention in any way.
[0031] Figure 1 The flowchart of an embodiment of the closed-loop training and capability assessment method based on multi-agent collaboration of the present invention is shown. Figure 3 This illustrates the interaction logic between a multi-agent framework and the civil aviation industry training process. (Reference) Figure 1 and Figure 3 The specific implementation steps of the method in this embodiment are detailed below.
[0032] Step 1: The demand identification agent identifies the training subjects and target positions from the training needs, and the capability assessment agent generates a target capability profile based on the job requirements of the target position.
[0033] Each specialized task capability agent (such as the requirement identification agent and capability assessment agent in this step) together with the overall control and orchestration agent constitutes a multi-agent collaborative framework. The task dependencies and verification rules between the various specialized task capability agents are established based on the overall control and orchestration agent. During the operation of the multi-agent collaborative framework, the overall control and orchestration agent performs consistency and / or confidence verification on the output results of each agent, and triggers retry, review, or manual intervention processes when the verification fails.
[0034] In this step, the demand identification agent is built based on domain model distillation techniques or lightweight large language models (LLMs), and internally includes: System prompt: Define its role as "Training Needs Analysis Expert", limit its input to trigger events and knowledge base update logs, and output to training needs objects in a standardized format; Rule parsing engine: combines hard logic rules (such as qualification expiration time threshold judgment) with the semantic understanding ability of large models; Standard message interface: Encapsulates a communication module that receives trigger events and outputs standard message objects containing training task identifiers, employee anonymity identifiers, job positions, and machine model identifiers.
[0035] The demand identification agent automatically identifies the target job based on triggering events / factors (cycled plans such as annual, quarterly, monthly, or qualification-based training plans, business events, job changes, new models / equipment being put into operation, changes to regulations or operation manuals, qualification expiration, operational events, competency indicators falling below thresholds, abnormal operations, special projects initiated by management personnel, etc.), target job positions, knowledge base changes, and employees' historical competency profiles. The identification process involves: the demand identification agent analyzing trigger signals (such as job change events in the HR system, cycled plans in the training system, or abnormal operation logs in the operation monitoring system) to extract key fields (such as employee ID, job code, and changed regulation number); then, by searching the "job-regulation-competency" mapping matrix, it identifies the target job position corresponding to the change event and the affected employees, thereby accurately analyzing and extracting the training recipients and target job positions.
[0036] The capability assessment agent incorporates a multi-source feature fusion operator, responsible for converting multimodal behavioral data such as answers, operation step sequences, and simulator results into unified training behavioral features. It then uses a built-in weighted decay calculation engine to calculate the capability value of each corresponding item based on the multimodal training behavioral data, calculating the capability value of the [number]th [item / item]. Item Ability Value The formula is:
[0037] in, For the first Class 1 evaluation data corresponds to the first The weight of each capability; For the first Standardized scores in class assessment data; For the first The credibility coefficient of the evaluation data; This is the time decay coefficient.
[0038] Then, based on the quantitative calculation results of the above-mentioned capability values using the decision-making big data model and the prompt words, a target capability profile is generated. This profile includes multiple job capability requirements and their corresponding target capability levels. The target capability profile is constructed based on the job position, machine type, responsibilities, qualification level, and operating scenario. Each job capability... This includes: capability item identifiers (such as specific skill / regulation codes), target capability levels, capability weights, mandatory thresholds, corresponding knowledge nodes (bound to regulations / manual nodes in the knowledge base), corresponding operational behaviors, and corresponding assessment methods (such as theoretical assessments, simulator exercises, etc.).
[0039] After receiving the target job identifier, the competency assessment agent initiates a query to the structured knowledge base (or knowledge graph) to retrieve the "job competency nodes" bound to the job, as well as their corresponding "legal clause nodes" and "operational procedure nodes." The retrieved job skill requirements (such as emergency response, standardized operation, and mastery of specific regulations) are then mapped to competency items. Each capability item is assigned a corresponding target level requirement and a threshold for achieving it, and these are ultimately combined and encapsulated into a target capability profile vector. ,in Indicates the first Job competency requirements. These requirements are subject to dynamic updates based on changes in regulations, job standards, or aircraft models.
[0040] Figure 2 The diagram showcases the static architecture and dynamic interaction overview of the entire system, including the connections between the overall control and orchestration agent, various specialized task agents, the aviation-specific knowledge base, and the aviation personnel capability profile database. The overall control and orchestration agent, as the core of the overall collaborative scheduling, is responsible for task orchestration and scheduling within the multi-agent collaborative framework. During operation, it performs consistency and / or confidence checks on the output results of each specialized task capability agent. If the check fails, it triggers a retry, review, or manual intervention process.
[0041] The knowledge review and update agent is responsible for verifying the candidate knowledge units for compliance with regulations, version consistency, source credibility, and duplication. It then updates the verified knowledge into the structured knowledge base and establishes traceable relationships.
[0042] Aviation knowledge base, question-and-answer knowledge base, and thesaurus (structured knowledge base) are used to store hierarchical knowledge structures or knowledge graphs constructed from standard knowledge units such as regulations, manuals, job standards, cases and courses, providing authoritative knowledge evidence for course generation and knowledge retrieval.
[0043] The knowledge retrieval agent is responsible for retrieving the corresponding knowledge unit from the structured knowledge base along the knowledge graph topology path based on the difference between the target capability item and the current capability profile, and returning a set of knowledge evidence with source and version information to the course generation agent.
[0044] The course-generating agent generates personalized training task packages based solely on a set of knowledge evidence that has passed validity verification, and establishes a binding relationship between each training item and its corresponding knowledge node, target capability item, and assessment item.
[0045] The training implementation agent is responsible for delivering training to trainees (such as through digital human interaction or by pushing task packages) and collecting multimodal training behavior data in real time throughout the training interaction process.
[0046] When the training path planning agent determines that the preset training objectives have not been achieved, it is responsible for regenerating a supplementary training path based on the capability gap and according to preset priorities (such as directly related knowledge nodes, prerequisite knowledge nodes, high-risk compulsory content, etc.).
[0047] The case extraction agent is responsible for extracting typical cases and experiences from multimodal training behavior data and text materials submitted by trainees, performing text segmentation, entity recognition, conflict identification and desensitization processing, and generating candidate knowledge units.
[0048] The competency assessment agent is responsible for generating target competency profiles based on job requirements, and for evaluating and updating the current competency profiles and competency gaps of trainees by integrating multimodal training behavior data and using operators such as weight, credibility, and time decay.
[0049] The aviation personnel competency profile database is used to store and dynamically maintain the historical and current competency profile vectors and competency assessment data of practitioners, which can be retrieved, compared and updated in real time by the competency assessment intelligent agent.
[0050] Figure 2 The dashed arrows shown illustrate the collaborative control and verification relationships of the central control and orchestration agent. The dashed arrows originating from the central control and orchestration agent point to the various specialized task capability agents, representing the central control and orchestration agent's unified scheduling, command issuance, and output verification control over the entire system. Overall control orchestration intelligent agent Knowledge retrieval agent / Course generation agent / Training implementation agent / Training path planning agent / Case extraction agent / Competency assessment agent / Knowledge review and update agent: The central control and orchestration agent acts as the core of the scheduling mechanism, distributing task orchestration instructions and scheduling information to various specialized task capability agents via standard message objects. Simultaneously, each agent returns its processing output to the central control and orchestration agent, which then performs consistency, evidence integrity, or confidence level checks. If a check fails, a retry, review, or manual intervention process is triggered.
[0051] Figure 2 The solid arrows in the diagram illustrate the specific data flow routes between the knowledge base, the user profile database, and various agents throughout the entire training process: 1. Data interaction between the knowledge base and knowledge retrieval, generation, and review processes. Based on the needs of capability gaps, the knowledge retrieval agent retrieves and extracts relevant professional regulations, manuals, cases, and other standard knowledge unit data from a structured knowledge base.
[0052] The knowledge retrieval agent transmits a set of trusted knowledge evidence with source identifiers and version information to the course generation agent.
[0053] The knowledge review and update agent updates and writes the candidate knowledge units that have passed the review into the structured knowledge base, realizing the automatic feedback and evolution of the knowledge base.
[0054] 2. Data interaction between course generation and training implementation
[0055] The course generation agent will transmit the assembled personalized training task package, which includes the binding relationships between "training content - knowledge nodes - competency items - assessment items", to the training implementation agent.
[0056] 3. Training process data collection and case extraction
[0057] The training implementation agent will transmit multimodal training behavior data, including trainees' answers, operation sequence, simulator results, and submitted text materials, collected during the training process, to the case extraction agent.
[0058] After segmenting, identifying entities, identifying conflicts, and desensitizing multimodal data and text materials, the case extraction agent generates "candidate knowledge units" which are then transmitted to the knowledge review and update agent.
[0059] 4. Data interaction between competency assessment and competency profile database
[0060] The training implementation agent synchronously transmits the collected multimodal training behavior data to the capability assessment agent for quantitative calculation of capability values.
[0061] The capability assessment agent retrieves the historical current capability profile vector of the training subject from the capability profile library; after calculating the updated capability value based on multimodal data, the capability assessment agent stores the newly generated current capability profile vector and updates the capability profile library.
[0062] 5. Data flow during cyclical retraining (closed loop for unmet goals)
[0063] The course generation agent provides the initially generated course and capability architecture information to the training path planning agent as a reference for path planning.
[0064] When the competency assessment determines that the competency is not met, the training path planning agent regenerates a "supplementary training path" with priority ranking based on the competency gap, and transmits it to the training implementation agent (or re-issues it to the course generation agent by the central control and arrangement agent for repackaging and implementation), and the supplementary training is carried out again.
[0065] Step 2: The capability assessment agent obtains the current capability profile of the trainee, the course generation agent combines the knowledge content in the structured knowledge base to generate a personalized training task package based on the difference between the target capability profile and the current capability profile, and the training implementation agent conducts training to the trainee and collects multimodal training behavior data during the training process.
[0066] The process of building a structured knowledge base includes: Acquire text and image data, including regulations, manuals, job standards, cases, courses, and historical training data, and perform format conversion, paragraph segmentation, duplicate content removal, sensitive information processing, version identification, expiration date identification, and source marking on the data to form standard knowledge units; Entities, attributes, and relationships are extracted from standard knowledge units to establish a hierarchical knowledge structure or knowledge graph that includes nodes for legal provisions, operating procedures, job competencies, equipment models, risk events, training courses, assessment questions, and case studies. Each knowledge unit is assigned a source identifier, version identifier, validity period identifier, credibility identifier, and access permission identifier.
[0067] When training needs are triggered, the competency assessment agent directly retrieves the trainee's existing historical and current competency profiles from the aviation personnel competency profile database. , where each parameter ( ) represents the training recipients (practitioners) in the first The current quantitative capability value for a specific job position.
[0068] Each parameter At the underlying level, the following specific meanings and attributes are mapped: 1. Corresponding job competency items (competency indicators): parameter Corresponding to the target job competency profile The same number of the ability items For example: specific emergency response capabilities, mastery of standardized operating procedures, and ability to troubleshoot system faults of a certain type of machine / equipment.
[0069] 2. Comprehensive reflection of multi-source assessment data (data sources) ): parameter The value is the first one generated by this practitioner in history. It is derived from a comprehensive calculation of various assessment data. Specifically, it includes: Theoretical assessment results ( Historical answer scores and test records; Simulation operation and scenario handling results ( Simulator drill operation sequence and key emergency options; Training process behavioral data ( ): History study time, number of times answers were modified, interaction history, etc.; Material writing quality ( ): Professionalism assessment of written materials such as summary reports and answer explanations; Actual runtime event data ( ): Records of abnormal operations and safety incidents during real daily operations or simulation training.
[0070] parameter The quantification formula is:
[0071] Therefore, the parameters in the historical current capability profile The specific value is determined and explained by the following four factors: Data weighting coefficient : No. Class 1 evaluation data for the first The relative importance of each ability assessment item. For example, the parameter "practical skills training ability". Weights of simulator operation data The weight will be significantly higher than that of theoretical answer data; Standardized score : Practitioners in the first The baseline score is the result of normalizing / standardizing the original performance in the assessment data.
[0072] Credibility coefficient Assess the reliability and completeness of the data source. Data generated from high-credibility scenarios (such as live-fire simulators and official exercises) has a credibility coefficient close to 1; data generated from low-credibility or incomplete evidence scenarios has a lower credibility coefficient.
[0073] Time decay coefficient (Core Forgetting Curve Parameters): This is the decay factor based on the time span between the time the assessment data occurred and the present. Because core skills in high-security industries such as aviation are characterized by being "easily forgotten if not used for a long time," the longer the assessment data was generated, the greater the time decay coefficient. The smaller the value, the better the historical ability value. It decays naturally over time, thus truly reflecting the current state of a person's ability to maintain their capabilities.
[0074] parameter It is usually set to a normalized value (e.g.) or (continuous real numbers).
[0075] Each parameter When used, it will be compared with the "must-achieve threshold" in the target capability profile. If If the parameter is not specified, it indicates a deficiency in that ability, and retraining will be triggered immediately; if the parameter is not specified... It belongs to the "high-risk capability item". Once the value is lower than the safety red line, the high-risk veto logic will be triggered.
[0076] After training implementation or historical operation, the competency assessment agent collects multi-source data, including theoretical assessments, situational responses, simulated operations, training process behaviors, document writing quality, historical training results, and actual operational events, and uses quantitative formulas. Real-time calculation of scores for each ability item to generate a current ability profile vector. The current capability profile vector It can be updated in real time.
[0077] The course-generating intelligent agent is built based on model distillation and vertical application of large models, ensuring that the generated content strictly adheres to a structured knowledge base.
[0078] The course generation agent includes: a knowledge input interface (RAG core): responsible for receiving the trusted "knowledge evidence set" returned by the knowledge retrieval agent; and a lesson plan and arrangement prompt engine: containing built-in course arrangement logic prompts to constrain the model to arrange knowledge content according to "theoretical learning". Case Analysis Situational Questions and Answers Simulation Exercise Assemble the closed-loop logic of "assessment and comparison"; Structured generator: Force the generated learning materials, exercises and supplementary learning rules into standardized JSON / XML format, and complete the four-element mapping binding of "training content-knowledge node-ability item-assessment item".
[0079] Capability assessment: Intelligent agent's computational target capability profile Compared with the current capability profile difference vector (Right now ), identify target competency items that are not up to standard or have gaps.
[0080] The knowledge retrieval agent retrieves knowledge units from the structured knowledge base that correspond to the target competency items where gaps exist, and returns a set of knowledge evidence with source and version information to the course generation agent.
[0081] The course-generating agent generates training content only based on knowledge evidence that has passed validity verification, preventing the generation of wild content / illusions; and establishes a binding mapping relationship between each generated training content and the corresponding knowledge node, target ability item and assessment item, thereby packaging and combining them into a personalized training task package.
[0082] The process of generating personalized training task packages includes: The knowledge retrieval agent retrieves knowledge units corresponding to the target capability items in the discrepancies from the structured knowledge base and returns a set of knowledge evidence with source and version information to the course generation agent; The course-generating agent generates training content based solely on the set of knowledge evidence that has passed validity verification, and establishes a binding relationship between each piece of training content and the corresponding knowledge node, target capability item, and assessment item.
[0083] The knowledge retrieval intelligent agent is built upon a hybrid retrieval architecture based on large-scale model / vector retrieval and structured query (Cypher / SQL), including: Semantic encoder (Embedding Engine) and vector database interface: Transform the differences in input capability items and knowledge requirements into vector representations; Graph database query operators: For hierarchical knowledge structures or knowledge graphs, a built-in graph query statement generator can be used to query along "capability nodes". Perform precise graph traversal retrieval based on the graph topology path corresponding to "knowledge nodes / rules / operational procedures / cases"; Knowledge filtering and evidence encapsulation module: The built-in rule engine is responsible for verifying the source identifier, version validity and access permission identifier of the search results, and encapsulating them into a knowledge evidence set with metadata (source, version and validity period).
[0084] When building the structured knowledge base, a mapping relationship between "job competency items, knowledge nodes, training content and assessment items" has been pre-established, and a knowledge graph / hierarchical structure containing textual logical relationships such as "applicable to, based on, prior to, associated with risks, corresponding competencies, corresponding jobs, corresponding models, corresponding training content, verified by cases, replaced or updated" has been constructed.
[0085] During retrieval, the knowledge retrieval agent uses a capability gap vector. The target competency items that did not meet the standards or have discrepancies Using the index, we follow the logical edges of "corresponding capabilities" and "based on / related risks" in the graph to locate the directly related and pre-related knowledge nodes, and then retrieve the corresponding standard knowledge units.
[0086] Standard knowledge units are formed as follows: Raw text and image data from civil aviation / industry regulations, manuals, job standards, case studies, courses, and historical training data are acquired; these data undergo format conversion (word2vec, OCR), paragraph segmentation, duplicate content removal, sensitive information processing, version identification, validity period identification, and source tagging. The knowledge unit includes the following attribute fields: Each standard knowledge unit contains source, version, applicable job position, applicable aircraft model, applicable scenario, and validity period, as well as source identifier, version identifier, validity period identifier, credibility identifier, and access permission identifier.
[0087] The knowledge evidence set is a collection of standard knowledge units that meet the requirements of specific competency items, retrieved from a structured knowledge base. Each piece of knowledge evidence in this set not only includes the knowledge text (such as rules and regulations, manual operation steps, and case explanations), but also forcibly includes metadata information such as source identifier (such as Article Y of Chapter X of the "XX Operation Manual"), version identifier (such as version V2.1 in 2026), validity period identifier (whether it is within the shelf life) and credibility identifier, which are used for "validity verification" of the intelligent agent generated in subsequent courses.
[0088] After receiving the knowledge evidence set, the course-generating agent first verifies the validity period (not expired) and version (the latest effective version) of the evidence. Training content is generated solely based on valid knowledge evidence; the model is strictly prohibited from incorporating external unknown knowledge or making arbitrary additions, thus eliminating "illusions" from the outset. When generating theoretical learning, case studies, scenario-based Q&A, simulations, exercises, and assessment questions, the course-generating agent generates a structured four-tuple data structure: [Training Content ID, Knowledge Node ID, Target Capability ID, Assessment Item ID]. For example, an assessment question (assessment item) about "emergency landing procedures" is explicitly bound to "Manual Item 3.2 (knowledge node)" and "Emergency Response Capability (target capability item)" in the knowledge base. This ensures that if a student answers the question incorrectly, the system can accurately trace back to the specific knowledge node and the unmet capability item. The above describes the specific process by which the course-generating agent generates training content solely based on valid knowledge evidence and establishes a binding relationship between each piece of training content and its corresponding knowledge node, target capability item, and assessment item.
[0089] Personalized training packages typically include the following: theoretical learning content, case study content, scenario-based Q&A content, simulated operation content, practice questions and assessment questions, training sequence, training duration, achievement criteria, and supplementary learning rules.
[0090] For courses with relatively fixed knowledge paths and low requirements for expert experience expansion, the training implementation agent interacts with trainees and conducts online / offline instruction in the form of a digital human (digital human courseware generation and interaction). For other courses, personalized training task packages are pushed to trainees for interactive Q&A and simulated operation exercises. Throughout the training interaction process, a multimodal data collection module (such as the case extraction agent linked with the training implementation module) is deployed in parallel in the background to monitor and capture multi-dimensional behavior logs and submitted materials of trainees on the learning end, answer interface, simulator device, and interactive terminal in real time.
[0091] Multimodal training behavior data during the training process includes: course completion status, answer content, answer duration, number of answer modifications, operation step sequence, simulator operation results, scenario handling selection, written materials submitted by employees (such as summaries and answer explanations), intelligent question and answer interaction records, and instructor evaluation data.
[0092] Step 3: The capability assessment agent evaluates and updates the current capability profile of the trainee based on multimodal training behavior data; calculates the capability gap between the updated current capability profile and the target capability profile; if it is determined that the preset training objective has not been achieved, the training path planning agent generates a supplementary training path based on the capability gap, and the overall control and orchestration agent re-triggers the generation and implementation of the training task. If the preset training objectives are determined to be met, the case extraction agent extracts candidate knowledge from the multimodal training behavior data and the materials submitted by the trainees. The knowledge review and update agent verifies and reviews the candidate knowledge and updates the approved candidate knowledge into the structured knowledge base.
[0093] Multimodal training behavior data refers to the multimodal data previously collected by the training implementation agent during the training process. This includes course completion status, answer content, answer duration, number of answer modifications, operation step sequences, simulator operation results, scenario handling selections, submitted written materials, intelligent question-and-answer interaction records, and instructor evaluation data. This heterogeneous multimodal data is transformed into unified training behavior characteristics, and a mapping relationship is established between these unified training behavior characteristics and the training subjects (employees), training tasks, structured knowledge base nodes, and target competency items.
[0094] As mentioned earlier, the competency assessment agent assesses the competencies of each job position. (corresponding to the first) (Each capability item), call the quantization evaluation operator to calculate the updated current capability value according to the following formula. :
[0095] Standardized computation: computational training subjects in the first Standardized scores in class assessment data ; Weighted and credibility-weighted: multiplied by the corresponding number of data points in this category. Specific weights of each ability And the credibility coefficient of the data source. (For example, the credibility of hands-on simulation is higher than that of purely theoretical multiple-choice questions.) Time decay correction: multiply by the time decay factor This allows historical performance data from a long time ago to naturally decay, highlighting recent training performance.
[0096] Calculate each item Then, the updated current capability profile vector is generated by combining and encapsulating the data. And it is updated and stored in the aviation personnel competency profile database in a synchronized manner.
[0097] Update the current ability profile vector With target capability profile vector Perform difference calculations to obtain the capability gap vector. (Including individual item differences) ).
[0098] According to the gap vector The evaluation data will determine whether the preset training objectives have been achieved. If any of the following conditions are met, the training is deemed unsuccessful: Current capability value of any required capability item Below the corresponding preset threshold; The total weighted capability gap exceeds a set threshold. If a high-risk ability item is not passed (one vote veto); The completeness of the assessment evidence or the confidence level of the competence assessment are below the set threshold.
[0099] If the capability assessment agent is deemed not to meet the standard, it will output a specific list of the unmet capabilities and an analysis of the reasons for the errors, and send it to the training path planning agent to generate a supplementary training path; if the capability is deemed to meet the standard, it will trigger the subsequent knowledge extraction and feedback process.
[0100] refer to Figure 4 When it is determined that the trainee has not reached the preset training objectives and retraining is triggered, the method in this embodiment distinguishes between the retraining process for an individual trainee and the retraining process for a class group: Individual trainee retraining process: When only individual trainees (such as trainee A and trainee B) fail to meet the competency requirements after training, the competency assessment agent makes a judgment based on the trainee's multimodal learning process during training. The course generation agent, in conjunction with the competency assessment agent, regenerates personalized supplementary training content and organizes retraining. After retraining, the competency profile is compared and judged again. If the trainee meets the requirements, the training process ends. If the trainee still fails to meet the requirements, the cycle continues until the maximum number of training rounds is reached.
[0101] Class-based retraining process: If multiple students fail to improve their skills (e.g., the overall improvement in skills after training is below the preset threshold), the system determines that the training content or the instructor's teaching method does not meet expectations. In this case, retraining is not triggered for individual students, but for the entire class: the course generating agent and the ability assessment agent regenerate / adjust the overall training content, or the teaching agent adjusts the teaching method, or the non-agent instructor is replaced and the training is reorganized.
[0102] The training path planning agent is constructed as follows: First, it has a built-in graph search operator that can trace back the preceding knowledge nodes and dependency trees in the knowledge graph based on the unmet capabilities in the capability gap vector. Then, it has a built-in rule priority engine that encapsulates 5 levels of hard priority reordering rules and a built-in path orchestration generator that reassembles the priority-ordered knowledge nodes and course units into a sequential supplementary training path, outputting a standard message object containing preceding task identifiers and training task identifiers.
[0103] The training path planning agent receives the capability gap vector output by the capability assessment agent. After the assessment report (including unmet skill items and incorrect answers), knowledge nodes and training units are reselected and prioritized according to a strict 5-level priority system: Priority 1 (highest): Knowledge nodes directly related to the unmet competency items (directly targeting weaknesses); Priority 2: Identify the underlying knowledge nodes that led to the error (trace the root cause and fill in the missing dependencies). Priority 3: Case nodes similar to error scenarios (reinforce training through similar variant cases); Priority 4: Training formats with historically proven effectiveness (combining learner preferences with historically high-yield teaching methods); Priority 5: Required content with high risk level (ensure that high-risk compliance items are reviewed first).
[0104] The planning agent will reorder and reassemble the knowledge node chain according to the above priorities, encapsulate it as a supplementary training path, and send it to the master control orchestration agent.
[0105] The overall control and orchestration agent, together with the various specialized task capability agents, constitute a multi-agent collaborative framework, serving as the central hub and commander of the entire system.
[0106] The overall control and orchestration agent is built on an architecture of "state machine + workflow orchestration engine + verification arbitrator": Workflow orchestration and state tracker: Maintaining state transitions throughout the entire closed-loop training lifecycle (requirements identification) Task generation Training Implementation Competency Assessment (Closed-loop supplementary training / feedback to the warehouse) Standard Message Bus: Responsible for distributing and converting standard message objects among various professional task capability agents (including demand identification agent, capability assessment agent, course generation agent, training implementation agent, training path planning agent, case extraction agent, and knowledge review and update agent); agents communicate with each other through standard message objects, which include at least several of the following: training task identifier, employee anonymity identifier, job position and model identifier, input data address, required knowledge scope, output data type, confidence level, reference basis, preceding task identifier, and processing status.
[0107] Consistency and confidence verification engine: It has a built-in verification rule library, which compares the output results reported by each agent with rules and determines the confidence threshold. When the verification fails, it automatically triggers the retry, review or manual intervention process.
[0108] When the master control and orchestration agent receives a decision message from the capability assessment agent indicating that the preset training objectives have not been met, and a supplementary training path submitted by the training path planning agent, it marks the status as triggering the supplementary training loop.
[0109] The master control and orchestration agent sends a standard message object with supplementary training paths and prerequisite task identifiers to the course generation agent, which then triggers the course generation agent to generate personalized supplementary training task packages based solely on the knowledge nodes in the supplementary paths.
[0110] After the course generation agent completes the task package, the overall control and orchestration agent performs consistency and confidence checks on its output content. After the checks pass, the overall control and orchestration agent issues execution instructions to the training implementation agent, which then pushes supplementary training content to the trainees and restarts the collection of multimodal training behavior data to achieve closed-loop control.
[0111] If the preset training objectives are determined to have been achieved, the case extraction agent extracts candidate knowledge from the multimodal training behavior data and the materials submitted by the trainees, verifies and reviews the candidate knowledge, and updates the structured knowledge base with the approved candidate knowledge, specifically including: The case extraction agent performs text segmentation (dividing long summaries or answer records into semantically independent paragraphs and sentence groups), entity recognition (extracting entities such as regulations, equipment / models, fault phenomena, and handling steps from the text) and relationship extraction (identifying logical relationships between entities, such as "fault phenomenon A") on the written materials submitted by the trainees. lead to Risk Event B, Handling Steps C in accordance with The manual includes clause D), distinguishing between factual and opinion content (separating objective factual descriptions, such as "operated XX switch" from subjective opinions / lessons learned), identifying conflicting content (comparing the extracted logic with the existing logic in the knowledge base and marking semantically contradictory or conflicting content), anonymization (masking or anonymizing sensitive privacy information such as names, specific machine numbers, and employee anonymity identifiers), and candidate knowledge unit generation (encapsulating the structured content, entity relationship triples, and case summaries processed above into standardized "candidate knowledge units"). Candidate knowledge units are draft standard knowledge units that have not been formally written into the knowledge base. Their content mainly consists of the following parts: knowledge text / core content (extracted standardized handling procedures, typical fault case analysis, excellent operating experience, or test question analysis), structured entities and triple relationships (extracted [entity 1-relationship-entity 2] logical links, such as associated equipment nodes, regulatory nodes, and risk event nodes), and metadata tags (including source identifiers such as which trainee / training task it was extracted from, suggested version number, applicable position and machine model, suggested credibility identifier, and anonymization status identifier).
[0112] The knowledge review and update agent performs checks on candidate knowledge units for regulatory consistency, version consistency, source credibility, and duplication. Candidate knowledge units that pass the checks are written into a structured knowledge base, and traceable relationships are established with the original training cases, trainees' competency items, and knowledge sources. The specific process is as follows.
[0113] Establish connections with knowledge sources: Attach source identifiers and citation basis attributes to knowledge nodes, and connect them to the original training record identifier, regulation clause node, or authoritative manual version corresponding to the source through "derived from / based on" edges; Establish connections with original training cases: Create logical relationship edges in the graph that are "verified by the case / derived from the case", and bind the newly written knowledge nodes to the original test questions / exercise scenario case nodes used in this training; Establish connections with trainees' competency items: Link new knowledge nodes to corresponding "job competency nodes" (e.g., ...). This establishes a mapping relationship between "improvement / corresponding ability" so that when other trainees have gaps in this ability, the system can accurately call upon the feedback case along the traceable relationship chain.
[0114] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.
[0115] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the invention.
[0116] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.
[0117] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.
[0118] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.
[0119] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A closed-loop training and capability assessment method based on multi-agent collaboration, characterized in that the method... include: The demand identification agent identifies the training subjects and target positions from the training needs, and the capability assessment agent generates a target capability profile based on the job requirements of the target positions. The ability assessment agent obtains the current ability profile of the trainee, the course generation agent combines the knowledge content in the structured knowledge base to generate a personalized training task package based on the difference between the target ability profile and the current ability profile, and the training implementation agent conducts training on the trainee and collects multimodal training behavior data during the training process. The capability assessment agent evaluates and updates the current capability profile of the training subject based on the multimodal training behavior data; Calculate the capability gap between the updated current capability profile and the target capability profile; If it is determined that the preset training objective has not been achieved, the training path planning agent will generate a supplementary training path based on the capability gap, and the overall control and scheduling agent will re-trigger the generation and implementation of the training task. If the preset training objective is determined to be achieved, the case extraction agent extracts candidate knowledge from the multimodal training behavior data and the materials submitted by the trainees. The knowledge review and update agent verifies and reviews the candidate knowledge and updates the approved candidate knowledge into the structured knowledge base.
2. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The required identification agent, the capability assessment agent, the course generation agent, the training implementation agent, the training path planning agent, the case extraction agent, and the knowledge review and update agent constitute a professional task capability agent, which together with the overall control and orchestration agent forms a multi-agent collaborative framework. The task dependencies and verification rules between the various professional task capability agents are established based on the overall control and orchestration agent. During the operation of the multi-agent collaborative framework, the overall control and orchestration agent performs consistency and / or confidence verification on the output results of each agent, and triggers a retry, review, or manual intervention process when the verification fails.
3. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The target competency profile includes multiple job competency requirements and corresponding target competency levels.
4. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The construction process of the structured knowledge base includes: Acquire text and image data, including regulations, manuals, job standards, cases, courses, and historical training data, and perform format conversion, paragraph segmentation, duplicate content removal, sensitive information processing, version identification, validity period identification, and source marking on the data to form standard knowledge units; Entities, attributes, and relationships are extracted from the standard knowledge units to establish a hierarchical knowledge structure or knowledge graph that includes nodes for legal provisions, operating procedures, job competencies, equipment models, risk events, training courses, assessment questions, and case studies. Each knowledge unit is assigned a source identifier, version identifier, validity period identifier, credibility identifier, and access permission identifier.
5. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 2, characterized in that, The various intelligent agents communicate with each other through standard message objects, which include at least several of the following: training task identifier, employee anonymity identifier, job position and model identifier, input data address, required knowledge scope, output data type, confidence level, reference basis, prerequisite task identifier, and processing status.
6. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The training requirements are generated by the requirement identification agent based on demand-triggered conditions or periodic trigger conditions, wherein: The triggering conditions include at least one of the following: job change, new equipment or system put into operation, change of regulations or operation manual version, qualification about to expire, occurrence of job-related event, a certain ability indicator in the ability profile falling below a preset threshold, abnormal operation, and management personnel initiating a special task; The periodic triggering conditions include: annual, quarterly, monthly, or qualification periodic training plans.
7. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The process of generating the personalized training task package includes: The knowledge retrieval agent retrieves knowledge units corresponding to the target capability items in the differences from the structured knowledge base, and returns a set of knowledge evidence with source and version information to the course generation agent; The course-generating agent generates training content based solely on the set of knowledge evidence that has passed validity verification, and establishes a binding relationship between each piece of training content and the corresponding knowledge node, target ability item, and assessment item.
8. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The multimodal training behavior data includes: At least several of the following: course completion status, answer content, answer duration, number of answer revisions, sequence of operation steps, simulator operation results, scenario handling selection, written materials, intelligent question and answer interaction records, and instructor evaluation data; The multimodal training behavior data is converted into unified training behavior features and associated with training subjects, training tasks, knowledge nodes, and competency items.
9. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The capability assessment agent calculates the first [value] based on the multimodal training behavior data. Item Ability Value The formula is: in, For the first Class 1 evaluation data corresponds to the first The weight of each capability; For the first Standardized scores in class assessment data; For the first The credibility coefficient of the evaluation data; This is the time decay coefficient.
10. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, Situations where the preset training objectives have not been achieved include: If the current capability value of any required capability item is lower than the corresponding threshold, the total weighted capability gap exceeds the set threshold, a high-risk capability item is not passed, the completeness of the assessment evidence is lower than the set threshold, or the confidence level of the capability assessment is lower than the set threshold; The supplementary training path is regenerated by the training path planning agent according to the following priorities: knowledge nodes directly associated with the unmet competency items, prerequisite knowledge nodes that lead to errors, case nodes similar to the error scenarios, training methods that have historically proven effective, and mandatory content with high risk levels.
11. The closed-loop training and capability assessment method based on multi-agent collaboration according to claim 1, characterized in that, The step of verifying and reviewing the candidate knowledge, and updating and writing the approved candidate knowledge into the structured knowledge base includes: The case extraction agent performs text segmentation, entity recognition, relation extraction, distinction between factual and opinion content, identification of conflicting content, desensitization processing, and generation of candidate knowledge units on the text materials submitted by the training subjects. The knowledge review and update agent performs legal consistency, version consistency, source credibility, and duplication checks on the candidate knowledge units, and writes the candidate knowledge units that pass the checks into the structured knowledge base, while establishing traceable associations with the original training cases, training subjects' competency items, and knowledge sources.
