Student training full-period management system based on artificial intelligence

Abstract

The invention discloses a student training full-period management system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the system comprises a student state sensing unit which is used for collecting multi-modal data, carrying out the multi-modal fusion analysis, and obtaining the comprehensive learning input degree; the knowledge graph construction unit is used for carrying out correlation degree calculation and establishing connection between the corresponding knowledge entities when the correlation degree is greater than a preset correlation degree threshold value so as to generate a cross-domain knowledge graph; the ability portrait evaluation unit is used for performing dynamic update analysis in combination to obtain the current mastery degree of the students on the knowledge points; and the learning path planning unit is used for carrying out learning readiness analysis and recommendation utility analysis and generating a personalized learning path. According to the invention, through multi-modal perception, cross-domain map construction and dynamic path planning, accurate and closed-loop management of student states, knowledge structures and ability development is realized, and the efficiency and effect of personalized training are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a student training full-cycle management system based on artificial intelligence. BACKGROUND

[0002] In the current information age, with the rapid development of frontier technologies such as artificial intelligence and big data, the education and training industry is undergoing profound changes. In order to meet the growing demand for multi-skilled compound talents, using technical means to improve training efficiency and achieve personalized teaching has become a core development trend. Against this background, various intelligent learning systems and platforms have emerged, which usually use knowledge graph technology to build the knowledge system of courses, analyze students' learning behavior data to establish a preliminary user portrait, and recommend appropriate learning resources based on the portrait. These systems try to structure complex teaching content and digitize implicit learning process, aiming to provide learners with more targeted learning experience beyond the traditional one-size-fits-all model, thereby promoting innovation and upgrading of talent training mode.

[0003] However, the existing technology still faces a series of deep-seated challenges and bottlenecks when applied to complex, especially cross-disciplinary, compound talent training scenarios. First, in the evaluation of students' learning state, the existing technology mostly relies on single-dimensional data such as online duration, answer accuracy, or students' subjective feedback, which makes it difficult to accurately and objectively understand students' real cognitive input and emotional state, leading to biased assessment of learning effectiveness. Second, in the construction of knowledge system, the knowledge graph created by the existing technology is mostly limited to a single discipline, forming independent "knowledge islands", lacking effective methods to explore and quantify the deep connections between knowledge in different fields, and being difficult to support real cross-disciplinary learning and innovation ability training. Third, due to the lack of accurate perception of students' state and the fragmentation of knowledge network, the existing system often fails to dynamically and truly reflect the subtle changes in students' knowledge mastery level when building students' ability portrait, for example, the system cannot distinguish between high scores obtained by guessing and achievements after real understanding. Finally, these deficiencies collectively lead to low efficiency of personalized learning path recommendation, with the system either recommending content that is too advanced for students to absorb due to lack of prerequisite knowledge, or recommending content that is too simple and cannot form effective challenges, failing to optimize learning efficiency and effectiveness. SUMMARY

[0004] The present application aims to provide a student training full-cycle management system based on artificial intelligence, which solves the problems in the background art.

[0005] To solve the above technical problems, the present application provides a student training full-cycle management system based on artificial intelligence, comprising: a student state perception unit for collecting multi-modal data of students and performing multi-modal fusion analysis to obtain comprehensive learning input degree;

[0006] A knowledge graph construction unit is configured to calculate the correlation degree of knowledge entities and establish a connection between corresponding knowledge entities when the correlation degree is greater than a preset correlation degree threshold to generate a cross-domain knowledge graph.

[0007] A capability portrait evaluation unit is configured to perform dynamic update analysis based on historical mastery degree, task performance and comprehensive learning input degree to obtain the current mastery degree of students on knowledge points.

[0008] A learning path planning unit is configured to perform learning preparation degree analysis and recommendation utility analysis based on the cross-domain knowledge graph and the current mastery degree to generate a personalized learning path.

[0009] Preferably, the multi-modal fusion analysis process is as follows:

[0010] The facial expression video stream, voice signal and system interaction log of the student are collected, and the facial expression video stream, voice signal and system interaction log are set as multi-modal data; the multi-modal data is processed to obtain visual modal score, auditory modal score and interactive modal score; the visual modal score, auditory modal score and interactive modal score are weighted and summed to obtain the comprehensive learning input degree.

[0011] Preferably, the visual modal score is obtained by processing the facial expression video stream through a preset convolutional neural network model; the auditory modal score is obtained by processing the voice signal through a preset recurrent neural network model; and the interactive modal score is obtained based on statistical analysis of the system interaction log.

[0012] Preferably, the correlation degree calculation process is as follows:

[0013] The entity description text and professional literature corpus of the knowledge entity are obtained; the entity description text is input into a preset language model for processing to obtain semantic similarity; the professional literature corpus is statistically analyzed to obtain context co-occurrence degree; and the semantic similarity and context co-occurrence degree are fused and weighted to obtain the correlation degree.

[0014] Preferably, the dynamic update analysis process is as follows:

[0015] The historical mastery degree of the student, the task performance score related to the knowledge points and the comprehensive learning input degree generated by the student state perception unit are acquired; the task performance score is multiplied by the comprehensive learning input degree to obtain the ability growth value; the historical mastery degree forgetting decay effect and the ability growth value are combined to perform update rate weighting processing to obtain the current mastery degree.

[0016] Preferably, the learning preparation degree analysis process is as follows:

[0017] Based on the cross-domain knowledge graph, a set of prerequisite knowledge points required for learning content is acquired; the current mastery degree of the student to each knowledge point in the set of prerequisite knowledge points is acquired; the current mastery degree of each knowledge point in the set of prerequisite knowledge points is subjected to average value processing to obtain a learning preparation degree factor of the learning content; when the learning content has no prerequisite knowledge requirement, the learning preparation degree factor is set to a preset value.

[0018] Preferably, the recommendation utility analysis process is as follows:

[0019] The importance of each knowledge point included in the learning content, the current mastery degree of the student to each knowledge point and the difficulty coefficient of the learning content are acquired; based on the importance and the current mastery degree, a knowledge gain value is calculated; the knowledge gain value is divided by the difficulty coefficient to obtain a knowledge gain efficiency; the learning preparation degree factor is multiplied by the knowledge gain efficiency to obtain the recommendation utility of the learning content.

[0020] Preferably, the generation process of the individualized learning path is as follows:

[0021] The recommendation utilities of all candidate learning contents are calculated, and all candidate learning contents are subjected to descending order sorting processing according to the recommendation utilities to generate an individualized learning path.

[0022] Compared with the prior art, the present application has the following beneficial effects:

[0023] 1. By establishing a multi-modal fusion perception mechanism, multi-dimensional information such as facial expressions, voice tones and system interaction behaviors of the student can be processed in parallel, and the heterogeneous data is converted into a continuous quantitative comprehensive learning input degree index, thereby realizing fine description of the entire state spectrum of the student from high immersion to complete distraction, providing a reliable basis for subsequent all individualized adjustments, and significantly improving the accuracy of the decision of the entire training system.

[0024] 2、The application proposes a knowledge association quantitative method fusing semantics and context, which can automatically explore and calculate the connection strength between knowledge entities in different fields, connect originally isolated discipline knowledge into a precise cross-field knowledge network, and clearly reveal the deep internal relationship between knowledge, which provides an indispensable knowledge network foundation for compound talent training, and makes it possible to plan learning paths across discipline boundaries.

[0025] 3、By combining the real-time learning investment of students, task performance and knowledge forgetting law, a dynamic evolving student ability portrait model is constructed, and when planning the learning path, the learning preparation degree evaluation is introduced, that is, whether the students have mastered the prerequisite knowledge necessary for learning new content is judged first, and this preparation degree evaluation and content recommendation mechanism ensures the scientificity and accessibility of the recommended path, enables students to learn efficiently in a step-by-step manner, and maximizes the individual learning efficiency and success rate. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings;

[0027] Figure 1 The logical block diagram of the system of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0029] Embodiment 1:

[0030] Please refer to Figure 1The application provides a student training full-cycle management system based on artificial intelligence, which comprises a student state sensing unit, a knowledge graph construction unit, an ability portrait evaluation unit and a learning path planning unit.

[0031] The student training full-cycle management system based on artificial intelligence provided by the embodiment aims to overcome the challenges of the prior art in the training of compound talents, such as difficulty in cross-domain knowledge fusion, inaccuracy in student ability portrait modeling and low efficiency in personalized learning path recommendation. The system forms an intelligent, personalized and logically closed training management process through the collaborative work of the student state sensing unit, the knowledge graph construction unit, the ability portrait evaluation unit and the learning path planning unit. The four units are closely linked, starting from the accurate quantification of the student state, to the deep construction of the knowledge system, to the dynamic evaluation of the personal ability, and finally outputting a highly adaptive learning path to realize precise management of the student growth throughout the whole cycle.

[0032] Embodiment 2

[0033] The multi-modal fusion analysis process is as follows: collecting facial expression video stream, voice signal and system interaction log of the student, and setting the facial expression video stream, voice signal and system interaction log as multi-modal data; processing the multi-modal data to obtain visual modal score, auditory modal score and interactive modal score respectively; performing weighted sum processing on the visual modal score, auditory modal score and interactive modal score to obtain the comprehensive learning input degree;

[0034] The visual modal score is obtained by processing the facial expression video stream through a preset convolutional neural network model; the auditory modal score is obtained by processing the voice signal through a preset recurrent neural network model; and the interactive modal score is obtained based on statistical analysis of the system interaction log;

[0035] In the present embodiment, the multi-modal fusion analysis process performed by the learner state perception unit has the technical gist of continuous quantitative evaluation of the learner's learning engagement, rather than traditional discrete state classification, thereby providing higher-precision input for subsequent analysis modules; the process specifically manifests as follows: the learner's facial expression video stream, speech signal, and interaction log with the learning system, such as answer response time, data query frequency, etc., are collected by the camera, microphone, and system background respectively as multi-modal data sources; for these data, the preset models called by the system all have their specific structural considerations: a preset convolutional neural network model is used to process the facial expression video stream, which can effectively capture facial micro-expression changes for emotion classification due to its powerful spatial feature extraction capability, thereby generating a visual modality score σ v ; specifically, the convolutional neural network model can adopt the classic ResNet-18 architecture and load the pre-trained weights on large-scale facial expression public datasets such as AffectNet, and then fine-tune a small amount of labeled data collected by the present system to adapt to specific training scenarios; the final output layer of the model is modified to a single neuron, which outputs a continuous value in the [0,1] interval through the Sigmoid activation function, i.e., the visual modality score σ v ; a preset recurrent neural network model is used to process the speech signal, which is good at capturing temporal information and can analyze the emotional fluctuations related to engagement from the learner's tone, speed, etc., to generate an auditory modality score σ a ; specifically, the recurrent neural network model can adopt a long short-term memory network structure with an attention mechanism; the input features are the mel-frequency cepstral coefficients extracted from each frame of the original speech signal, and the network learns the time series variation of the speech features, and the final output is also processed through the Sigmoid activation function to obtain the auditory modality score σ a ; at the same time, by statistically analyzing the interaction log of the learner with the learning system, the behavior participation of the learner can be quantified to obtain an interaction modality score σ i ; the statistical analysis specifically includes: calculating a series of quantifiable behavior indicators such as the learner's mouse click frequency, page average dwell time, data query times, and note function usage frequency per unit time; after normalizing these indicators respectively, a comprehensive interaction modality score σ i ; is fitted through linear weighting or training a simple logistic regression model; all these scores σ are normalized to ensure dimensional consistency; on this basis, a weighted sum formula is used to calculate the comprehensive learning engagement;

[0036] The theoretical basis of the formula is derived from the comprehensive observation of human learning behavior, that is, effective learning is a unified embodiment of the synergy of multiple sensory channels and behavior patterns. The technical motivation for creating this formula is to break through the bottleneck of single-modal sentiment analysis and to build a continuous quantitative index that can more truly and comprehensively reflect the cognitive and emotional involvement of learners by fusing multi-dimensional information.

[0037]

[0038] Ψ represents the final calculated comprehensive learning input degree, which is a dimensionless scalar in the interval [0, 1];

[0039] i represents different data modalities, such as vision, hearing, and interaction;

[0040] σ i represents the normalized score under modality i, which is derived from the visual modality score σ v obtained by processing facial expression video streams with a convolutional neural network model, the auditory modality score σ a obtained by processing speech signals with a recurrent neural network model, and the interaction modality score σ i obtained based on statistical analysis of interaction logs;

[0041] ω i represents the weight coefficient of modality i, which is automatically calibrated on a sample data set labeled with true input degree through a machine learning optimization algorithm, aiming to minimize the deviation between the calculated Ψ value and the true label, and the sum of all weights is 1, i.e. ∑ω i = 1;

[0042] The machine learning optimization algorithm can be constructed as a constrained linear regression problem; on a small-scale sample data set artificially labeled with true learning input degree (for example, subjective scores of 1-10 given by education psychology experts after reviewing the video, then normalized to the interval [0, 1]), the objective function is to minimize the mean square error between the predicted comprehensive learning input degree Ψ and the true labeled input degree Ψ true : minimize ∑(Ψ-Ψ true ) 2 , while satisfying the constraints ω i ≥ 0 and ∑ω i = 1, the optimal weight combination {ω v , ω a , ω i} is obtained by solving the optimization algorithm such as gradient descent;

[0043] The formula is used by the student state perception unit to calculate the comprehensive learning input degree Ψ; compared with the binary judgment of attention or distraction in the prior art, the continuous value Ψ output by the method can more subtly depict the entire state spectrum from complete disengagement to high immersion; this accurate input degree value will be the core basis for subsequent ability portrait evaluation correction, directly affecting the calculation result of the mastery degree, thereby converting passive perception of the student state into active correction of the student ability evaluation, realizing the logical association and data flow between the first key modules in the system.

[0044] Embodiment 3:

[0045] The correlation degree calculation process is as follows: obtaining the entity description text of the knowledge entity and the professional literature corpus; inputting the entity description text into a preset language model for processing to obtain the semantic similarity; performing statistical analysis on the professional literature corpus to obtain the context co-occurrence degree; and performing fusion coefficient weighting processing on the semantic similarity and the context co-occurrence degree to obtain the correlation degree.

[0046] In the embodiment, the correlation degree calculation process performed by the knowledge graph construction unit has the core contribution of proposing a method capable of quantifying the connection strength between knowledge entities in different fields, thereby effectively solving the technical pain point that cross-field knowledge is difficult to fuse; the process starts with obtaining two kinds of core data: the descriptive text of each knowledge entity, and a large-scale professional literature corpus; in order to clarify the internal logic and understand the knowledge from the abstract semantic level, the entity description text is input into a pre-trained language model, such as a BERT model based on the Transformer architecture; such a model has deep language understanding ability through pre-training on a large amount of text, and can convert the text into a high-dimensional embedding vector; by calculating the cosine similarity between these vectors, the semantic similarity S i between the knowledge entities e j and e ij is obtained; at the same time, in order to explore the association from the specific application context, the system performs statistical analysis on the professional literature corpus, and obtains the context co-occurrence degree C ij of the two knowledge entities by calculating the normalized frequency of their appearance in the same context window, such as a sentence or a paragraph.

[0047] Specifically, the statistical analysis can use the point mutual information algorithm to calculate the context co-occurrence degree; first, a sliding window with a size of N (for example, N = 10) is set, the entire professional literature corpus is traversed, and the number of times P(i,j) that the knowledge entities e i and e j appear together in the same window, and the number of times P(i) and P(j) that they appear independently are counted; the context co-occurrence degree C ij can be calculated as:

[0048]

[0049] C ij represents the semantic similarity between knowledge entities e i and e j ;

[0050] P(i,j) represents the co-occurrence degree of knowledge entities e i and e j in the same window;

[0051] P(i) and P(j) represent the independent occurrence times of knowledge entities e i and e j , respectively;

[0052] and the results are normalized;

[0053] The design idea of this formula draws on the hybrid similarity calculation theory in the field of knowledge engineering, and its fundamental motivation lies in the fact that single-dimensional similarity judgment often has bias; for example, relying only on semantics may incorrectly associate two concepts that are similar in theory but have nothing to do in application, and relying only on co-occurrence may ignore knowledge that is logically closely connected although it does not frequently co-occur in text; therefore, this technical solution attempts to combine abstract semantic information and specific contextual information to more comprehensively and accurately discover deep potential associations between knowledge;

[0054] R ij = αS ij + (1-α)C ij

[0055] R ij represents the final association degree between knowledge entities e i and e j ;

[0056] S ij represents the semantic similarity between two entities, and its value is calculated by cosine similarity from the embedding vectors output by the pre-trained language model;

[0057] C ij represents the context co-occurrence degree, which is the normalized frequency obtained after statistical analysis of a large-scale professional literature corpus;

[0058] α is a fusion coefficient that can be adjusted according to the knowledge attributes;

[0059] To realize the systematization of this adjustment, one of the following methods can be used: one is to pre-label the theoretical or application type category label for each knowledge entity in the knowledge base, and the system automatically applies the preset α value (such as 0.7 and 0.3 respectively) according to the label; the second is to automatically find the α value that can make the F1 score of the correlation degree calculation optimal through grid search and other optimization methods on a verification set containing known associated and non-associated knowledge pairs;

[0060] The formula is used by the knowledge graph construction unit to calculate the correlation degree R between any two knowledge entities ij ; when the calculated R ij value exceeds a preset correlation degree threshold θ R , the system establishes a connection edge between the two knowledge entities; the setting logic of the threshold θ R here is not arbitrarily specified, but is determined based on the precision and recall rate balance analysis of historical data, aiming to maximize the number of effective correlation discoveries while controlling the proportion of false correlations within an acceptable range; the specific determination method is as follows: first, manually construct a test set containing positive samples (confirmed correlation) and negative samples (confirmed non-correlation); then, traverse all possible values of θ R in a reasonable interval (such as [0.1, 0.9]) with a step of 0.05; for each value, calculate its precision and recall rate on the test set, and calculate the F1 score:

[0061]

[0062] F1 is the F1 score, used to balance the precision and recall rate;

[0063] Precision is the precision;

[0064] Recall is the recall rate;

[0065] Finally, the θ R that makes the F1 score reach the maximum value is selected as the preset correlation degree threshold;

[0066] Through repeated execution of this process, the system can automatically construct a large and precise structure of cross-domain knowledge graph; this graph not only contains the knowledge dependency relationship within the same field, but more importantly, reveals the bridge between different domain knowledge, providing an indispensable knowledge network foundation for cultivating compound talents.

[0067] Embodiment 4:

[0068] The dynamic updating analysis process is as follows: obtaining the historical mastery degree of the trainee, the task performance score related to the knowledge point, and the comprehensive learning input degree generated by the trainee state perception unit; multiplying the task performance score and the comprehensive learning input degree to obtain the ability growth value; combining the forgetting decay effect of the historical mastery degree and the ability growth value to perform updating rate weighting processing to obtain the current mastery degree;

[0069] In the embodiment, the core of the dynamic updating analysis process performed by the ability portrait evaluation unit is to accurately model the learning process as a dynamic evolution process affected by the internal learning state and the external time effect; the process obtains three key inputs: the mastery degree of the trainee on the knowledge point at the last time, the task performance score related to the knowledge point in the current learning activity, and the real-time comprehensive learning input degree calculated by the aforementioned trainee state perception unit;

[0070] The derivation of the formula is logically an application and extension of the previous comprehensive learning input degree formula, and its theoretical basis is the innovative combination of the classic Ebbinghaus forgetting curve model and the cognitive load theory in educational psychology; the forgetting curve model inspires the simulation of the natural decay of knowledge over time in the formula, and the cognitive load theory suggests that the learning effect depends not only on the task result, but also on the mental state in the learning process; the technical motivation is to create a dynamic model that can truly reflect the changes in ability, avoiding the one-sidedness of the traditional evaluation method that equates a high score in a test to complete mastery of the ability;

[0071] p k (t)=(1-λ)p k (t-1)+λ·π k (t)·Ψ(t)p k (t) represents the mastery degree of the trainee on the knowledge point e k at the current time t, which is a dimensionless scalar;

[0072] p k (t-1) is the mastery degree of the trainee at the last time;

[0073] λ is the learning updating rate, which is a configurable parameter, and its source is set according to the length and nature of the training period, for example, a higher λ value is set for short-term crash training to achieve rapid ability updating, and a lower value is set for long-term consolidation learning; for example, for a one-week short-term crash training, a higher updating rate such as λ = 0.4 can be set to quickly reflect the improvement of the trainee's ability; while for a course lasting a semester, a lower updating rate such as λ = 0.1 can be set to smooth the short-term fluctuations in the learning process and more stably evaluate the long-term mastery;

[0074] In this formula, the (1-λ) term is the forgetting factor, which implicitly embodies the effect of Ebbinghaus forgetting curve: after each round of update, the mastery p k (t-1) is not completely preserved, but decays by a proportion of (1-λ), which simulates the natural forgetting process of knowledge over time, while new learning achievements are supplemented by a proportion of λ; π k (t) is the performance score of the learner in the current learning task related to the knowledge point e k , such as online tests, programming assignments, etc., which is also normalized; Ψ(t) is exactly the comprehensive learning input degree of the learner in performing this task, calculated by the first formula;

[0075] This formula is used by the ability profile evaluation unit to dynamically update the mastery p k (t) of each knowledge point of the learner; the key design here is the multiplication relationship between task performance π k (t) and learning input degree Ψ(t); this design embodies a core custom rule: only excellent performance in a state of sufficient input is recognized by the system as a real and effective ability improvement; for example, a learner may guess and get the right answer in a test (high π k (t)), but if the system judges through multi-modal perception that he was in a distracted state at the time (low Ψ(t)), his final ability growth value will also be reduced accordingly; this mechanism greatly improves the authenticity and credibility of the ability profile, ensuring that the evaluation results accurately reflect the learner's real inner ability, rather than superficial task scores; this successfully couples the learning state perception and ability evaluation modules, establishing a direct data flow and logical dependence.

[0076] Embodiment 5:

[0077] The learning preparation degree analysis process is as follows: based on the cross-domain knowledge graph, the set of prerequisite knowledge points required for learning content is obtained; the current mastery of each knowledge point in the set of prerequisite knowledge points is obtained; the current mastery of each knowledge point in the set of prerequisite knowledge points is averaged to obtain the learning preparation degree factor of the learning content; when the learning content has no prerequisite knowledge requirement, the learning preparation degree factor is set to a preset value;

[0078] The recommendation utility analysis process is as follows: the importance of each knowledge point included in the learning content, the current mastery of each knowledge point by the learner, and the difficulty coefficient of the learning content are obtained; based on the importance and current mastery, the knowledge gain value is calculated; the knowledge gain value is divided by the difficulty coefficient to obtain the knowledge gain efficiency; the learning preparation degree factor is multiplied by the knowledge gain efficiency to obtain the recommendation utility of the learning content;

[0079] The personalized learning path is generated by calculating the recommendation utility of all candidate learning contents, sorting all candidate learning contents in descending order of the recommendation utility, and generating the personalized learning path.

[0080] In the final stage of the embodiment, the learning path planning unit integrates the results of all the aforementioned modules, and generates a truly personalized learning path for the learner through a hierarchical and logically rigorous analysis process; this process first explicitly assesses the learner's learning readiness for each candidate learning content to ensure the accessibility of the recommended content;

[0081] The learning readiness for a new knowledge essentially depends on the average mastery of all its prerequisite knowledge; the motivation for setting up this formula is to solve the common skipping problem in learning recommendation, that is, the system directly recommends a senior knowledge point without thinking when it finds that the learner is weak in this knowledge point, but ignores the fact that the learner may not even master its foundation, thus leading to ineffective recommendation and damaging the learner's confidence;

[0082]

[0083] Φ j is the learning readiness factor of the learner for the learning content c j ;

[0084] P j is the set of prerequisite knowledge points required for the learning content c j , which is not manually specified but directly and automatically provided by the dependency relationship edges in the previously constructed knowledge graph;

[0085] p m (t) is the mastery degree of the learner at the current time t for a certain prerequisite knowledge point e m , which is derived from the calculation results of the third mastery degree updating formula;

[0086] |P j | is the total number of prerequisite knowledge points;

[0087] In special cases, if a learning content c j has no prerequisite knowledge requirement, its learning readiness factor Φ j is set to a preset value of 1;

[0088] The readiness factor Φ j calculated by this formula becomes a key prerequisite constraint indicator; in the subsequent utility calculation, a lower Φ j value will significantly lower the final recommendation priority of a learning content, thus effectively avoiding the recommendation of content to the learner for which the learner does not have the learning foundation;

[0089] After the readiness is calculated, the system then calculates the final recommendation utility of each learning content, which is a comprehensive evaluation combining readiness, knowledge gain and learning cost;

[0090] The derivation of this formula follows a core principle, which is to maximize the learning efficiency adjusted by readiness; its technical motivation lies in creating a more intelligent decision-making model than traditional recommendation systems; traditional models may only focus on what the learner lacks to fill in, while this model further considers what to learn under the premise of the learner's current readiness to maximize the effectiveness; it considers the three dimensions of benefit (knowledge gain), cost (difficulty) and feasibility (readiness) in a comprehensive manner;

[0091]

[0092] U j is the final recommendation utility of learning content c j ;

[0093] Φ j is the learning readiness factor calculated in the previous step;

[0094] C j represents the set of knowledge points contained in learning content c j ; w k is the importance weight of knowledge point e k , which can be set by domain experts;

[0095] w k is the importance weight of knowledge point e k ;

[0096] Specifically, it can be scored by multiple domain experts on the importance of knowledge points in the interval [1, 10], and then normalized, for example, to ensure that the importance weight of all knowledge points in the same learning content c j is 1, that is, p k (t) is the current mastery of knowledge point e k by the learner; p k (t) is the current mastery of knowledge point e k by the learner, derived from the third formula; therefore, the sum term in the numerator represents the total knowledge gain value that the learner can obtain after completing the learning content;

[0097] D j in the denominator is the learning content c jdifficulty coefficient of the whole, which is a normalized dimensionless scalar; its source can be pre-set by experts based on teaching experience, or dynamically determined by statistical analysis of historical student learning behavior data such as average completion time, first attempt failure rate, etc. and mapping to a standardized difficulty interval such as [1, 5], ensuring dimensional consistency; the formula divides the total knowledge gain value by the difficulty coefficient to obtain the knowledge gain efficiency, and then multiplies it by the learning preparation degree factor to finally obtain the comprehensive recommended utility;

[0098] Specifically, the difficulty coefficient D j can be dynamically calculated as follows: from historical data, obtain the average completion time T j , the first attempt failure rate F avg,j and the number of help H rate,j of all students who have learned the content c avg,j ; normalize these indicators in all learning content to obtain T norm,j , F norm,j and H norm,j ; the difficulty coefficient D j can be obtained by weighted sum of these normalized indicators:

[0099] D j =β1T norm,j +β2F norm,j +β3H norm,j

[0100] D j is the overall difficulty coefficient of the learning content c j ;

[0101] β1,β2,β3 are preset weights, and β1+β2+β3=1;

[0102] T norm,j , F norm,j , H norm,j are the values obtained by normalizing the average completion time, the first attempt failure rate and the number of help in the historical data, respectively

[0103] The result can be further mapped to a standardized difficulty interval such as [1, 5];

[0104] The structure of the formula clearly reflects its design logic; the preparation degree factor Φ j is multiplied by the subsequent knowledge gain efficiency part, which means that even if the potential knowledge gap of a topic is huge, if the mastery of its prerequisite knowledge is insufficient (Φ j is very low), its final recommended utility U jThe learning content recommendation list will also be significantly inhibited; this mechanism ensures that the top of the recommendation list is always the content that the student both needs and has the ability to efficiently absorb; based on this, the generation process of the personalized learning path is the recommendation utility U j The calculation is performed, and the global descending order is sorted according to the U j value; the personalized learning path generated for the student is logically coherent, has clear priorities, and is the most efficient; this not only improves learning efficiency and experience, but also realizes the whole-cycle intelligent management from state perception to path planning through a complete data and logic loop, and exhibits the significant progress of the present application over the prior art.

[0105] Compared with the prior art, the student training whole-cycle management system based on artificial intelligence disclosed by the present application has made significant technical progress through its unique unitized collaborative design and innovative algorithm model, which is embodied in the following aspects:

[0106] The primary difference and progress of the present scheme is that the student state perception unit is established, solving the problem of single dimension and insufficient precision of the prior art in evaluating the learning state of the student; the prior art usually relies on the task performance or subjective feedback of the student to evaluate the learning effect, and this method cannot penetrate the surface behavior and understand the real cognitive and emotional investment of the student; the multi-modal fusion analysis process in the present scheme uses a pre-set convolutional neural network model to process facial expression video streams, and uses a pre-set recurrent neural network model to analyze speech signals, and combines the statistics of system interaction logs to realize parallel and deep perception of visual, auditory and interactive behavior;

[0107] The comprehensive learning input degree formula applied in the above process is:

[0108]

[0109] Ψ is the comprehensive learning input degree;

[0110] i is the mode;

[0111] ω i is the mode weight;

[0112] σ i is the mode score;

[0113] It is not an isolated mathematical construct, but a direct mathematical expression of the technical step of weighting and summing the mode scores described in the specification; the practical significance of this formula is to condense multiple heterogeneous and nonlinear state indicators into a quantitative and continuous input degree scalar Ψ; this enables the subsequent ability profile evaluation unit to introduce a correction factor reflecting the true learning quality when updating the student's mastery degree; for example, the formula applied in the ability dynamic update analysis process:

[0114] p k (t)=(1-λ)p k (t-1)+λ·π k (t)·Ψ(t)p k is the mastery degree;

[0115] t is time;

[0116] λ is the update rate;

[0117] π k is the task performance;

[0118] Ψ is the input degree;

[0119] The logical deduction thereof explicitly depends on the output of the previous Ψ value; when a student's task performance score π k (t) is high, but its input degree Ψ(t) is extremely low, the ability growth value λ·π k (t)·Ψ(t) will be significantly weakened; this precisely discriminates the false performance obtained by relying on short-term memory or guessing, ensuring the authenticity and accuracy of the student's ability portrait, which is difficult for the prior art to achieve;

[0120] Secondly, the present scheme effectively overcomes the barrier of cross-domain knowledge fusion difficulty in complex talent training scenarios through the knowledge graph construction unit; the knowledge system constructed by the prior art is mostly an independent tree structure within the field, and the horizontal association between knowledge is seriously lacking; the correlation degree calculation process proposed by the present scheme has the correlation degree formula applied thereto:

[0121] R ij =αS ij +(1-α)C ij

[0122] R ij is the correlation degree;

[0123] α is the fusion coefficient;

[0124] S ij is the semantic similarity;

[0125] C ij is the context co-occurrence degree;

[0126] has a clear derivable path; as detailed in the specification, the semantic similarity S ij is obtained by processing the entity description text through a preset language model, and the context co-occurrence degree C ij can be calculated as:

[0127]

[0128] C ijrepresenting knowledge entity e i and e j ;

[0129] P(i,j) represents the number of times knowledge entities e i and e j co-occur in the same window;

[0130] P(i) and P(j) represent the number of times knowledge entities e i and e j independently appear respectively;

[0131] Considering the range of this value, it needs to be normalized to be integrated with semantic similarity S ij , for example, a Sigmoid function can be used to map it to the interval [0, 1]:

[0132]

[0133] The C ij′ used in subsequent calculations is the normalized one, and this formula is a direct embodiment of the weighting of the fusion coefficient of the two kinds of metrics;

[0134] The value of this formula lies in that it can discover and quantify the deep connection between different knowledge entities; taking the cultivation of a financial technology talent as an example, the existing technology may separate his financial knowledge from programming skills; while this scheme can find through calculation that there is a high correlation R ij between the option pricing model in finance and the Monte Carlo simulation in computer science, because they frequently co-occur in professional literature and have semantic similarities in solving randomness problems; based on this calculation result, the system establishes a connection between the two and generates a cross-domain knowledge graph; the construction of such a graph makes knowledge no longer an island, providing a structured foundation for subsequent planning of learning paths that can cross disciplinary boundaries;

[0135] Finally, this scheme solves the core pain points of the existing technology of single personalized recommendation path and low efficiency through the learning path planning unit; the existing recommendation system often only focuses on the knowledge short board of students, while ignoring their learning readiness, resulting in recommended content often being too difficult to be effectively absorbed; the learning readiness analysis and recommendation utility analysis of this scheme form a logically rigorous decision-making process; the learning readiness factor formula is:

[0136]

[0137] Φ j is the readiness factor;

[0138] P j is the set of prerequisite knowledge points;

[0139] p m is the mastery degree, t is time;

[0140] The derivation logic is rooted in the average value processing of the current mastery degree of each knowledge point in the set of prerequisite knowledge points described in the specification, which accurately evaluates the feasibility of the learner's learning new content;

[0141] On this basis, the recommendation utility formula is:

[0142]

[0143] U j is the recommendation utility;

[0144] Φ j is the preparation degree factor;

[0145] C j is the knowledge point set;

[0146] w k is the importance;

[0147] p k is the mastery degree;

[0148] D j is the difficulty coefficient;

[0149] Its composition is the logical collection and sublimation of all the calculation results in the previous sequence; it not only considers the ratio of knowledge gain w k (1-p k (t)) and difficulty D j , i.e. learning efficiency, but also uses the preparation degree factor Φ j to weight and adjust this efficiency; continuing the example of the financial technology talent, even if the system judges that the Monte Carlo simulation mastery degree p k (t) of the learner is very low, and the knowledge gap is huge, but if the mastery degree of the prerequisite knowledge of probability theory is also very low, resulting in the preparation degree factor Φ j tends to zero, then the recommendation utility U j of the Monte Carlo simulation will be greatly reduced; the system will turn to recommend the learning content of probability theory foundation with high priority; this mechanism ensures that the generation process of the personalized learning path not only makes up for the knowledge gap, but also follows the cognitive rule of gradual progress, and realizes the maximization of learning efficiency and success rate;

[0150] In summary, the present application is not a single technical point improvement, but a complete set of closed-loop systems connected with each other and data-driven from learner state perception, knowledge graph construction, ability portrait evaluation to learning path planning, which realizes intelligent, precise and efficient management of the whole cycle of learner training, and has essential differences and significant progress compared with the prior art.

[0151] The above merely describes the preferred embodiments of the present application, but is not intended to limit the present application in other forms. Any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes, and apply to other fields. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solution content of the present application still falls within the protection scope of the present application.

Claims

1. A student training full-cycle management system based on artificial intelligence, characterized in that, include: The learner status perception unit is used to collect learners' multimodal data and perform multimodal fusion analysis to obtain the overall learning engagement. The knowledge graph construction unit is used to calculate the correlation between knowledge entities and establish connections between corresponding knowledge entities when the correlation is greater than a preset correlation threshold, so as to generate a cross-domain knowledge graph. The competency profile assessment unit is used to dynamically update and analyze the student's current mastery of the knowledge points by combining historical mastery, task performance, and overall learning engagement. The learning path planning unit is used to perform learning readiness analysis and recommendation utility analysis based on cross-domain knowledge graphs and current mastery levels, and generate personalized learning paths.

2. The AI-based full-cycle student training management system according to claim 1, characterized in that, The multimodal fusion analysis process is as follows: The system collects video streams of trainees' facial expressions, voice signals, and system interaction logs, and sets these data as multimodal data. The multimodal data is processed to obtain visual modality scores, auditory modality scores, and interactive modality scores. The visual modality scores, auditory modality scores, and interactive modality scores are then weighted and summed to obtain the overall learning engagement.

3. The AI-based full-cycle student training management system according to claim 2, characterized in that, The visual modality score is obtained by processing the facial expression video stream using a preset convolutional neural network model; The auditory modality score is obtained by processing the speech signal through a pre-defined recurrent neural network model; the interactive modality score is obtained based on statistical analysis of the system's interaction logs.

4. The AI-based full-cycle student training management system according to claim 1, characterized in that, The correlation calculation process is as follows: The process involves acquiring entity description texts of knowledge entities and a corpus of professional literature; inputting the entity description texts into a pre-defined language model for processing to obtain semantic similarity; performing statistical analysis on the corpus of professional literature to obtain context co-occurrence; and then weighting the semantic similarity and context co-occurrence using a fusion coefficient to obtain the relevance.

5. The AI-based full-cycle student training management system according to claim 1, characterized in that, The dynamic update analysis process is as follows: The system obtains the student's historical mastery level, the score of task performance related to the knowledge points, and the comprehensive learning engagement level generated by the student's status perception unit; it multiplies the task performance score by the comprehensive learning engagement level to obtain the ability growth value; and it combines the forgetting decay effect of historical mastery level with the ability growth value, and performs update rate weighting to obtain the current mastery level.

6. The AI-based full-cycle student training management system according to claim 1, characterized in that, The learning readiness analysis process is as follows: Based on cross-domain knowledge graphs, obtain the set of prerequisite knowledge points required for the learning content; obtain the current mastery level of students on each knowledge point in the prerequisite knowledge point set; average the current mastery level of each knowledge point in the prerequisite knowledge point set to obtain the learning readiness factor of the learning content. When the learning content does not require prior knowledge, the learning readiness factor is set to the preset value.

7. The AI-based full-cycle student training management system according to claim 6, characterized in that, The recommendation utility analysis process is as follows: The importance of each knowledge point in the learning content, the student's current mastery of each knowledge point, and the difficulty coefficient of the learning content are obtained; based on the importance and current mastery, the knowledge gain value is calculated. Divide the knowledge gain value by the difficulty coefficient to obtain the knowledge gain efficiency. Multiplying the learning readiness factor by the knowledge gain efficiency yields the recommended utility of the learning content.

8. The AI-based full-cycle student training management system according to claim 7, characterized in that, The process of generating the personalized learning path is as follows: The recommendation utility of all candidate learning content is calculated, and all candidate learning content is sorted in descending order according to the recommendation utility to generate a personalized learning path.

Images