An intelligent education evaluation method and system based on learning behavior analysis

Abstract

This invention relates to the field of intelligent education and learning behavior analysis technology, specifically to an intelligent education assessment method and system based on learning behavior analysis. The method includes: acquiring multidimensional micro-behavioral sequence data and the current interaction time benchmark of an interactive terminal; extracting spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features; calculating the behavioral information entropy sequence using a time-series sliding window; mapping this to internal and external cognitive load states via a Hidden Markov Model; calculating the probability time point of future energy exhaustion using a power exhaustion risk prediction model based on the current interaction time benchmark; generating a learning sustainability index based on the time difference; and triggering an adaptive intervention command when the index falls below a preset safety threshold, otherwise maintaining the current output strategy. This invention achieves real-time monitoring of learning sustainability, early identification of cognitive overload, and dynamic control of the pace of teaching output.

Description

Technical Field

[0001] This invention relates to the field of intelligent education and learning behavior analysis technology, specifically to an intelligent education assessment method and system based on learning behavior analysis. Background Technology

[0002] Currently, intelligent education assessment mechanisms are usually based on the accuracy of answering questions, online time, video completion rate, or changes in performance over time. This approach makes it difficult to continuously identify micro-behavioral changes, cognitive load accumulation, and the risk of motivation exhaustion during the learning process. When students have already shown a significant decline in learning motivation, decreased interaction activity, or unstable learning efficiency, the platform often cannot intervene accurately in advance, reducing the timeliness and effectiveness of learning status assessment and teaching regulation. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent educational assessment method and system based on learning behavior analysis. This aims to improve upon existing assessment mechanisms that rely solely on results such as answer accuracy or online time, which struggle to continuously identify micro-behavioral changes, cognitive load accumulation, and the risk of motivation exhaustion. This results in an inability to intervene accurately and proactively before learning motivation declines or efficiency becomes unstable, reducing the timeliness and effectiveness of status assessment and teaching regulation. The purpose of this invention is achieved through the following technical solution:

[0004] On the one hand, this invention provides an intelligent educational assessment method based on learning behavior analysis, comprising the following steps:

[0005] Acquire multidimensional microscopic behavioral sequence data and current interaction time reference of the interactive terminal, and extract spatiotemporal trajectory features, temporal rhythm features and physiological mapping features from the multidimensional microscopic behavioral sequence data;

[0006] Based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, multiple behavioral information entropy values ​​are calculated according to a time series sliding window to construct a behavioral information entropy value sequence.

[0007] The behavioral information entropy value sequence is processed by a pre-set Hidden Markov Model to map the behavioral information entropy value sequence to internal cognitive load state and external cognitive load state; combined with the internal cognitive load state and external cognitive load state, as well as the current interaction time base, the future power exhaustion probability time point is calculated by a pre-set power exhaustion risk prediction model.

[0008] The system calculates the time difference between the future probability of power depletion and the current interaction time base, generates a learning sustainability index based on the time difference, and triggers an adaptive intervention command to the interactive terminal to change its task delivery or multimedia data output status when the learning sustainability index is lower than a preset safety threshold. When the learning sustainability index is higher than or equal to the preset safety threshold, the current output strategy is maintained.

[0009] Preferably, the spatiotemporal trajectory features include touch movement trajectory data and pressure intensity data; the temporal rhythm features include interaction time variance data and multimedia control frequency data; and the physiological mapping features include error correction modification number data and operation dwell time data.

[0010] Preferably, based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, multiple behavioral information entropy values ​​are calculated according to a time series sliding window to construct a behavioral information entropy value sequence, including: fusing spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features to generate a multimodal feature vector;

[0011] The multimodal feature vectors are segmented according to the time series sliding window, and the combined values ​​of the multimodal feature vectors are mapped to pre-declared discrete behavior set states. The proportion of each discrete behavior set state in the time series sliding window is calculated to calculate the probability distribution of each discrete behavior set state in the time series sliding window.

[0012] Based on the probability distribution, multiple behavioral information entropy values ​​are calculated by solving the negative sum of the product of the probability distribution and its logarithm, and then spliced ​​together according to the time series to form a sequence of behavioral information entropy values.

[0013] Preferably, the behavioral information entropy value sequence is processed by a preset Hidden Markov Model to map the behavioral information entropy value sequence to an internal cognitive load state and an external cognitive load state, including: using the behavioral information entropy value sequence as the observation sequence input of the preset Hidden Markov Model; and decoding the observation sequence through a preset state transition matrix and a preset observation probability matrix in the preset Hidden Markov Model.

[0014] Output the hidden state sequence corresponding to the observation sequence. Each hidden state in the hidden state sequence is mapped to a preset two-dimensional state space. The two dimensions of the two-dimensional state space are decoupled and resolved into internal cognitive load state and external cognitive load state.

[0015] Preferably, by combining the internal cognitive load state and the external cognitive load state, as well as the current interaction time reference, a preset energy exhaustion risk prediction model is used to calculate the future energy exhaustion probability time point, including: obtaining the forgetting curve decay parameter and emotional exhaustion parameter extracted from the historical interaction data of the interactive terminal; and inputting the internal cognitive load state, the external cognitive load state, the forgetting curve decay parameter and the emotional exhaustion parameter into the preset energy exhaustion risk prediction model.

[0016] Using the survival analysis algorithm in the preset power exhaustion risk prediction model, the probability distribution curve of the interaction activity decline gradient of the interactive terminal being greater than the preset gradient threshold is calculated under the internal cognitive load state and the external cognitive load state. From the probability distribution curve, the time node that reaches the preset probability threshold is extracted as the future power exhaustion probability time point.

[0017] Preferably, the adaptive intervention command is triggered to the interactive terminal, including: generating cognitive overload warning information when the learning sustainability index is lower than a preset safety threshold; mapping the internal cognitive load state and the external cognitive load state to the corresponding load quantification values ​​respectively; determining the dominant load type by comparing the magnitude of the quantification values ​​of the internal cognitive load state and the external cognitive load state, and using it as the attribution result for attribution analysis of the cognitive overload warning information.

[0018] In response to the attribution results, an adaptive intervention instruction is generated, which includes blocking the current task data transmission from the interactive terminal and switching to a preset buffered multimodal data output strategy.

[0019] On the other hand, the present invention provides an intelligent education assessment system based on learning behavior analysis, including: a feature extraction module, an entropy calculation module, a state mapping module, a risk prediction module, and an assessment intervention module;

[0020] The feature extraction module is used to acquire multidimensional microscopic behavioral sequence data and current interaction time reference of the interactive terminal, and extract spatiotemporal trajectory features, temporal rhythm features and physiological mapping features from the multidimensional microscopic behavioral sequence data;

[0021] The entropy calculation module is used to calculate multiple behavioral information entropy values ​​based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, according to a time series sliding window, in order to construct a behavioral information entropy value sequence;

[0022] The state mapping module is used to process the sequence of behavioral information entropy values ​​through a preset hidden Markov model, so as to map the sequence of behavioral information entropy values ​​into internal cognitive load state and external cognitive load state.

[0023] The risk prediction module is used to combine the internal cognitive load state and the external cognitive load state, as well as the current interaction time base, and calculate the probability time point of future power exhaustion through a preset power exhaustion risk prediction model.

[0024] The assessment and intervention module is used to calculate the time difference between the future probability of power depletion and the current interaction time base. Based on the time difference, a learning sustainability index is generated. If the learning sustainability index is lower than a preset safety threshold, an adaptive intervention command is triggered to the interactive terminal to change its task delivery or multimedia data output status. If the learning sustainability index is higher than or equal to the preset safety threshold, the current output strategy is maintained.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] This invention extracts spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features from multidimensional microscopic behavioral sequence data of interactive terminals, and integrates and calculates behavioral information entropy value sequences. This can transform microscopic operations into continuously trackable observation states, effectively capture fine-grained fluctuations in learning behavior, and improve the real-time performance and stability of intelligent evaluation.

[0027] The behavioral information entropy sequence is decoded using a pre-defined Hidden Markov Model, mapping and decoupling it into internal and external cognitive load states. This approach identifies potential state transition processes, effectively distinguishes the additional costs caused by different loads, avoids misjudging abnormal root causes, and improves the targeting of subsequent intervention strategies.

[0028] By combining the forgetting curve decay parameter and the emotional exhaustion parameter, a motivation exhaustion risk prediction model is used to calculate the probability time point of future motivation exhaustion. By incorporating a learning sustainability index for pre-emptive identification and triggering adaptive intervention instructions, early warnings can be issued before a significant decline in interaction activity, solving the problem of severe lag in intervention actions in traditional mechanisms and improving the individual adaptability of the prediction. Attached Figure Description

[0029] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0030] Figure 1 This is a flowchart illustrating an intelligent educational assessment method based on learning behavior analysis according to the present invention.

[0031] Figure 2 This is a schematic diagram of a module of an intelligent education assessment system based on learning behavior analysis according to the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1 A smart education assessment method based on learning behavior analysis includes the following steps: acquiring multidimensional micro-behavioral sequence data and current interaction time reference of the interactive terminal, and extracting spatiotemporal trajectory features, temporal rhythm features and physiological mapping features from the multidimensional micro-behavioral sequence data;

[0034] Based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, multiple behavioral information entropy values ​​are calculated according to a time series sliding window to construct a behavioral information entropy value sequence.

[0035] The behavioral information entropy value sequence is processed by a pre-defined hidden Markov model to map the behavioral information entropy value sequence into internal cognitive load state and external cognitive load state.

[0036] By combining the internal and external cognitive load states, as well as the current interaction time base, the probability time point of future power exhaustion is calculated through a preset power exhaustion risk prediction model.

[0037] The system calculates the time difference between the future probability of power exhaustion and the current interaction time base. Based on this time difference, it generates a learning sustainability index. If the learning sustainability index is lower than a preset safety threshold, it triggers an adaptive intervention command to the interactive terminal to change its task delivery or multimedia data output status. If the learning sustainability index is higher than or equal to the preset safety threshold, it maintains the current output strategy. The preset safety threshold is determined based on the weighted quotient of the average duration of historical similar subject tasks and the current user's emotional tolerance limit, and its value range is set in the interval [0.2, 0.3].

[0038] This embodiment provides an intelligent education assessment mechanism based on learning behavior analysis. Specifically, the mechanism is deployed in a tablet-based online learning platform for senior high school students to conduct real-time assessments of the continuous learning process during evening self-study sessions. The platform does not judge the learning status solely based on the accuracy of answers, online time, or video playback completion rate. Instead, it continuously estimates whether the cognitive load is accumulating to an unsustainable level by focusing on the micro-operational changes of students on interactive terminals and intervenes before a sharp decline in motivation occurs.

[0039] Specifically, the interactive terminal can be a tablet computer, a laptop computer, or an all-in-one teaching terminal. The multi-dimensional micro-behavioral sequence data collected by the terminal side includes at least touch trajectory, click or press action, time spent answering questions, video pause and playback records, error correction logs, and page dwell records. The current interaction time reference is preferably the session clock issued by the platform server, or it can be a unified timestamp after local time synchronization on the terminal. Its role is to provide a consistent time reference for subsequent time series slicing, risk prediction, and intervention timing control.

[0040] During the feature extraction stage, the system does not directly use the original coordinate stream or original logs, but organizes them into three types of features that can reflect the real learning state: spatiotemporal trajectory features are used to represent the stability of the movement of fingers, mice or styluses in the interface and the clarity of the target; temporal rhythm features are used to represent whether the learning action maintains a relatively stable rhythm, or whether there is a fluctuating pattern of alternating high and low interaction frequency and frequent interruptions.

[0041] Although physiological mapping features do not directly collect physiological sensor data, they reflect students' hesitation, fatigue, stagnation and mental exhaustion through phenomena such as repeated error correction and excessive pauses. The engineering basis behind this is that when cognitive resources are insufficient, interface operations usually no longer maintain linearity and efficiency, but will show external signs such as back-and-forth dragging, repeated pauses, modifications and additions, and loss of regularity in playback control.

[0042] During the entropy construction phase, the system observes the short-term changes of the above multimodal features according to the time series sliding window; here, the behavioral information entropy can be understood as a quantitative representation of the degree of behavioral organization within a certain time window: when the variance of the behavioral feature distribution is lower than the first preset variance threshold, it is determined that the behavioral pattern meets the preset concentration index, and the corresponding entropy value is within the first preset threshold range.

[0043] If students begin to exhibit irregular operations, nonlinear interactions, frequent interface switching, or repeated modifications, the distribution of behaviors within the window will become significantly more dispersed, and the entropy value will increase. Conversely, if students enter a state of almost mechanical repetitive clicking and repetitive execution of similar basic interactive tasks, they may also exhibit another anomaly, namely, overly simplistic behavior lacking effective cognitive processing. Therefore, this entropy value sequence can serve as a correlation feature parameter for changes in learning states, transforming the originally discrete and heterogeneous micro-operations into a continuously trackable state observation sequence.

[0044] During the state mapping phase, the system inputs the sequence of behavioral information entropy values ​​into a pre-defined hidden Markov model; the role of this model is not to describe the individual action itself, but to identify the underlying cognitive state transitions behind the action sequence.

[0045] The hidden state here is preferably constructed as a two-dimensional state space. One dimension corresponds to the internal cognitive load, reflecting the difficulty of understanding the learning content itself, the conceptual span, and the depth of reasoning; the other dimension corresponds to the external cognitive load, reflecting additional consumption such as interface interference, task switching additional load, and unfriendly presentation methods. The reason for adopting this approach is that the root causes of frequent pausing and repeated modification may be completely different: one is constrained by the high complexity of the knowledge itself, and the other is that the teaching interface interferes with effective processing. If the above two are not effectively decoupled, the accuracy of subsequent adaptive intervention strategies will be significantly reduced.

[0046] During the risk prediction phase, the system combines the current decoded internal cognitive load state, external cognitive load state, and current interaction time benchmark, and calls the preset motivation exhaustion risk prediction model to output the future motivation exhaustion probability time point. This time point is not simply the remaining learning time, but represents a high-risk moment when students experience a significant decline in interactive activity, a marked decrease in their willingness to learn actively, and a decline in knowledge processing efficiency, given that the current load structure continues. The reason for using this prediction method is that the decline in learning motivation is cumulative and has a lag. Before conventional outcome indicators show a decline trend, the system is already in a critical state of exhaustion. If only outcome score changes are used as the intervention trigger condition, it usually leads to a significant lag in teaching regulation.

[0047] During the assessment and control phase, the system calculates the time difference between the future probability of power exhaustion and the current interaction time benchmark, and generates a learning sustainability index accordingly. Specifically, the learning sustainability index is obtained by normalizing the time difference by dividing it by the estimated total time of the current subject task. This index is used to characterize the real-time quantitative result of the remaining time margin before the current dynamic learning state reaches the predicted instability threshold.

[0048] When the index is below the preset safety threshold, it indicates that the learning buffer space is insufficient. Continuing to push high-pressure tasks at the original pace is likely to trigger cognitive overload and emotional resistance. Therefore, the system issues an adaptive intervention instruction. When the index is above or equal to the preset safety threshold, it indicates that the current learning state is still in the sustainable range. The platform maintains the current output strategy to avoid prematurely interrupting the normal learning flow.

[0049] As an exception handling mechanism, in actual deployment, it is also necessary to handle abnormal data and boundary conditions. If the terminal is offline for a short time, the touch device fails, or the log is missing, resulting in the inability to extract a certain type of feature stably, the system can temporarily continue to generate the degraded behavior information entropy sequence based on the remaining available features and give a confidence assessment label.

[0050] If data is missing across multiple consecutive windows, resulting in insufficient confidence in the state mapping, the proactive warning of the risk prediction model will be paused, and only the current output strategy will be maintained or the system will be switched to conservative mode to prevent the collection anomaly from being mistaken for the learning anomaly. If a student has not operated for a long time but the page remains online, the system can also combine the foreground dwell status, screen on / off status, or input focus status to distinguish between valid dwell time and invalid dwell time without input or output, so as to avoid data bias in the learning sustainability index.

[0051] For example, in a key high school's senior year math intensive preparation platform in a certain city, students enter the topic of comprehensive application of derivatives at 7 pm during evening self-study; within the first 20 minutes, the terminal side records that the variance of the touch trajectory distribution is lower than the preset trajectory dispersion threshold and the frequency of video pauses is lower than the preset frequency threshold. Based on this, the system extracts stable spatiotemporal trajectory features and temporal rhythm features, and the behavioral information entropy sequence remains within the normal range.

[0052] As the topic progressed to the stage of complex proof problems involving the monotonicity of functions and the discussion of parameters, students began to repeatedly go back and forth between pages, watch video replays more often, and spend significantly longer periods modifying the same problem. The system identified this change as a decrease in the degree of behavioral organization and found through state mapping that their internal cognitive load continued to increase, while their external cognitive load also increased due to multiple levels of interface jumps.

[0053] Based on this, the risk prediction module judged that if the current task intensity is maintained, students will enter a high-risk zone of energy exhaustion in a short period of time. Therefore, the platform generates a warning of declining learning sustainability index in advance and prepares to implement buffer intervention before reaching the safety threshold, rather than waiting for students to drop out of the course or make continuous mistakes before making passive adjustments.

[0054] The purpose of this step is to transform the assessment method in traditional education platforms, which is characterized by delayed results and late intervention, into an assessment method that is characterized by continuous process and proactive risk assessment. This will enable real-time monitoring of learning sustainability, early identification of cognitive overload, and dynamic control of the pace of teaching output.

[0055] As one embodiment of the present invention, the spatiotemporal trajectory features include touch movement trajectory data and pressure intensity data; the temporal rhythm features include interaction time variance data and multimedia control frequency data; the physiological mapping features include error correction modification number data and operation dwell time data; the physiological mapping features are essentially the mental resource occupancy state deduced in reverse through the pauses and repetitions of micro-interactions, rather than direct bioelectrical signals.

[0056] This embodiment provides a mechanism for refining and extracting multidimensional micro-behavioral features. Specifically, in the aforementioned online learning scenario of evening self-study, in order to avoid misjudging the student's status based on a single interaction indicator, the system divides the terminal behavior into three categories: spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, each corresponding to different sources of learning anomalies.

[0057] Specifically, touch movement trajectory data reflects the student's target orientation on the interface; if the student clearly knows which tool to select, which section of analysis to view, or which question to switch to next, their trajectory usually shows a trajectory that is close to the preset shortest interaction distance and a relatively direct movement trend.

[0058] Conversely, if students repeatedly wander in the same area, approach the target and then retreat, or frequently jump between areas, it usually indicates that they have hesitation, confusion or increased task search burden at the cognitive level; the pressure intensity data is used to reflect the changes and stability of the force during operation; although the pressure intensity cannot be directly equated with emotion, under the same equipment and similar posture, a sudden change rate of pressure intensity exceeding the preset change rate threshold often corresponds to a decrease in operational control caused by tension, impatience or fatigue, so it can be used as a supplementary dimension of spatiotemporal trajectory characteristics;

[0059] Furthermore, the interaction time variance data is used to describe whether the time spent on actions such as answering questions, turning pages, annotating, and reviewing is stable; when cognitive activity is in a good state, the processing time of students on similar tasks is usually in a relatively continuous range; when the cognitive resource consumption reaches the preset saturation threshold, it is manifested by the duration of the first action sequence exceeding the first preset duration threshold, and the duration of the adjacent second action sequence being lower than the second preset duration threshold, and it is determined that the overall interaction rhythm parameters exceed the preset variance range;

[0060] Multimedia control frequency data mainly includes statistics on the occurrence density of control actions such as pause, playback, speed switching, volume adjustment, and interface collapse and expansion. More of these actions are not necessarily worse, but if the rate of change of their control frequency within a preset time window exceeds a preset growth threshold, and the variance of the action distribution is higher than a preset regularity discrete threshold, it usually indicates that students are experiencing a blockage in absorbing the content, or that the organization of interface information has created additional interference for them.

[0061] Furthermore, the number of corrections and modification attempts and the duration of operation pauses are used as physiological mapping features. Here, physiological mapping does not refer to directly measuring physiological signals such as heart rate and skin conductance, but rather indirectly reflecting mental exertion and psychological burden by utilizing highly correlated explicit results of operations. For example, if a student repeatedly deletes and rewrites intermediate steps in the same answer box, it often indicates that the information in their working memory is unstable. If the time spent on key buttons, core questions, or explanation areas exceeds a preset time threshold, it may indicate that the student is at a cognitive blockade point, has delayed judgment, or has difficulty allocating attentional resources. Therefore, these two types of data can be used to supplement the deep load changes that are difficult to reveal by trajectory and rhythm alone.

[0062] As an anomaly handling mechanism, the different hardware capabilities of different terminals may lead to differences in the features that can be acquired; if the terminal does not support pressure intensity sampling, the pressure duration, the change in the contact area, or the click jitter amplitude can be used as alternative inputs.

[0063] If students use an external keyboard or non-touchscreen device, mouse pauses, cursor backspace, backspace frequency, and focus switching in the editing area can replace some touchscreen features. If there is little multimedia content, resulting in sparse multimedia control frequency data, the system can reduce its weight and retain it only as auxiliary evidence of anomalies rather than as the primary criterion. Regarding the number of corrections and the duration of dwell time, if there is network lag or page loading anomalies, the platform must first filter out pseudo-dwell times caused by system performance issues to avoid incorrectly mapping device problems as learning load.

[0064] For example, during the same senior high school mathematics topic study process, after students began to engage in training on highly difficult comprehensive problems, the system detected that their touch trajectory was no longer concentrated in the question input area and the draft area, but switched back and forth between the question stem, condition prompts and video analysis entry; at the same time, the fluctuation variance of the pressure intensity increased significantly compared to the previous time window, local operations were more frequent and the interaction time interval was shortened.

[0065] The time spent answering questions changed from being relatively balanced to fluctuating significantly. Video playback and pausing actions increased in a short period of time, and multiple deletions and rewrites occurred in the solution box for the same question. The time spent in the analysis area also increased significantly. Based on this, the system did not simply judge that the student was not paying attention, but was able to identify that the student was in a high-load processing stage and prepare discriminative underlying features for subsequent entropy value modeling.

[0066] The purpose of this step is to build a data foundation that is closer to the real learning state by using micro-behavioral characteristics from different sources and with different mechanisms, so as to achieve distinguishable observation of different abnormal forms such as confusion, mechanical repetition of interfaces, and cognitive fatigue.

[0067] As one embodiment of the present invention, multiple behavioral information entropy values ​​are calculated according to a time series sliding window based on spatiotemporal trajectory features, temporal rhythm features and physiological mapping features to construct a behavioral information entropy value sequence, including: fusing spatiotemporal trajectory features, temporal rhythm features and physiological mapping features to generate a multimodal feature vector;

[0068] The multimodal feature vectors are segmented according to a time series sliding window. The combined values ​​of the multimodal feature vectors are mapped to pre-declared discrete behavior set states. The proportion of each discrete behavior set state within the time series sliding window is calculated to determine the probability distribution of each discrete behavior set state within the time series sliding window. Based on the probability distribution, multiple behavior information entropy values ​​are calculated by solving the negative sum of the product of the probability distribution and its logarithm, and then concatenated according to the time series to form a sequence of behavior information entropy values.

[0069] This embodiment provides a mechanism for constructing a behavioral information entropy value sequence. Specifically, in the aforementioned main scenario, simply observing the trajectory, rhythm, and number of modifications may still result in local normality but overall instability. Therefore, it is necessary to fuse multi-source features and analyze their organization degree according to time windows to form a behavioral information entropy value sequence that can continuously reflect changes in the learning state.

[0070] Specifically, the system first aligns the spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features collected under the same time reference to generate a multimodal feature vector. This vector does not emphasize the number of dimensions, but rather whether the cross-section of behavior at the same moment is complete. For example, within a short time segment, it records not only the degree of deviation of the touch trajectory, but also the frequency of playback control, the length of editing pauses, and the number of modification actions within that segment. Only in this way can the vector formed express the complex state of the student at that moment, who is both hesitating and frequently reviewing the footage, and repeatedly correcting errors.

[0071] Based on this, the system is segmented according to a time series sliding window; each window can cover a continuous learning time, and adjacent windows are allowed to overlap in order to capture gradual changes in state rather than just abrupt changes; the focus of processing within the window is not to calculate the absolute size of a certain feature, but to statistically analyze whether the behavioral distribution pattern presented by the multimodal feature vector is concentrated, dispersed, or changes suddenly.

[0072] In the specific implementation of the algorithm mapping, the behavioral combinations within the time sliding window can be clustered into several discrete baseline states based on preset rules, such as stable answering, back-and-forth searching, frequent revisiting, long-term stagnation, and repeated modification. Then, the proportion structure of these states in the window can be observed. If most behaviors within a window are concentrated in stable answering, then the window exhibits strong organization. If multiple states do not have a clear dominant position and switch frequently, it indicates that the learning behavior within the window is chaotic or unbalanced.

[0073] Furthermore, the system calculates the behavioral information entropy value based on the probability distribution within the window; in specific engineering operations, the system first maps different combinations of multimodal feature vector values ​​within the sliding window to pre-declared discrete behavioral set states, assuming the total number of independent event units collected within the sliding window is... Statistical analysis through counting units The frequency at which the pre-declared discrete behavior set states are captured, further utilizing the first... Frequency of state capture for a discrete set of behaviors Divide by the total number of independent event units Calculate the first The proportion of local probability distributions occupied by a set of discrete behavior states in the current sliding window ,Right now ;

[0074] The system is based on the calculation formula Entropy value of extracted behavioral information The solution is determined by accumulating the result of the calculation. The probability distribution of the state of a set of behaviors This is achieved through the product of its logarithm, where... For summation, For the natural constant A logarithmic function with base 0;

[0075] Meanwhile, the system adds fault-tolerant boundary conditions at the engineering layer, where the statistical frequency of a certain type of behavior makes the probability distribution proportion... When the value is 0, a calculation anomaly will occur due to taking the logarithm. The calculation item is forcibly reduced to 0 so that the accumulation continues; its physical meaning is that the more disordered, scattered and lacking of a stable processing main line the behavior is, the higher the entropy value is usually; the more unified and regular the behavior is, the lower the entropy value is usually. However, if the entropy value is so low that it approaches the preset lower limit, it can also indicate that the interactive terminal has entered an inefficient single repetitive interaction mode.

[0076] To facilitate the public disclosure of the data flow process, a simplified data verification model can be used for illustration; assuming that the three consecutive windows are denoted as follows: , , ,in Most of the actions involved were consistent answering and a small amount of research. The video contains numerous replays, pauses, and alternations. The interior consists almost entirely of repetitive, similar basic interactive tasks performed at a fixed rhythm; The corresponding behavioral information entropy can fall within the normal stable range. Corresponding to higher entropy values, Another possible scenario is an abnormally low value; the system will process the data in chronological order. , , The results are spliced ​​together to form a sequence of behavioral information entropy values ​​that are updated as the learning process progresses;

[0077] The key advantage of this construction method is that it unifies heterogeneous behaviors that are originally difficult to compare directly into a representation framework based on the degree of window organization. In intelligent education application scenarios, the core feature that induces cognitive exhaustion in the system is often not a simple increase in the frequency of a single interactive action, but an imbalanced combination of multi-dimensional micro-behavioral features formed within a short time slice. The entropy sequence is just right as a compressed expression of this imbalanced combination and can provide stable input for subsequent hidden state decoding.

[0078] As an exception handling mechanism, the window length and sliding step size should not be fixed to be completely rigid; if the window is shorter than the first preset time length, occasional clicks or short replays are easily misjudged as sudden changes in state; if the window is longer than the second preset time length, it will mask the rapid increase in cognitive load; therefore, the platform can adaptively select window parameters according to the content type:

[0079] For short video explanation segments, a sliding window with a first preset time length is used; for the derivation stage of comprehensive questions, a sliding window with a second preset time length is used, wherein the second preset time length is greater than the first preset time length. If there are too few valid data points in a certain window, such as when a student leaves his seat for a short time or when the terminal network fluctuates and the logs are sparse, then the window can be marked as a low confidence window and smoothed out by the preceding and following windows, instead of directly outputting the entropy value that fluctuates drastically.

[0080] If multiple consecutive windows are of low confidence, the splicing and updating will be paused, and the generation of sequences will continue after the data is recovered. The system performs a data integrity poll every preset detection cycle. When the extractability rate of the multimodal feature vector exceeds 80% again, it will automatically switch back from the degraded mode to the full evaluation mode.

[0081] For example, during the learning process of challenging comprehensive derivative problems on the senior high school mathematics platform, the system integrates the student's touch frequency, pause and replay density, problem-solving box modification frequency, and question stem dwell time into a multimodal feature vector. In the early stage, the main state within the window is stable reading and rhythmic answering, so the corresponding behavioral information entropy value is relatively stable. In the middle stage, after encountering parameter discussion problems, the state within the window changes to a mix of replay, pause, modification, and interface switching, and the behavioral information entropy value increases significantly.

[0082] If students later switch to continuously and repeatedly performing the same basic interactive tasks without system intervention, the state within the window may again be concentrated into a highly repetitive single action, forming another type of anomaly; by splicing these window results over time, the platform can obtain a sequence of behavioral information entropy values ​​that can reflect the evolutionary trajectory of steady-chaotic-mechanical repetition.

[0083] The purpose of this step is to transform the short-term organization of multimodal micro-behaviors into a continuous and traceable sequence of states, thereby capturing fine-grained fluctuations in the learning process and providing a unified observational basis for the identification of potential cognitive loads.

[0084] As one embodiment of the present invention, a behavioral information entropy value sequence is processed using a preset hidden Markov model to map the behavioral information entropy value sequence into an internal cognitive load state and an external cognitive load state, including:

[0085] Discretize the behavioral information entropy value sequence into a level mapping, dividing the continuous entropy values ​​into four discrete observation states: low disorder, medium disorder, high disorder, and mechanical repetition according to the numerical range, and use the behavioral information entropy value sequence as the observation sequence input of the preset hidden Markov model.

[0086] The observation sequence is decoded using a preset state transition matrix and a preset observation probability matrix in a preset Hidden Markov Model; the hidden state sequence corresponding to the observation sequence is output, wherein each hidden state in the hidden state sequence is mapped to a preset two-dimensional state space, and the two dimensions of the two-dimensional state space are decoupled and analyzed into internal cognitive load state and external cognitive load state.

[0087] This embodiment provides a cognitive load state mapping mechanism based on a hidden Markov model. Specifically, after obtaining the behavioral information entropy value sequence in the previous stage, if we remain at the superficial level of judging the entropy value, it will be difficult to answer whether the current anomaly comes from the excessive difficulty of the content or from external factors such as the interface or task arrangement. Therefore, the system further introduces a hidden Markov model to perform latent state decoding on the entropy value sequence, thereby outputting the internal cognitive load state and the external cognitive load state.

[0088] Specifically, the sequence of behavioral information entropy values ​​is used as the input of the observation sequence into the model. Here, the observation sequence can be understood as the changes in the complexity of the student's behavior on the surface, while what really needs to be identified is the implicit cognitive state behind it. Hidden Markov models are suitable for this type of problem because the learning state itself has continuous evolution: since the cognitive state of the learner has continuous evolution, it will not abruptly change from a state of extremely low load to a state of exhaustion without transition at a single time point, but will go through a Markov state transition process consisting of multiple potential nodes such as a stable cognitive stage, a stage of mild information processing obstruction, a stage of high cognitive load, and a stage of persistent interaction disorder. The preset state transition matrix in the model is used to express the evolutionary inertia between these stages, and the preset observation probability matrix is ​​used to express how the entropy value usually corresponds to different hidden states.

[0089] Furthermore, in order to distinguish the load from different sources, this embodiment designs the hidden state as a two-dimensional state space; the first dimension represents the internal cognitive load, which mainly reflects the difficulty of the knowledge itself, the degree of conceptual coupling, and the length of the reasoning chain; the second dimension represents the external cognitive load, which mainly reflects the presentation method of the question, the organization of the interface, the overhead of resource switching, and additional consumption such as multimedia interruption; the data flow is explained by constructing a data state inference model of the two-dimensional state space.

[0090] Assuming that the observed entropy patterns for three consecutive time periods are stable—increasing—continuously increasing, the hidden state sequence that may be obtained after model decoding is not a single label, but a combination structure similar to internal medium / external low internal high / external medium internal high / external high.

[0091] The system decouples the two-dimensional state space, outputting the internal cognitive load state and the external cognitive load state respectively, for subsequent risk prediction and intervention attribution. At the specific architecture execution level, the decoding and mapping of the above combined structure cannot rely on pure computational exhaustion; the classic Viterbi algorithm can be used to achieve the above state decoding and parsing process.

[0092] The system predefines an enumeration space for the two-dimensional state space. For example, it divides the two cognitive load dimensions into low, medium, and high nodes. After the two types of parameters are combined, a Cartesian product is performed to calculate the cross-calculation, thereby forming nine sets of composite hidden points. When the observation sequence slice advances to each observation step, the algorithm first performs a probability prediction connection on the state retention value of the previous time span to the current composite node. The method is to multiply the historical best cumulative probability and the connection coefficient of the preset state transition matrix stored in the model, and then immediately multiply it with the emission probability value of the current observed abnormal entropy value generated for the composite node in the preset observation probability matrix. The algorithm retains the highest value combination leading to each candidate position and eliminates the state transition path with low confidence.

[0093] Once the entire continuous sequence has passed evaluation, the system performs dimensionality reduction on the full cross-section of the retained optimal composite sequence. That is, it performs reverse indexing to obtain labels based on the aforementioned Cartesian combination matrix, decouples and strips the merging attributes of each state node, and transforms them into an internal cognitive load state that simply represents the source of cognitive difficulty, and an extraction record that represents the external cognitive load state that represents external interference.

[0094] This decoupling has clear educational engineering significance. If the system identifies high internal cognitive load and low external cognitive load, it indicates that the user of the interactive terminal is mainly constrained by the structural complexity of the knowledge itself. In this case, it is appropriate to trigger instructions to reduce the span of knowledge concepts, increase the breakdown of example tasks, or supplement prior knowledge. If it identifies high external cognitive load and internal cognitive load that has not reached the threshold, it indicates that the presentation of the multimedia interface induces additional cognitive overhead. In this case, the system prioritizes executing instructions to optimize the interface view hierarchy, block unnecessary animation interactions, and limit the frequency of cross-screen jumps, rather than simply lowering the task difficulty threshold. It can be seen that hidden state decoding is not only a data analysis step, but also directly determines the correctness of the intervention path.

[0095] As an anomaly handling mechanism, if the observation sequence is too short, the model may not be able to stably determine the state transition trend. In this case, the system can delay the output of the state results and decode them after accumulating to the minimum window size. If the current entropy value sequence shows an abnormal jump but network jitter, page reload, or application switching is detected at the same time, the observation data can be marked as abnormal data affected by external disturbances and will not directly participate in state mapping, thus avoiding misidentifying system performance fluctuations as changes in cognitive load. For newly connected users or users with a small number of samples, if the individual model has not been fully calibrated, the group prior parameters can be used for initial decoding first, and then the state transition and observation distribution parameters can be gradually updated as historical interactions increase.

[0096] For example, in the aforementioned senior high school mathematics platform, after students enter the discussion of high-difficulty parameter questions, the behavioral information entropy value rises continuously for multiple windows. If only the absolute value of the entropy value is used for judgment, the system can only identify that the disorder of behavior exceeds the normal threshold range. After further decoding by the Hidden Markov Model, the system finds that the students in the previous stage were mainly in an internal high / external low state, indicating that their core difficulty lies in the large reasoning span.

[0097] However, during the continued learning process, because the question analysis page requires frequent switching between the question stem, answer, and knowledge point cards, the hidden state evolves into an internal high / external high state; the decoding result of this observation sequence indicates that the interactive object is not facing a single reasoning logic obstacle, but is in a dual high-load state of processing highly complex content and frequent interface interaction; based on this, the system can provide a more accurate load structure for subsequent risk prediction, rather than uniformly attributing all anomalies to the student's lack of ability;

[0098] The purpose of this step is to elevate observable behavioral disorder to interpretable load structure identification, thereby decomposing the sources of learning difficulties and laying a clearer causal foundation for subsequent prediction and targeted intervention.

[0099] As one embodiment of the present invention, combining the internal cognitive load state and the external cognitive load state, as well as the current interaction time reference, a preset energy exhaustion risk prediction model is used to calculate the future energy exhaustion probability time point, including: obtaining the forgetting curve decay parameter and the emotional exhaustion parameter extracted from the historical interaction data of the interactive terminal; and inputting the internal cognitive load state, the external cognitive load state, the forgetting curve decay parameter and the emotional exhaustion parameter into the preset energy exhaustion risk prediction model.

[0100] Using the survival analysis algorithm in the preset power depletion risk prediction model, the probability distribution curve of the interaction activity decrease gradient of the interactive terminal being greater than the preset gradient threshold is calculated under the internal cognitive load state and the external cognitive load state.

[0101] From the probability distribution curve, the time point at which the preset probability threshold is reached is extracted as the future probability time point of energy depletion; whereby the preset probability threshold is defined as the cumulative probability critical point at which the interaction activity drops to 50% of the initial baseline value. This time point represents the logical prediction position at which students enter an irreversible period of fatigue decline at both the physiological and psychological levels.

[0102] This embodiment provides a mechanism for predicting the risk of energy exhaustion. Specifically, after obtaining the internal and external cognitive load states, if the system only judges that the current load is high, it still cannot answer a more critical question, namely, how long can the student continue under the current task structure. To solve this problem, this embodiment further introduces the forgetting curve decay parameter and emotional exhaustion parameter related to historical interactions, and estimates the probability time point of future energy exhaustion through a preset energy exhaustion risk prediction model.

[0103] Specifically, the forgetting curve decay parameter is derived from the trend of changes in the retention of knowledge after learning and the consolidation efficiency after repeated exposure in students' historical learning data; its physical meaning is to reflect the speed at which students retain new knowledge and the intensity of their need for consolidation; if a student forgets similar knowledge points quickly, it means that they need more spaced repetition and buffering during high-intensity learning, otherwise the new input content will be covered by subsequent information before it is solidified, which can easily cause additional cognitive burden.

[0104] Emotional exhaustion parameters are derived from long-standing behavioral patterns in historical interactions, such as passive completion, increased procrastination, dropping out of classes midway, and decreased activity after continuous nighttime study. These parameters are used to characterize students' emotional resilience and tolerance limits when facing continuous task pressure. Both types of parameters describe how the same workload can produce completely different consequences when it falls on different individuals.

[0105] Furthermore, the system inputs the current internal cognitive load state, external cognitive load state, forgetting curve decay parameters, and emotional exhaustion parameters into a preset motivation exhaustion risk prediction model. The model adopts the idea of ​​survival analysis, and instead of giving a simple binary judgment, it outputs a probability distribution curve of a significant decrease in interaction activity in the future. Here, the gradient of the decrease in interaction activity is greater than the preset gradient threshold, which can be understood as students showing a significant trend of less operation, less response, less correction, actively terminating learning, or switching to low-quality mechanical interaction in a short period of time. Its essence is that the learning motivation is accelerating from a sustainable state to an unstable state.

[0106] The data flow of the survival analysis algorithm in the pre-defined energy exhaustion risk prediction model follows the following structural decomposition: The system acquires input features such as internal cognitive load status, external cognitive load status, forgetting curve decay parameters, and emotional exhaustion parameters, concatenates them in an equal-scale form, and performs dimensional alignment processing to construct a covariate vector corresponding to individual behavior. The model internally establishes a risk recursive baseline curve based on the full historical sample, which naturally varies over time and is called the baseline risk ratio. ;

[0107] To reflect the degree to which input covariates influence the direction of individualized distortion, the system obtains the current comprehensive risk ratio for this student based on rules. The regular expression is:

[0108] In the formula, the system introduces a time variable to represent the inference and prediction. For the constructed covariate vector The regression coefficient vector embedded in the database during model training to balance the weighting of various covariates. Execute based on transpose operator symbol Matrix operations are used to calculate the aforementioned current comprehensive risk ratio. , Represented by natural constant An exponential function with base 0, and a vector of regression coefficients. The model was obtained in advance by training a Cox proportional hazards model on a large-scale sample of desensitized learning behaviors, and was used to quantify the contribution weights of intrinsic and extrinsic loads to the rate of power exhaustion.

[0109] Then, by integrating the comprehensive risk ratio for specific time units that may be experienced in the future and calculating the decay effect of the cumulative term, a probability distribution curve is formed, which shows that the gradient of the interaction activity of the interactive terminal is greater than a preset gradient threshold under both internal and external cognitive load states. When the above probability distribution curve continues to decline with the future projection of the horizontal axis to a probability threshold line set by the preset conditions, the horizontal axis time coordinate mark where the occurrence point coincides is extracted and directly used as the probability time point of the student's energy exhaustion.

[0110] The reason for adopting this mechanism is that learning motivation exhaustion is not triggered by a single moment, but is affected by the current workload, historical forgetting patterns, and emotional tolerance. For two students who are in the same high internal / medium external state, if one student has shown faster forgetting and weaker emotional recovery ability in similar past topics, the time period for them to reach the preset probability of motivation exhaustion will be shortened accordingly. If the other student has good prior knowledge and high resilience, they may still be able to maintain effective learning for a longer period of time under the same workload. Therefore, by introducing historical parameters, risk prediction can be improved from a group average judgment to an individualized and continuous judgment.

[0111] As an exception handling mechanism, if a student is using the platform for the first time or there is insufficient historical interaction data, the system can first use the baseline parameters of the same grade, subject, and ability level group as a temporary substitute, and gradually update the individual parameters in subsequent learning.

[0112] If historical data is complete but there are recent situations such as exam week, abnormal work and rest schedule, or equipment replacement, which may cause the emotional exhaustion parameters to deviate from the current state, the platform can limit the influence of historical parameters and prioritize correction based on the behavioral trends of the most recent few days.

[0113] If the current interactive activity level drops abnormally due to network failure, background operation, or unified suspension of classes, such non-learning factors should be removed from the risk prediction event set to avoid misjudging external downtime as power exhaustion.

[0114] For example, in the same senior high school math intensive training platform, the system discovered that a student is currently under a high internal / high external workload structure. Combined with their historical data from the past month, it was found that this student forgets complex function topics relatively quickly and often avoids interaction and quits early after 45 minutes of continuous high-pressure learning. Based on this, the system uses a risk prediction model to generate a risk distribution for a future period, identifying a high probability that the student will enter a period of rapid decline in interactive activity early at the current pace. Therefore, this moment is designated as the probability point of future exhaustion. In contrast, another student with a more solid foundation and greater historical resilience, even with a similar current workload, will have their corresponding time point later. Thus, the platform can implement different intervention times for different individuals, rather than uniformly requiring all students to rest or reduce difficulty after a fixed period.

[0115] The purpose of this step is to transform static load identification into dynamic exhaustion time prediction, thereby enabling early quantification of when instability will occur and providing a time-based basis with individual differences for the generation of the sustainability index.

[0116] As one embodiment of the present invention, triggering an adaptive intervention command to an interactive terminal includes: generating a cognitive overload warning message when the learning sustainability index is lower than a preset safety threshold; mapping the internal cognitive load state and the external cognitive load state to corresponding load quantification values ​​respectively; determining the dominant load type by comparing the magnitudes of the quantification values ​​of the internal cognitive load state and the external cognitive load state, which serves as the attribution result for attribution analysis of the cognitive overload warning message; and generating an adaptive intervention command in response to the attribution result, wherein the adaptive intervention command includes blocking the current task data transmission from the interactive terminal and switching to a preset buffered multimodal data output strategy.

[0117] This embodiment provides an adaptive intervention mechanism based on the learning sustainability index. Specifically, in the aforementioned process, the system is already able to predict the risk of future energy depletion. However, if only prediction is completed without control, the probability of students entering an unstable state cannot be truly reduced. Therefore, this embodiment further stipulates that when the learning sustainability index is lower than a preset safety threshold, the platform generates a cognitive overload warning and selects different buffering intervention strategies according to the dominant load type.

[0118] Specifically, if the learning sustainability index falls below the safety threshold, it means that students are too close to the predicted high-risk moment of energy exhaustion, and continuing to maintain the current task output pace will significantly increase the probability of instability. At this time, the system will first generate a cognitive overload warning message. This warning message can be sent to the teacher management terminal, parent terminal, or student terminal, or it can be used as a background scheduling signal for internal use by the platform. The focus of the warning is not to simply indicate a decrease in activity, but to clearly indicate the risk of overload and the need to adjust the teaching output method.

[0119] The system compares the values ​​of internal and external cognitive load states to determine the dominant load type and uses it as the attribution result for the cognitive overload warning information. If the value of the internal cognitive load state is higher than that of the external cognitive load state, it indicates that the main problem comes from the knowledge itself being too difficult, the reasoning span being too large, or insufficient prior knowledge. If the value of the external cognitive load state is higher than that of the internal cognitive load state, the main problem comes from task presentation, interface switching, redundant prompts, or improper multimedia organization.

[0120] Based on different attributions, the system generates different adaptive intervention instructions; the so-called blocking the current task data transmission does not completely stop learning, but temporarily stops pushing high-pressure tasks, long-chain reasoning questions or complex information streams to avoid further stacking of input when the load is already close to the limit; the so-called switching to the preset buffered multimodal data output strategy can include one or more of the following methods: switching long logic explanations to step-by-step diagrams, breaking down comprehensive questions into short chain sub-tasks, compressing the number of information layers displayed on the interface at the same time, switching pure text derivation to diagrammatic explanation, or inserting buffer content such as short debriefing, low-interference review and rest guidance;

[0121] This control method has a clear mechanistic basis: continuing to add tasks of the same or even higher intensity under cognitive overload often does not improve learning efficiency, and thus causes the cognitive load parameter to overflow the preset threshold and trigger negative feedback of interaction activity; while providing a buffer output that matches the dominant load before overload occurs can help students restore processing order, reduce ineffective consumption, and thus extend effective learning time; especially in long-term high-pressure scenarios such as the senior year of high school, timely and targeted buffering is more critical than post-event remediation.

[0122] As an exception handling mechanism, if the learning sustainability index is close to the threshold but has not fallen below it, the system can first issue a light warning, such as reducing the speed of question push and reducing pop-up prompts, without immediately blocking the task flow to avoid excessive interference with the normal sprint state; if the internal cognitive load and external cognitive load are close and it is impossible to clearly determine the dominant type, a hybrid buffering strategy can be adopted, such as simultaneously reducing the content span and interface complexity; if students are in special modes such as exam simulations or timed assessments that should not be interrupted, the system can only record the risk and delay intervention, and then perform buffered output after submission; if the teacher has manually locked the teaching pace, the platform can move the intervention authority to the teacher's confirmation to balance the boundaries of automatic control and teaching management.

[0123] For example, in the aforementioned senior high school mathematics platform, after students continuously studied complex derivative topics, their learning sustainability index fell below the safe threshold. The system then generated a cognitive overload warning and, through attribution analysis, found that their internal cognitive load was significantly higher than their external cognitive load, indicating that the main problem was not the interface design, but rather that the concept span had exceeded the current sustainable processing range. Therefore, the platform stopped issuing new comprehensive and difficult problems and switched to a buffered multimodal output consisting of step-by-step prompt cards, key concept diagrams, and short-term consolidation exercises.

[0124] If the system identifies a higher external cognitive load in another student, the platform will prioritize compressing the interface information hierarchy, closing unnecessary side resources, reducing frequent jumps between videos and questions, and maintaining the knowledge difficulty unchanged. In the above closed-loop intervention control, the regulation mechanism of the intelligent education assessment system does not execute indiscriminate task interruption or global difficulty reduction, but rather performs targeted cognitive consumption suppression and dynamic adaptation of multimodal data output strategies based on the identified dominant load type.

[0125] The purpose of this step is to transform the risk prediction results into actionable teaching control measures, thereby achieving pre-overload intervention, cause-based measures, and adaptive buffer control of learning pace.

[0126] As one embodiment of the present invention, it includes the following modules: a feature extraction module, used to acquire multidimensional microscopic behavior sequence data of the interactive terminal and the current interaction time reference, and extract spatiotemporal trajectory features, temporal rhythm features and physiological mapping features from the multidimensional microscopic behavior sequence data;

[0127] The entropy calculation module is used to calculate the entropy values ​​of multiple behavioral information based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, according to a time series sliding window, in order to construct a sequence of behavioral information entropy values.

[0128] The state mapping module is used to process the sequence of behavioral information entropy values ​​through a preset hidden Markov model, so as to map the sequence of behavioral information entropy values ​​into internal cognitive load state and external cognitive load state.

[0129] The risk prediction module is used to combine the internal cognitive load state and the external cognitive load state, as well as the current interaction time base, and calculate the probability time point of future power exhaustion through a preset power exhaustion risk prediction model.

[0130] The assessment and intervention module is used to calculate the time difference between the probability of future energy exhaustion and the current interaction time base. Based on the time difference, a learning sustainability index is generated, and if the learning sustainability index is lower than a preset safety threshold, an adaptive intervention instruction is triggered. The preset safety threshold is determined based on the weighted quotient of the average duration of historical similar subject tasks and the current user's emotional tolerance limit. Its value range is set in the interval [0.2, 0.3], which is used to represent that the remaining sustainable learning time is less than 30% of the total expected task duration, and to maintain the current output strategy if the learning sustainability index is higher than or equal to the preset safety threshold.

[0131] Please see Figure 2 This embodiment provides an intelligent education assessment system based on learning behavior analysis. Specifically, the system can be deployed in an integrated architecture of cloud teaching platform + terminal acquisition plugin + teacher management backend, or in a local teaching environment composed of school local server and student terminal. The system consists of feature extraction module, entropy calculation module, state mapping module, risk prediction module and assessment intervention module. Each module can be implemented by software program, or by software and terminal-side acquisition hardware working together.

[0132] Specifically, the feature extraction module is used to collect multi-dimensional micro-behavioral sequence data from the interactive terminal and obtain the current interaction time reference; this module is preferably set between the terminal-side acquisition service and the cloud processing service; the terminal side is responsible for recording low-level logs such as touch events, page jumps, playback control, and input editing, and performing preliminary noise reduction; the cloud or edge nodes are responsible for unifying timestamps, aligning session segments, and extracting spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features;

[0133] After receiving the above features, the entropy calculation module analyzes the organization of multimodal behavior using a sliding window and outputs a sequence of behavioral information entropy values. The state mapping module further decodes the sequence of behavioral information entropy values ​​into internal cognitive load state and external cognitive load state based on a preset hidden Markov model. The risk prediction module combines forgetting and emotion-related parameters extracted from historical interactions to output the probability time point of future energy depletion. The assessment and intervention module generates a learning sustainability index based on the time difference between this time point and the current interaction time benchmark, and decides whether to maintain the current output strategy or trigger adaptive intervention.

[0134] In terms of engineering implementation, the modules can be decoupled through message queues or event buses; the feature extraction module can continuously report compressed behavioral summaries without transmitting the entire original trajectory, thereby reducing network load and privacy exposure; the entropy calculation module and state mapping module can run on edge nodes for rapid response during the course; the risk prediction module and assessment and intervention module can run in the cloud to combine more complete historical data and strategy library for unified decision-making; for school-wide deployment scenarios, all modules can also be integrated into the school's education server, allowing the teacher management backend to read the sustainable distribution of the student population to assist in classroom pacing and after-class task design.

[0135] As an exception handling mechanism, the system should also have fault degradation capability at the module level; if the feature extraction module is unable to collect the pressure intensity due to terminal permission restrictions, the remaining available features can still continue to enter subsequent modules.

[0136] If the entropy calculation module detects that there is insufficient valid data in the input window, it can output a low confidence flag and request supplementary data collection.

[0137] If the state mapping module is temporarily unavailable due to unsuccessful loading of model parameters, the system will at least retain the ability to monitor abnormal behavior, but will suspend the dominant load attribution; if the risk prediction module lacks historical data, the assessment and intervention module can use conservative thresholds for temporary control.

[0138] If communication between the assessment intervention module and the content distribution module fails, the platform should maintain a safe output strategy by default, such as reducing the intensity of task pushes instead of continuously sending high-pressure content, in order to avoid the risk from being amplified due to the interruption of the control link.

[0139] For example, in the aforementioned senior high school evening self-study math platform, the student terminal first reports touch back and forth, playback operations, modification frequency, and dwell behavior through the feature extraction module; the entropy calculation module generates a sequence of behavior information entropy values ​​according to a continuous time window;

[0140] The state mapping module identifies that the student has shifted from internally average / externally low to internally high / externally high; the risk prediction module, combining the student's historical forgetting rate and recent fatigue trend, determines that there is a high risk of motivation exhaustion in the near future; the assessment and intervention module generates a low learning sustainability index based on this and issues an intervention instruction to the content engine to stop pushing new high-difficulty comprehensive questions and instead send low-interference review content and diagrammatic explanations; if the terminal network is interrupted at this time, resulting in the loss of some logs, the system automatically enters a degraded mode, maintaining only conservative output and not continuing to perform fine-grained attribution that strongly relies on real-time data;

[0141] The purpose of this step is to implement the aforementioned methodology into a deployable, interconnected, and degradeable system architecture, thereby achieving closed-loop support for the entire chain of learning behavior collection, status assessment, risk prediction, and teaching intervention.

[0142] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any conventional modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent education evaluation based on learning behavior analysis, characterized in that, The method includes the following steps: Acquire multidimensional microscopic behavioral sequence data and current interaction time reference of the interactive terminal, and extract spatiotemporal trajectory features, temporal rhythm features and physiological mapping features from the multidimensional microscopic behavioral sequence data; Based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, multiple behavioral information entropy values ​​are calculated according to a time series sliding window to construct a behavioral information entropy value sequence. The behavioral information entropy value sequence is processed by a pre-defined hidden Markov model to map the behavioral information entropy value sequence into internal cognitive load state and external cognitive load state. By combining the internal and external cognitive load states, as well as the current interaction time base, the probability time point of future power exhaustion is calculated through a preset power exhaustion risk prediction model. Calculate the time difference between the future power exhaustion probability time point and the current interaction time base, generate a learning sustainability index based on the time difference, and trigger an adaptive intervention command to the interactive terminal to change its task distribution or multimedia data output state when the learning sustainability index is lower than the preset safety threshold. When the learning sustainability index is higher than or equal to the preset safety threshold, maintain the current output strategy. Based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, multiple behavioral information entropy values ​​are calculated according to a time series sliding window to construct a behavioral information entropy value sequence, including: By fusing spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, a multimodal feature vector is generated. The multimodal feature vectors are segmented according to the time series sliding window, and the combined values ​​of the multimodal feature vectors are mapped to pre-declared discrete behavior set states. The proportion of each discrete behavior set state in the time series sliding window is calculated to calculate the probability distribution of each discrete behavior set state in the time series sliding window. Based on the probability distribution, the negative sum of the product of the probability distribution and its logarithm is calculated. When a certain probability value in the probability distribution is 0, the term of the product of that probability value and its logarithm is set to 0, and the accumulation continues, thereby solving multiple behavioral information entropy values. The behavioral information entropy sequence is processed using a pre-defined Hidden Markov Model to map the sequence to internal and external cognitive load states, including: The sequence of behavioral information entropy values ​​is used as the input of the observation sequence of the pre-defined Hidden Markov Model; The observation sequence is decoded using the preset state transition matrix and preset observation probability matrix in the preset Hidden Markov Model; Output the hidden state sequence corresponding to the observation sequence. Each hidden state in the hidden state sequence is mapped to a preset two-dimensional state space. The two dimensions of the two-dimensional state space are decoupled and resolved into internal cognitive load state and external cognitive load state. Combining internal and external cognitive load states, along with the current interaction time baseline, a pre-defined power exhaustion risk prediction model is used to calculate the future probability time points of power exhaustion, including: Obtain the forgetting curve decay parameters and emotional exhaustion parameters extracted from historical interaction data of interactive terminals; Input the internal cognitive load state, external cognitive load state, forgetting curve decay parameter and emotional exhaustion parameter into the preset energy exhaustion risk prediction model; Using the survival analysis algorithm in the preset power depletion risk prediction model, the probability distribution curve of the interaction activity decrease gradient of the interactive terminal being greater than the preset gradient threshold is calculated under the internal cognitive load state and the external cognitive load state. From the probability distribution curve, the time point at which the preset probability threshold is reached is extracted as the future probability time point of power depletion.

2. The intelligent educational assessment method based on learning behavior analysis according to claim 1, characterized in that, Spatiotemporal trajectory features include touch movement trajectory data and pressure intensity data; temporal rhythm features include interaction time variance data and multimedia control frequency data; physiological mapping features include error correction and modification number data and operation dwell time data.

3. The intelligent educational assessment method based on learning behavior analysis according to claim 1, characterized in that, Trigger adaptive intervention commands to the interactive terminal, including: When the learning sustainability index falls below a preset safety threshold, a cognitive overload warning message is generated. The internal cognitive load state and the external cognitive load state are mapped to corresponding load quantification values ​​respectively. The dominant load type is determined by comparing the magnitude of the quantification values ​​of the internal cognitive load state and the external cognitive load state, which serves as the attribution result for the attribution analysis of cognitive overload warning information. In response to the attribution results, an adaptive intervention instruction is generated, which includes blocking the current task data transmission from the interactive terminal and switching to a preset buffered multimodal data output strategy.

4. An intelligent education assessment system based on learning behavior analysis, comprising the intelligent education assessment method based on learning behavior analysis as described in any one of claims 1 to 3, characterized in that, Includes the following modules: The feature extraction module is used to acquire multidimensional microscopic behavioral sequence data and the current interaction time reference of the interactive terminal, and extract spatiotemporal trajectory features, temporal rhythm features and physiological mapping features from the multidimensional microscopic behavioral sequence data; The entropy calculation module is used to calculate the entropy values ​​of multiple behavioral information based on spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features, according to a time series sliding window, in order to construct a behavioral information entropy value sequence. Specifically, it is used to: fuse spatiotemporal trajectory features, temporal rhythm features, and physiological mapping features to generate multimodal feature vectors; segment the multimodal feature vectors according to the time series sliding window, map the combined values ​​of the multimodal feature vectors to pre-declared discrete behavioral set states, and calculate the proportion of each discrete behavioral set state within the time series sliding window to calculate the probability distribution of each discrete behavioral set state within the time series sliding window; Based on the probability distribution, the negative sum of the product of the probability distribution and its logarithm is calculated. When a certain probability value in the probability distribution is 0, the term of the product of that probability value and its logarithm is set to 0, and the accumulation continues, thereby solving multiple behavioral information entropy values. The state mapping module is used to process the behavioral information entropy value sequence through a preset hidden Markov model, so as to map the behavioral information entropy value sequence into internal cognitive load state and external cognitive load state; specifically, it is used to: take the behavioral information entropy value sequence as the observation sequence input of the preset hidden Markov model; and decode the observation sequence through the preset state transition matrix and preset observation probability matrix in the preset hidden Markov model. Output the hidden state sequence corresponding to the observation sequence. Each hidden state in the hidden state sequence is mapped to a preset two-dimensional state space. The two dimensions of the two-dimensional state space are decoupled and resolved into internal cognitive load state and external cognitive load state. The risk prediction module combines internal and external cognitive load states with the current interaction time baseline to calculate the probability time point of future energy exhaustion using a preset energy exhaustion risk prediction model. Specifically, it is used to: acquire forgetting curve decay parameters and emotional exhaustion parameters extracted from historical interaction data of the interactive terminal; input the internal and external cognitive load states, forgetting curve decay parameters, and emotional exhaustion parameters into the preset energy exhaustion risk prediction model; calculate the probability distribution curve of the interaction activity decrease gradient of the interactive terminal being greater than a preset gradient threshold under the internal and external cognitive load states using the survival analysis algorithm in the preset energy exhaustion risk prediction model; and extract the time point of reaching the preset probability threshold from the probability distribution curve as the probability time point of future energy exhaustion. The evaluation and intervention module is used to calculate the time difference between the future probability time of power exhaustion and the current interaction time base. Based on the time difference, a learning sustainability index is generated. If the learning sustainability index is lower than a preset safety threshold, an adaptive intervention command is triggered to the interactive terminal to change its task delivery or multimedia data output status. If the learning sustainability index is higher than or equal to the preset safety threshold, the current output strategy is maintained.

Images