Teaching evaluation method and system based on multi-agent collaborative evaluation and closed loop

Abstract

The application discloses a teaching evaluation method and system based on multi-agent collaborative evaluation closed loop, and belongs to the technical field of education informatization and artificial intelligence. The method comprises the following steps: collecting multi-source evaluation data of school teachers, enterprise tutors and student self-evaluation and mutual evaluation; constructing a multi-dimensional index system comprising theoretical knowledge, industrial practical ability, team cooperation ability and innovative thinking; adopting a fuzzy comprehensive evaluation method and a D-S evidence theory cascade two-stage fusion algorithm to fuse data, and dynamically adjusting the weights of each evaluation subject; establishing a correlation matrix of knowledge points and evaluation indexes, triggering an optimization suggestion generation engine when the evaluation result is lower than a threshold value, retrieving a course content optimization scheme from a knowledge graph / case library and pushing the scheme; iteratively executing the above steps to drive continuous optimization of course content, practical projects and training modes; and adopting a reinforcement learning algorithm to self-learn and optimize strategies. The application realizes the connection of evaluation and industrial demand and the self-adaptive evolution of the teaching system, and improves the talent training quality.

Description

Technical Field

[0001] This application relates to the fields of educational informatization and artificial intelligence technology, specifically to a teaching evaluation system and method based on a multi-subject collaborative evaluation closed loop, which is particularly suitable for the evaluation and continuous improvement of compound application-oriented talents in applied chemistry, intelligent manufacturing and other majors in higher engineering education. Background Technology

[0002] In the process of developing this application, the inventors discovered at least the following problems in the prior art: With industrial upgrading and economic transformation, society's demand for highly qualified, application-oriented talents is becoming increasingly urgent. Traditional teaching evaluation mainly relies on teachers' subjective observations and final exam scores, which suffers from problems such as a single evaluation subject, one-sided dimensions, and delayed feedback. It is difficult to truly reflect students' comprehensive abilities and cannot drive the optimization of the teaching system in a timely manner.

[0003] In recent years, student behavior analysis systems based on Learning Management System (LMS) data have emerged. For example, Chinese patent document CN120912038A discloses an artificial intelligence-based teaching evaluation method that generates student profiles and dynamically adjusts learning strategies by collecting learning behavior data. However, such systems still primarily rely on internal school data, lacking enterprise perspectives and student self-awareness perspectives. Evaluation dimensions are mostly limited to knowledge mastery, failing to address comprehensive qualities such as industry practice skills and teamwork. Furthermore, feedback optimization remains at the teaching plan level, unable to drive fundamental iterations of course content and training models. Chinese patent document CN121119453A further proposes a knowledge-ability-quality triple-map evaluation method, employing model predictive control to optimize teaching interventions. However, its data source is still mainly internal school records, without incorporating enterprise mentor evaluations and student self-evaluations and peer evaluations. Moreover, the optimization targets are limited to teaching execution levels such as class rescheduling, without addressing iterations of the course content itself.

[0004] Therefore, there is an urgent need for an intelligent teaching evaluation scheme that can integrate evaluations from schools, enterprises, and students, quantify multi-dimensional capabilities such as industry practice, and form an "evaluation-feedback-optimization" closed loop to drive the continuous evolution of the teaching system. Summary of the Invention

[0005] This application aims to address the problems of existing teaching evaluations, such as a single evaluation subject, incomplete dimensions, and shallow feedback optimization. It provides a teaching evaluation system and method based on a multi-subject collaborative evaluation and education closed loop, which can achieve deep integration of industry and education, accurate assessment of abilities, and adaptive iteration of the teaching system.

[0006] To achieve the above objectives, this application provides the following technical solution: a teaching evaluation method based on a multi-subject collaborative evaluation closed loop, comprising the following steps: S1: multi-source data collection; S2: multi-dimensional ability quantification; S3: multi-source heterogeneous data fusion; S4: knowledge point association and optimization suggestion generation; S5: iterative closed loop and training program evolution; S6: reinforcement learning self-optimization.

[0007] This application also provides a teaching evaluation system based on a multi-subject collaborative evaluation closed loop, including at least one processor; and at least one memory, wherein the memory stores computer-executable instructions; when the instructions are executed by the at least one processor, the system enables the following modules to be implemented: a multi-source data acquisition module, a multi-dimensional ability quantification module, a multi-source data fusion module, a knowledge point association and optimization suggestion generation module, an iterative closed-loop control module, and a reinforcement learning self-optimization module.

[0008] During the actual operation of the system, the front-end interactive interface displays the progress of multi-source evaluation data collection, the comprehensive ability score after two-level fusion, the knowledge point-indicator correlation heatmap, and the triggering and push status of optimization suggestions in real time in the form of dashboards, radar charts, push lists, and other icons. This enables teaching administrators, teachers, and corporate mentors to intuitively perceive the dynamic effect of the evaluation closed loop and ensure that the system is "operable, monitorable, and interactive".

[0009] In the method of this application, step S1 collects multi-source evaluation data, including school teacher evaluation data (classroom performance, homework grades, lab reports), corporate mentor evaluation data (automatically collected through the corporate internship platform interface, including project completion, technical compliance, etc.), and student self-evaluation and peer evaluation data (collected via mobile terminal after project completion). Outlier detection and cleaning are performed on the collected data (outliers are removed using the 3σ principle, and missing values ​​are imputed multiple times).

[0010] Step S2 constructs a multi-dimensional evaluation index system, including at least the following indicators: theoretical knowledge mastery, industry practice ability, teamwork ability, and innovative thinking. Industry practice ability is quantified through a pre-defined multi-level indicator tree (first-level indicators include "enterprise project completion rate," "technical specification compliance," and "on-site problem-solving efficiency"; second-level indicators include "project schedule deviation rate," "deliverable quality score," and "customer feedback satisfaction," etc.). Teamwork ability is calculated using social network analysis methods based on interactive data from the collaboration platform (degree centrality, proximity centrality, betweenness centrality, network density, and reciprocity), and the entropy weight method is used to integrate the centrality index with the mutual score to obtain the teamwork ability score.

[0011] Specifically, the calculation method for the technical specification compliance is as follows: Technical Specification Compliance = (Code Specification Score × w1 + Document Completeness × w2 + Process Compliance Rate × w3), where w1, w2, and w3 are preset weights, each defaulting to 1 / 3; the code specification score is obtained automatically using a static code analysis tool to obtain a percentage score; the document completeness score is based on the document chapter coverage rate; and the process compliance rate is calculated based on the enterprise standard process checklist. The calculation method for the on-site problem-solving efficiency is as follows: On-site Problem-Solving Efficiency = 0.4 × (1 - Normalized Average Problem Response Time) + 0.4 × Problem Solving Rate on the First Attempt + 0.2 × Problem Reporting Compliance Score; where the Normalized Average Problem Response Time = (Actual Response Time - Minimum Response Time) / (Maximum Response Time - Minimum Response Time); the problem solving rate on the first attempt = number of problems solved on the first attempt / total number of problems; and the problem reporting compliance score is assessed by the enterprise mentor based on the report format and content completeness. The calculation method for the theoretical knowledge mastery is as follows: Theoretical knowledge mastery = Classroom test score × 0.5 + Homework completion rate score × 0.2 + Classroom interactive question answering accuracy rate × 0.3; where the classroom test score is a standardized test score, the homework completion rate score = (actual submissions / required submissions) × 100, and the classroom interactive question answering accuracy rate is calculated through the classroom response system. The calculation method for the innovative thinking index is as follows: Innovative thinking index = Project scheme novelty evaluation score × 0.5 + Open-ended question answering score × 0.3 + Patent or paper output score × 0.2; the novelty evaluation is based on a comprehensive score from enterprise mentors and school teachers based on the originality and feasibility of the scheme, the open-ended question answering score combines semantic similarity with expert scoring, and the patent or paper output score is assigned according to the application or publication level.

[0012] Step S3 employs a two-level fusion algorithm: The first level utilizes fuzzy comprehensive evaluation to determine the evaluation factor set U = {theoretical knowledge mastery, industry practice ability, teamwork ability, innovative thinking} and the comment set V = {excellent, good, average, poor}. The membership function transforms each indicator value into a fuzzy matrix R. The weights of each evaluation subject are dynamically determined using the Analytic Hierarchy Process (AHP), with the collaboration quality index output from social network analysis serving as a weight correction factor. The fuzzy comprehensive evaluation vector B = W is synthesized using a weighted average M(·,+) operator. R. The second level uses the fuzzy comprehensive evaluation results as input to the basic probability assignment function of the Dempster-Shaffer evidence theory. It then uses the Dempster synthesis rule to fuse multi-source evidence, obtaining the fused confidence interval, and defuzzifies it to obtain the final score. When teacher evaluations conflict with enterprise evaluations, the weight of unreliable evidence is automatically reduced using the conflict coefficient.

[0013] Step S4 establishes a correlation matrix between course knowledge points / skill points and evaluation indicators (through syllabus analysis or historical data correlation calculation). When the comprehensive evaluation result corresponding to a certain knowledge point is lower than a preset threshold (the preset threshold is dynamically set according to historical evaluation data), the optimization suggestion generation engine is triggered to retrieve matching course content optimization schemes (e.g., adding enterprise cases, updating experimental projects) from the knowledge graph / case library and push them to the teacher's end. Specifically, a graph matching algorithm based on the knowledge graph is used to perform subgraph isomorphic matching between the course content knowledge graph and the optimization case graph in the case library, returning the optimization scheme with the highest similarity; or a recommendation algorithm based on collaborative filtering is used to calculate the cosine similarity between the current class feature vector and the historical class feature vectors, recommending optimization schemes adopted by similar classes; when the two recommendation results conflict, they are weighted and fused according to the historical adoption rate.

[0014] Step S5 iteratively executes the above steps, driving continuous optimization of course content, practical projects, and training models (e.g., if a certain ability indicator is low for two consecutive semesters, a revised draft of the training program is automatically generated). Simultaneously, the evaluation results of the next semester after implementing the optimization suggestions are compared with the baseline before optimization. If the improvement is less than expected, the threshold is adjusted and the optimization process is retried. Failure cases are stored in the anomaly database for negative sample training of the reinforcement learning model.

[0015] Step S6 employs a reinforcement learning algorithm to self-learn the closed-loop optimization strategy, aiming to maximize the improvement of students' comprehensive abilities. It dynamically adjusts the trigger threshold and recommendation strategy of the optimization suggestions. The reinforcement learning uses Q-learning. The state space includes course type, current average industry practice ability score, teamwork ability score, and the execution cost of historical optimization solutions. The action space includes four discrete actions and their combinations: "add enterprise cases," "introduce simulation experiments," "adjust project difficulty," and "increase teamwork training." The reward function is defined as the improvement in students' comprehensive ability score in the next semester minus the intervention cost, and a long-term discount factor γ=0.9 is introduced. Actions are selected through an ε-greedy strategy to update the Q-table, and the optimal strategy is learned through multiple iterations.

[0016] One of the above technical solutions has the following advantages or beneficial effects: 1. Multi-party collaborative evaluation for a more comprehensive and objective assessment: For the first time, the evaluation of corporate mentors, student self-evaluation and peer evaluation and school teacher evaluation are deeply integrated. The data of corporate mentors is automatically collected through the interface of the corporate internship platform, and the data of student self-evaluation and peer evaluation is collected through mobile terminals. Anomaly detection and cleaning technology is used to ensure data quality, forming a three-party evaluation community of "school, enterprise and student". This overcomes the one-sidedness of evaluation by a single academic authority and makes the evaluation results closer to the actual industry and students' self-awareness.

[0017] 2. Multidimensional competency quantification, seamless alignment of evaluation with industry needs: A multi-level quantitative indicator system is constructed, encompassing industry practical skills, teamwork abilities, and more. Multi-level indicator trees and social network analysis are used to transform abstract competencies into calculable indicators, aligning talent evaluation standards with industry needs. In particular, indicators such as centrality and network density obtained from social network analysis objectively reflect students' roles and contributions in collaboration, compensating for the subjectivity of traditional peer evaluation.

[0018] 3. Closed-Loop Feedback Drives the Evolution of the Teaching System: By establishing a knowledge point-evaluation indicator correlation matrix and an optimization suggestion generation engine, evaluation results are directly applied to the iteration of course content, practical projects, and training models, forming an adaptive closed loop of "evaluation-feedback-optimization." When the evaluation result is lower than a preset threshold, a matching optimization solution is automatically retrieved from the case library and pushed to the teacher. If the result is lower than the preset threshold for multiple consecutive periods, a revised draft of the training program is automatically generated, enabling the teaching system to continuously evolve based on industry changes and student performance.

[0019] 4. Reinforcement learning self-optimization, continuous system evolution: The reinforcement learning algorithm is introduced to learn the closed-loop optimization strategy. With the goal of maximizing the improvement of students' comprehensive abilities, the trigger threshold and recommendation strategy are dynamically adjusted, so that the system can continuously improve itself based on the historical optimization effect, avoiding the limitations of static thresholds and fixed recommendation strategies.

[0020] 5. Deep Synergy of Technical Means and Unexpected Technical Effects: The aforementioned technical means are not simply superimposed; they exhibit close synergistic effects. The collaborative features output by social network analysis are not only used to quantify team collaboration capabilities but also serve as input for dynamic weight adjustment, making the weight of corporate mentors in the fusion algorithm more reflective of actual collaboration performance. The knowledge point-evaluation index correlation matrix accurately maps evaluation results to knowledge points, providing targeted optimization suggestions. Reinforcement learning adaptively adjusts preset thresholds and recommendation strategies based on historical optimization effects, enabling the system to continuously evolve. Furthermore, the front-end visual demonstration of the operational effects makes the abstract evaluation fusion process and optimization suggestion generation process intuitively presented, enhancing the system's credibility and user experience. This facilitates real-time monitoring and intervention by teaching administrators, truly achieving a workable and perceptible intelligent teaching evaluation closed loop. Ablation experiments show (see detailed implementation methods) that, compared to solutions that remove the fusion algorithm or reinforcement learning, the complete solution of this application increases students' industry practice capabilities by more than 22% and shortens the course iteration cycle by 34.8%, achieving unexpected technical effects. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a teaching evaluation method based on a multi-subject collaborative evaluation closed loop, according to an exemplary embodiment. Figure 2This is a schematic diagram of the functional module structure of a teaching evaluation system based on a multi-subject collaborative evaluation closed loop, according to an exemplary embodiment. Figure 3 This is a schematic diagram illustrating a multi-level indicator tree for industrial practice capabilities according to an exemplary embodiment; Figure 4 This is a schematic diagram of the main interface of the system in this embodiment of the application; Figure 5 yes Figure 4 Schematic diagram of the data acquisition and monitoring area on the left; Figure 6 yes Figure 4 A schematic diagram of the comprehensive evaluation results display area in the central region; Figure 7 yes Figure 4 The right-hand optimization suggestion summary panel Figure 8 This is an icon-based schematic diagram of the interface for pushing comprehensive evaluation results and optimization suggestions in the embodiments of this application; Figure 9 yes Figure 8 A schematic diagram of the correlation matrix between knowledge points and evaluation indicators; Figure 10 yes Figure 8 A schematic diagram of the recommended optimization scheme; Figure 11 This is an evolutionary trend diagram of the iterative closed-loop control and cultivation scheme in the embodiments of this application; Figure 12 yes Figure 11 A schematic diagram of the semester-long variation curve of comprehensive ability dimension; Figure 13 yes Figure 11 A schematic diagram of the Q-value convergence curve in reinforcement learning. Detailed Implementation

[0022] To more clearly illustrate the technical features of this application, the following detailed description, in conjunction with the accompanying drawings, provides specific embodiments and examples to implement different structures of this application. To simplify the disclosure, the components and arrangements of specific examples are described below. Of course, these are merely examples and are not intended to limit the scope of this application.

[0023] Example 1: Method Flow like Figure 1 As shown in the embodiment of this application, a teaching evaluation method based on a multi-subject collaborative evaluation closed loop includes the following steps: S1, Multi-source evaluation data collection: Collect evaluation data from school teachers, corporate mentors, and students' self-evaluation and peer evaluation data, and perform outlier detection and cleaning.

[0024] Specifically, the corporate mentor evaluation data is automatically collected through the corporate internship platform interface, while the student self-evaluation and peer evaluation data are collected via mobile terminals after the project ends. When the corporate mentor evaluation data is automatically collected through the corporate internship platform interface, a timed data collection task is set up. Based on the evaluation cycle, the API is automatically called to obtain project completion scores, technical compliance scores, and text comments. Natural language processing is used to perform sentiment analysis on the comments, extracting keywords as auxiliary evaluation factors. When the student self-evaluation and peer evaluation data are collected via mobile terminals, a scale design based on project response theory is adopted. The questionnaire questions are adaptively adjusted according to the project difficulty, and consistency checks are performed on extreme responses after collection. Data with response times below a preset threshold or with obviously contradictory options are marked as invalid and re-sent.

[0025] The outlier detection and cleaning process specifically includes: identifying outliers using the 3σ principle, marking and removing data that exceed the mean ± 3 standard deviations; identifying outliers using box plots, marking data that are below the first quartile minus 1.5 interquartile ranges or above the third quartile plus 1.5 interquartile ranges as outliers; and filling in the missing values ​​after removal using multiple imputation, where multiple imputation generates multiple imputation datasets based on a chain equation and takes the average as the final imputation value.

[0026] Specifically, the calculation method for the technical specification compliance is as follows: Technical Specification Compliance = (Code Specification Score × w1 + Document Completeness × w2 + Process Compliance Rate × w3), where w1, w2, and w3 are preset weights, each defaulting to 1 / 3; the code specification score is obtained automatically using a static code analysis tool to obtain a percentage score; the document completeness score is based on the document chapter coverage rate; and the process compliance rate is calculated based on the enterprise standard process checklist. The calculation method for the on-site problem-solving efficiency is as follows: On-site Problem-Solving Efficiency = 0.4 × (1 - Normalized Average Problem Response Time) + 0.4 × Problem Solving Rate on the First Attempt + 0.2 × Problem Reporting Compliance Score; where the Normalized Average Problem Response Time = (Actual Response Time - Minimum Response Time) / (Maximum Response Time - Minimum Response Time); the problem solving rate on the first attempt = number of problems solved on the first attempt / total number of problems; and the problem reporting compliance score is assessed by the enterprise mentor based on the report format and content completeness. The calculation method for the theoretical knowledge mastery is as follows: Theoretical knowledge mastery = Classroom test score × 0.5 + Homework completion rate score × 0.2 + Classroom interactive question answering accuracy rate × 0.3; where the classroom test score is a standardized test score, the homework completion rate score = (actual submissions / required submissions) × 100, and the classroom interactive question answering accuracy rate is calculated through the classroom response system. The calculation method for the innovative thinking index is as follows: Innovative thinking index = Project scheme novelty evaluation score × 0.5 + Open-ended question answering score × 0.3 + Patent or paper output score × 0.2; the novelty evaluation is based on a comprehensive score from enterprise mentors and school teachers based on the originality and feasibility of the scheme, the open-ended question answering score combines semantic similarity with expert scoring, and the patent or paper output score is assigned according to the application or publication level.

[0027] The multiple imputation method specifically includes: establishing a prediction model to regress each missing variable, estimating model parameters using non-missing data, generating imputed values ​​based on parameter simulation, repeating the process 5 to 10 times to form multiple complete datasets, and merging the analysis results using Rubin's rule to obtain the final imputed value.

[0028] S2, Multidimensional capability quantification: Construct a multidimensional evaluation index system that includes theoretical knowledge mastery, industrial practice ability, teamwork ability, and innovative thinking; among them, industrial practice ability is quantified through a pre-set multi-level index tree.

[0029] Specifically, the multi-level indicator tree includes at least two-level indicators such as "enterprise project completion rate", "technical specification compliance rate", and "on-site problem-solving efficiency" and their three-level quantitative indicators; team collaboration ability is calculated based on the interaction data of the collaboration platform through social network analysis methods, including calculating the centrality, density and reciprocity of the interaction network of students on the collaboration platform, and weighting these indicators with the mutual score to obtain the team collaboration ability score.

[0030] The aforementioned industry practice capabilities are quantified through a multi-level indicator tree, specifically including: The primary indicator is industry practice capability; the secondary indicators include "enterprise project completion rate", "technical specification compliance rate" and "on-site problem solving efficiency"; among the tertiary indicators, "enterprise project completion rate" includes "project schedule deviation rate", "deliverable quality score" and "customer feedback satisfaction". The "Technical Specifications Compliance" score includes "Code Standards Score", "Document Completeness" and "Process Compliance Rate". The "On-site Problem Solving Efficiency" category includes "Average Problem Response Time", "First-Time Problem Resolution Rate", and "Property Management of Problem Reporting".

[0031] The "Project Schedule Deviation Rate" is calculated using the following formula: Project Schedule Deviation Rate = |Actual Working Hours - Planned Working Hours| / Planned Working Hours × 100%; the "Deliverable Quality Score" is calculated by the enterprise mentor scoring each item on the deliverable checklist and then averaging the scores; the "Customer Feedback Satisfaction" is obtained by analyzing customer comments through natural language processing, and then normalizing the difference between positive and negative sentiment scores using a sentiment dictionary to obtain the satisfaction score.

[0032] The social network analysis method specifically includes: Collect data on students' interaction frequency, collaborative document editing records, task assignment response times, and discussion forum posts on the collaborative platform to construct a directed weighted interaction network; Calculate the degree centrality, proximity centrality, and betweenness centrality of each node, as well as the overall density and reciprocity metrics of the network; After standardizing the centrality index using Z-score, the weight of each centrality index is determined using the entropy weight method, and the weighted sum with the mutual score is used to obtain the team collaboration ability score.

[0033] The entropy weight method for determining weights specifically includes: constructing a centrality index matrix, calculating the information entropy of each index (the smaller the information entropy, the larger the weight), and calculating the weight coefficient using the information entropy, where the information entropy is... The weighting coefficient is ,in, Let be the information entropy value of the j-th indicator, and n be the number of evaluation objects. Let i be the weight of the i-th evaluation object on the j-th indicator. , Let be the value of the j-th centrality metric (such as degree centrality, closeness centrality, etc.) for the i-th student. To Take the natural logarithm. When At that time, it was stipulated To avoid mathematical undefinedness; Let be the weight coefficient of the j-th indicator, which is the centrality index weight ultimately used for weighted aggregation. Let be the difference coefficient (also known as redundancy) of the j-th indicator. The larger the difference coefficient, the more obvious the difference between different students, and the higher the weight should be assigned to the indicator. The number of centrality metrics. The sum of the difference coefficients of all indicators is used for normalization, so that the sum of the weights of each indicator is 1.

[0034] The mastery of theoretical knowledge is calculated by weighting course test scores, homework completion rate, and accuracy rate of classroom interaction answers. The innovative thinking index is comprehensively evaluated by assessing the novelty of project solutions, scores for answering open-ended questions, and the output of patents or papers.

[0035] S3, Multi-source heterogeneous data fusion: A two-level fusion algorithm combining fuzzy comprehensive evaluation method and DS evidence theory is used to fuse the multi-source evaluation data collected in step S1 to generate comprehensive evaluation results for individual students and groups; wherein, the team collaboration ability score obtained from the social network analysis is used as one of the inputs for weight adjustment, and the weights of each evaluation subject are dynamically adjusted according to the course type and industry needs through the analytic hierarchy process or the entropy weight method.

[0036] Specifically, the two-level fusion algorithm includes: The first level of fusion adopts the fuzzy comprehensive evaluation method, establishes the evaluation factor set U={u1,u2,u3,u4} and the comment set V={v1,v2,v3,v4}, transforms each index value into a fuzzy matrix R through the membership function, uses the analytic hierarchy process to determine the initial weight of each evaluation subject, and uses the weighted average type M(·,+) operator to synthesize the fuzzy comprehensive evaluation vector B; The second-level fusion uses the membership degree of each comment in B as the basic probability assignment function of the DS evidence theory. It uses the Dempster synthesis rule to fuse three types of evidence: school teachers, corporate mentors, and student self-evaluation and peer evaluation, to obtain the fused confidence interval. Then, it uses the maximum confidence principle to defuzzify and obtain the comprehensive evaluation score.

[0037] The membership function adopts a trapezoidal distribution. Four threshold parameters are set for positive indicators, and a descending half-trapezoidal distribution is adopted for negative indicators. The membership degree relative to each rating level is calculated based on the actual value of the indicator.

[0038] The Dempster synthesis rule introduces a conflict coefficient k. When the conflict coefficient is greater than a set threshold, an improved weighted allocation strategy is adopted: the conflicting evidence is assigned a smaller weight before synthesis, so as to avoid fusion distortion caused by serious disagreements between the evaluation subjects.

[0039] The weights of each evaluation subject are dynamically adjusted based on course type and industry needs using the analytic hierarchy process (AHP) or entropy weighting method. Specifically, for theoretical courses, the initial weight for school teachers is set at 0.6, for industry mentors at 0.2, and for student self-evaluation and peer evaluation at 0.2; for practical training courses, the initial weight for school teachers is set at 0.3, for industry mentors at 0.5, and for student self-evaluation and peer evaluation at 0.2. Each semester, the weights are recalculated using the AHP based on changes in industry needs, and the collaboration quality index obtained from social network analysis is used as a weight adjustment factor in the entropy weighting method calculation to dynamically optimize the weight allocation.

[0040] The collaboration quality index is calculated by weighting network density and reciprocity indicators obtained through social network analysis. When the collaboration quality index is lower than the historical average, the weight coefficient of the team collaboration dimension in the corporate mentor evaluation is increased in the entropy weight method, forming a dynamic linkage between the weight of the evaluation subject and the team collaboration performance.

[0041] S4, Knowledge Point Association and Optimization Suggestion Generation: Establish an association matrix between course knowledge points / skill points and evaluation indicators. The association matrix is ​​constructed by analyzing the course syllabus or calculating the correlation coefficient between historical evaluation data and the mastery of knowledge points. When the comprehensive evaluation result corresponding to a certain knowledge point is lower than a preset threshold, the optimization suggestion generation engine is triggered. The optimization suggestion generation engine uses a graph matching algorithm based on knowledge graphs and / or a recommendation algorithm based on collaborative filtering to retrieve matching course content optimization schemes from the case library and push them to the teacher's end. Specifically, the establishment of the correlation matrix includes: Extract the list of knowledge points and skill points from the course syllabus, and establish an initial mapping between each knowledge point and multidimensional evaluation indicators; Collect evaluation data from previous semesters, calculate the Pearson correlation coefficient between the score of each knowledge point and the scores of each indicator. If the correlation coefficient is greater than 0.6 and the p-value is less than 0.05, it is determined to be a strong correlation and the correlation matrix is ​​updated. The correlation matrix is ​​dynamically adjusted based on the evaluation data each semester.

[0042] The score for the knowledge point is calculated as follows: a weighted average of the correctness of questions related to the knowledge point in classroom tests, assignments, and lab reports, and adjusted based on the quality of responses to the knowledge point during classroom interactions.

[0043] The optimization suggestion generation engine employs a knowledge graph-based graph matching algorithm, specifically including: Construct a knowledge graph of course content, with nodes representing knowledge points, teaching methods, and case resources, and edges representing relationships; construct a graph of case library optimization solutions, with nodes representing optimization measures, applicable scenarios, and implementation effects. The subgraph isomorphic matching algorithm is used to calculate the similarity between the current course graph and the graphs of each solution in the case library. The top three optimization solutions with the highest similarity are returned and the expected results are marked.

[0044] The optimization suggestion generation engine employs a collaborative filtering-based recommendation algorithm, specifically including: Extract the feature vector of the current class, including average ability score, ability distribution variance, and industry demand matching degree; Calculate the cosine similarity between the feature vector of the current class and the feature vectors of historical classes. Select the top 5 historical classes with the highest similarity, extract their adopted optimization schemes, and recommend them after weighting and sorting them according to the improvement of effect after adoption.

[0045] When the recommendation results output by the knowledge graph-based graph matching algorithm and the collaborative filtering-based recommendation algorithm are inconsistent, a weighted fusion strategy is adopted: the weights are dynamically allocated according to the recommendation adoption rate and effect improvement rate of the two algorithms in historical scenarios, and the top three of the two recommendation lists are merged according to their weights as the final recommendation scheme.

[0046] S5, Iterative Closed Loop and Evolution of Training Program: Iteratively execute the above steps to drive continuous optimization of course content, practical projects and training models; among them, when two or more competency dimension indicators are below the preset threshold for two consecutive semesters, a draft revision of the training program is automatically generated and pushed to the teaching management review end for review.

[0047] Specifically, the continuous optimization of the driving force for course content, practical projects, and training models includes: The evaluation process is automatically triggered at the end of each semester, generating evaluation reports for each ability dimension and knowledge point. If a score in a certain ability dimension is below the preset threshold for two consecutive semesters, a revised draft containing suggestions for adjusting course hours, replacing practical projects, and optimizing teaching methods will be automatically generated and pushed to the teaching management review end for review. Once approved, the course syllabus and teaching plan will be automatically updated, and the revised content will be synchronized to the teaching resource library.

[0048] The preset threshold is dynamically set based on changes in industry demand and historical evaluation data: by statistically analyzing the distribution of scores for each capability dimension over the past three years, the 25th percentile is used as the base threshold, and adjustments are made in conjunction with feedback from industry demand surveys. When the priority of industry demand increases, the threshold for the corresponding capability dimension increases accordingly.

[0049] After driving the continuous optimization of course content, practical projects and training models, it also includes: comparing the evaluation results of the next semester after the implementation of optimization suggestions with the baseline before optimization. If the improvement is lower than the expected improvement threshold, the threshold adjustment mechanism is triggered to lower the trigger threshold for the next round of optimization, and the optimization scheme is marked as an "inefficient scheme" and stored in the negative sample library for reinforcement learning model training.

[0050] S6, Reinforcement Learning Self-Optimization: The reinforcement learning algorithm is used to learn the closed-loop optimization strategy on its own, with the goal of maximizing the improvement of students' comprehensive abilities, and dynamically adjusting the trigger threshold and recommendation strategy of the optimization suggestions.

[0051] Specifically, the reinforcement learning algorithm is based on Q-learning or deep Q-networks, defining the state space as the course configuration parameters and the current ability distribution of students, the action space as optional optimization suggestions, and the reward function as the improvement value of students' comprehensive ability in the next semester minus the intervention cost.

[0052] When the reinforcement learning algorithm is based on Q-learning, the state space includes: course type code, current average industry practice ability score, current average team collaboration ability score, historical optimization scheme execution cost, and student ability distribution dispersion; the action space includes: four discrete actions such as "adding enterprise cases", "introducing simulation experiments", "adjusting project difficulty", and "increasing team collaboration training", as well as pairwise combined actions; the reward function is defined as the improvement value of students' comprehensive ability score in the next semester minus the weighted sum of intervention cost and execution time.

[0053] The Q-learning algorithm employs an ε-greedy strategy for action selection, with an initial exploration rate ε = 0.9, which gradually decreases to 0.1 as the number of iterations increases; the learning rate α = 0.01, and the discount factor γ = 0.95; the Q-table uses state-action pairs as indices and updates the Q-value using the Bellman equation, with the following formula: .

[0054] The formula for calculating the comprehensive ability improvement value in the reward function is: Comprehensive ability improvement value = (Industry practice ability improvement value × weight 1 + Team collaboration ability improvement value × weight 2 + Theoretical knowledge mastery improvement value × weight 3), where the weights are dynamically adjusted according to industry needs; intervention costs include teacher training time, curriculum resource development costs, and implementation cycle converted into class hours.

[0055] The deep Q-network is used to process high-dimensional state spaces. It employs an experience replay mechanism and a target network technique. The experience replay pool size is 10,000, and the target network is updated every 500 steps to eliminate sample correlation and improve learning stability.

[0056] The deep Q-network has a three-layer fully connected neural network structure. The input layer dimension corresponds to the state space dimension, the number of hidden layer neurons are 128 and 64 respectively, the output layer dimension corresponds to the action space size, the activation function is ReLU, the optimizer is Adam, and the loss function is mean squared error.

[0057] The trigger threshold and recommendation strategy for dynamically adjusting optimization suggestions specifically include: the reinforcement learning agent outputs the recommendation strategy and threshold adjustment amount for the next cycle based on historical optimization results; the system updates the trigger threshold based on the adjustment amount; and uses the recommendation strategy as the preference input for the optimization suggestion generation engine to achieve an adaptive closed loop of evaluation-optimization strategy.

[0058] Preferably, the teaching evaluation method based on multi-subject collaborative evaluation closed loop further includes the following steps: constructing a knowledge graph-driven teaching optimization case library, storing the effect feedback after the implementation of optimization suggestions as new cases in the case library, and updating the association weight between optimization measures and effects in the knowledge graph to form a continuously evolving teaching optimization knowledge library.

[0059] This embodiment achieves seamless integration of evaluation with industry needs and adaptive evolution of the teaching system, significantly improving the quality of talent cultivation.

[0060] Example 2: System Architecture This embodiment uses the training of applied chemistry professionals as an example for illustration, but this application is not limited to this profession.

[0061] This application provides a teaching evaluation system based on a multi-subject collaborative evaluation closed loop, comprising: At least one processor; and At least one memory, wherein computer-executable instructions are stored in the memory; When the instruction is executed by the at least one processor, the system performs the following: Figure 2 The module shown: Multi-source data acquisition module: Connects to the school's academic affairs system, enterprise internship management platform, and student mobile APP via API interface to collect teacher evaluation data (classroom performance, homework grades, lab reports), enterprise mentor evaluation data (project progress, technical specification scores, comments), and student self-evaluation / peer evaluation data (collected through questionnaires). The acquisition module has a built-in data cleaning unit, which uses the 3σ principle to identify outliers and uses multiple imputation methods to fill in missing values.

[0062] The multidimensional capability quantification module is used to construct a multidimensional evaluation index system that includes theoretical knowledge mastery, industry practice ability, teamwork ability, and innovative thinking; among them, industry practice ability is quantified through a preset multi-level index tree. Multidimensional capability quantification module: Includes a built-in indicator library that can be dynamically configured. For example... Figure 3 As shown, the industry practice ability is divided into two secondary indicators: "Enterprise Project Completion Rate," "Technical Specification Compliance Rate," and "On-site Problem Solving Efficiency." The "Enterprise Project Completion Rate" is further quantified using three tertiary indicators: "Project Schedule Deviation Rate," "Deliverable Quality Score," and "Customer Feedback Satisfaction." Project Schedule Deviation Rate = (Actual Working Hours - Planned Working Hours) / Planned Working Hours; Deliverable Quality Score is given by enterprise mentors (0-100 points); Customer Feedback Satisfaction is obtained through NLP analysis of customer comments to obtain an emotional score (0-1). Teamwork ability is assessed by analyzing student interaction data on collaborative platforms such as GitLab and Trello, calculating degree centrality, proximity centrality, and betweenness centrality, and then using the entropy weight method to weight the centrality index with the interaction score to obtain the teamwork ability score.

[0063] Multi-source data fusion module: Employs a two-level fusion algorithm. The first level is fuzzy comprehensive evaluation, determining the evaluation factor set U = {theoretical knowledge mastery, industry practice ability, teamwork ability, innovative thinking}, and the comment set V = {excellent, good, average, poor}. The membership function transforms each indicator value into a fuzzy matrix R. The weights of each evaluation subject are dynamically determined using the Analytic Hierarchy Process (AHP): for theoretical courses, teachers have a weight of 0.6, enterprises 0.2, and students 0.2; for practical courses, teachers 0.3, enterprises 0.5, and students 0.2. The collaboration quality index output from social network analysis is used as a weight correction factor in the entropy weight method calculation. The fuzzy comprehensive evaluation vector B = W is synthesized using the weighted average type M(·,+) operator. R. The second level uses the fuzzy comprehensive evaluation results as input to the basic probability allocation function of the DS evidence theory. It then uses the Dempster synthesis rule to fuse multi-source evidence, obtaining the fused confidence interval, and defuzzifies it to obtain the final score. When teacher evaluations conflict with enterprise evaluations, the weight of unreliable evidence is automatically reduced using the conflict coefficient.

[0064] The knowledge point association and optimization suggestion generation module pre-constructs a course knowledge graph, associating knowledge points of the "Chemical Reaction Engineering" course with indicators such as "Process Design Ability." When the average score of a class's "Process Design Ability" falls below a threshold of 0.6, the engine retrieves matching optimization solutions from the case library, such as "adding Aspen simulation training" or "introducing real-world enterprise process cases," and pushes them to the instructor's mobile device. If the score falls below the threshold for two consecutive semesters, a draft revision of the curriculum plan is automatically generated, suggesting adjustments to course hours or the addition of elective courses. The construction of the association matrix is ​​based on calculating the Pearson correlation coefficient between the scores of each knowledge point and each indicator using data from previous semesters. If the correlation coefficient is >0.6, a strong correlation is considered, and the matrix is ​​updated.

[0065] Iterative closed-loop control module: A scheduled task is set up to trigger a new round of evaluation and optimization at the end of each semester, and the implementation status of optimization suggestions is fed back to the system. The evaluation results of the following semester after the optimization suggestions are implemented are compared with the baseline before optimization. If the improvement is lower than expected, the threshold is adjusted and the optimization process is retried. Failed cases are stored in an anomaly database for negative sample training of the reinforcement learning model.

[0066] The reinforcement learning self-optimization module employs the Q-learning algorithm. The state space s∈S includes course type, current average industry practice ability score, teamwork ability score, and historical optimization scheme execution costs. The action space a∈A includes four discrete actions and their combinations: "adding enterprise cases," "introducing simulation experiments," "adjusting project difficulty," and "increasing teamwork training." The reward function R is defined as the increase in students' comprehensive ability score in the next semester minus the intervention cost (e.g., the increase in class hours). Actions are selected using an ε-greedy strategy to update the Q-table. Through multiple iterations, the optimal strategy is learned, and the threshold and recommended strategy are dynamically adjusted.

[0067] This embodiment solves the technical problems of traditional teaching evaluation, such as subjectivity, singularity, and delayed feedback, by executing specific algorithms such as fuzzy comprehensive evaluation, DS evidence theory, knowledge graph matching, and reinforcement learning through a computer system. Furthermore, it adopts technical means such as data collection, fusion processing, and intelligent recommendation to improve the objectivity of evaluation and the evolution efficiency of the teaching system.

[0068] Example 3: Comparative Experimental Data To verify the technical effectiveness of this application, a comparative experiment was conducted in two parallel classes (30 students each) of the same core course ("Chemical Process Design") in the Applied Chemistry major of a university. The experimental group used the system of this application for teaching evaluation and optimization, while the control group used existing technology (based on the method described in Chinese Patent CN121119453A). The experiment lasted for an entire teaching semester, during which multiple rounds of optimization suggestions were implemented and iterated based on the evaluation results. The final data collected at the end of the semester are as follows (see Table 1).

[0069] Table 1 Comparison of teaching effects between the experimental group and the control group

[0070] Furthermore, ablation experiments show that when the fusion algorithm is removed and only the weighted summation is retained, the improvement in industrial application capability decreases to 11.2%; when reinforcement learning is removed and only the static threshold is retained, the course iteration cycle is extended to 58 days. These results fully demonstrate that there is a significant synergistic effect among the various technical features of this application, rather than a simple additive effect, achieving unexpected technical results.

[0071] Example 4: System Operation Results and Visualization To more clearly illustrate the workflow and expected effects of the technical solution of this application in practical applications, the following description is provided in conjunction with an exemplary system operation interface and simulation data. These interface layouts, chart styles, and data curves are merely examples and do not constitute a limitation on the scope of protection of this application.

[0072] This embodiment is combined with the appendix Figures 4 to 13 This section provides a detailed explanation of the interactive interface and data display of the system during actual deployment and operation.

[0073] like Figure 4 As shown in Figure 5, the system's main interface adopts a responsive dashboard layout. Figure 4 The left side is the data acquisition and monitoring area, which displays the real-time data transmission progress from the school teacher's terminal, the enterprise mentor interface, and the student self-evaluation and peer evaluation mobile terminal. A circular progress bar and status indicator lights (green / yellow / red) are used to indicate the effective acquisition ratio of each data source. (See Figure 6.) Figure 4 The central section is the comprehensive evaluation results display area, dynamically presenting the class's average theoretical knowledge mastery, industry practice ability, teamwork ability, and innovative thinking scores in the form of a radar chart. Historical best-to-best value curves are overlaid around the radar chart. (See Figure 7.) Figure 4 The right side displays a summary panel of optimization suggestions, showing the number and type of optimization suggestions currently triggered. This interface also provides "Export Detailed Report" and "Optimization Scheme Push History" buttons, allowing users to drill down to view individual student evaluation details. This main interface intuitively reflects the complete chain of the system's "multi-subject collaborative evaluation and education" process, from data collection to fusion output, demonstrating excellent operability and real-time feedback.

[0074] like Figure 8 As shown, when the comprehensive evaluation result corresponding to a certain knowledge point (such as "chemical process design") is lower than a preset dynamic threshold (e.g., 0.65), the system automatically triggers the optimization suggestion generation engine. A floating prompt box pops up on the interface, as shown below. Figure 9 As shown, the algorithm displays the indicators with high correlation coefficients in the knowledge point-evaluation indicator correlation matrix (e.g., the correlation coefficient for "Enterprise Project Completion" under Industry Practice Capability is 0.73), and uses a knowledge graph-based graph matching algorithm to output a list of recommended optimization solutions. Figure 10 As shown, the icons present three recommended solutions in card format: "Increase the library of real-world enterprise process cases (expected improvement of 15%)", "Introduce Aspen simulation training (expected improvement of 22%)", and "Conduct cross-disciplinary team collaboration workshops (expected improvement of 12%)". Each solution includes an "Adopt" button and a historical adoption rate label. Teachers can push solutions to the course instructors with one click and record the implementation status. Figure 8 and Figure 10The interface also displays a small window for monitoring the conflict coefficient of the fusion algorithm. When the conflict coefficient between teacher evaluations and enterprise evaluations exceeds 0.3, the system automatically lowers the weight of conflicting evidence and outputs a prompt. This interface clearly demonstrates the core triggering and decision-making processes in the "evaluation-feedback-optimization" closed loop.

[0075] like Figure 11 As shown, the system provides a trend dashboard for the continuous evolution of the training program. The horizontal axis represents the semester (6 consecutive semesters), and the vertical axis represents the comprehensive ability score (out of 100). Figure 12 As shown, the graph illustrates the semester-long trends in four dimensions: theoretical knowledge, industry practice, teamwork, and innovative thinking. The time points for each optimization intervention are marked below the curves (e.g., "adding enterprise cases," "revising the draft curriculum plan"). When two or more competency dimensions fall below a preset threshold for two consecutive semesters (dynamically calculated by the system based on historical data, indicated by red dashed lines in the graph), the system automatically generates a draft revision of the curriculum plan (including adjustments to course hours and suggestions for replacing practical projects) and pushes it to the teaching management review panel for approval. Once approved, the system automatically updates the course syllabus and synchronizes it to the teaching resource library. Figure 11 It also includes a small plot of the reinforcement learning Q-value convergence curve, such as... Figure 13 As shown, the convergence of the reward function after multiple iterations is illustrated, demonstrating the effectiveness of the system's self-optimization strategy. This visualization panel allows teaching administrators to intuitively assess the closed-loop optimization effect, forming a virtuous cycle of "evaluation-decision-evolution."

[0076] Actual operational tests show that, after two complete cycles, the adoption rate of optimization suggestions by teachers in classes using the system increased from the initial 67% to 89%, and the improvement rate of industry practice ability was 2.3 times faster than the control group. The average response time of the system front-end is less than 0.8 seconds, and it supports concurrent online data collection and fusion for 500 users, demonstrating high practicality and stability. The above operational results fully verify that the teaching evaluation method and system proposed in this application are not only theoretically sound but also possess excellent operability and visual interactive capabilities, effectively supporting the continuous improvement of large-scale, interdisciplinary, and application-oriented talent training.

[0077] This embodiment demonstrates the evaluation fusion, optimization push, and evolution process of the cultivation scheme through a visual operation interface, making the system have significant effects of being operable, monitorable, and interactive.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation methods of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A teaching evaluation method based on a multi-subject collaborative assessment closed loop, characterized in that, Includes the following steps: Collect school teacher evaluation data, corporate mentor evaluation data, and student self-evaluation and peer evaluation data, and perform outlier detection and cleaning; Construct a multi-dimensional evaluation index system that includes theoretical knowledge mastery, industry practice ability, teamwork ability, and innovative thinking; among them, industry practice ability is quantified through a pre-set multi-level index tree. A two-level fusion algorithm combining fuzzy comprehensive evaluation method and DS evidence theory is used to fuse the collected data and generate comprehensive evaluation results for individual students and groups. Establish a correlation matrix between course knowledge points / skill points and evaluation indicators. The correlation matrix is ​​constructed by analyzing the course syllabus or calculating the correlation coefficient between historical evaluation data and the mastery of knowledge points. When the comprehensive evaluation result corresponding to a certain knowledge point is lower than a preset threshold, an optimization suggestion generation engine is triggered. The optimization suggestion generation engine uses a graph matching algorithm based on knowledge graphs and / or a recommendation algorithm based on collaborative filtering to retrieve matching course content optimization schemes from the case library and push them to the teacher. The preset threshold is dynamically set according to historical evaluation data. The above steps are executed iteratively to drive the continuous optimization of course content, practical projects, and training models. When two or more competency indicators are below the preset threshold for two consecutive semesters, a draft revision of the training program is automatically generated and pushed to the teaching management review end for review. A reinforcement learning algorithm is used to learn the closed-loop optimization strategy, aiming to maximize the improvement of students' comprehensive abilities, and dynamically adjust the trigger threshold and recommendation strategy of the optimization suggestions.

2. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 1, characterized in that, The outlier detection and cleaning specifically includes: Outliers were identified using the 3σ principle, and data exceeding the mean ± 3 standard deviations were marked as outliers and removed. Outliers were identified using box plots, and data that were below the first quartile minus 1.5 times the interquartile range or above the third quartile plus 1.5 times the interquartile range were marked as outliers. The missing values ​​after removal are filled using a multiple imputation method, which generates multiple imputation datasets based on a chain equation and takes the average as the final imputation value.

3. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 1, characterized in that, The multi-level indicator tree includes at least two-level indicators such as "enterprise project completion rate", "technical specification compliance rate", and "on-site problem-solving efficiency" and their three-level quantitative indicators. Team collaboration ability is calculated based on the interaction data of the collaboration platform using social network analysis methods. This includes calculating the centrality, density, and reciprocity indicators of the interaction network of students on the collaboration platform, and weighting these indicators with the mutual score to obtain the team collaboration ability score.

4. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 3, characterized in that, The aforementioned industry practice capabilities are quantified through a multi-level indicator tree, specifically including: The primary indicator is industry practice capability; the secondary indicators include "enterprise project completion rate", "technical specification compliance rate" and "on-site problem-solving efficiency"; among the tertiary indicators, "enterprise project completion rate" includes "project schedule deviation rate", "deliverable quality score" and "customer feedback satisfaction". The "Technical Specifications Compliance" score includes "Code Standards Score", "Document Completeness" and "Process Compliance Rate". "On-site problem-solving efficiency" includes "average problem response time", "problem resolution rate on the first attempt" and "problem reporting standardization".

5. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 4, characterized in that, The team collaboration ability score obtained by the social network analysis method is used as one of the inputs for weight adjustment. The weights of each evaluation subject are dynamically adjusted according to the course type and industry needs using the analytic hierarchy process (AHP) or the entropy weight method. The two-level fusion algorithm specifically includes: The first level of fusion adopts the fuzzy comprehensive evaluation method, establishes the evaluation factor set U={u1,u2,u3,u4} and the comment set V={v1,v2,v3,v4}, transforms each index value into a fuzzy matrix R through the membership function, uses the analytic hierarchy process to determine the initial weight of each evaluation subject, and uses the weighted average type M(·,+) operator to synthesize the fuzzy comprehensive evaluation vector B; The second-level fusion uses the membership degree of each comment in B as the basic probability assignment function of the DS evidence theory. It uses the Dempster synthesis rule to fuse three types of evidence: school teachers, corporate mentors, and student self-evaluation and peer evaluation, to obtain the fused confidence interval. Then, it uses the maximum confidence principle to defuzzify and obtain the comprehensive evaluation score.

6. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 1, characterized in that, The establishment of the correlation matrix specifically includes: Extract the list of knowledge points and skill points from the course syllabus, and establish an initial mapping between each knowledge point and multidimensional evaluation indicators; Collect evaluation data from previous semesters, calculate the Pearson correlation coefficient between the score of each knowledge point and the scores of each indicator. If the correlation coefficient is greater than 0.6 and the p-value is less than 0.05, it is determined to be a strong correlation and the correlation matrix is ​​updated. The correlation matrix is ​​dynamically adjusted based on the evaluation data each semester.

7. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 1, characterized in that, The continuous optimization of the driving force for course content, practical projects, and training models specifically includes: The evaluation process is automatically triggered at the end of each semester, generating evaluation reports for each ability dimension and knowledge point. If a score in a certain ability dimension is below the preset threshold for two consecutive semesters, a revised draft containing suggestions for adjusting course hours, replacing practical projects, and optimizing teaching methods will be automatically generated and pushed to the teaching management review end for review. Once approved, the course syllabus and teaching plan will be automatically updated, and the revised content will be synchronized to the teaching resource library.

8. The teaching evaluation method based on a multi-subject collaborative assessment closed loop according to any one of claims 1-7, characterized in that, The reinforcement learning algorithm is based on Q-learning or deep Q-networks. The state space is defined as the course configuration parameters and the current ability distribution of students, the action space is the optional optimization suggestions, and the reward function is the improvement value of students' comprehensive ability in the next semester minus the intervention cost.

9. The teaching evaluation method based on a multi-subject collaborative assessment closed loop as described in claim 8, characterized in that, When the reinforcement learning algorithm is based on Q-learning, the state space includes: course type code, current average industry practice ability score, current average team collaboration ability score, historical optimization scheme execution cost, and student ability distribution dispersion; the action space includes: four discrete actions such as "add enterprise cases", "introduce simulation experiments", "adjust project difficulty", and "add team collaboration training", as well as pairwise combined actions; the reward function is defined as the improvement value of students' comprehensive ability score in the next semester minus the weighted sum of intervention cost and execution time.

10. A teaching evaluation system based on a multi-subject collaborative assessment closed loop, characterized in that, include: At least one processor; as well as At least one memory, wherein computer-executable instructions are stored in the memory; When the instruction is executed by the at least one processor, the system implements the following modules: The multi-source data acquisition module is used to collect school teacher evaluation data, corporate mentor evaluation data, and student self-evaluation and peer evaluation data, and to perform outlier detection and cleaning. The multidimensional capability quantification module is used to construct a multidimensional evaluation index system that includes theoretical knowledge mastery, industry practice ability, teamwork ability, and innovative thinking; among them, industry practice ability is quantified through a preset multi-level index tree. The multi-source data fusion module is used to fuse the collected data using a two-level fusion algorithm that combines fuzzy comprehensive evaluation method with DS evidence theory, and generate comprehensive evaluation results for individual students and groups. The knowledge point association and optimization suggestion generation module is used to establish an association matrix between course knowledge points / skill points and evaluation indicators. The association matrix is ​​constructed by analyzing the course syllabus or calculating the correlation coefficient between historical evaluation data and the mastery of knowledge points. When the comprehensive evaluation result corresponding to a certain knowledge point is lower than a preset threshold, the optimization suggestion generation engine is triggered. The optimization suggestion generation engine uses a graph matching algorithm based on knowledge graphs and / or a recommendation algorithm based on collaborative filtering to retrieve matching course content optimization schemes from the case library and push them to the teacher's end. The preset threshold is dynamically set according to historical evaluation data. The iterative closed-loop control module is used to drive the continuous optimization of course content, practical projects and training models. When two or more competency indicators are below the preset threshold for two consecutive semesters, a draft revision of the training program is automatically generated and pushed to the teaching management review terminal for review. The reinforcement learning self-optimization module is used to learn closed-loop optimization strategies using reinforcement learning algorithms, with the goal of maximizing the improvement of students' comprehensive abilities, and dynamically adjusting the trigger threshold and recommendation strategy of optimization suggestions.

Images