Safety training model, method and system based on capability evaluation dynamic weight optimization model, and computer medium

Abstract

The invention provides a safety training model, method and system based on a capability evaluation dynamic weight optimization model, and a computer medium, and the method specifically comprises the steps: building a capability-knowledge-course three-dimensional map model, building a three-dimensional mapping relation, and determining an initial weight; fusing, optimizing, adjusting and triggering subjective and objective weights dynamically based on student behavior data, generating a personalized curriculum recommendation sequence through a reinforcement learning algorithm, forming a state space by an ability evaluation vector and a knowledge mastery degree vector, and taking a curriculum resource set as an action space; the short-term learning efficiency and the long-term post assessment passing rate are used as reward functions to construct a recommendation strategy, meanwhile, the invention further provides a computer medium achieved based on the scheme, and the method can provide safety capability assessment means and training courses which are high in pertinence and good in accuracy for chemical enterprise employees.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of chemical safety training model system and method, and particularly relates to a safety training model, method and system based on dynamic weight optimization model of ability evaluation, and a computer medium. BACKGROUND

[0002] Safety production is of great importance in high-risk industries such as chemical industry, and the cultivation of professional talents and the improvement of safety ability are the key to ensure safety production. The chemical industry uses a variety of dangerous chemicals and has complex processes, and once an accident occurs, the harm is great. With the continuous progress of technology in these industries, production equipment is becoming increasingly complex, and the safety skills of employees are also becoming higher and higher. Safety production training is a basic work in the production process of chemical enterprises. With the continuous optimization and upgrading of China's industrial structure, the level of mechanization, automation, digitization and intelligence of the chemical industry will continue to improve, and the production device is more complex, the process operation is more dangerous, and the safety hidden danger is more concealed and complex. It is urgent to invest more professional, high-level, and high-receiving ability talents into the chemical industry to promote the stable and orderly development of the entire chemical industry.

[0003] At present, there are still significant deficiencies in the establishment and evaluation of safety ability models. In terms of model establishment, existing researches are mostly targeted at specific industries or positions, with poor universality and difficulty in direct application to other industries. In terms of safety ability evaluation, the evaluation methods are often complex and not accurate enough, which cannot effectively meet the needs of enterprises for rapid and accurate evaluation of employees' safety skills. Some scholars have pointed out that the current miners have accident-prone characteristics based on accident analysis, and have constructed a safety competence model of behavior characteristics by using text analysis, personnel interviews, questionnaire surveys and statistical analysis, and combining qualitative and quantitative analysis methods for job competence. Some scholars have proposed to use the iceberg model theory and combine principal component analysis to construct an employee safety competence hypothesis model based on the consideration of the work characteristics of power enterprises, providing a theoretical basis for personnel selection of enterprises. Some scholars have proposed to analyze the safety behavior of power plant main operators to construct a human cognitive model and an evaluation model based on multi-level matter element analysis in view of the problem of the proportion of human factors in accidents under the automation trend of power plants. In the same year, some scholars analyzed the relationship between the influencing factors of safety ability, and then proposed a safety ability modeling method based on the questionnaire survey results, and obtained the key influencing factors of safety ability of workshop operators. DEMATEL

[0004] In terms of safety ability evaluation, some scholars have proposed to avoid the subjectivity of the evaluation process by establishing a safety ability evaluation system based on the concept and connotation of the safety ability of mechanical processing workshop operators. AHP ​In view of the defects of the fuzzy judgment in the law, the safety ability hierarchical model is selected based on the fuzzy hierarchical weight method. Some scholars propose a three-factor model of construction personnel safety ability based on competency theory and safety theory, combined with literature research, questionnaire survey and behavior event interview, and establish a safety ability evaluation system combined with structural equation method. Some scholars propose to obtain the relationship between safety atmosphere and subway construction personnel safety by using questionnaire to investigate personnel and applying SPSS and AMOS software to verify and correct the research data to obtain the safety relationship model. In the same year, some scholars proposed to establish a safety ability evaluation index system based on tree-type hierarchical structure, and obtained that the safety ability of air traffic control units is formed by the mutual influence of personnel, equipment, environment, management and transportation in the production link.

[0005] At present, there are still problems and shortcomings in the quantity, structure, quality, specialty, training and protection of the talent team construction in the chemical industry, which is not compatible with the current demand of the talent team in the chemical industry, and there are still many problems and difficulties, for example, 1) the number of chemical professionals is relatively insufficient, and the business quality needs to be improved, 2) the chemical industry pays insufficient attention to talent training, on the one hand, the training lacks pertinence and practicality, and cannot meet the needs of enterprise safety production; on the other hand, the training courses are not systematic, the training methods are single, and there is a lack of unified, scientific and diversified training methods; 3) the safety skills of chemical employees lack effective evaluation, and it is difficult to establish an effective evaluation system.

[0006] Therefore, in view of the actual problems existing at present, how to provide safety ability evaluation means and training courses with strong pertinence and good precision for chemical enterprise employees, and train high-quality chemical talents for chemical enterprises, is of great significance to the safety production level of the chemical industry. SUMMARY

[0007] Technical scheme: In order to solve the above technical problems, the safety training model based on the dynamic weight optimization model of ability evaluation provided by the application, the model includes an ability evaluation module, a dynamic weight calculation module, a reinforcement learning recommendation module and a real-time feedback module; The ability evaluation module is configured to extract the three-dimensional indexes of basic ability, professional ability and management ability required by the employee's post based on the post competency model, generate a personalized ability evaluation vector combined with the context characteristics of the employee, and the ability evaluation vector contains the label level and initial weight of each ability dimension; The dynamic weight calculation module is configured to build a three-dimensional correlation relationship of "ability-knowledge-course", and to optimize and adjust the weight, the weight optimization includes establishing an ability-knowledge mapping matrix, a knowledge-course mapping relationship and a subjective and objective weight fusion; The reinforcement learning recommendation module uses the ability assessment vector and the knowledge mastery vector to form the state space. The course resources are considered as an action space, with short-term learning efficiency (the ratio of learning duration to mastery) and long-term job performance evaluation pass rate serving as the reward function. Based on the reinforcement learning framework and policy gradient algorithm, a recommendation strategy is constructed, and the course recommendation order is given by dynamically optimizing the Q-value iteration. The real-time feedback module is configured to collect student behavior data and contextual features in real time and trigger a set mechanism. The model uses a three-dimensional graph to structurally link job competencies, professional knowledge systems, and training course resources, forming a dynamically evolving competency development path.

[0008] As an improvement, the competency assessment module selects structural equation modeling analysis for index selection. It uses structural equation modeling to verify the safety competency of practitioners, indirectly reflecting previously unobservable latent variables, discovering the relationship between latent variables, reflecting the relationship between manifest and latent variables, setting evaluation indicators, and establishing a model of the relationship between variables. The structural equation model includes a measurement model and a structural model; specifically... (1) (2) (3) In equations (1)-(3), In the model, it is represented as the first... j One exogenous manifest variable, Represented as the first j One endogenous explicit variable Represented as the first i An exogenous latent variable, Represented as the first i , One endogenous latent variable, Represented as exogenous manifest variables In exogenous latent variables Factor loading coefficient matrix on Λy ij Indicates endogenous manifest variables In endogenous latent variables Factor loading coefficient matrix on Represented as exogenous manifest variables The error; Representing endogenous latent variables and The path coefficient matrix between them; Representing exogenous latent variables Endogenous latent variables The impact; Represents the residual term; Indicates endogenous potential The error.

[0009] As an improvement, the exogenous explicit variables include exam scores, training attendance rates, simulation operation scores, and fault handling speed; the endogenous explicit variables include emergency drill scores, accident handling scores, safety behavior self-assessment, accident incidence rate, and safety inspection scores; the endogenous latent variables include emergency response capabilities, safety attitudes, and safety performance; and the exogenous latent variables include safety knowledge and operational skills.

[0010] As an improvement, specific ways to configure the "ability-knowledge-curriculum" three-dimensional relationship include: (1) Establish a capability-knowledge mapping matrix and determine the initial association weights between capability dimensions and knowledge units based on expert knowledge graphs and historical training data mining; (2) Establish the knowledge-course mapping relationship, and determine the coverage weight of knowledge units and course resources by analyzing the course outline, marking the knowledge points of test questions, and back-inferring the learning effect; (3) Integrate subjective and objective weights and set up a dynamic adjustment formula. Subjective weight Based on the Analytic Hierarchy Process AHP Alternatively, the Delphi method can be used to determine objective weights. Based on students' answer accuracy, learning time, and simulated operation scores, dynamic calculations are performed using a data association algorithm, with weighting adjustment factors. The decay factor decreases exponentially with increasing training data volume; , t For the number of iterations, This is the attenuation coefficient, which has a value greater than 0 and is used to control the weight adjustment factor. The rate at which the training data decays with increasing iterations is, in safety capability assessment, where λ can be dynamically adjusted based on the amount of accident case data. When historical accident data is limited (e.g., ...), the decay rate can be adjusted when there is little historical accident data (e.g., ...). t <10), Higher, preferred AHP Determine the weight of security knowledge; once enough data is accumulated (e.g.) t >50), Approaching 0, the model mainly calculates weights based on objective data such as trainee operation scores and accident handling records.

[0011] As an improvement, in the reinforcement learning recommendation module, the state space... ,in Indicates the first i Evaluation values ​​for each capability dimension Indicates the first jMastery of each knowledge unit; reward function in , These are the weighting coefficients. These are the assessment values ​​for the ability dimensions before and after learning, respectively. For study time, The pass rate for job performance evaluation.

[0012] As an improvement, the contextual features include at least one of the following: years of service, job type, position type, position risk level, and historical training records; the reinforcement learning framework includes deep learning. Q network DQN Policy gradient algorithm PolicyGradient At least one; the behavioral data includes the accuracy rate of answering questions, learning time, interactive operation scores, and number of video replays.

[0013] As an improvement, the established mechanisms include: (1) Update the importance of knowledge nodes by collecting real-time behavioral data on student operations, assessment scores, and knowledge mastery. Utilize this collected data to... PageRank The algorithm dynamically adjusts the node weights in the knowledge point network, updates the importance of knowledge nodes, establishes a safety capability assessment system for chemical industry practitioners, and establishes a safety training model based on the dynamic weight optimization model of capability assessment. It comprehensively considers factors such as employees' psychological qualities, business and technical capabilities, and safe operation skills, providing a basis for safety assessment, pre-training assessment, and post-training effectiveness verification. (2) Abnormal knowledge node detection: When the standard deviation of the student's answer accuracy rate is >0.3, or the set knowledge points for detection are consecutive N When the error rate in answering questions exceeds the threshold, a weight adjustment and a reconstruction of the course resource association relationship are triggered. (3) The weight adjustment factor is adaptive. The ratio of subjective and objective weight fusion is dynamically adjusted according to the data sample size. In the initial sample size, the expert experience weight is given priority, and the proportion of data-driven weight is gradually increased as the data accumulates.

[0014] As a specific embodiment of the present invention, a safety training method based on a capability assessment dynamic weight optimization model is also provided, the method comprising: (1) Construct a three-dimensional graph model of competence-knowledge-course. In the three-dimensional graph model, the competence dimension is decomposed based on the job competency model and the weight is quantified. The knowledge dimension uses knowledge graph technology to construct a knowledge point network. The course dimension annotates the training resource library with metadata. (2) Dynamically adjust subjective and objective weights based on student behavior data. Determine the initial weights by establishing a capability-knowledge mapping matrix and a knowledge-course mapping relationship, and adopt a subjective and objective weight fusion formula. Perform weight optimization, where exponentially decays with the increase of training data amount; (3) Generating a personalized course recommendation sequence through a reinforcement learning algorithm, taking the ability evaluation vector and the knowledge mastery vector as the state space, the course resource set as the action space, and the short-term learning efficiency and the long-term post examination pass rate as the reward function to construct a recommendation strategy.

[0015] As a specific embodiment of the present application, the present application provides a safety training system based on a dynamic weight optimization model of ability evaluation, which works based on the above safety training model, including an initialization stage, a running stage, and an optimization iteration stage, specifically: (1) Initialization stage: (1.1) The ability evaluation module inputs the individualized ability evaluation vector containing the level of each ability dimension and the initial weight, and outputs the individualized ability evaluation vector of the dimension label level and the initial weight generated by the post competency model and the optimized "ability-knowledge-course" mapping weight ; (1.2) The exogenous and endogenous manifest variables are screened out through the structural equation model; (1.3) The ability evaluation module transmits the evaluation value of each ability dimension in the ability evaluation vector containing the initial mastery of each knowledge unit （K 1 ,K 2 ,…,K n ） to the reinforcement learning recommendation module as the initial value of the state space of the reinforcement learning ; (1.4) The dynamic weight optimization module sends the initial weight of the "ability-knowledge-course" three-dimensional correlation relationship constructed to the reinforcement learning recommendation module, including the initial weight data of the ability-knowledge mapping matrix and the knowledge-course mapping relationship; (2) Running stage: (2.1) The reinforcement learning recommendation module calls the latest "ability-knowledge-course" three-dimensional correlation weight data through the dynamic weight optimization module in real time through the interface request, which is used for recommendation strategy calculation; (2.2) The dynamic weight optimization module returns the fused subjective and objective weights calculated by the reinforcement learning recommendation module , wherein , is the decay coefficient; (2.3) The reinforcement learning recommendation module pushes the personalized course recommendation sequence generated based on the reinforcement learning framework and the policy gradient algorithm to the real-time feedback module, and the sequence is sent to the user through the interface request after being fused with the real-time feedback dataQ Value iteration and dynamic optimization; (2.4) The real-time feedback module transmits student behavior data and features back to the reinforcement learning recommendation module in real time; the real-time feedback module transmits student behavior data and assessment data to the ability assessment module, triggering the update of the ability assessment vector and knowledge mastery vector; the real-time feedback module sends student operation behavior data, assessment score data, and knowledge mastery data to the dynamic weight optimization module, triggering the update of knowledge node importance, and adjusting the weight of knowledge point network nodes through the PageRank algorithm; when the abnormal knowledge node detection conditions are met, the standard deviation of the answer accuracy rate is >0.3 or the set detection knowledge point's consecutive N-time answer error rate exceeds the threshold, the weight adjustment and course resource association relationship reconstruction are triggered. (3) Optimization and iteration stage: (3.1) The updated capability assessment vector, including the latest assessment values ​​of each capability dimension and the knowledge mastery vector, is sent to the dynamic weight optimization module to trigger weight optimization and adjustment. (3.2) The dynamic weight optimization module pushes the optimized three-dimensional correlation weight data of "ability-knowledge-course" to the reinforcement learning recommendation module to update the recommendation strategy; (3.3) The ability assessment module passes the updated ability assessment vector and knowledge mastery vector to the reinforcement learning recommendation module to update the reinforcement learning state space S; (3.4) Real-time feedback module: Provides assessment values ​​of students' abilities before and after learning. Study time and job performance evaluation pass rate P Used to calculate the reward function The reinforcement learning recommendation module is based on this. Q Value iteration optimization recommendation strategy; (3.5) Dynamic weight optimization module: Synchronize the change information of the weight adjustment factor ∝, and the real-time feedback module adjusts the data collection frequency and the threshold parameters of the triggering mechanism accordingly; when it is detected that the student's ability has significantly improved after learning a certain type of course, the change in the ability assessment value exceeds the set threshold γ or there is no significant improvement or it is lower than the threshold. δ At the same time, the real-time feedback module sends course effectiveness evaluation data to assist the dynamic weight optimization module in deeply calibrating the relationship between "ability-knowledge-course".

[0016] As another specific embodiment of the present invention, a computer medium is also provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps of the security training method based on a capability assessment dynamic weight optimization model as described above.

[0017] Beneficial effects: The safety training model and method based on a capability assessment dynamic weight optimization model proposed in this invention have the following advantages compared with conventional methods: (1) The model in this invention includes four modules. Through closed-loop information interaction of the ability assessment module, dynamic weight optimization module, reinforcement learning recommendation module, and real-time feedback module, it achieves dynamic adaptation of the entire process of "assessment-recommendation-learning-feedback-optimization". Compared with conventional static recommendation models, it can more accurately capture learners' ability shortcomings and knowledge gaps. Specifically, it is reflected in: The real-time data interaction between the competency assessment module and the dynamic weight optimization module enables the three-dimensional relationship between "competency-knowledge-course" to be dynamically adjusted as the learner's competency changes, thus avoiding recommendation bias caused by the fixed weights in conventional models. The high-frequency information flow between the reinforcement learning recommendation module and the real-time feedback module can quickly iterate the recommendation strategy based on the student's real-time learning behavior (such as the accuracy of answering questions, the number of times the video is watched, etc.), so that the course recommendation is upgraded from "batch push" to "personalized adaptation", and the short-term learning efficiency (time to mastery ratio) is improved by more than 30%. The real-time feedback module transmits data bidirectionally to the competency assessment module and the dynamic weight optimization module, triggering updates to the importance of knowledge nodes and detection of abnormal knowledge nodes. This ensures that the model can respond promptly to learners' learning bottlenecks. Compared to the periodic updates of conventional models, the problem response speed is improved by 50%, and the pass rate for long-term job assessments is increased by more than 20%.

[0018] (2) The multi-module collaborative interaction mechanism of the present invention solves the pain points of "disconnect between assessment and recommendation" and "feedback lagging behind the learning process" in conventional training models, and transforms the ability development path from "preset fixed route" to "dynamic evolution trajectory", which significantly improves the utilization efficiency of training resources and the matching accuracy of job competency.

[0019] (3) The "ability-knowledge-course" three-dimensional graph model constructed in this invention is based on the ability dimension, which is decomposed and weighted according to the job competency model. The knowledge dimension uses knowledge graph technology to construct a knowledge point network. The course dimension uses metadata annotation for the training resource library.

[0020] (4) The present invention annotates multi-dimensional metadata for courses, and realizes quantitative association with knowledge and ability through a "knowledge-course mapping matrix". In existing systems, ability, knowledge and courses are mostly unidirectionally associated with "ability → course", and the association relationship is fixed. The present invention forms a dynamic closed loop among the three: ability shortcomings drive knowledge weight adjustment, knowledge blind spots trigger course recommendation, course learning effect feeds back into ability assessment, and weight changes with the amount of data through W. final The formula is automatically optimized. Furthermore, this invention offers advantages such as improved accuracy, strong dynamic adaptability, optimized resource utilization, and enhanced interpretability. Attached Figure Description

[0021] Figure 1 A dynamic weight optimization logic diagram for the present application.

[0022] Figure 2 A three-dimensional atlas construction flowchart for the present application.

[0023] Figure 3 A training course system part screenshot of embodiment 2 of the present application.

[0024] Figure 4 System architecture diagram example 1 of the present application.

[0025] Figure 5 System architecture diagram example 2 of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely below, so that the people skilled in the art can better understand the advantages and features of the present application, and the protection scope of the present application can be defined more clearly. The described embodiments of the present application are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the people skilled in the art without creative labor fall within the protection scope of the present application.

[0027] A safety training model, method and system of a capability evaluation dynamic weight optimization model, and a computer medium are provided in the present application, which are described and introduced in detail below.

[0028] (I) A safety training model of a capability evaluation dynamic weight optimization model is provided

[0029] The model of the present application is applied to a chemical enterprise safety training system, and the model comprises a capability evaluation module, a dynamic weight calculation module, a reinforcement learning recommendation module and a real-time feedback module.

[0030] (1.1) Capability evaluation module In the capability evaluation module of the present application, the specific embodiments adopted include using DACUM(Developing A Curriculum) Post task analysis is performed, the professional ability description required by the post of the chemical enterprise is specific and clear, which provides necessary and accurate basic information for the development of the training course system, and provides very necessary knowledge and skill information for the construction of the professional course system.

[0031] Further, the module of the application adopts a structural equation model to verify the model of the safety ability of the practitioner, indirectly reflects the latent variable which cannot be observed originally, finds the relationship between the latent variables, reflects the relationship between the manifest variable and the latent variable, sets the evaluation index, and establishes the relationship model between the variables.

[0032] The structural equation model of the application includes a measurement model and a structural model, and the measurement equation and the structural equation corresponding to the two models are as follows: (1) Wherein, represents an exogenous manifest variable, which is represented as an examination score, a training attendance rate, a simulation operation score, and a fault handling speed in the model. represents an exogenous latent variable, which is represented as safety knowledge, operation skill, and the like in the model. represents a factor loading coefficient matrix of the exogenous manifest variable on the exogenous latent variable , for example, represents the contribution degree of the safety examination score to the safety knowledge ; represents an error of the exogenous manifest variable , for example, represents the part of the safety examination score which is not explained by the safety knowledge.

[0033] (2) represents an endogenous manifest variable, which is represented as an emergency drill score, an accident handling score, a safety behavior self-evaluation, an accident occurrence rate, and a safety inspection score in the model. represents an endogenous latent variable, which is represented as emergency ability, safety attitude, safety performance, and the like in the model. represents a factor loading coefficient matrix of the endogenous manifest variable Y on the endogenous latent variable η , for example, represents the representativeness of the emergency drill score to the emergency ability . represents an error of the endogenous manifest variable , for example, ε 1 represents the part of the emergency drill score which is not explained by the emergency ability.

[0034] The structural model reflects the relationship between latent variables, and the path diagram can represent the structural relationship. The mathematical expression of the structural model is as follows: (3) denotes the path coefficient matrix between endogenous latent variables, for example denotes the direct influence of emergency capability on safety performance ; denotes the influence of exogenous latent variables on endogenous latent variables, for example denotes the direct influence of safety knowledge on emergency capability ; denotes the residual term, reflecting the part of η that is not explained, for example is the emergency capability residual explained by the model.

[0035] (1.2) Dynamic weight calculation module As shown in Figure 1 , it is a dynamic weight optimization logic diagram of the dynamic weight calculation module of the present application, configured to build a "ability-knowledge-course" three-dimensional correlation relationship, and to optimize and adjust the weight. The weight optimization includes establishing an ability-knowledge mapping matrix, a knowledge-course mapping relationship, and subjective and objective weight determination, fusion, and triggering.

[0036] Example 1 1.2.a Build a "ability-knowledge-course" three-dimensional correlation relationship Select 100 reaction kettle operators in a chemical enterprise as samples, and build a "ability-knowledge-course" three-dimensional correlation relationship for their "safety operation ability" training, and track the dynamic weight changes in a 3-month training period.

[0037] Table 1 Schematic diagram of building a "ability-knowledge-course" three-dimensional correlation relationship dimension element ability dimension safety operation ability (C1), emergency disposal ability (C2) equipment maintenance ability (C3) knowledge unit reactor pressure control specification (K1), dangerous chemical leakage emergency process (K2), safety valve calibration standard (K3), equipment inspection specification (K4) course resources course A (《reaction kettle safety operation specification》), course B (《dangerous chemical leakage emergency drill》), course C (《special equipment calibration technology》) 1.2.b Ability-knowledge mapping matrix construction and weight optimization (1) Initial weight determination (expert subjective weight W expert ) Invite 5 senior senior engineers and safety experts (more than 10 years of experience) to score the "ability-knowledge" correlation degree (1-5 points, 5 points for strong correlation), and normalize the average value to obtain W expert .

[0038] Table 2 Data after initial weight determination ability dimension K1 reactor pressure control K2 dangerous chemical leakage emergency K3 safety valve calibration safety operation ability (C1) 0.8 0.3 0.7 emergency disposal ability (C2) 0.2 0.9 0.4 (2) Extract the training data of the past 1 year (80 trainees) for this post, calculate the Pearson correlation coefficient of "knowledge unit mastery improvement" and "ability assessment value improvement", and get W_data after normalization.

[0039] Table 3 Data of W_data after normalization ability dimension K1 reactor pressure control K2 dangerous chemical leakage emergency K3 safety valve calibration safety operation ability (C1) 0.75 (correlation coefficient 0.72) 0.25 (correlation coefficient 0.21) 0.65 (correlation coefficient 0.63) emergency disposal ability (C2) 0.15 (correlation coefficient 0.13) 0.85 (correlation coefficient 0.81) 0.35 (correlation coefficient 0.32) (3) Subjective and objective weight fusion Initial stage (t=5, iterate 5 times, small amount of data).

[0040] Take λ=0.1 (decay coefficient), then ∝=e^(-0.1×5)=0.606, the weight after fusion.

[0041] Table 4 Weight data after fusion in initial stage ability dimension K1 reactor pressure control K2 dangerous chemical leakage emergency K3 safety valve calibration safety operation ability (C1) 0.606×0.8+0.394×0.75≈0.78 0.606×0.3+0.394×0.25≈0.28 0.606×0.7+0.394×0.65≈0.68 emergency disposal ability (C2) 0.606×0.2+0.394×0.15≈0.18 0.606×0.9+0.394×0.85≈0.88 0.606×0.4+0.394×0.35≈0.38 Data accumulation stage (t=30, iterate 30 times, sufficient amount of data).

[0042] The weight after fusion (more dependent on objective data).

[0043] Table 5 Weight data after fusion in data accumulation stage ability dimension K1 reactor pressure control K2 dangerous chemical leakage emergency K3 safety valve calibration safety operation ability (C1) 0.05×0.8+0.95×0.75≈0.75 0.05×0.3+0.95×0.25≈0.25 0.05×0.7+0.95×0.65≈0.65 1.2.c Knowledge—Course Mapping Relationship Construction and Weight Optimization (1) Initial coverage weight (expert evaluation) Table 6 Expert evaluation of course coverage on knowledge unit (0-1) knowledge unit course A (operation specification) course B (emergency drill) course C (calibration technology) K1 reactor pressure control 0.9 (core coverage) 0.2 (secondary involvement) 0.3 (related mention) K2 dangerous chemical leakage emergency 0.3 (related mention) 0.9 (core coverage) 0.1 (not involved) K3 safety valve calibration 0.2 (secondary involvement) 0.1 (not involved) 0.9 (core coverage) (2) Dynamic adjustment (based on learning effect backtracking) Through the real-time feedback module, the trainee data is collected: after learning course A, the K1 knowledge unit mastery improvement rate is 70% (the historical average is 50%), then the coverage weight of course A on K1 is adjusted to 0.95; A trainee has made mistakes in K2 knowledge unit for 3 times in a row (error rate 70%> threshold 50%), triggering the coverage weight of course B on K2 from 0.9 to 0.98, and increasing the recommended priority of course B.

[0044] 1.2.d Trigger mechanism and optimization effect Abnormal knowledge node detection: in the 2nd month of training, it is found that the standard deviation of K1 (hazardous chemical leakage emergency process) answer correctness is 0.35>0.3, and the real-time feedback module triggers the dynamic weight calculation part.

[0045] Adjust the mapping weight of K and emergency disposal ability (C2) (from 0.88 to 0.92).

[0046] Reconstruct the association between courses B and K, and increase the proportion of course B (from 20% to 30%).

[0047] Effect verification: 3 months later, the post students.

[0048] The average assessment value of safety operation ability (C1) is increased by 40% (25% for regular training).

[0049] The average assessment value of emergency disposal ability (C2) is increased by 55% (30% for regular training).

[0050] The passing rate of post examination is increased from 65% to 88%.

[0051] See Figure 2 The construction process diagram of the ability-knowledge-course three-dimensional graph model is shown in the figure, and the three dimensions include the ability dimension, the knowledge dimension and the course dimension.

[0052] Further, the knowledge dimension is a knowledge point network constructed by using knowledge graph technology and mapped with the ability, and the knowledge units such as "fire and explosion prevention principle" and "dangerous chemical storage specification" and the knowledge relationship such as "fire and explosion prevention principle" being the prerequisite knowledge of "gas characteristics" are defined.

[0053] Further, the course dimension is metadata annotation of the training resource library, covering course length, difficulty, form and other attributes, and the matching knowledge points and the coverage range of the ability target include designing a three-dimensional graph model framework, defining the ability dimension, decomposing the basic ability, professional ability and management ability based on the post competency model, and quantifying through the weight; defining the knowledge dimension, constructing the knowledge point network mapped with the ability, and structuring by using the knowledge graph technology; defining the course dimension, and metadata annotation of the training resource library.

[0054] In Figure 2 , the ability-knowledge-course three-dimensional graph model is constructed through specific embodiments, specifically: the course dimension is established through professional course system and general course system, the professional course system is usually output by using work task analysis, and needs to be strongly associated with the work scene of employees, and employees need to learn what they do.

[0055] Based on the post ability characteristics, the specific content of each post knowledge item is summarized, the behavior description of each post ability item is refined, the behavior characteristics of different levels are described respectively, the transformation of post competence from abstract concept to specific content is realized, and the specific content of each post knowledge and ability is analyzed, refined, integrated and summarized to form a training course, a course outline is researched and developed, and a chemical practitioner training course library is constructed.

[0056] Based on the training course library, the post competence is taken as the basis, the training target is taken as the basis, the modular course management is taken as the idea, the courses are layered and classified, the courses are reasonably distributed and combined according to the characteristics and needs of employees at different development stages, the logical relationship between the courses is combed, the transformation from the course library to the course system is realized, and finally a multi-level and multi-scene safety skill training course system of "post-competence-course" is established.

[0057] Embodiment 2

[0058] See Figure 3 , which is a part of the specific implementation mode screenshot, wherein (1) the letters A, B and C represent the mastery degree: A represents knowing; B represents independent operation; and C represents guiding others; (2) the numbers represent the period: 1 represents a period of 1 year; 2 represents a period of 2 years; and 3 represents a period of 3 years; (3) the training mode: M1 means classroom teaching; M2 means classroom teaching + examination; M3 means meeting or self-study; and M4 means actual operation.

[0059] (1.3) Reinforcement learning recommendation module and real-time feedback module In the reinforcement learning recommendation module in the present application, the state space is composed of the ability evaluation vector and the knowledge mastery degree vector, the course resource set is the action space, the short-term learning efficiency is the ratio of the learning time and the mastery degree, and the long-term post examination pass rate is the reward function. Based on the reinforcement learning framework and the policy gradient algorithm, the recommendation strategy is constructed, the recommendation order is given by Q value iteration dynamic optimization course; The state space is , wherein represents the evaluation value of the i-th ability dimension, i represents the mastery degree of the i-th knowledge unit; The reward function j , are weight coefficients, are the ability dimension evaluation values before and after learning, is the learning time, is the post examination pass rate. Embodiment 3

[0060] Embodiment 3

[0061] ​Set state space S (ability + knowledge), define ability dimension (3): C1 = safe operation ability (0-10 points), C2 = emergency disposal ability (0-10 points), C3 = equipment maintenance ability (0-10 points). Knowledge unit (4): K1 = reaction kettle pressure control (0-1 mastery), K2 = dangerous chemical leakage emergency (0-1), K3 = safety valve calibration (0-1), K4 = equipment inspection specification (0-1).

[0062] Further, obtain the initial state S0: S0 = [C1 = 5.5, C2 = 4.2, C3 = 6.8; K1 = 0.5, K2 = 0.3, K3 = 0.6, K4 = 0.7].

[0063] Further, the action space A (course resources): Course A: “Reaction Kettle Pressure Control Operation” (related to C1, K1); Course B: “Dangerous Chemical Leakage Emergency Drill” (related to C2, K2); Course C: “Safety Valve Calibration Standard” (related to C3, K3).

[0064] Then, the reward function R (ω1 = 0.5, ω2 = 0.5):

[0065] Where Δt : learning duration (hours); P : post assessment pass rate (0-1).

[0066] The core role of real-time feedback part: data collection: learning duration, correct answer rate, knowledge mastery degree change, examination result.

[0067] Trigger adjustment: when the knowledge unit mastery degree improvement rate <30% or the answer error rate >50%, real-time push warning to the reinforcement learning module.

[0068] The following will introduce and explain the above technical content through a complete process.

[0069] 1. Initial recommendation (based on state S0) Reinforcement learning module calculates initial Q value (Q(S0,A)), policy gradient algorithm selects the optimal course: Q(S0,A) = 2.8, Q(S0,B) = 3.5, Q(S0,C) = 2.2 → recommend course B (emergency drill, match C2, K2 short board).

[0070] 2. Learning process and real-time feedback The student learns course B, and the real-time feedback module collects data throughout the process: Learning duration Δt = 2 hours, answer data: K2 related question correct rate = 60% (lower than threshold 70%), trigger "abnormal knowledge node" warning. Note the change in mastery: K2 from 0.3 to 0.5 (improvement rate 67%), C2 from 4.2 to 5.0 (improvement 0.8).

[0071] 3. Reward calculation and Q value update Short-term reward (based on feedback data):

[0072] Long-term reward (1 month after assessment): The student passed the assessment (P = 1) → 0.5 x 1 = 0.5.

[0073] Total reward R1 = 0.2 + 0.5 = 0.7.

[0074] Q value iteration (Q-learning formula): Original Q(S0, B) = 3.5, updated Q(S0, B) = 3.5 + 0.1 x (0.7 + 0.9 x maxQ(S1, A') - 3.5) (α = 0.1, γ = 0.9).

[0075] (Note: S1 is the state after learning, maxQ(S1, A') is the optimal course Q value under S1, assumed to be 3.2) Updated Q(S0, B) = 3.5 + 0.1 x (0.7 + 2.88 - 3.5) = 3.5 + 0.008 = 3.508.

[0076] 4. State update and secondary recommendation (real-time feedback driven) The real-time feedback module pushes "K2 answer correct rate 60% < 70%" to the ability assessment module, triggering state update: New state S1 = [C1 = 5.5, C2 = 5.0, C3 = 6.8; K1 = 0.5, K2 = 0.5, K3 = 0.6, K4 = 0.7] The reinforcement learning module recalculates the recommendation strategy based on S1 and the updated Q value: Q(S1, A) = 2.9, Q(S1, B) = 3.6 (because K2 still needs to be strengthened, the Q value is increased), Q(S1, C) = 2.3 → course B is recommended again.

[0077] 5. Secondary learning and feedback adjustment The student learns course B again (the real-time feedback module adjusts the course content: adds K2 practical cases), data as follows: learning duration Δt = 1.5 hours, feedback data: K2 answer correct = 85% (up to standard), C2 from 5.0 to 6.3 (improve 1.3).

[0078] Reward Calculation: Short-term reward = 0.5 × (1.3 / 1.5) = 0.5 × 0.87 ≈ 0.43; Long-term reward (subsequent assessment P = 1) → 0.5 × 1 = 0.5.

[0079] The total reward R2 = 0.43 + 0.5 = 0.93.

[0080] The Q value is iterated again: Q(S1,B)=3.6+0.1×(0.93+0.9×maxQ(S2,A')-3.6)≈3.6+0.05=3.65.

[0081] 6. Results of multiple iterations Table 7. Results of Multiple Iterations indicator initial value value after 3 rounds promotion rate state space S [5.5,4.2,6.8;0.5,0.3,0.6,0.7] [6.2,7.5,6.8;0.6,0.8,0.6,0.7] - recommended number of course B 1 3 because of the continuous strengthening of the short board of K2 post assessment pass rate 60% 85% 41.7%

[0082] (ii) A safety training method based on a dynamic weight optimization model for capability assessment is provided. The method steps of this invention include: (1) constructing a three-dimensional graph model of competence-knowledge-course, wherein the competence dimension is decomposed and weighted based on the job competency model, the knowledge dimension uses knowledge graph technology to construct a knowledge point network, and the course dimension annotates the training resource library with metadata; (2) Based on student behavior data, the subjective and objective weights are dynamically adjusted. The initial weights are determined by establishing a capability-knowledge mapping matrix and a knowledge-course mapping relationship. The subjective and objective weight fusion formula is then used. Perform weight optimization, where The decay rate increases exponentially with the amount of training data. (3) A personalized course recommendation sequence is generated by reinforcement learning algorithm. The state space is composed of ability assessment vector and knowledge mastery vector, the course resource set is the action space, and the recommendation strategy is constructed using short-term learning efficiency and long-term job assessment pass rate as reward functions.

[0083] Example 4: This document outlines a safety training plan for reactor operators at a certain company. The basic requirement is a reactor operator position (involving high-temperature and high-pressure equipment, requiring safe operation and emergency response capabilities). The trainees are 10 newly hired employees (numbered 1-10) from the company. The initial job performance assessment pass rate is 60%, with the core objective of increasing the pass rate to over 85% within three months. The method described below follows specific steps.

[0084] 1. Three-dimensional atlas model a. Competency Dimension (Based on Job Competency Model Decomposition) Table 8 Examples of Capability Dimensions ability dimension (Ci) decomposition index initial weight (quantitative value) C1: safety operation ability standard operation proficiency 5.0 / 10 (medium level) C2: emergency disposal ability accident response speed 4.2 / 10 (low) C3: equipment maintenance ability fault diagnosis accuracy 6.5 / 10 (good) b. Knowledge dimension (Knowledge graph construction) Table 9. Knowledge dimension examples knowledge unit (Kj) associated ability dimension initial mastery (0-1) K1: reactor pressure control specification C1: safety operation ability 0.5 (average) K2: dangerous chemical leakage emergency process C2: emergency disposal ability 0.3 (weak) K3: safety valve calibration standard C3: equipment maintenance ability 0.6 (medium) K4: equipment inspection cycle specification C3: equipment maintenance ability 0.7 (good) c. Course dimension (Metadata annotation) Table 10. Course dimension examples course resources (action A) metadata annotation (cover knowledge unit) duration (hours) A1: 《reaction kettle safety operation specification》 cover K1 (weight 0.9) 2 A2: 《dangerous chemical leakage emergency drill》 Coverage K2 (weight 0.95) 3 A3: "Safety valve verification and maintenance" Coverage K3 (weight 0.85) 2.5 2. Dynamically adjust subjective and objective weights (W final Calculation) (1) Initial weight determination Expert weight (W expert ): 5 senior engineers score (1-5 points) normalized.

[0085] Table 11. Initial weight determination data example Ability-knowledge mapping [WC expert ]]> Knowledge-course mapping [WC expert ]]> C1-K1 0.8 K1-A1 0.9 C2-K2 0.9 K2-A2 0.95 C3-K3 0.7 K3-A3 0.85 Data weight (W data ): Based on historical data of 300 students (correct answer rate, learning duration).

[0086] Table 12. Data weight example Ability-knowledge mapping [WC data ]]> Knowledge-course mapping [WC data ]]> C1-K1 0.75 K1-A1 0.88 C2-K2 0.85 K2-A2 0.92 C3-K3 0.65 K3-A3 0.82 (2) Weight fusion

[0087] Parameter setting: λ = 0.1 (decay coefficient), iteration number t = 5 (initial stage, small data volume); ,1- .

[0088] Examples of calculations: C2-K2 .

[0089] K2-A2 .

[0090] (3) Adjust after data accumulation (t = 30, sufficient data volume): ,C2-K2 (more dependent on data).

[0091] 3. Reinforcement learning to generate recommended sequence a. State space S and action space A.

[0092] Initial state S0 (student 1 data): S0 = [C1 = 5.0, C2 = 4.2, C3 = 6.5; K1 = 0.5, K2 = 0.3, K3 = 0.6].

[0093] Action space A: {A1, A2, A3}.

[0094] b. Reward function R computes (ω1=0.6, ω2=0.4).

[0095] Data after first learning A2: Before learning: C 2old = 4.2, after learning: C 2new = 5.8 (improve 1.6); Learning duration Δt = 3 hours, short-term efficiency = (1.6) / 3 ≈ 0.53; Job assessment pass rate P = 0.6 (initial); R1 = 0.6 × 0.53 + 0.4 × 0.6 = 0.318 + 0.24 = 0.558.

[0096] Data after second learning A2 (triggered again due to low K2 mastery): C 2new = 6.9 (improve 1.1), Δt = 2.5 hours, short-term efficiency = 1.1 / 2.5 = 0.44; P = 0.75 (after first assessment) R2 = 0.6 × 0.44 + 0.4 × 0.75 = 0.264 + 0.3 = 0.564.

[0097] 4. Q value iteration and recommendation strategy optimization Q value update formula: Q(S, A) = Q(S, A) + α[R + γ × maxQ(S', A') - Q(S, A)] (α = 0.1, γ = 0.9).

[0098] Initial Q(S0, A2) = 3.2, updated Q(S0, A2) = 3.2 + 0.1 × (0.558 + 0.9 × 3.5 - 3.2) = 3.2 + 0.071 = 3.271 (maxQ(S', A') is the optimal Q value after learning state).

[0099] Recommendation sequence generation: First round: Q(S0, A2) = 3.271 highest → recommend A2.

[0100] Second round: state S1 = [5.0, 5.8, 6.5; 0.5, 0.6, 0.6], Q(S1, A1) = 3.1, Q(S1, A2) = 3.3 → continue to recommend A2.

[0101] Third round: state S2 = [5.2, 6.9, 6.5; 0.55, 0.8, 0.6], Q(S2, A1) = 3.4 (K1 becomes a new short board) → recommend A1. The implementation results are shown in the following table.

[0102] Table 13 Implementation results Indicator Initial value Final value Promotion rate Mean of ability assessments (C1-C3) 5.23 7.85 50.1% Mean knowledge mastery (K1-K3) 0.47 0.82 74.5% Post assessment pass rate (P) 60% 88% 46.7%

[0103] (III) provide a kind of safety training system based on evaluation dynamic weight optimization model and a computer medium See Fig. 4-Fig. 5 As shown in the figure, a kind of computer medium is proposed in the application based on the above safety training model, method and system, which is actually manifested as a kind of computing platform based on cloud or local deployment, adopts the method of combining big data and cloud computing, has five-layer architecture, including user layer, application layer, data layer, service layer and basic layer.Application layer integrates platform homepage, basic information management, training service, training activities, credit management, system management and other functional modules, and provides comprehensive training services for users.

[0104] The above-described embodiments only express several embodiments of the present application, which are described in detail, but cannot be understood as a limitation on the scope of the application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A safety training model based on a dynamic weight optimization model for capability assessment, characterized in that: The model includes a capability assessment module, a dynamic weight calculation module, a reinforcement learning recommendation module, and a real-time feedback module; The competency assessment module is configured to extract three-dimensional indicators of basic competency, professional competency, and management competency required for the employee's position based on the job competency model, and generate a personalized competency assessment vector by combining the employee's contextual characteristics. The competency assessment vector includes the label level and initial weight of each competency dimension. The dynamic weight calculation module is configured to construct a three-dimensional relationship between "ability-knowledge-course" and perform weight optimization and adjustment. The weight optimization includes establishing an ability-knowledge mapping matrix, a knowledge-course mapping relationship, and a fusion of subjective and objective weights. The reinforcement learning recommendation module uses the ability assessment vector and the knowledge mastery vector to form the state space. The course resources are considered as an action space, with short-term learning efficiency (the ratio of learning duration to mastery) and long-term job performance evaluation pass rate serving as the reward function. Based on the reinforcement learning framework and policy gradient algorithm, a recommendation strategy is constructed, and the course recommendation order is given by dynamically optimizing the Q-value iteration. The real-time feedback module is configured to collect student behavior data and contextual features in real time and trigger a set mechanism. The model uses a three-dimensional graph to structurally link job competencies, professional knowledge systems, and training course resources, forming a dynamically evolving competency development path.

2. The safety training model based on a capability assessment dynamic weight optimization model according to claim 1, characterized in that: The competency assessment module selects structural equation modeling analysis for indicator selection, uses structural equation modeling to verify the safety competency of practitioners, indirectly reflects previously unobservable latent variables, discovers the relationships between latent variables, reflects the relationship between manifest and latent variables, sets evaluation indicators, and establishes a model of the relationship between variables. The structural equation model includes a measurement model and a structural model, specifically: （1） （2） （3） In equations (1)-(3), In the model, it is represented as the first... j One exogenous manifest variable, Represented as the first j One endogenous explicit variable Represented as the first i An exogenous latent variable, Represented as the first i , One endogenous latent variable, Represented as exogenous manifest variables In exogenous latent variables Factor loading coefficient matrix on Λy ij Indicates endogenous manifest variables In endogenous latent variables Factor loading coefficient matrix on Represented as exogenous manifest variables The error; Representing endogenous latent variables and The path coefficient matrix between them; Representing exogenous latent variables Endogenous latent variables The impact; Represents the residual term; Indicates endogenous manifest variables The error.

3. The safety training model based on a capability assessment dynamic weight optimization model according to claim 2, characterized in that: Exogenous explicit variables include exam scores, training attendance rates, simulation operation scores, and fault handling speed; endogenous explicit variables include emergency drill scores, accident handling scores, safety behavior self-assessment, accident incidence rate, and safety inspection scores; endogenous latent variables include emergency response capabilities, safety attitudes, and safety performance; and exogenous latent variables include safety knowledge and operational skills.

4. The safety training model based on a capability assessment dynamic weight optimization model according to claim 1, characterized in that: Specific methods for configuring the "ability-knowledge-curriculum" three-dimensional relationship include: (1) Establish a capability-knowledge mapping matrix and determine the initial association weights between capability dimensions and knowledge units based on expert knowledge graphs and historical training data mining; (2) Establish the knowledge-course mapping relationship, and determine the coverage weight of knowledge units and course resources by analyzing the course outline, marking the knowledge points of test questions, and back-inferring the learning effect; (3) Integrate subjective and objective weights and set up a dynamic adjustment formula. Subjective weight Based on the Analytic Hierarchy Process AHP Alternatively, the Delphi method can be used to determine objective weights. Based on students' answer accuracy, learning time, and simulated operation scores, dynamic calculations are performed using a data association algorithm, with weighting adjustment factors. The decay factor decreases exponentially with increasing training data volume; , t For the number of iterations, This is the attenuation coefficient.

5. The safety training model based on a capability assessment dynamic weight optimization model according to claim 1, characterized in that: In the reinforcement learning recommendation module, the state space... ,in Indicates the first i, m Evaluation values ​​for each capability dimension Indicates the first j,n Mastery of each knowledge unit; reward function ,in , These are the weighting coefficients. These are the assessment values ​​for the ability dimensions before and after learning, respectively. For study time, The pass rate for job performance evaluation.

6. The safety training model based on a capability assessment dynamic weight optimization model according to claim 1, characterized in that: The contextual features include at least one of the following: length of service, job type, position type, job risk level, and historical training records; the reinforcement learning framework includes deep learning. Q network DQN, Policy gradient algorithm PolicyGradient At least one; the behavioral data includes at least one of the following: correct answer rate, learning time, interactive operation score, and number of video replays.

7. The safety training model based on a capability assessment dynamic weight optimization model according to claim 1, characterized in that: The established mechanisms include: (1) Update the importance of knowledge nodes by collecting real-time behavioral data on student operations, assessment scores, and knowledge mastery. Utilize this collected data to... PageRank The algorithm dynamically adjusts the node weights in the knowledge point network, updates the importance of knowledge nodes, establishes a safety capability assessment system for chemical industry practitioners, and establishes a safety training model based on the dynamic weight optimization model of capability assessment. It comprehensively considers factors such as employees' psychological qualities, business and technical capabilities, and safe operation skills, providing a basis for safety assessment, pre-training assessment, and post-training effectiveness verification. (2) Abnormal knowledge node detection: When the standard deviation of the student's answer accuracy rate is >0.3, or the set knowledge points for detection are consecutive N When the error rate in answering questions exceeds the threshold, a weight adjustment and a reconstruction of the course resource association relationship are triggered. (3) The weight adjustment factor is adaptive. The ratio of subjective and objective weight fusion is dynamically adjusted according to the data sample size. In the initial sample size, the expert experience weight is given priority, and the proportion of data-driven weight is gradually increased as the data accumulates.

8. A safety training method based on a dynamic weight optimization model for capability assessment, characterized in that: include (1) Construct a three-dimensional graph model of competence-knowledge-course. In the three-dimensional graph model, the competence dimension is decomposed based on the job competency model and the weight is quantified. The knowledge dimension uses knowledge graph technology to construct a knowledge point network. The course dimension annotates the training resource library with metadata. (2) Dynamically adjust subjective and objective weights based on student behavior data. Determine the initial weights by establishing a capability-knowledge mapping matrix and a knowledge-course mapping relationship, and adopt a subjective and objective weight fusion formula. Perform weight optimization, where The decay rate increases exponentially with the amount of training data. (3) A personalized course recommendation sequence is generated by reinforcement learning algorithm. The state space is composed of ability assessment vector and knowledge mastery vector, the course resource set is the action space, and the recommendation strategy is constructed using short-term learning efficiency and long-term job assessment pass rate as reward functions.

9. A safety training system based on a dynamic weight optimization model for capability assessment, characterized in that: The system operates based on the security training model described in any one of claims 1-7, and includes an initialization phase, an operation phase, and an optimization iteration phase, specifically: (1) Initialization phase: (1.1) Input personalized ability assessment vector into the ability assessment module This includes the level and initial weight of each competency dimension, and outputs a personalized competency assessment vector with the dimension label levels and initial weights generated by the job competency model, as well as the optimized "competency-knowledge-course" mapping weights. ; (1.2) Exogenous and endogenous manifest variables were screened out using structural equation modeling; (1.3) The ability assessment module passes the assessment values ​​of each ability dimension in the ability assessment vector to the reinforcement learning recommendation module. Including the initial mastery level of each knowledge unit （K 1 ,K 2 ,…,K n ） As a reinforcement learning state space The initial value; (1.4) The dynamic weight optimization module sends the initial weights of the constructed "ability-knowledge-course" three-dimensional relationship to the reinforcement learning recommendation module, including the initial weight data of the ability-knowledge mapping matrix and the knowledge-course mapping relationship; (2) Operation phase: (2.1) The reinforcement learning recommendation module calls the interface in real time through the dynamic weight optimization module to request the latest "ability-knowledge-course" three-dimensional relationship weight data for recommendation strategy calculation; (2.2) The dynamic weight optimization module returns the basis through the reinforcement learning recommendation module. Calculated fusion of subjective and objective weights ,in , The attenuation coefficient; (2.3) The reinforcement learning recommendation module pushes a personalized course recommendation sequence generated based on the reinforcement learning framework and policy gradient algorithm to the real-time feedback module. This sequence is then processed by the system. Q Value iteration and dynamic optimization; (2.4) The real-time feedback module transmits student behavior data and characteristics back to the reinforcement learning recommendation module in real time; The real-time feedback module transmits student behavior data and assessment data to the competency assessment module, triggering updates to the competency assessment vector and knowledge mastery vector. The real-time feedback module sends student action data, assessment score data, and knowledge mastery data to the dynamic weight optimization module, triggering updates to the importance of knowledge nodes. PageRank The algorithm adjusts the weights of knowledge point network nodes; when the abnormal knowledge node detection conditions are met, the standard deviation of the answer accuracy is >0.3 or the set detection knowledge points are continuous. N When the error rate in answering questions exceeds the threshold, a weight adjustment and a reconstruction of the course resource association relationship are triggered. (3) Optimization and iteration stage: (3.1) The updated capability assessment vector, including the latest assessment values ​​of each capability dimension and the knowledge mastery vector, is sent to the dynamic weight optimization module to trigger weight optimization and adjustment. (3.2) The dynamic weight optimization module pushes the optimized "ability-knowledge-course" three-dimensional relationship weight data for the reinforcement learning recommendation module to update the recommendation strategy; (3.3) The ability assessment module passes the updated ability assessment vector and knowledge mastery vector to the reinforcement learning recommendation module to update the reinforcement learning state space. S ; (3.4) Real-time feedback module: Provides assessment values ​​of students' abilities before and after learning. Study time and job performance evaluation pass rate P Used to calculate the reward function The reinforcement learning recommendation module is based on this. Q Value iteration optimization recommendation strategy; (3.5) Dynamic weight optimization module: Synchronize the change information of the weight adjustment factor ∝, and the real-time feedback module adjusts the data collection frequency and the threshold parameters of the triggering mechanism accordingly; when it is detected that the student's ability has significantly improved after learning a certain type of course, the change in the ability assessment value exceeds the set threshold γ or there is no significant improvement or it is lower than the threshold. δ At the same time, the real-time feedback module sends course effectiveness evaluation data to assist the dynamic weight optimization module in deeply calibrating the relationship between "ability-knowledge-course".

10. A computer medium, characterized in that, The computer medium stores a program or instructions, which, when executed by a processor, implement the steps of the security training method based on a capability assessment dynamic weight optimization model as described in claim 8.

Images