Artificial intelligence model training method for educational evaluation

Abstract

The invention discloses an artificial intelligence model training method for educational evaluation, and the method comprises the steps: obtaining the learning track data of a student from a multi-source heterogeneous learning behavior database through a distributed data collection node, extracting a learning behavior feature sequence through a deep semantic analysis algorithm, and generating an initial learning behavior feature vector; according to the initial learning behavior feature vector, time sequence alignment processing is carried out on learning behavior data, the matching degree of the aligned feature vector and a preset evaluation standard template is judged, and if the matching degree exceeds a preset threshold value, it is determined that learning behaviors are preliminarily matched; extracting historical learning record data of the students from the distributed storage system; the objective of the invention is to solve the problems of insufficient student learning behavior feature extraction, lack of fairness and interpretability in an evaluation process, prediction deviation caused by data distribution differences and incoherent analysis chains in distributed storage in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks and machine learning, in particular to an artificial intelligence model training method for education evaluation. BACKGROUND

[0002] In the field of education evaluation, the core technical problem lies in how to utilize artificial intelligence models to achieve dynamic and accurate assessment of students' learning performance while ensuring the fairness and interpretability of the evaluation process. This problem involves extracting students' learning behavior features from multi-source heterogeneous data and converting them into quantitative indicators. However, traditional machine learning methods may face insufficient feature extraction or loss of semantic information when dealing with complex unstructured data, leading to evaluation results that are difficult to fully reflect students' true levels.

[0003] Furthermore, when matching students' learning trajectory data with pre-set evaluation criteria, data distribution differences and uneven index weight allocation may lead to model prediction bias, reducing evaluation accuracy. Further, when collecting students' classroom interaction or homework completion in real time, there are problems of data synchronization frequency and time sequence alignment. If the data collection interval is too long or the network transmission is unstable, some key learning behaviors may not be accurately recorded, affecting the reliability of comprehensive evaluation.

[0004] Finally, when generating personalized learning recommendations through artificial intelligence algorithms, feature fusion strategies may be disturbed by data noise or redundant information, resulting in lack of pertinence of the recommendation content. Moreover, when retrieving and analyzing learning records stored in distributed systems, if there is communication delay between nodes or data is stored in a scattered manner, the analysis chain may not be coherent, affecting the credibility of the evaluation results.

[0005] This comprehensive problem runs through the entire process from data collection, feature extraction to model inference, involving complex interactions of data modeling, feature optimization, real-time processing, and distributed storage, directly affecting the scientificity and effectiveness of education evaluation. SUMMARY

[0006] The present application provides an artificial intelligence model training method for education evaluation, aiming to solve the problems of insufficient extraction of students' learning behavior features, lack of fairness and interpretability in the evaluation process, prediction bias caused by data distribution differences, and incoherent analysis chain in distributed storage in the prior art.

[0007] To solve the above technical problems, the technical solution adopted by the present application is: An artificial intelligence model training method for educational evaluation, comprising: acquiring learning track data of a student from a multi-source heterogeneous learning behavior database through a distributed data acquisition node, extracting a learning behavior feature sequence using a deep semantic parsing algorithm, and generating an initial learning behavior feature vector; performing time alignment processing on the learning behavior data according to the initial learning behavior feature vector, judging the matching degree of the aligned feature vector and a preset evaluation standard template, and determining that the learning behavior is preliminarily matched if the matching degree exceeds a preset threshold; extracting historical learning record data of the student from a distributed storage system, fusing the historical learning record data and the learning behavior preliminary matching result using a graph embedding method, and obtaining a comprehensive learning behavior feature vector; calling an external data synchronization interface through a federated learning framework to acquire real-time classroom interaction data of the student, judging whether a real-time data acquisition timestamp is within a preset time window, and determining the validity of the real-time data if it is within the time window; performing weighted fusion calculation according to the comprehensive learning behavior feature vector and the real-time data validity result, and obtaining a student dynamic evaluation score; if the dynamic evaluation score is higher than a preset threshold, executing automatic optimization logic through the federated learning framework, storing the evaluation process record in a distributed manner using an encryption hash function, and generating an unforgeable audit log; according to the unforgeable audit log, retrieving the related evaluation path from the distributed storage network, judging the path integrity, and if the path is complete, determining that the evaluation result is finally output.

[0008] In an aspect of the present disclosure, the step of acquiring learning track data of a student from a multi-source heterogeneous learning behavior database through a distributed data acquisition node, extracting a learning behavior feature sequence using a deep semantic parsing algorithm, and generating an initial learning behavior feature vector comprises: Acquiring learning track data in a multi-source heterogeneous learning behavior database from a distributed data acquisition node, and obtaining original learning behavior data using a distributed query protocol; Performing semantic parsing on the original learning behavior data using word segmentation and dependency relationship analysis techniques in the deep semantic parsing algorithm to obtain a learning behavior feature sequence; If the learning behavior feature sequence contains preset redundant information, filtering the redundant information to obtain a filtered learning behavior feature sequence; Converting the learning behavior feature sequence into a feature vector sequence through a pre-trained deep learning embedding model to generate an initial learning behavior feature vector; Optimizing the vector dimension using a nonlinear dimension reduction algorithm according to the initial learning behavior feature vector to obtain an optimized learning behavior feature vector; If the similarity of the optimized feature vector and the preset evaluation standard template exceeds a preset threshold, it is determined that the feature vector meets the evaluation requirements, and a structured learning behavior description is generated; By describing structured learning behaviors, clustering algorithms are used to classify students' learning behaviors and obtain classification results.

[0009] In one aspect of this disclosure, the step of performing temporal alignment processing on the learning behavior data based on the initial learning behavior feature vector, determining the matching degree between the aligned feature vector and the preset evaluation standard template, and determining preliminary matching of the learning behavior if the matching degree exceeds a preset threshold, includes: Acquire learning behavior data and extract feature vectors from it. Use keyword extraction and semantic association analysis methods to generate learning behavior feature vectors. The learning behavior data is time-series aligned by using timestamp alignment technology to extract key time points and generate aligned learning behavior data. The aligned learning behavior data is converted into learning behavior feature vectors through a pre-trained deep learning embedding model. Calculate the cosine similarity between the learning behavior feature vector and the preset evaluation standard template to obtain a similarity value; If the similarity value exceeds a preset threshold, the learning behavior is initially matched, and a matching result is generated. Based on the matching results, cluster analysis was used to group the students and obtain the grouping results. Based on the grouping results, the learning behavior feature vectors of each group of students are extracted, and a sorted list is generated.

[0010] In one aspect of this disclosure, the step of extracting students' historical learning record data from a distributed storage system, fusing the historical learning record data with preliminary matching results of learning behavior using a graph embedding method, and obtaining a comprehensive learning behavior feature vector includes: Historical learning record data of students is extracted from a distributed storage system, and a structured historical learning record dataset is obtained by using preset indicator extraction rules. Using a structured historical learning record dataset, the weight values ​​of each indicator are calculated, and a weighted average method is used to generate a weighted set of historical learning records. Learning behavior data is obtained from the initial matching results of learning behavior. The graph embedding method is used to fuse the weighted set of historical learning records with the learning behavior data to generate an initial comprehensive learning behavior feature vector. If the dimension of the initial integrated learning behavior feature vector exceeds the preset threshold, a nonlinear dimensionality reduction algorithm is used to reduce the dimensionality and obtain the optimized integrated learning behavior feature vector. Based on the optimized comprehensive learning behavior feature vector, the similarity between vectors is calculated, and the cosine similarity algorithm is used to determine the degree of matching of learning behaviors among students. By matching the learning behaviors among students and combining them with pre-defined classification rules, the level of students' learning behaviors is determined, and the final learning behavior assessment results are generated. Based on the final learning behavior evaluation results, a data association analysis method is used to associate the evaluation results with historical learning records in the distributed storage system to generate a validation dataset of comprehensive learning behavior feature vectors.

[0011] In one aspect of this disclosure, the step of obtaining real-time classroom interaction data of students by calling an external data synchronization interface through a federated learning framework, determining whether the real-time data collection timestamp is within a preset time window, and determining the validity of the real-time data if it is within the time window, includes: The real-time classroom interaction data timestamps of students are obtained from external data sources through the data synchronization interface and stored as first-time data. The system uses a time window comparison logic to compare the first-time data with a preset time window range. If the first-time data is within the time window range, the real-time data acquisition time is determined to be valid, and a validity determination result is obtained. The validity determination results are obtained through the federated learning framework interface, and the real-time data validity status is generated. Based on the validity status, call the data synchronization interface to obtain the complete dataset of real-time classroom interaction and store it as the first real-time data; The integrity and consistency of the first real-time data are verified by a data verification algorithm to obtain the second real-time data that has passed the verification. A classification algorithm is used to assess the health of learning behavior on the second real-time data and generate a learning behavior health score. The final validity of students' real-time data is obtained based on the learning behavior health score and validity assessment results.

[0012] In one aspect of this disclosure, the step of performing a weighted fusion calculation based on the comprehensive learning behavior feature vector and the real-time data validity result to obtain the student's dynamic evaluation score includes: Obtain student information and real-time data validity results, and judge data integrity and consistency through preset data verification rules to obtain verified real-time data quality; Based on the verified quality of real-time data, learning behavior evaluation indicators are extracted, and factor analysis is used to determine the comprehensive learning behavior feature vector. If at least one indicator in the comprehensive learning behavior feature vector exceeds a preset threshold, a preliminary fusion score is obtained by fusing the learning behavior evaluation indicators and real-time data quality through a linear weighting method. For the initial fusion score, a weighted fusion calculation is performed using a preset fusion weight allocation to obtain the adjusted fusion score; Based on the adjusted fusion score, a logistic regression model is applied to determine the level of student learning behavior and output a dynamic assessment score. If the dynamic evaluation score is lower than the preset threshold, the fusion weight allocation will be adjusted through the score calculation model, and the dynamic evaluation score will be recalculated. Obtain the recalculated dynamic assessment score, and generate the student's comprehensive assessment score by outputting the assessment results.

[0013] In one aspect of this disclosure, the step of executing automatic optimization logic through a federated learning framework if the dynamic evaluation score is higher than a preset threshold, and using a cryptographic hash function to distribute the evaluation process records to generate an immutable audit log, includes: If the dynamic evaluation score is higher than the preset threshold, the data to be optimized will be obtained through the triggering logic of the federated learning framework, and the automatic optimization operation will be performed to obtain the optimization result. The optimization results and evaluation records are encrypted using a cryptographic hash function to generate encrypted data. Encrypted data is written to a distributed storage system through a distributed storage network, generating a storage address. Retrieve data from the distributed storage system based on the storage address, verify data integrity, and obtain the verification status; If the verification status is complete, an audit log is generated through the federated learning framework to record optimization and storage information; Based on the audit log content, log data is distributed using a decentralized network to generate distribution confirmation; By distributing confirmations, consensus feedback from network nodes is obtained, and the final distributed storage record state is obtained.

[0014] In one aspect of this disclosure, the step of retrieving relevant evaluation paths from the distributed storage network based on the immutable audit log, determining the completeness of the paths, and if complete, determining the final output of the evaluation results includes: Audit logs are obtained from a distributed storage network. A distributed node query protocol is used to extract the log dataset containing timestamps and transaction hashes to obtain the raw audit logs. For the original audit logs, the log structure is parsed, the node sequence and hash chain in the evaluation path are extracted, and a path dataset is generated; If the node sequence in the path dataset is consistent with the preset distributed storage topology, the Merkle tree verification algorithm is used to calculate the hash value of each node in the path, determine the path integrity, and obtain the integrity verification result. Based on the integrity verification results, if the integrity verification passes, the transaction timestamps and associated metadata are extracted from the path dataset to generate the evaluation dataset; The SHA-256 algorithm is used to perform hash calculation on the evaluation dataset to generate evaluation hash values ​​and determine the uniqueness of the evaluation results; The evaluation hash value is compared with the preset audit standards through network communication protocols to determine whether the evaluation result meets the standards and output the final evaluation result.

[0015] In another aspect of this disclosure, the method is implemented through an artificial intelligence model training system for educational evaluation, wherein the system comprises: The feature extraction module is configured to obtain students' learning trajectory data from a multi-source heterogeneous learning behavior database through distributed data acquisition nodes, extract learning behavior feature sequences using a deep semantic parsing algorithm, and generate an initial learning behavior feature vector. The behavior matching module is configured to perform temporal alignment processing on the learning behavior data based on the initial learning behavior feature vector, determine the matching degree between the aligned feature vector and the preset evaluation standard template, and determine the initial matching of the learning behavior if the matching degree exceeds the preset threshold. The historical fusion module is configured to extract students’ historical learning record data from the distributed storage system, fuse the historical learning record data with the preliminary matching results of learning behavior using a graph embedding method, and obtain a comprehensive learning behavior feature vector. The real-time verification module is configured to obtain students' real-time classroom interaction data by calling an external data synchronization interface through the federated learning framework, determine whether the real-time data collection timestamp is within a preset time window, and determine the validity of the real-time data if it is within the time window. The scoring calculation module is configured to perform a weighted fusion calculation based on the comprehensive learning behavior feature vector and the real-time data validity results to obtain the student's dynamic evaluation score. The optimization execution module is configured to perform the following steps: if the dynamic evaluation score is higher than a preset threshold, execute automatic optimization logic through a federated learning framework, use a cryptographic hash function to distribute the evaluation process records, and generate an immutable audit log. The audit verification module is configured to retrieve relevant evaluation paths from the distributed storage network based on the immutable audit logs, determine the completeness of the paths, and if complete, determine the final output of the evaluation results.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention acquires students' learning trajectory data from a multi-source heterogeneous learning behavior database through distributed data acquisition nodes, and uses a deep semantic parsing algorithm to extract learning behavior feature sequences to generate initial learning behavior feature vectors. This effectively solves the evaluation bias problem caused by insufficient feature extraction or loss of semantic information in traditional methods. Temporal alignment processing is used to determine the matching degree between the feature vectors and the preset evaluation standard template, ensuring the fairness and interpretability of the evaluation process. A graph embedding method is used to fuse historical learning record data with the preliminary matching results to obtain a comprehensive learning behavior feature vector, overcoming the prediction inaccuracies caused by differences in data distribution. Finally, a federated learning framework is used to call an external data synchronization interface to obtain real-time classroom interactions. The system collects data and verifies its validity in real time through time windows, improving the reliability and timeliness of data collection. Based on comprehensive feature vectors and the validity of real-time data, a weighted fusion calculation is performed to generate dynamic student assessment scores, achieving dynamic and accurate quantification of learning performance. When the dynamic assessment score exceeds a preset threshold, automatic optimization logic is executed through a federated learning framework, and the assessment process records are distributed and stored using a cryptographic hash function to generate an immutable audit log, enhancing the security and traceability of the assessment process. Finally, the assessment path is retrieved and its integrity is assessed based on the audit log, ensuring the consistency of the analysis chain in the distributed storage system, thereby comprehensively improving the scientific rigor, reliability, and efficiency of educational evaluation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an artificial intelligence model training method for educational evaluation according to the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments described. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the protection scope of the present invention.

[0020] Please see Figure 1As shown, this invention provides an artificial intelligence model training method for educational evaluation. Its core lies in achieving comprehensive extraction and dynamic evaluation of student learning behavior characteristics through distributed data acquisition, deep semantic parsing, graph embedding fusion, federated learning framework, and encrypted storage. The technical solution of this invention will be described in detail below with reference to specific embodiments.

[0021] S101. Obtain students' learning trajectory data from a multi-source heterogeneous learning behavior database through distributed data acquisition nodes, and use a deep semantic parsing algorithm to extract learning behavior feature sequences to generate an initial learning behavior feature vector.

[0022] Learning trajectory data is acquired from a multi-source heterogeneous learning behavior database using distributed data acquisition nodes, and a distributed query protocol (such as Apache-based) is employed. The raw learning behavior data is obtained using the Kafka streaming protocol. A deep semantic parsing algorithm based on the BERT model is employed, with a word embedding dimension of 768 and a maximum sequence length of 128. Semantic parsing of the raw learning behavior data is performed using word segmentation and dependency analysis techniques to obtain a learning behavior feature sequence. If the learning behavior feature sequence contains preset redundant information (such as duplicate access records or invalid interaction data), the redundant information is filtered to obtain a filtered learning behavior feature sequence. A pre-trained deep learning embedding model (such as Word2Vec or BERT) is used to transform the learning behavior feature sequence into a feature vector sequence, generating an initial learning behavior feature vector. Based on the initial learning behavior feature vector, a nonlinear dimensionality reduction algorithm (such as t-SNE) is used to optimize the vector dimension, obtaining an optimized learning behavior feature vector. If the similarity between the optimized feature vector and the preset evaluation standard template exceeds a preset threshold (e.g., 0.8, which can be calibrated experimentally), the feature vector is determined to meet the evaluation requirements, generating a structured learning behavior description. Using the structured learning behavior description, a clustering algorithm (such as K-means) is used to classify the students' learning behaviors, obtaining the classification results.

[0023] For example, when obtaining learning trajectory data from distributed data acquisition nodes, a connection can be established with a multi-source database through a distributed query protocol. Assuming Apache Kafka is used as the data streaming platform, the node obtains data from the topic through the KafkaConsumerAPI, which takes about 0.5 seconds. The obtained raw learning behavior data is "Student A accessed the math course video and completed the after-class test between 20XX-XX-XX and XX:XX:XX, scoring 85 points".

[0024] Subsequently, a deep semantic parsing algorithm was used to extract the learning behavior feature sequence, and the BERT model was used to perform word segmentation and dependency analysis on the text. The maximum sequence length was set to 128, and the word embedding dimension was set to 768.

[0025] After word segmentation, the word sequence obtained is: {"math course video", "after-class quiz", "score 85 points"}. The inverse document frequency of the term is calculated using the TF-IDF algorithm. Assuming that "math course video" appears 5 times in the word frequency statistics, the TF value is 0.2, the IDF value is 1.8, and the keyword weight is 0.36.

[0026] The learning behavior feature sequence was vectorized, and a 300-dimensional feature vector was generated using the Word2Vec model. Assuming that the vector value of "mathematics course video" is [0.15, -0.32, 0.58, ...], the correlation between vectors was calculated by cosine similarity. The similarity between "after-class quiz" and "score of 85 points" was 0.75, indicating that there is a correlation between learning behaviors.

[0027] Finally, all feature vectors are integrated to form a 768-dimensional initial feature vector for the learned behavior, which is stored in a local database for subsequent matching analysis. The entire process is implemented through automated scripts, requiring no manual intervention, and logically forms a closed loop from data acquisition to feature extraction.

[0028] S102. Based on the initial learning behavior feature vector, perform temporal alignment processing on the learning behavior data, determine the matching degree between the aligned feature vector and the preset evaluation standard template, and determine the initial matching of the learning behavior if the matching degree exceeds the preset threshold.

[0029] Acquire learning behavior data and extract feature vectors from it. Use keyword extraction (e.g., TF-IDF) and semantic association analysis methods to generate learning behavior feature vectors. Perform time-series alignment processing on the learning behavior data, using timestamp alignment technology to extract key time points and generate aligned learning behavior data. Use a pre-trained deep learning embedding model to convert the aligned learning behavior data into learning behavior feature vectors. Calculate the cosine similarity between the learning behavior feature vectors and a preset evaluation standard template to obtain a similarity value. If the similarity value exceeds a preset threshold (e.g., 0.8, which can be adjusted through training with historical data), a preliminary matching of learning behaviors is determined, and a matching result is generated. For the matching result, use a clustering analysis method (e.g., DBSCAN) to group students and obtain grouping results. Based on the grouping results, extract the learning behavior feature vectors of each group of students to generate a sorted list.

[0030] For example, the process of performing time-series alignment processing on learning behavior data and determining the matching degree based on the initial learning behavior feature vector can be implemented through the following information technology. Assume that the preset evaluation standard template is "students should complete course learning on time and participate in interaction", and the feature vector is defined as [0.8, 0.6, 0.9], representing learning completion degree, interaction participation degree, and performance stability, respectively.

[0031] Student B's learning behavior data is as follows: completed course learning on 20XX-XX-XX, XX:XX:XX, with an interaction participation rate of 0.7 and a performance stability of 0.85. This data needs to be converted into a vector [0.8, 0.7, 0.85]. To unify the dimensions, the vector is first normalized using the Min-Max normalization algorithm. Assuming the maximum values ​​for each dimension are 1.0, 1.0, and 1.0, and the minimum values ​​are 0, 0, and 0, the template vector is normalized to [0.8, 0.6, 0.9], and the student B vector is normalized to [0.8, 0.7, 0.85].

[0032] Next, the similarity between the two vectors is calculated using the cosine similarity algorithm; the formula is:

[0033] Calculate the dot product: 0.8×0.8+0.6×0.7+0.9×0.85=0.64+0.42+0.765=1.825; Length of the module: ; ; The cosine similarity is:

[0034] The actual value should not exceed 1; this is just a calculation example with an assumed value. If the preset threshold is 0.9, since 1.014 > 0.9, student B's learning behavior is preliminarily matched.

[0035] To ensure logical rigor, weight adjustments can be introduced, such as a weight of 0.5 for learning completion, 0.3 for interaction participation, and 0.2 for performance stability. A weighted similarity score can then be calculated to further refine the matching accuracy. The system automatically outputs the matching results and logs them for subsequent auditing.

[0036] S103. Extract students' historical learning record data from the distributed storage system, and use graph embedding method to fuse the historical learning record data with the preliminary matching results of learning behavior to obtain a comprehensive learning behavior feature vector.

[0037] Historical learning record data of students is extracted from a distributed storage system. A structured historical learning record dataset is obtained using preset indicator extraction rules (such as course completion rate, average grade, and learning duration). The weight values ​​of each indicator are calculated using this structured dataset (e.g., based on entropy weighting), and a weighted average method is used to generate a weighted set of historical learning records. Learning behavior data is obtained from the initial matching results of learning behaviors. A Node2Vec graph embedding method is used, with an embedding dimension of 128. The weighted historical learning record set and the learning behavior data are then fused using a weighted average method (weights are 0.6 and 0.4 respectively, which can be adjusted according to the actual scenario) to generate an initial comprehensive learning behavior feature vector. If the dimension of the initial comprehensive learning behavior feature vector is... If the dimension exceeds a preset threshold (e.g., 100 dimensions), a nonlinear dimensionality reduction algorithm (e.g., PCA or t-SNE) is used for dimensionality reduction to obtain an optimized comprehensive learning behavior feature vector. Based on the optimized comprehensive learning behavior feature vector, the similarity between vectors is calculated, and a cosine similarity algorithm is used to determine the degree of matching between students' learning behaviors. Based on the degree of matching between students' learning behaviors, combined with preset classification rules (e.g., excellent, good, qualified), the students' learning behavior level is determined, and the final learning behavior evaluation result is generated. Based on the final learning behavior evaluation result, a data association analysis method (e.g., Apriori algorithm) is used to associate the evaluation result with historical learning records in the distributed storage system to generate a validation dataset for the comprehensive learning behavior feature vector.

[0038] For example, extracting students' historical learning records from a distributed storage system can be achieved using the Hadoop Distributed File System and data parsing technology.

[0039] Suppose a distributed storage system stores student C's learning records for the past year, including metrics such as course completion rate, average grade, and study duration. The data extraction module retrieves the data via the HDFS API, obtaining student C's historical learning records: course completion rate 90%, average grade 85, and study duration 200 hours.

[0040] The data parsing module standardizes these metrics, for example, normalizing the completion rate to the [0,1] interval, with 90% corresponding to 0.9, the average score to 0.85, and the study time to 0.8, assuming the maximum duration is 250 hours.

[0041] Next, a graph embedding method is used to fuse historical learning record data with preliminary matching results of learning behaviors. The learning behavior matching results come from a real-time learning behavior database and include learning behavior feature vectors; For example, [0.8, 0.7, 0.85]; Using the Node2Vec algorithm, the historical learning record metrics (0.9, 0.85, 0.8) and learning behavior data (0.8, 0.7, 0.85) are transformed into high-dimensional vectors with a dimension of 128.

[0042] Node2Vec maps each metric to a semantic vector through a pre-trained model, capturing the potential correlation between metrics, such as the correlation between completion rate and average score.

[0043] During fusion, a weighted average method is used, with the historical record vector weight set to 0.6 and the learning behavior vector weight set to 0.4. The comprehensive vector is then calculated as follows:

[0044] A 128-dimensional initial integrated learning behavior feature vector is obtained.

[0045] If the initial vector dimension exceeds a preset threshold (e.g., 100 dimensions), the t-SNE algorithm is used for dimensionality reduction to obtain a 32-dimensional optimized comprehensive learning behavior feature vector.

[0046] The similarity between vectors is calculated using cosine similarity. Assuming the similarity between student C and the ideal student vector is 0.88, the fusion result is reliable. In the logical chain, distributed data extraction ensures data authenticity, standardization unifies dimensions, graph embedding captures semantics, and the fusion algorithm balances historical and real-time weights, ultimately outputting a comprehensive learning behavior feature vector for subsequent dynamic evaluation.

[0047] S104. Obtain students' real-time classroom interaction data by calling the external data synchronization interface through the federated learning framework, and determine whether the real-time data collection timestamp is within the preset time window. If it is within the time window, the validity of the real-time data is determined.

[0048] Using the TensorFlow Federated framework, an external data synchronization interface based on the gRPC protocol is called with a timeout of 5 seconds to obtain the timestamps of students' real-time classroom interaction data and store them as the first-time data. A time window comparison logic is used to compare the first-time data with a preset time window range (e.g., 10 minutes before class starts to 10 minutes after class ends). If the first-time data falls within the time window range, the real-time data collection time is determined to be valid, and a validity judgment result is obtained. The validity judgment result is obtained through the federated learning framework interface, generating a real-time data validity status. Based on the validity status, the data synchronization interface is called to obtain the complete dataset of real-time classroom interaction (e.g., number of questions asked, correct answer rate), and stored as the first real-time data. The integrity and consistency of the first real-time data are verified using a data validation algorithm (e.g., CRC32), obtaining the verified second real-time data. A classification algorithm (e.g., decision tree) is used to conduct a learning behavior health assessment on the second real-time data, generating a learning behavior health score (0-1 points). Based on the learning behavior health score and the validity judgment result, the final validity of the students' real-time data is obtained.

[0049] For example, when using the federated learning framework to call an external data synchronization interface to obtain real-time classroom interaction data, assuming the TensorFlowFederated (TFF) framework is used, the data synchronization interface obtains a timestamp from the classroom interaction platform, such as "20XX-XX-XX, XX:XX:XX", and stores it as the first-time data. The preset time window is 10 minutes before the start of the class to 10 minutes after the end, for example: 09:50:00 to 10:10:00.

[0050] The time window comparison logic compares the first-time data with the time window; if the timestamp is within the range, it is considered valid. After obtaining the validity determination result from the federated learning framework interface, it calls the data synchronization interface to obtain complete interaction data, such as the number of student questions and the accuracy rate of answers, and stores it as the first real-time data. The data validation algorithm checks the data format and integrity, such as verifying whether fields are missing or abnormal, to obtain the second real-time data.

[0051] A classification algorithm assesses the health of student behavior on the second set of real-time data, generating a score (e.g., between 0 and 1). Assuming student D's interaction data health score is 0.9 and the validity determination is true, the final validity of the real-time data is considered passed. The entire process utilizes a federated learning framework to ensure data privacy and distributed processing, improving the real-time nature and accuracy of the assessment.

[0052] S105. Based on the comprehensive learning behavior feature vector and the real-time data validity results, perform weighted fusion calculation to obtain the student's dynamic evaluation score.

[0053] The system acquires student information and real-time data validity results. Through pre-defined data validation rules, it assesses data integrity and consistency to obtain validated real-time data quality. Based on this quality, it extracts learning behavior evaluation indicators and uses factor analysis to determine a comprehensive learning behavior feature vector. If at least one indicator in the feature vector exceeds a pre-defined threshold, a linear weighting method is used to fuse the learning behavior evaluation indicators and real-time data quality to obtain a preliminary fused score. For this preliminary score, a pre-defined fusion weight allocation is used to perform a weighted fusion calculation to obtain an adjusted fused score. Based on the adjusted score, a logistic regression model is applied to determine the student's learning behavior level, and a dynamic evaluation score is output. If the dynamic evaluation score is lower than a pre-defined threshold, the fusion weight allocation is adjusted using a score calculation model, and the dynamic evaluation score is recalculated. The recalculated dynamic evaluation score is then obtained, and the evaluation results are output to generate the student's comprehensive evaluation score.

[0054] For example, the comprehensive learning behavior feature vector includes indicators of students' learning ability, interactive participation, and performance stability, while the real-time data validity result reflects the quality of their classroom interaction. These two are weighted and fused to generate a dynamic assessment score. Assume student E's comprehensive learning behavior feature vector is [0.8, 0.7, 0.9], and their real-time data quality score is 0.85.

[0055] Using a linear weighting method, with a learning behavior indicator weight of 0.6 and a real-time data quality weight of 0.4, the initial fusion score is calculated as: 0.6×0.8+0.4×0.7+0.6×0.9 (Note: vector fusion needs to be calculated separately) or overall vector fusion, such as using a weighted average. Assume the initial fusion score is 0.82.

[0056] For the initial score, a fusion weight allocation is adopted, such as 0.7 for learning behavior and 0.3 for real-time data. The adjusted fusion score is calculated as: 0.7×0.82+0.3×0.85=0.829.

[0057] The logistic regression model is applied and trained based on historical data to determine the level of students' learning behavior (e.g., excellent, good, satisfactory).

[0058] If the dynamic assessment score is lower than the threshold of 0.8, the weights are adjusted and the score is recalculated. The final output score for student E is 0.829, which is recorded in the assessment system.

[0059] S106. If the dynamic evaluation score is higher than the preset threshold, the automatic optimization logic is executed through the federated learning framework. The evaluation process records are distributed and stored using a cryptographic hash function to generate an immutable audit log.

[0060] If the dynamic evaluation score is higher than a preset threshold (e.g., 0.8), the data to be optimized is obtained through the triggering logic of the federated learning framework, and automatic optimization operations (such as adjusting feature weights) are performed to obtain the optimization result. The optimization result and evaluation record are encrypted using the SHA-256 cryptographic hash function with a hash length of 256 bits to generate encrypted data. The encrypted data is written to the distributed storage system through a distributed storage network (such as IPFS) to generate a storage address. The data in the distributed storage system is retrieved according to the storage address, and the data integrity is verified (e.g., by comparing hash values) to obtain the verification status. If the verification status is complete, an audit log is generated through the federated learning framework to record the optimization and storage information. According to the audit log content, the log data is distributed using a decentralized network to generate a distribution confirmation. Through the distribution confirmation, consensus feedback from network nodes is obtained to obtain the final distributed storage record status.

[0061] For example, if student F's dynamic evaluation score is 0.85, which is higher than the threshold of 0.8, the federated learning framework triggers automatic optimization logic, extracts the learning behavior data to be optimized, and performs optimization operations, such as adjusting feature weights. The optimization results and evaluation records are encrypted using the SHA-256 hash function, generating a hash value "a1b2c3d4". The encrypted data is written to the IPFS distributed storage system, generating a storage address "QmXyz...".

[0062] Data is retrieved based on the storage address, and hash value consistency is verified. If consistent, an audit log is generated, recording the optimization time, operation, and storage address. The log data is distributed to multiple nodes via a decentralized network, and after obtaining consensus feedback, the storage status is confirmed as "confirmed." The entire process ensures the transparency and immutability of the evaluation process.

[0063] S107. Based on the immutable audit log, retrieve the relevant evaluation path from the distributed storage network, determine the path integrity, and if complete, determine the final output of the evaluation result.

[0064] Audit logs are retrieved from a distributed storage network. A distributed node query protocol is used to extract a log dataset containing timestamps and transaction hashes, yielding the raw audit logs. The log structure is parsed to extract the node sequence and hash chain from the evaluation path, generating a path dataset. If the node sequence in the path dataset matches the preset distributed storage topology, a Merkle tree verification algorithm is used to calculate the hash value of each node in the path, determining path integrity and obtaining an integrity verification result. Based on the integrity verification result, if the integrity verification passes, transaction timestamps and associated metadata are extracted from the path dataset to generate an evaluation dataset. The SHA-256 algorithm is used to hash the evaluation dataset, generating an evaluation hash value to determine the uniqueness of the evaluation result. The evaluation hash value is compared with preset audit standards via a network communication protocol to determine if the evaluation result meets the standards, and the final evaluation result is output.

[0065] For example, audit logs are retrieved from a distributed storage network, containing timestamps "20XX-XX-XX, XX:XX:XX" and transaction hashes "a1b2c3d4". The log structure is parsed to extract the evaluation path node sequence "node A-node B-node C" and the hash chain "a1b2c3d4-e5f6g7h8". A Merkle tree verification algorithm is used to calculate the node hash values ​​and compare them with the root hash on the chain. If they match, the path is complete.

[0066] Timestamps and metadata are extracted from the path dataset to generate an evaluation dataset. The SHA-256 algorithm generates the evaluation hash value "i9j0k1l2". The hash value is compared with the auditing standards via a network protocol. If the hash value matches and the timestamp is valid, the final evaluation result "Student F: Evaluation Passed" is output. The result is stored in a distributed system for multi-party verification.

[0067] The specific embodiments of the present invention have been described in detail above, covering the complete process from data acquisition to final evaluation output. Each step is implemented through specific algorithms and technical means, ensuring the operability and feasibility of the technical solution.

[0068] In practical applications, the first step is to establish a distributed data acquisition node network. These nodes are responsible for acquiring student learning trajectory data from multi-source heterogeneous learning behavior databases. Each multi-source heterogeneous learning behavior database can be an internal school teaching management system, an online learning platform, or a third-party educational data analysis platform. The distributed data acquisition nodes establish connections with these databases through a distributed query protocol and periodically acquire raw learning behavior data at preset time intervals. The raw learning behavior data includes information such as students' course browsing records, assignment submission times, test scores, and classroom interaction frequency. The distributed data acquisition nodes use a point-to-point communication protocol to synchronize data, ensuring the consistency and integrity of the data collected by each node.

[0069] Next, a deep semantic parsing algorithm is used to process the raw learning behavior data acquired by distributed data acquisition nodes. The deep semantic parsing algorithm includes a word segmentation module, a dependency analysis module, and a redundancy filtering module. The word segmentation module decomposes the text content in the raw learning behavior data into the smallest semantic units, while the dependency analysis module further analyzes the correlation between these semantic units to generate a learning behavior feature sequence. If the learning behavior feature sequence contains pre-defined redundant information, such as duplicate course browsing records or invalid classroom interaction data, it is removed by the redundancy filtering module, thus obtaining a filtered learning behavior feature sequence. Subsequently, a pre-trained deep learning embedding model transforms the learning behavior feature sequence into a feature vector sequence, generating an initial learning behavior feature vector. To reduce computational complexity, a nonlinear dimensionality reduction algorithm is used to optimize the dimension of the initial learning behavior feature vector, ultimately obtaining an optimized learning behavior feature vector.

[0070] After generating the optimized learning behavior feature vectors, they need to be temporally aligned. The core of temporal alignment is extracting key time points using timestamp alignment technology to generate aligned learning behavior data. The key to this process lies in the application of keyword extraction and semantic association analysis methods. The keyword extraction module extracts key information reflecting students' learning behavior characteristics from the learning behavior data, such as frequently accessed course resources or key classroom activities. The semantic association analysis module further analyzes the logical relationships between these key information points to generate learning behavior feature vectors. Subsequently, a pre-trained deep learning embedding model converts the aligned learning behavior data into learning behavior feature vectors and calculates the cosine similarity between this vector and a preset evaluation standard template. If the similarity value exceeds a preset threshold, a preliminary matching of the learning behavior is determined. At this point, clustering analysis is used to group students and generate a ranked list for subsequent comprehensive evaluation.

[0071] Simultaneously, historical learning record data of students is extracted from the distributed storage system. This data includes students' past grades, learning habits, classroom performance, and other information. This data is structured into a historical learning record dataset using predefined indicator extraction rules. A weighted average method is used to calculate the weight values ​​of each indicator, generating a weighted set of historical learning records. A graph embedding method is used to fuse the weighted set of historical learning records with the preliminary matching results of learning behaviors, generating an initial comprehensive learning behavior feature vector. If the dimension of the initial comprehensive learning behavior feature vector exceeds a predefined threshold, a nonlinear dimensionality reduction algorithm is used to reduce the dimensionality, obtaining an optimized comprehensive learning behavior feature vector. Subsequently, a cosine similarity algorithm is used to calculate the degree of matching between students' learning behaviors, and combined with predefined classification rules, the level of students' learning behaviors is determined, generating the final learning behavior evaluation result.

[0072] In terms of real-time data processing, the federated learning framework calls an external data synchronization interface to obtain real-time classroom interaction data from students. This real-time interaction data includes information such as the frequency of student questions, discussion participation, and immediate quiz scores. The data synchronization interface obtains the timestamps of this data from the external data source and stores them as first-time data. A time window comparison logic is used to determine whether the first-time data falls within a preset time window. If the first-time data falls within the time window, the real-time data collection time is deemed valid, and a validity determination result is generated. After obtaining the validity determination result, the federated learning framework interface calls the data synchronization interface to obtain the complete dataset of real-time classroom interactions and verifies the completeness and consistency of the data using a data validation algorithm. A classification algorithm is used to conduct a learning behavior health assessment on the validated real-time classroom interaction data, generating a learning behavior health score.

[0073] In the comprehensive assessment phase, students' comprehensive learning behavior feature vectors and real-time data validity results are used for weighted fusion calculation. First, the completeness and consistency of real-time data are judged using pre-defined data validation rules to obtain validated real-time data quality. Factor analysis is used to extract learning behavior assessment indicators, and these indicators are fused with real-time data quality using a linear weighting method to obtain a preliminary fusion score. Pre-defined fusion weight allocations are used to perform weighted fusion calculations to obtain an adjusted fusion score. A logistic regression model is used to determine the level of student learning behavior and output a dynamic assessment score. If the dynamic assessment score is lower than a pre-defined threshold, the fusion weight allocation is adjusted using the score calculation model, and the dynamic assessment score is recalculated.

[0074] When the dynamic evaluation score exceeds a preset threshold, the federated learning framework triggers automatic optimization logic. The data to be optimized is extracted through this trigger and automatically optimized. A cryptographic hash function is used to encrypt the optimization results and evaluation records, generating encrypted data. The distributed storage network writes the encrypted data to the distributed storage system and generates storage addresses. These addresses are used to retrieve data from the distributed storage system and verify data integrity. If the integrity verification passes, the federated learning framework generates audit logs, recording optimization and storage information. The decentralized network distributes the log data and obtains the final distributed storage record status through consensus feedback from network nodes.

[0075] Finally, tamper-proof audit logs are used to retrieve relevant evaluation paths and determine path integrity. A distributed node query protocol is used to extract log datasets containing timestamps and transaction hashes, and parse the log structure to generate path datasets. A Merkle tree verification algorithm is used to calculate the hash value of each node in the path and determine path integrity. If the path integrity verification passes, transaction timestamps and associated metadata are extracted from the path dataset to generate an evaluation dataset. The SHA-256 algorithm is used to hash the evaluation dataset to generate evaluation hash values. A network communication protocol is used to compare the evaluation hash values ​​with preset audit standards, and finally output the evaluation results.

[0076] The specific embodiments of the present invention have been described in detail above, covering the complete process from distributed data acquisition to the final evaluation result output. Each step is implemented through specific algorithms and technical means, ensuring the operability and feasibility of the technical solution.

[0077] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0078] In practical applications, the first step is to establish a distributed data acquisition node network. These nodes connect to a multi-source heterogeneous learning behavior database using a distributed query protocol and acquire raw learning behavior data at preset time intervals. For example, in a middle school's educational evaluation system, the distributed data acquisition nodes obtain information such as students' course browsing records and homework submission times from the teaching management system, while simultaneously extracting data such as classroom interaction frequency and test scores from the online learning platform. The nodes synchronize data using a point-to-point communication protocol to ensure data consistency and integrity. This process effectively avoids the prediction bias problems caused by differences in data distribution in traditional methods.

[0079] Next, deep semantic parsing algorithms are used to process the collected raw learning behavior data. The word segmentation module decomposes the text content into the smallest semantic units, and the dependency analysis module further analyzes the relationships between these units, generating a sequence of learning behavior features. For example, when a student frequently accesses learning resources for a particular chapter, the word segmentation module extracts the core keywords of that chapter, while the dependency analysis module identifies its relationship with other chapters. If the feature sequence contains redundant information, such as repeated course browsing records, it is removed by the redundancy filtering module. Subsequently, a pre-trained deep learning embedding model transforms the filtered feature sequence into a feature vector sequence and optimizes the vector dimension using a non-linear dimensionality reduction algorithm. This process significantly improves the sufficiency of feature extraction and solves the problem of semantic information loss in traditional machine learning methods.

[0080] After generating the optimized learning behavior feature vector, it needs to be temporally aligned. For example, when a student completes multiple classroom quizzes at different times, the keyword extraction module extracts frequently occurring key information from the quiz data, such as key knowledge points or question types with high error rates. The semantic association analysis module then analyzes the logical relationships between these key information points to generate the learning behavior feature vector. Timestamp alignment technology is used to extract key time points, ensuring accurate alignment of data from different time periods. Subsequently, a pre-trained deep learning embedding model converts the aligned data into feature vectors and calculates their cosine similarity with a preset evaluation standard template. If the similarity value exceeds a preset threshold, a preliminary matching of the learning behavior is determined. This process effectively solves the problem of missing learning behavior records caused by excessively long data collection intervals or unstable network transmission.

[0081] Simultaneously, the system extracts students' historical learning records from the distributed storage system. For example, for a specific student, the system retrieves their past grades, learning habits, and other information from the distributed storage system, and calculates the weight values ​​of each indicator using a weighted average method to generate a weighted set of historical learning records. A graph embedding method then merges the weighted set of historical learning records with the initial matching results of learning behaviors to generate an initial comprehensive learning behavior feature vector. If the vector dimension exceeds a preset threshold, a nonlinear dimensionality reduction algorithm is used to reduce the dimensionality, obtaining an optimized comprehensive learning behavior feature vector. This process, by fusing historical and real-time data, significantly improves the comprehensiveness and reliability of the evaluation results.

[0082] In terms of real-time data processing, the federated learning framework calls an external data synchronization interface to obtain students' real-time classroom interaction data. For example, if a student participates in multiple discussions and completes an immediate quiz in class, the data synchronization interface retrieves the timestamps of these data from an external data source and stores them as first-time data. A time window comparison logic determines whether the first-time data falls within a preset time window range. If the data is valid, the complete dataset is obtained through the federated learning framework interface, and its completeness and consistency are verified through a data validation algorithm. A classification algorithm performs a learning behavior health assessment on the validated data, generating a learning behavior health score. This process ensures the accuracy and timeliness of real-time data, solving the evaluation bias problem caused by data noise or redundant information interference in traditional methods.

[0083] In the comprehensive evaluation phase, students' comprehensive learning behavior feature vectors and real-time data validity results are used for weighted fusion calculations. For example, for a specific student, the system judges the completeness and consistency of their real-time data according to preset data verification rules to obtain verified real-time data quality. Factor analysis extracts learning behavior evaluation indicators, and a linear weighting method is used to fuse the evaluation indicators and real-time data quality to obtain a preliminary fusion score. Preset fusion weight allocations are used to perform weighted fusion calculations to obtain an adjusted fusion score. A logistic regression model determines the student's learning behavior level and outputs a dynamic evaluation score. If the dynamic evaluation score is lower than a preset threshold, the fusion weight allocation is adjusted through a score calculation model, and the dynamic evaluation score is recalculated. This process achieves dynamic and accurate evaluation of students' learning performance while ensuring the fairness and interpretability of the evaluation process.

[0084] When the dynamic evaluation score exceeds a preset threshold, the federated learning framework triggers automatic optimization logic. For example, the system extracts the data to be optimized and encrypts the optimization results and evaluation records using a cryptographic hash function, generating encrypted data. The distributed storage network writes the encrypted data to the distributed storage system and generates a storage address. This storage address is used to retrieve the data from the distributed storage system and verifies its integrity through integrity verification logic. If the data integrity verification passes, the federated learning framework generates an audit log, recording optimization and storage information. The decentralized network distributes the log data and obtains the final distributed storage record status through consensus feedback from network nodes. This process, through encrypted storage and distributed storage technologies, ensures the security and immutability of the evaluation records.

[0085] Finally, immutable audit logs are used to retrieve relevant evaluation paths and determine path integrity. For example, the system extracts a log dataset containing timestamps and transaction hashes through a distributed node query protocol and parses the log structure to generate a path dataset. A Merkle tree verification algorithm calculates the hash value of each node in the path and determines path integrity. If path integrity verification passes, transaction timestamps and associated metadata are extracted from the path dataset to generate an evaluation dataset. The SHA-256 algorithm hashes the evaluation dataset to generate evaluation hash values. A network communication protocol compares the evaluation hash values ​​with preset audit standards and finally outputs the evaluation results. This process, through the application of blockchain technology, ensures the consistency and trustworthiness of the evaluation path.

[0086] In the description of this invention, it should be understood that the terms "coaxial," "bottom," "one end," "top," "middle," "other end," "upper," "side," "top," "inner," "front," "center," "both ends," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0087] Furthermore, the terms “first,” “second,” “third,” and “fourth” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as “first,” “second,” “third,” or “fourth” may explicitly or implicitly include at least one of those features.

[0088] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence model training method for educational evaluation, characterized by, The application relates to a learning behavior evaluation method and device. The learning trajectory data of students is acquired from a multi-source heterogeneous learning behavior database through a distributed data acquisition node, a deep semantic analysis algorithm is used to extract a learning behavior feature sequence, and an initial learning behavior feature vector is generated; According to the initial learning behavior feature vector, the learning behavior data is subjected to time sequence alignment processing, the matching degree of the feature vector after alignment and a preset evaluation standard template is judged, and if the matching degree exceeds a preset threshold, it is determined that the learning behavior is preliminarily matched; The historical learning record data of students is extracted from a distributed storage system, a graph embedding method is used to fuse the historical learning record data and the learning behavior preliminary matching result, and a comprehensive learning behavior feature vector is obtained; Through a federal learning framework, an external data synchronization interface is called to acquire real-time classroom interaction data of students, whether the real-time data acquisition timestamp is within a preset time window is judged, and if the real-time data acquisition timestamp is within the preset time window, the real-time data validity is determined; According to the comprehensive learning behavior feature vector and the real-time data validity result, weighted fusion calculation is carried out, and a student dynamic evaluation score is obtained; If the dynamic evaluation score is higher than a preset threshold, an automatic optimization logic is executed through the federal learning framework, an encryption hash function is used to store the evaluation process record in a distributed manner, and an unalterable audit log is generated; According to the unalterable audit log, the related evaluation path is retrieved from the distributed storage network, the path integrity is judged, and if the path integrity is complete, the evaluation result is finally output. 2.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The step of acquiring the learning trajectory data of students from the multi-source heterogeneous learning behavior database through the distributed data acquisition node, using the deep semantic analysis algorithm to extract the learning behavior feature sequence, and generating the initial learning behavior feature vector comprises the following steps: The learning trajectory data in the multi-source heterogeneous learning behavior database is acquired from the distributed data acquisition node, and a distributed query protocol is used to obtain the original learning behavior data; The original learning behavior data is subjected to semantic analysis by using the word segmentation and dependency relationship analysis technology in the deep semantic analysis algorithm, and the learning behavior feature sequence is obtained; If the learning behavior feature sequence contains preset redundant information, the redundant information is filtered, and the filtered learning behavior feature sequence is obtained; The learning behavior feature sequence is converted into a feature vector sequence through a pre-trained deep learning embedding model, and the initial learning behavior feature vector is generated; According to the initial learning behavior feature vector, a nonlinear dimension reduction algorithm is used to optimize the vector dimension, and the optimized learning behavior feature vector is obtained; If the similarity of the optimized feature vector and the preset evaluation standard template exceeds a preset threshold, it is determined that the feature vector meets the evaluation requirement, and a structured learning behavior description is generated; Through the structured learning behavior description, a clustering algorithm is used to classify the learning behavior of students, and a classification result is obtained. 3.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The step of acquiring the learning trajectory data of students from the multi-source heterogeneous learning behavior database through the distributed data acquisition node, using the deep semantic analysis algorithm to extract the learning behavior feature sequence, and generating the initial learning behavior feature vector comprises the following steps: The learning behavior data is acquired and a feature vector is extracted therefrom, a keyword extraction and semantic correlation analysis method is used to generate a learning behavior feature vector; The learning behavior data is processed for time sequence alignment, a timestamp alignment technology is used to extract key time points, and aligned learning behavior data is generated; The aligned learning behavior data is converted into a learning behavior feature vector through a pre-trained deep learning embedding model; The cosine similarity of the learning behavior feature vector and a preset evaluation standard template is calculated to obtain a similarity value; If the similarity value exceeds a preset threshold, it is determined that the learning behavior is preliminarily matched, and a matching result is generated; For the matching result, a clustering analysis method is used to group students to obtain a grouping result; According to the grouping result, the learning behavior feature vector of each group of students is extracted to generate a ranking list. 4.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The step of extracting the student's historical learning record data from the distributed storage system, using a graph embedding method to fuse the historical learning record data and the learning behavior preliminary matching result, and obtaining a comprehensive learning behavior feature vector, comprises: The student's historical learning record data is extracted from the distributed storage system, and a preset index extraction rule is used to obtain a structured historical learning record data set; The weight values of each index are calculated through the structured historical learning record data set, and a weighted average method is used to generate a weighted historical learning record set; The learning behavior data is obtained from the learning behavior preliminary matching result, and a graph embedding method is used to fuse the weighted historical learning record set and the learning behavior data to generate an initial comprehensive learning behavior feature vector; If the dimension of the initial comprehensive learning behavior feature vector exceeds a preset threshold, a nonlinear dimension reduction algorithm is used for dimension reduction processing to obtain an optimized comprehensive learning behavior feature vector; According to the optimized comprehensive learning behavior feature vector, the similarity between vectors is calculated, and a cosine similarity algorithm is used to determine the learning behavior matching degree between students; Through the learning behavior matching degree between students, combined with a preset classification rule, the learning behavior level of the student is judged to generate a final learning behavior evaluation result; According to the final learning behavior evaluation result, a data correlation analysis method is used to associate the evaluation result with the historical learning record in the distributed storage system to generate a verification data set of the comprehensive learning behavior feature vector. 5.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The step of calling an external data synchronization interface through a federated learning framework to obtain real-time classroom interaction data of a student, and determining whether the real-time data collection timestamp is within a preset time window, if within the time window, determining the validity of the real-time data, comprises: The real-time classroom interaction data timestamp of the student is obtained from the external data source through the data synchronization interface and stored as first time data; A time window comparison logic is used to compare the first time data with a preset time window range, if the first time data is within the time window range, it is determined that the real-time data collection time is valid, and an effectiveness determination result is obtained; The effectiveness determination result is obtained through the federated learning framework interface to generate a real-time data validity state; According to the validity state, the data synchronization interface is called to obtain a complete data set of real-time classroom interaction, and the complete data set is stored as first real-time data; The first real-time data is verified for integrity and consistency by a data verification algorithm, and second real-time data that passes the verification is obtained; A learning behavior health score is generated by using a classification algorithm to perform learning behavior health assessment on the second real-time data; According to the learning behavior health score and the validity determination result, the final validity of the real-time data of the student is obtained. 6.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The step of performing weighted fusion calculation according to the comprehensive learning behavior feature vector and the real-time data validity result to obtain a student dynamic evaluation score comprises: Obtaining student information and real-time data validity result, judging data integrity and consistency through a preset data verification rule, and obtaining verified real-time data quality; According to the verified real-time data quality, extracting learning behavior evaluation indexes, and determining a comprehensive learning behavior feature vector by using a factor analysis method; If at least one index in the comprehensive learning behavior feature vector exceeds a preset threshold value, then the learning behavior evaluation indexes and the real-time data quality are fused by using a linear weighting method to obtain a preliminary fusion score; For the preliminary fusion score, a preset fusion weight distribution is used to perform weighted fusion calculation to obtain an adjusted fusion score; According to the adjusted fusion score, a logic regression model is applied to judge the learning behavior level of the student, and a dynamic evaluation score is outputted; If the dynamic evaluation score is lower than a preset threshold value, then the fusion weight distribution is adjusted by using a score calculation model, and the dynamic evaluation score is recalculated; The recalculated dynamic evaluation score is obtained, and a student comprehensive evaluation score is generated through evaluation result output. 7.The artificial intelligence model training method for educational assessment according to claim 1, wherein: The step of performing automatic optimization logic by using a federated learning framework if the dynamic evaluation score is higher than a preset threshold value, and using an encryption hash function to perform distributed storage on evaluation process records to generate an unforgeable audit log comprises: If the dynamic evaluation score is higher than a preset threshold value, then the data to be optimized is obtained by using the trigger logic of the federated learning framework, and an automatic optimization operation is performed to obtain an optimization result; The optimization result and the evaluation record are encrypted by using an encryption hash function to generate encrypted data; The encrypted data is written into a distributed storage system through a distributed storage network to generate a storage address; The data in the distributed storage system is obtained according to the storage address, and the data integrity is verified to obtain a verification state; If the verification state is complete, then an audit log is generated by using the federated learning framework to record the optimization and storage information; According to the content of the audit log, the log data is distributed by using a decentralized network to generate a distribution confirmation; Through the distribution confirmation, the consensus feedback of the network node is obtained, and the final distributed storage record state is obtained. 8.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The step of retrieving a related evaluation path from a distributed storage network according to the unforgeable audit log, judging the path integrity, and determining the final output of the evaluation result if the path is complete comprises: The audit log is obtained from the distributed storage network, a distributed node query protocol is used to extract a log data set containing a timestamp and a transaction hash, and an original audit log is obtained; For the original audit log, the log structure is parsed, the node sequence and hash chain in the evaluation path are extracted, and a path data set is generated; If the node sequence in the path data set is consistent with the preset distributed storage topology structure, the Merkle tree verification algorithm is used to calculate the hash value of each node in the path, the path integrity is judged, and the integrity verification result is obtained; According to the integrity verification result, if the integrity verification is passed, the transaction timestamp and associated metadata are extracted from the path data set to generate an evaluation data set; SHA-256 algorithm is used to calculate the hash value of the evaluation data set, and the uniqueness of the evaluation result is determined; Through the network communication protocol, the evaluation hash value is compared with the preset audit standard to determine whether the evaluation result meets the standard, and the final evaluation result is output. 9.The artificial intelligence model training method for educational evaluation according to claim 1, wherein: The method is implemented by an artificial intelligence model training system for educational evaluation, wherein the system comprises: A feature extraction module configured to obtain student learning trajectory data from a multi-source heterogeneous learning behavior database through a distributed data collection node, extract learning behavior feature sequences using a deep semantic parsing algorithm, and generate an initial learning behavior feature vector; A behavior matching module configured to perform time sequence alignment processing on learning behavior data according to the initial learning behavior feature vector, judge the matching degree of the aligned feature vector and the preset evaluation standard template, and determine the preliminary matching of learning behavior if the matching degree exceeds the preset threshold; A history fusion module configured to extract student historical learning record data from a distributed storage system, fuse the historical learning record data and the preliminary matching result of learning behavior using a graph embedding method, and obtain a comprehensive learning behavior feature vector; A real-time verification module configured to call an external data synchronization interface through a federated learning framework to obtain student real-time classroom interaction data, judge whether the real-time data collection timestamp is within a preset time window, and determine the validity of the real-time data if it is within the time window; A score calculation module configured to perform weighted fusion calculation according to the comprehensive learning behavior feature vector and the real-time data validity result to obtain a student dynamic evaluation score; An optimization execution module configured to execute automatic optimization logic through a federated learning framework if the dynamic evaluation score is higher than a preset threshold, use an encryption hash function to distribute the evaluation process record, and generate an unalterable audit log; An audit verification module configured to retrieve the relevant evaluation path from the distributed storage network according to the unalterable audit log, judge the path integrity, and determine the final output of the evaluation result if it is complete.

Images