Artificial intelligence-based course quality dynamic evaluation and teaching research feedback optimization method

Abstract

The application belongs to the technical field of teaching and research, and in particular to a course quality dynamic evaluation and teaching and research feedback optimization method based on artificial intelligence. In view of the problems of lack of consistency processing, and not supporting independent interaction and joint editing, the following scheme is proposed, which comprises a data acquisition module, an AI evaluation module, a teaching and research feedback module, a cloud storage module and a terminal interaction module. The data acquisition module is used for collecting course full-dimension dynamic data, and a standardized data set is generated through multi-source data fusion technology. The full-dimension dynamic data comprises classroom real-time interaction data and stage homework / examination score data, end-to-end automatic generation from voice instructions to entity printing is realized, and efficiency is improved. The multi-IP role and scene semantic coordination, style uniformity are guaranteed. Multi-window touch interaction supports personalized real-time adjustment. Through a fusion algorithm, seams and color differences are eliminated, and high-quality images with consistent style and coordinated illumination are generated.

Description

Technical Field

[0001] This invention relates to the field of teaching and research technology, and in particular to a method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence. Background Technology

[0002] Chinese patent application number 202510965425.4 discloses a method, apparatus, and storage medium for evaluating the quality of online accounting courses. The method includes: collecting target accounting course information and user feedback information; setting subjective user ratings based on the user feedback; collecting accounting principles text in real time; extracting the target course synchronization rate based on the accounting principles text and the course's accounting principles text; setting a course timeliness score based on the synchronization rate; extracting practical characteristics of the target accounting course based on user practical project results; identifying weak chapters in the course based on these practical characteristics; setting a practical score for the course; constructing a quality score for the target accounting course; and obtaining the evaluation result of the target accounting course. This invention effectively improves the evaluation efficiency of online accounting courses.

[0003] However, the online accounting course quality assessment method, device and storage medium also have some problems. For example, it cannot achieve end-to-end automated generation from voice commands to printable cultural IP derivatives, lacks semantic coordination and style consistency processing for multiple IP characters and scenes, and does not support users to independently interact with and jointly edit multiple characters and accessories in multiple display areas. Summary of the Invention

[0004] Given the lack of consistent processing and the absence of support for independent interaction and collaborative editing in the existing technologies, this invention proposes an AI-based method for dynamic evaluation of course quality and optimization of teaching and research feedback.

[0005] The present invention proposes an AI-based method for dynamic evaluation of course quality and optimization of teaching and research feedback, which includes a data acquisition module, an AI evaluation module, a teaching and research feedback module, a cloud storage module, and a terminal interaction module.

[0006] The data acquisition module is used to collect dynamic data from all dimensions of the course and generate a standardized dataset through multi-source data fusion technology. The dynamic data from all dimensions includes real-time classroom interaction data, periodic homework / exam score data, time-series teaching behavior data, dynamic student satisfaction survey data, peer review data, and course resource usage data.

[0007] The AI ​​evaluation module is based on a multi-algorithm fusion framework. It performs dynamic evaluation, feature mining, and problem attribution on standardized datasets, and outputs quantitative evaluation results, weak link labels, and core cause analysis reports.

[0008] The teaching and research feedback module is used to generate personalized optimization plans based on the evaluation results and cause analysis, combined with the teaching and research case library, track the entire process of plan execution and dynamically adjust the feedback strategy.

[0009] The cloud storage module is used for distributed storage of course data, evaluation model parameters, teaching and research case library, feedback execution records and cross-disciplinary adaptation templates, and supports incremental updates and secure retrieval.

[0010] The terminal interaction module includes a data visualization display area, an evaluation result analysis area, a feedback scheme editing area, and a collaborative interaction area. It provides human-computer interaction and multi-person collaboration interfaces, allowing teachers to view data, adjust schemes, and share teaching and research results. It achieves end-to-end automated generation from voice commands to physical printing, improving efficiency; ensures semantic coordination and style consistency among multiple IP roles and scenes; supports personalized real-time adjustments for multi-window touch interaction; and eliminates seams and color differences through fusion algorithms to generate high-quality images with consistent style and coordinated lighting.

[0011] Preferably, the data acquisition module acquires data through three methods: API interface to connect to the teaching management system, real-time collection of classroom environment data by sensors, and input of subjective evaluation data by smart terminals. The data acquisition time interval can be configured from 1 minute to 10 minutes and can be adjusted as needed.

[0012] The data transmission adopts dual encryption of blockchain encryption protocol and HTTPS. Blockchain nodes are used to record data collection, transmission and modification logs to ensure that the data cannot be tampered with. The communication method is one of 5G cellular network, Wi-Fi 6 and wired Ethernet, and the transmission delay is ≤50ms.

[0013] The terminal interaction module can be a web page of the teaching management platform, a smart tablet, a dedicated teaching and research terminal, or a mobile teaching APP. It supports multi-terminal data synchronization and hierarchical permission management. Administrators have the operation permissions to configure different roles. Double encryption ensures the security and immutability of data transmission. Multiple communication methods are adapted to different usage scenarios to meet the requirements of high-speed and stable transmission. Multi-terminal support and hierarchical permission management improve system compatibility and usage flexibility.

[0014] Preferably, the method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence includes the following steps:

[0015] S1: Multi-dimensional data collection and preprocessing: The data collection module acquires raw data from all dimensions of the course, and sequentially performs data desensitization, data cleaning, standardization, outlier handling and data fusion operations to generate a structured dataset and send it to the AI ​​evaluation module;

[0016] S2: Feature Extraction and Vector Encoding: The AI ​​evaluation module performs multi-dimensional feature extraction on the structured dataset, retains key features through feature selection algorithms, and converts them into feature vectors that the model can recognize through vector encoding, thus constructing a feature space for course quality evaluation.

[0017] S3: Dynamic evaluation model training and inference: The AI ​​evaluation module calls the training samples and model parameters in the cloud storage module, performs real-time training and inference based on the multi-algorithm fusion model, and outputs a percentage score of course quality, weak link labels and problem impact factor matrix.

[0018] S4: Problem Attribution Analysis: The AI ​​assessment module uses association rule mining and causal inference algorithms to locate the core causes of weaknesses, generate a problem attribution report that includes a description of the cause, a quantitative value of the degree of impact, and supporting data, and sends it to the teaching and research feedback module.

[0019] S5: Personalized Teaching and Research Feedback Generation: The teaching and research feedback module combines problem attribution reports, cloud-based teaching and research case libraries, and the subject characteristics of the course to generate adaptive teaching and research optimization plans through reinforcement learning recommendation algorithms, and pushes them to the terminal interaction module.

[0020] S6: Feedback Scheme Interaction and Collaborative Adjustment: Teachers can view the optimization scheme through the terminal interaction module, manually adjust the scheme details, support multi-person collaborative editing in the teaching and research team, and the terminal system records all adjustment operations, operators and timestamps, and synchronizes them to the cloud storage module;

[0021] S7: Implementation and Data Tracking: The teaching and research feedback module tracks the implementation progress of the optimization plan in real time, collects course dynamic data after the implementation of the plan at preset time intervals, and sends it to the AI ​​evaluation module for secondary evaluation;

[0022] S8: Evaluation Result Comparison and Feedback Iteration: The AI ​​evaluation module uses statistical analysis methods to compare course quality data before and after the implementation of the plan, and generates an effect evaluation report. The teaching and research feedback module updates the feedback strategy parameters based on the report through reinforcement learning algorithms, forming a closed-loop iterative mechanism of "evaluation-feedback-implementation-re-evaluation". The standardized processing and closed-loop iterative mechanism throughout the entire process enable dynamic tracking and continuous optimization of course quality, ensuring the consistency and effectiveness of evaluation and feedback.

[0023] Preferably, the specific logical steps of S1 are as follows:

[0024] S101: Data anonymization uses the k-anonymity algorithm to process students' personal identity information to ensure that the data complies with legal requirements;

[0025] S102: Data cleaning employs a missing value imputation algorithm and duplicate data removal rules. Missing value imputation uses a weighted interpolation method based on K-nearest neighbors (KNN), with the weight calculation formula as follows: ,in This represents the Euclidean distance between data points. As a smoothing coefficient, duplicate data are identified through a combination of hash verification and content similarity comparison.

[0026] S103: Data standardization uses the Z-score standardization formula: ,in The original data, The mean of the data. Let be the standard deviation of the data. After standardization, the mean of the data is 0 and the variance is 1.

[0027] S104: Outlier handling uses the IsolationForest algorithm, with a set outlier detection threshold. When data points have abnormal scores When an outlier is detected, continuous data is corrected and discrete data is removed.

[0028] S105: Data fusion adopts a weighted average fusion algorithm, and the fusion formula is as follows: ,in For data from the k-th type of data source, The credibility weight of the k-th class of data. ;

[0029] S106: The processed dataset is packaged into JSON format, with timestamps, data source tags, and data credibility scores added. It is then sent to the AI ​​evaluation module through a secure channel. Precise data cleaning and outlier handling algorithms improve the quality of the dataset, providing reliable data support for subsequent evaluations and reducing noise interference with the results.

[0030] Preferably, the specific logical steps of S2 are as follows:

[0031] S201: A multi-layer feature extraction strategy is adopted. Numerical features are extracted by statistical calculation to extract distribution features. Textual features are extracted by TF-IDF algorithm and BERT model to extract semantic features. Behavioral features are extracted by time series analysis to extract trend and periodic features. Video features are extracted by YOLO object detection algorithm to extract features such as teacher body language and student concentration.

[0032] S202: Feature selection is performed using ANOVA and mutual information calculations, retaining features that are more correlated with course quality than a threshold. Eliminate redundant features based on their characteristics.

[0033] S203: The text feature vector is optimized using a BERT-based semantic encoding function. The encoding formula is as follows: ,in For text data, This is the encoded 768-dimensional high-dimensional semantic feature vector;

[0034] S204: Principal Component Analysis (PCA) is used to reduce the dimensionality of the high-dimensional eigenvectors, retaining principal components with a cumulative variance contribution rate ≥90%, and generating a course quality assessment eigenvector with a unified dimension. Multi-layer feature extraction and dimensionality reduction optimization balance feature integrity and model computational efficiency. BERT semantic encoding enhances text feature representation capabilities and improves evaluation accuracy.

[0035] Preferably, the multi-algorithm fusion model of S3 includes a Transformer time series prediction sub-model, a random forest evaluation sub-model, and a gradient boosting tree XGBoost optimization sub-model. The division of labor among the sub-models is as follows: the Transformer sub-model processes the teaching time series data and outputs the trend evaluation value; the random forest sub-model processes the classification feature data and outputs the quality level label; the XGBoost sub-model performs weighted optimization on the output results of the former two.

[0036] During model training, an adaptive weight allocation algorithm is used to dynamically adjust the weights of each sub-model. The weight calculation formula is as follows: ,in For the first Cross-validation accuracy of individual sub-models For the first The generalization ability score of each sub-model; the number of iterations for model training is set to 100 to 500, the convergence condition is that the loss function value is ≤0.001, and the loss function adopts mean squared error (MSE); the final evaluation result is output through weighted voting during the inference stage, the course quality score adopts a percentage system of 0-100 points, the weak link label is generated based on the preset classification system, the problem influence factor is calculated by the analytic hierarchy process (AHP), and the value range is [0,1]. Multiple algorithms are integrated and adaptive weight allocation is used to adapt to different types of course data, improve the generalization ability and result credibility of the evaluation model, and the quantitative indicators make the evaluation more intuitive.

[0037] Preferably, the specific logical steps of S4 are as follows:

[0038] S401: The Apriori association rule mining algorithm is used to mine the association between weak links and potential influencing factors, with the minimum support set to... The minimum confidence level is set to Output the top-5 strong association rules;

[0039] S402: By combining Bayesian networks with the Do-calculus causal inference framework, a causal graph of "influencing factors-weak links" is constructed, the average causal effect ACE value of each factor is calculated, and the strength of causal association is quantified.

[0040] S403: Based on the causal effect value and the confidence of the association rule, a weighted ranking is performed with weights set to 0.6 and 0.4 respectively, and the top 3 core causes are selected;

[0041] S404: Generates a problem attribution report that includes a description of the cause, a quantitative value of the degree of impact, supporting data sources, and improvement priorities. Improvement priorities are divided into three levels: high, medium, and low. By combining association rule mining and causal inference, the core causes are accurately located, and the graded rectification priorities clarify the direction of improvement, thereby improving the efficiency of teaching and research rectification.

[0042] Preferably, the specific logical steps of S5 are as follows:

[0043] S501: The teaching and research feedback module retrieves teaching and research case libraries from the cloud using a cosine similarity algorithm. The similarity calculation formula is as follows: Match successful cases with a similarity ≥ 0.7 to the current problem attribution report;

[0044] S502: The Deep Reinforcement Learning Recommendation Algorithm (DQN) is used to generate personalized optimization schemes. The state space of the agent consists of the course quality feature vector and the problem cause vector, the action space consists of the set of teaching and research improvement measures, and the reward function is the quality improvement rate after the scheme is implemented.

[0045] S503: The plan includes suggestions for adjusting teaching methods, a list of supplementary resources, a teaching and research activity plan, effectiveness verification indicators, and an implementation timetable. The effectiveness verification indicators include a quantitative indicator of a quality score improvement of ≥10 points and a qualitative indicator of a student satisfaction improvement of ≥15%.

[0046] S504: Perform a feasibility assessment on the optimization scheme. The assessment formula is as follows: ,in The suitability of the solution is rated on a scale of 0-100. Resource availability is scored from 0 to 100. Execution cost is scored from 0 to 100. These are the weighting coefficients;

[0047] S505: When The solution is pushed directly when the score is 60 or less than 80. When the score is less than 60, the solution parameters are adjusted by a genetic algorithm and then pushed. When the score is less than 60, cases are retrieved again and solutions are generated. Case matching and reinforcement learning recommendation are combined to generate personalized solutions with strong adaptability and high feasibility. A multi-dimensional scoring mechanism ensures the implementation of the solutions.

[0048] Preferably, the specific logical steps of S8 are as follows:

[0049] S801: The AI ​​assessment module uses a paired-samples t-test to compare course quality scores before and after the implementation of the scheme and calculate the improvement. And the significance of the difference was assessed using the effect size Cohen's d. The time was judged as a significant improvement;

[0050] S802: Collect teachers' evaluations of the feedback plan using a 5-point Likert scale, combined with the improvement rate. Generate a feedback score for the effectiveness of the proposed solution. The scoring formula is as follows: The average score of teacher evaluations;

[0051] S803: The teaching and research feedback module updates the reward function parameters of the reinforcement learning algorithm based on the effect score. When EffectScore ≥ 0.8, the reward value of the corresponding action is increased; when EffectScore ≤ 0.5 < 0.8, the reward value remains unchanged; when EffectScore < 0.5, the reward value of the corresponding action is decreased.

[0052] S804: Synchronize the updated model parameters to the cloud storage module, and record the iteration log, including iteration time, parameter changes and effect improvement. The iteration cycle is consistent with the course cycle. Statistical verification and quantitative evaluation of effects are carried out to objectively measure the effectiveness of feedback, strengthen the learning iteration optimization recommendation strategy, and continuously improve the pertinence and effectiveness of teaching and research feedback.

[0053] Preferably, the system supports cross-disciplinary course quality comparison analysis, and adjusts the evaluation model parameters through a domain adaptation algorithm. The domain adaptation algorithm adopts the Domain Adaptive Network (DANN) in transfer learning, and minimizes the cross-disciplinary data distribution differences through a gradient inversion layer.

[0054] The cloud storage module regularly updates the teaching and research case library using an incremental learning algorithm. The update cycle is one month. Cases must undergo three levels of review before being added to the library: AI effect verification, expert review, and practical verification.

[0055] The teaching and research feedback module has a built-in course quality early warning mechanism. When the course quality score is below 60 points for two consecutive evaluation cycles or the single drop is ≥20 points, a first-level early warning is automatically triggered, an emergency rectification plan is pushed and the teaching and research management department is notified. When the score is between 60 and 70 points, a second-level early warning is triggered, a regular optimization plan is pushed, cross-disciplinary adaptation algorithms expand the system's applicability, triple review ensures the quality of the case library, and the hierarchical early warning mechanism realizes risk prevention and control in advance, helping to rapidly improve teaching quality.

[0056] The beneficial effects of this invention are:

[0057] It achieves end-to-end automated generation from voice commands to physical printing, improving efficiency; ensures semantic coordination and style consistency among multiple IP roles and scenes; supports personalized real-time adjustments through multi-window touch interaction; and eliminates seams and color differences through fusion algorithms to generate high-quality images with consistent style and coordinated lighting. Attached Figure Description

[0058] Figure 1 This is a system block diagram proposed in this invention;

[0059] Figure 2 This is the timing diagram proposed in this invention. Detailed Implementation

[0060] The present invention will be further explained below with reference to specific embodiments.

[0061] Reference Figure 1-2 , Example

[0062] This embodiment proposes a method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence, including a data acquisition module, an AI evaluation module, a teaching and research feedback module, a cloud storage module, and a terminal interaction module;

[0063] The data acquisition module is used to collect dynamic data from all dimensions of the course and generate standardized datasets through multi-source data fusion technology. The dynamic data from all dimensions includes real-time classroom interaction data, periodic assignment / exam score data, time-series data of teaching behavior, dynamic survey data of student satisfaction, peer review data, and course resource usage data.

[0064] The AI ​​assessment module is based on a multi-algorithm fusion framework. It performs dynamic assessment, feature mining, and problem attribution on standardized datasets, and outputs quantitative assessment results, weak link labels, and core cause analysis reports.

[0065] The teaching and research feedback module is used to generate personalized optimization plans based on the evaluation results and cause analysis, combined with the teaching and research case library, to track the entire process of plan implementation and dynamically adjust the feedback strategy.

[0066] The cloud storage module is used for distributed storage of course data, evaluation model parameters, teaching and research case library, feedback execution records and cross-disciplinary adaptation templates, and supports incremental updates and secure retrieval;

[0067] The terminal interaction module includes a data visualization display area, an evaluation result analysis area, a feedback plan editing area, and a collaborative interaction area. It provides human-computer interaction and multi-person collaboration interfaces, supporting teachers to view data, adjust plans, and share teaching and research results.

[0068] The data acquisition module obtains data through three methods: API interface to connect to the teaching management system, real-time collection of classroom environment data by sensors, and input of subjective evaluation data by smart terminals. The data acquisition time interval can be configured from 1 minute to 10 minutes and can be adjusted as needed.

[0069] Data transmission employs dual encryption using blockchain encryption protocol and HTTPS. Blockchain nodes are used to record data collection, transmission, and modification logs to ensure that the data is tamper-proof. The communication method is one of 5G cellular network, Wi-Fi 6, and wired Ethernet, with a transmission latency of ≤50ms.

[0070] The terminal interaction module can be a web page of the teaching management platform, a smart tablet, a dedicated teaching and research terminal, or a mobile teaching APP. It supports multi-terminal data synchronization and hierarchical permission management, and the administrator has the operation permissions to configure different roles.

[0071] Specifically, the following steps are included:

[0072] S1: Multi-dimensional data collection and preprocessing: The data collection module acquires raw data from all dimensions of the course, and sequentially performs data desensitization, data cleaning, standardization, outlier handling and data fusion operations to generate a structured dataset and send it to the AI ​​evaluation module;

[0073] The specific logical steps of S1 are as follows:

[0074] S101: Data anonymization uses the k-anonymity algorithm to process students' personal identity information to ensure that the data complies with legal requirements;

[0075] S102: Data cleaning employs a missing value imputation algorithm and duplicate data removal rules. Missing value imputation uses a weighted interpolation method based on K-nearest neighbors (KNN), with the weight calculation formula as follows: ,in This represents the Euclidean distance between data points. As a smoothing coefficient, duplicate data are identified through a combination of hash verification and content similarity comparison.

[0076] S103: Data standardization uses the Z-score standardization formula: ,in The original data, The mean of the data. Let be the standard deviation of the data. After standardization, the mean of the data is 0 and the variance is 1.

[0077] S104: Outlier handling uses the IsolationForest algorithm, with a set outlier detection threshold. When data points have abnormal scores When an outlier is detected, continuous data is corrected and discrete data is removed.

[0078] S105: Data fusion adopts a weighted average fusion algorithm, and the fusion formula is as follows: ,in For data from the k-th type of data source, The credibility weight of the k-th class of data. ;

[0079] S106: Encapsulate the processed dataset into JSON format, add timestamps, data source tags, and data credibility scores, and send it to the AI ​​evaluation module through a secure channel;

[0080] S2: Feature Extraction and Vector Encoding: The AI ​​evaluation module performs multi-dimensional feature extraction on the structured dataset, retains key features through feature selection algorithms, and converts them into feature vectors that the model can recognize through vector encoding, thus constructing a feature space for course quality evaluation.

[0081] The specific logical steps of S2 are as follows:

[0082] S201: A multi-layer feature extraction strategy is adopted. Numerical features are extracted by statistical calculation to extract distribution features. Textual features are extracted by TF-IDF algorithm and BERT model to extract semantic features. Behavioral features are extracted by time series analysis to extract trend and periodic features. Video features are extracted by YOLO object detection algorithm to extract features such as teacher body language and student concentration.

[0083] S202: Feature selection is performed using ANOVA and mutual information calculations, retaining features that are more correlated with course quality than a threshold. Eliminate redundant features based on their characteristics.

[0084] S203: The text feature vector is optimized using a BERT-based semantic encoding function. The encoding formula is as follows: ,in For text data, This is the encoded 768-dimensional high-dimensional semantic feature vector;

[0085] S204: Principal Component Analysis (PCA) is used to reduce the dimensionality of the high-dimensional eigenvectors, retaining principal components with a cumulative variance contribution rate ≥90%, and generating a course quality assessment eigenvector with a unified dimension. ;

[0086] S3: Dynamic evaluation model training and inference: The AI ​​evaluation module calls the training samples and model parameters in the cloud storage module, performs real-time training and inference based on the multi-algorithm fusion model, and outputs a percentage score of course quality, weak link labels and problem impact factor matrix.

[0087] The S3 multi-algorithm fusion model includes a Transformer time series prediction sub-model, a Random Forest evaluation sub-model, and a Gradient Boosting Tree XGBoost optimization sub-model. The division of labor among the sub-models is as follows: the Transformer sub-model processes the teaching time series data and outputs trend evaluation values; the Random Forest sub-model processes the classification feature data and outputs quality level labels; the XGBoost sub-model performs weighted optimization on the outputs of the former two.

[0088] During model training, an adaptive weight allocation algorithm is used to dynamically adjust the weights of each sub-model. The weight calculation formula is as follows: ,in For the first Cross-validation accuracy of individual sub-models For the first The generalization ability score of each sub-model; the number of iterations for model training is set to 100 to 500, the convergence condition is that the loss function value is ≤0.001, and the loss function adopts mean squared error (MSE); the final evaluation result is output through weighted voting during the inference stage, the course quality score adopts a percentage system of 0-100 points, the weak link label is generated based on the preset classification system, and the problem impact factor is calculated by the analytic hierarchy process (AHP), with a value range of [0,1].

[0089] S4: Problem Attribution Analysis: The AI ​​assessment module uses association rule mining and causal inference algorithms to locate the core causes of weaknesses, generate a problem attribution report that includes a description of the cause, a quantitative value of the degree of impact, and supporting data, and sends it to the teaching and research feedback module.

[0090] The specific logical steps of S4 are as follows:

[0091] S401: The Apriori association rule mining algorithm is used to mine the association between weak links and potential influencing factors, with the minimum support set to... The minimum confidence level is set to Output the top-5 strong association rules;

[0092] S402: By combining Bayesian networks with the Do-calculus causal inference framework, a causal graph of "influencing factors-weak links" is constructed, the average causal effect ACE value of each factor is calculated, and the strength of causal association is quantified.

[0093] S403: Based on the causal effect value and the confidence of the association rule, a weighted ranking is performed with weights set to 0.6 and 0.4 respectively, and the top 3 core causes are selected;

[0094] S404: Generate a problem attribution report that includes a description of the cause, a quantitative value of the degree of impact, supporting data sources, and improvement priorities. Improvement priorities are divided into three levels: high, medium, and low.

[0095] S5: Personalized Teaching and Research Feedback Generation: The teaching and research feedback module combines problem attribution reports, cloud-based teaching and research case libraries, and the subject characteristics of the course to generate adaptive teaching and research optimization plans through reinforcement learning recommendation algorithms, and pushes them to the terminal interaction module.

[0096] The specific logical steps of S5 are as follows:

[0097] S501: The teaching and research feedback module retrieves teaching and research case libraries from the cloud using a cosine similarity algorithm. The similarity calculation formula is as follows: Match successful cases with a similarity ≥ 0.7 to the current problem attribution report;

[0098] S502: The Deep Reinforcement Learning Recommendation Algorithm (DQN) is used to generate personalized optimization schemes. The state space of the agent consists of the course quality feature vector and the problem cause vector, the action space consists of the set of teaching and research improvement measures, and the reward function is the quality improvement rate after the scheme is implemented.

[0099] S503: The plan includes suggestions for adjusting teaching methods, a list of supplementary resources, a teaching and research activity plan, effectiveness verification indicators, and an implementation timetable. The effectiveness verification indicators include a quantitative indicator of a quality score improvement of ≥10 points and a qualitative indicator of a student satisfaction improvement of ≥15%.

[0100] S504: Perform a feasibility assessment on the optimization scheme. The assessment formula is as follows: ,in The suitability of the solution is rated on a scale of 0-100. Resource availability is scored from 0 to 100. Execution cost is scored from 0 to 100. These are the weighting coefficients;

[0101] S505: When When the score is 60 ≤ Score < 80, the solution parameters are adjusted using a genetic algorithm before being pushed; when the score < 60, cases are retrieved again and a solution is generated.

[0102] S6: Feedback Scheme Interaction and Collaborative Adjustment: Teachers can view the optimization scheme through the terminal interaction module, manually adjust the scheme details, support multi-person collaborative editing in the teaching and research team, and the terminal system records all adjustment operations, operators and timestamps, and synchronizes them to the cloud storage module;

[0103] S7: Implementation and Data Tracking: The teaching and research feedback module tracks the implementation progress of the optimization plan in real time, collects course dynamic data after the implementation of the plan at preset time intervals, and sends it to the AI ​​evaluation module for secondary evaluation;

[0104] S8: Evaluation Result Comparison and Feedback Iteration: The AI ​​evaluation module uses statistical analysis methods to compare course quality data before and after the implementation of the plan, and generates an effect evaluation report. The teaching and research feedback module updates the feedback strategy parameters based on the report through reinforcement learning algorithms, forming a closed-loop iterative mechanism of "evaluation-feedback-implementation-re-evaluation".

[0105] The specific logical steps of S8 are as follows:

[0106] S801: The AI ​​assessment module uses a paired-samples t-test to compare course quality scores before and after the implementation of the scheme and calculate the improvement. And the significance of the difference was assessed using the effect size Cohen's d. The time was judged as a significant improvement;

[0107] S802: Collect teachers' evaluations of the feedback plan using a 5-point Likert scale, ranging from 1 - very dissatisfied to 5 - very satisfied, and combine this with the improvement rate. Generate a feedback score for the effectiveness of the proposed solution. The scoring formula is as follows: The average score of teacher evaluations;

[0108] S803: The teaching and research feedback module updates the reward function parameters of the reinforcement learning algorithm based on the effect score. When EffectScore ≥ 0.8, the reward value of the corresponding action is increased; when EffectScore ≤ 0.5 < 0.8, the reward value remains unchanged; when EffectScore < 0.5, the reward value of the corresponding action is decreased.

[0109] S804: Synchronize the updated model parameters to the cloud storage module, and record the iteration log, including iteration time, parameter changes and performance improvement. The iteration cycle is consistent with the course cycle.

[0110] The system supports cross-disciplinary course quality comparison and analysis. It adjusts the evaluation model parameters through a domain adaptation algorithm, which adopts the Domain Adaptive Network (DANN) in transfer learning and minimizes the differences in cross-disciplinary data distribution through a gradient inversion layer.

[0111] The cloud storage module regularly updates the teaching and research case library using an incremental learning algorithm, with an update cycle of one month. Cases entering the library must undergo three rounds of review: AI effect verification, expert review, and practical verification.

[0112] The teaching and research feedback module has a built-in course quality early warning mechanism. When the course quality score is below 60 points for two consecutive evaluation cycles or the single drop is ≥20 points, a first-level early warning is automatically triggered, an emergency rectification plan is pushed and the teaching and research management department is notified; when the score is between 60 and 70 points, a second-level early warning is triggered and a regular optimization plan is pushed.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence, characterized in that, It includes a data acquisition module, an AI evaluation module, a teaching and research feedback module, a cloud storage module, and a terminal interaction module; The data acquisition module is used to collect dynamic data from all dimensions of the course and generate a standardized dataset through multi-source data fusion technology. The dynamic data from all dimensions includes real-time classroom interaction data, periodic homework / exam score data, time-series teaching behavior data, dynamic student satisfaction survey data, peer review data, and course resource usage data. The AI ​​evaluation module is based on a multi-algorithm fusion framework. It performs dynamic evaluation, feature mining, and problem attribution on standardized datasets, and outputs quantitative evaluation results, weak link labels, and core cause analysis reports. The teaching and research feedback module is used to generate personalized optimization plans based on the evaluation results and cause analysis, combined with the teaching and research case library, track the entire process of plan execution and dynamically adjust the feedback strategy. The cloud storage module is used for distributed storage of course data, evaluation model parameters, teaching and research case library, feedback execution records and cross-disciplinary adaptation templates, and supports incremental updates and secure retrieval. The terminal interaction module includes a data visualization display area, an evaluation result analysis area, a feedback scheme editing area, and a collaborative interaction area, providing human-computer interaction and multi-person collaboration interfaces to support teachers in viewing data, adjusting schemes, and sharing teaching and research results.

2. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 1, characterized in that, The data acquisition module obtains data through three methods: API interface to connect to the teaching management system, real-time collection of classroom environment data by sensors, and input of subjective evaluation data by smart terminals. The data acquisition time interval can be configured from 1 minute to 10 minutes and can be adjusted as needed. The data transmission adopts dual encryption of blockchain encryption protocol and HTTPS. Blockchain nodes are used to record data collection, transmission and modification logs to ensure that the data cannot be tampered with. The communication method is one of 5G cellular network, Wi-Fi 6 and wired Ethernet, and the transmission delay is ≤50ms. The terminal interaction module can be a web page of the teaching management platform, a smart tablet, a dedicated teaching and research terminal, or a mobile teaching APP. It supports multi-terminal data synchronization and hierarchical permission management, and the administrator has the operation permissions to configure different roles.

3. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 1, characterized in that, Includes the following steps: S1: Multi-dimensional data collection and preprocessing: The data collection module acquires raw data from all dimensions of the course, and sequentially performs data desensitization, data cleaning, standardization, outlier handling and data fusion operations to generate a structured dataset and send it to the AI ​​evaluation module; S2: Feature Extraction and Vector Encoding: The AI ​​evaluation module performs multi-dimensional feature extraction on the structured dataset, retains key features through feature selection algorithms, and converts them into feature vectors that the model can recognize through vector encoding, thus constructing a feature space for course quality evaluation. S3: Dynamic evaluation model training and inference: The AI ​​evaluation module calls the training samples and model parameters in the cloud storage module, performs real-time training and inference based on the multi-algorithm fusion model, and outputs a percentage score of course quality, weak link labels and problem impact factor matrix. S4: Problem Attribution Analysis: The AI ​​assessment module uses association rule mining and causal inference algorithms to locate the core causes of weaknesses, generate a problem attribution report that includes a description of the cause, a quantitative value of the degree of impact, and supporting data, and sends it to the teaching and research feedback module. S5: Personalized Teaching and Research Feedback Generation: The teaching and research feedback module combines problem attribution reports, cloud-based teaching and research case libraries, and the subject characteristics of the course to generate adaptive teaching and research optimization plans through reinforcement learning recommendation algorithms, and pushes them to the terminal interaction module. S6: Feedback Scheme Interaction and Collaborative Adjustment: Teachers can view the optimization scheme through the terminal interaction module, manually adjust the scheme details, support multi-person collaborative editing in the teaching and research team, and the terminal system records all adjustment operations, operators and timestamps, and synchronizes them to the cloud storage module; S7: Implementation and Data Tracking: The teaching and research feedback module tracks the implementation progress of the optimization plan in real time, collects course dynamic data after the implementation of the plan at preset time intervals, and sends it to the AI ​​evaluation module for secondary evaluation; S8: Evaluation Result Comparison and Feedback Iteration: The AI ​​evaluation module uses statistical analysis methods to compare course quality data before and after the implementation of the plan, and generates an effect evaluation report. The teaching and research feedback module updates the feedback strategy parameters based on the report through reinforcement learning algorithms, forming a closed-loop iterative mechanism of "evaluation-feedback-implementation-re-evaluation".

4. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 3, characterized in that, The specific logical steps of S1 are as follows: S101: Data anonymization uses the k-anonymity algorithm to process students' personal identity information to ensure that the data complies with legal requirements; S102: Data cleaning employs a missing value imputation algorithm and duplicate data removal rules. Missing value imputation uses a weighted interpolation method based on K-nearest neighbors (KNN), with the weight calculation formula as follows: ,in This represents the Euclidean distance between data points. As a smoothing coefficient, duplicate data are identified through a combination of hash verification and content similarity comparison. S103: Data standardization uses the Z-score standardization formula: ,in The original data, The mean of the data. Let be the standard deviation of the data. After standardization, the mean of the data is 0 and the variance is 1. S104: Outlier handling uses the IsolationForest algorithm, with a set outlier detection threshold. When data points have abnormal scores When an outlier is detected, continuous data is corrected and discrete data is removed. S105: Data fusion adopts a weighted average fusion algorithm, and the fusion formula is as follows: ,in For data from the k-th type of data source, Let k be the confidence weight of the k-th class of data. ; S106: Encapsulate the processed dataset into JSON format, add timestamps, data source tags, and data credibility scores, and send it to the AI ​​evaluation module through a secure channel.

5. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 3, characterized in that, The specific logical steps of S2 are as follows: S201: A multi-layer feature extraction strategy is adopted. Numerical features are extracted by statistical calculation to extract distribution features. Textual features are extracted by TF-IDF algorithm and BERT model to extract semantic features. Behavioral features are extracted by time series analysis to extract trend and periodic features. Video features are extracted by YOLO object detection algorithm to extract features such as teacher body language and student concentration. S202: Feature selection is performed using ANOVA and mutual information calculations, retaining features that are more correlated with course quality than a threshold. Eliminate redundant features based on their characteristics. S203: The text feature vector is optimized using a BERT-based semantic encoding function. The encoding formula is as follows: ,in For text data, This is the encoded 768-dimensional high-dimensional semantic feature vector; S204: Principal Component Analysis (PCA) is used to reduce the dimensionality of the high-dimensional eigenvectors, retaining principal components with a cumulative variance contribution rate ≥90%, and generating a course quality assessment eigenvector with a unified dimension. .

6. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 3, characterized in that, The S3 multi-algorithm fusion model includes a Transformer time series prediction sub-model, a random forest evaluation sub-model, and a gradient boosting tree XGBoost optimization sub-model. The division of labor among the sub-models is as follows: the Transformer sub-model processes teaching time series data and outputs trend evaluation values; the random forest sub-model processes classification feature data and outputs quality level labels; the XGBoost sub-model performs weighted optimization on the outputs of the former two. During model training, an adaptive weight allocation algorithm is used to dynamically adjust the weights of each sub-model. The weight calculation formula is as follows: ,in For the first Cross-validation accuracy of individual sub-models For the first The generalization ability score of each sub-model; the number of iterations for model training is set to 100 to 500, the convergence condition is that the loss function value is ≤0.001, and the loss function adopts mean squared error (MSE); the final evaluation result is output through weighted voting during the inference stage, the course quality score adopts a percentage system of 0-100 points, the weak link label is generated based on the preset classification system, and the problem impact factor is calculated by the analytic hierarchy process (AHP), with a value range of [0,1].

7. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 3, characterized in that, The specific logical steps of S4 are as follows: S401: The Apriori association rule mining algorithm is used to mine the association between weak links and potential influencing factors, with the minimum support set to... The minimum confidence level is set to Output the top-5 strong association rules; S402: By combining Bayesian networks with the Do-calculus causal inference framework, a causal graph of "influencing factors-weak links" is constructed, and the average causal effect ACE value of each factor is calculated to quantify the strength of causal association. S403: Based on the causal effect value and the confidence of the association rule, a weighted ranking is performed with weights set to 0.6 and 0.4 respectively, and the top 3 core causes are selected; S404: Generate a problem attribution report that includes a description of the cause, a quantitative value of the degree of impact, supporting data sources, and improvement priorities. Improvement priorities are divided into three levels: high, medium, and low.

8. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 3, characterized in that, The specific logical steps of S5 are as follows: S501: The teaching and research feedback module retrieves teaching and research case libraries from the cloud using a cosine similarity algorithm. The similarity calculation formula is as follows: Match successful cases with a similarity ≥ 0.7 to the current problem attribution report; S502: The Deep Reinforcement Learning Recommendation Algorithm (DQN) is used to generate personalized optimization schemes. The state space of the agent consists of the course quality feature vector and the problem cause vector, the action space consists of the set of teaching and research improvement measures, and the reward function is the quality improvement rate after the scheme is implemented. S503: The plan includes suggestions for adjusting teaching methods, a list of supplementary resources, a teaching and research activity plan, effectiveness verification indicators, and an implementation timetable. The effectiveness verification indicators include a quantitative indicator of a quality score improvement of ≥10 points and a qualitative indicator of a student satisfaction improvement of ≥15%. S504: Perform a feasibility assessment on the optimization scheme. The assessment formula is as follows: ,in The suitability of the solution is rated on a scale of 0-100. Resource availability is scored from 0 to 100. Execution cost is scored from 0 to 100. These are the weighting coefficients; S505: When When the score is 60 ≤ Score < 80, the solution parameters are adjusted using a genetic algorithm before being pushed; when the score < 60, cases are retrieved again and a solution is generated.

9. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 1, characterized in that, The specific logical steps of S8 are as follows: S801: The AI ​​assessment module uses a paired-samples t-test to compare course quality scores before and after the implementation of the scheme and calculate the improvement. And the significance of the difference was assessed by effect size. The time was judged as a significant improvement; S802: Collect teachers' evaluations of the feedback plan using a 5-point Likert scale, combined with the improvement rate. Generate a feedback score for the effectiveness of the proposed solution. The scoring formula is as follows: The average score of teacher evaluations; S803: The teaching and research feedback module updates the reward function parameters of the reinforcement learning algorithm based on the effect score. When EffectScore ≥ 0.8, the reward value of the corresponding action is increased; when EffectScore ≤ 0.5 < 0.8, the reward value remains unchanged; when EffectScore < 0.5, the reward value of the corresponding action is decreased. S804: Synchronize the updated model parameters to the cloud storage module, and record the iteration log, including iteration time, parameter changes and performance improvement. The iteration cycle is consistent with the course cycle.

10. The method for dynamic evaluation of course quality and optimization of teaching and research feedback based on artificial intelligence according to claim 3, characterized in that, The system supports cross-disciplinary course quality comparison and analysis, and adjusts the evaluation model parameters through a domain adaptation algorithm. The domain adaptation algorithm adopts the Domain Adaptive Network (DANN) in transfer learning and minimizes the cross-disciplinary data distribution differences through a gradient inversion layer. The cloud storage module regularly updates the teaching and research case library using an incremental learning algorithm. The update cycle is one month. Cases must undergo three levels of review before being added to the library: AI effect verification, expert review, and practical verification. The teaching and research feedback module has a built-in course quality early warning mechanism. When the course quality score is below 60 points for two consecutive evaluation cycles or the single drop is ≥20 points, a first-level early warning is automatically triggered, an emergency rectification plan is pushed and the teaching and research management department is notified; when the score is between 60 and 70 points, a second-level early warning is triggered and a regular optimization plan is pushed.

Images