Interactive intelligent education method and system of artificial intelligence and virtual reality

Abstract

The application relates to the technical field of virtual reality, in particular to an interactive intelligent education method and system of artificial intelligence and virtual reality, which comprises the following steps: opening a native signal collection channel to collect signals, and executing a personal baseline calibration process to establish a detection reference; when the cumulative attempt number of an interaction node reaches a specified number, performing a dimension transformation operation on a historical interaction record queue to reconstruct a two-dimensional time sequence matrix, extracting an objective statistical quantity to calculate a node stagnation depth index; comparing the node stagnation depth index with a trigger threshold value, and entering a support injection process to execute an injection operation of a corresponding level; and performing an update operation on a node-level mastery graph based on complete interaction records, and executing path re-planning. The application effectively reduces the probability of misjudging repeated trial-and-error exploration as a stagnation state.

Description

Technical Field

[0001] This application relates to the field of virtual reality technology, specifically to an interactive intelligent education method and system that combines artificial intelligence and virtual reality. Background Technology

[0002] With the widespread application of virtual reality technology in education, virtual reality learning sessions can provide learners with a highly immersive interactive learning environment, thereby significantly improving the training effect of complex operational skills. However, in practical applications, learners often experience cognitive stagnation when faced with complex atomic interaction tasks, repeatedly trying and failing at a certain interaction node or hesitating for a long time. How to accurately identify this micro-level cognitive stagnation and provide appropriate intervention and guidance has become a core technical challenge that current virtual reality intelligent education systems urgently need to solve.

[0003] Existing virtual reality education systems typically use simple rules based on fixed time thresholds to determine learner status. For example, if a learner stays in a scene for longer than a preset fixed duration, the system automatically pops up a global notification window or directly plays the complete instructional video. Some systems also combine standardized test scores after the session to conduct a post-session evaluation of the learner's overall learning effectiveness, using this as a basis for adjusting the difficulty of subsequent courses. These conventional methods mainly rely on macro-level statistical indicators and uniform judgment standards, attempting to cover the interaction process of all learners through standardized procedures.

[0004] However, the fixed time threshold method completely ignores the natural differences in the rhythm of individual learners' responses, and is very likely to misjudge learners who are slow to react but are still actively thinking as being in a stagnant state, thus triggering unnecessary intervention and interrupting their immersive experience. At the same time, the post-evaluation mechanism cannot provide accurate contextual support when learners encounter cognitive bottlenecks, and global pop-up prompts often disrupt the spatial coherence of virtual reality scenes, causing learners' attention to be distracted, making it difficult to truly solve the problem of cognitive weakness at a single interaction node. Summary of the Invention

[0005] To address the problems of existing technologies that fail to accurately identify cognitive stagnation and whose intervention methods disrupt immersion, this application provides an interactive intelligent education method and system based on artificial intelligence and virtual reality.

[0006] In a first aspect, this application provides an interactive intelligent education method based on artificial intelligence and virtual reality, comprising: opening parallel native signal acquisition channels to acquire interactive event streams and node response delays respectively; performing a personal baseline calibration process to summarize the baseline response delay values ​​at each difficulty level and the minimum observation window size into a personal delay baseline file to establish a detection benchmark; when the cumulative number of attempts of a certain interactive node reaches the number specified by the minimum observation window, performing a dimension transformation operation on the historical interaction record queue of the node to reconstruct it into a two-dimensional time series matrix; calculating a node stagnation depth index based on the two-dimensional time series matrix; comparing the node stagnation depth index with a preset trigger threshold; when the node stagnation depth index exceeds the trigger threshold, entering a scaffold injection process; determining the scaffold strength level according to the specific interval in which the node stagnation depth index falls; performing an injection operation of the corresponding level based on the complete interaction records of each node; performing an update operation on the node-level mastery map based on the updated node-level mastery map; performing path replanning based on the updated node-level mastery map; and selecting the scene sequence for the next session.

[0007] By enabling parallel native signal acquisition channels and executing personal baseline calibration procedures, a detection benchmark based on the learner's individual differences was established before the session officially began. This effectively avoided the systematic bias of misjudging learners with slower reaction times as being in a stagnant state, and provided a reliable personalized reference basis for subsequent node-level stagnation depth measurement.

[0008] Preferably, the step of opening parallel native signal acquisition channels to acquire interactive event streams and node response delays respectively, executing a personal baseline calibration process, and summarizing the baseline response delay values ​​and minimum observation window size under each difficulty level into a personal delay baseline file to establish a detection benchmark includes: determining the spatial affiliation of each interactive object's object identifier with the gaze landing point coordinates of the current frame, and accumulating the continuous dwell time on the object as a gaze dwell signal; recording all operation actions issued on each atomic interactive node in an event-driven manner as an interactive event stream; taking the time interval from the moment the task prompt is presented when entering an interactive node to the first arbitrary operation action as a node response delay; sequentially presenting several standard interactive nodes of known difficulty levels to the learner, acquiring all response delay values ​​under each difficulty level, calculating the distribution characteristics of multiple acquisition results under the same difficulty level, taking the median of the distribution as the baseline response delay value under the corresponding difficulty level, and summarizing the baseline response delay values ​​of all difficulty levels into a personal delay baseline file.

[0009] Preferably, the step of performing a dimensional transformation operation on the historical interaction record queue of the node to reconstruct a two-dimensional time series matrix includes: reconstructing the attempt records stored in the historical interaction record queue as a one-dimensional sequence into the two-dimensional time series matrix; the row index of the two-dimensional time series matrix corresponds to the attempt round, and the column index corresponds to the time slice within each attempt; the sampling interval of the time slice is taken as the median of the average single operation time consumption at the current difficulty level during the calibration phase; the value of each element in the two-dimensional time series matrix is ​​the total count of operation events that occurred in the corresponding time slice for the corresponding attempt round.

[0010] Preferably, the step of extracting objective statistics based on the two-dimensional time series matrix includes: extracting the failure rate, which is the ratio of the number of failed attempts to the total number of attempts; extracting the cross-attempt response delay trend coefficient, which is determined by the slope obtained by performing least squares linear regression on the response delay value sequence of each round of attempts; and extracting the intra-attempt hesitation peak value, wherein for each row of the two-dimensional time series matrix, the number of the longest segment among the consecutive segments with element values ​​of zero in that row is counted as the gap length of that row, the arithmetic mean of the gap lengths of all rows is taken, and then normalized by dividing by the actual effective number of segments in that row to obtain the intra-attempt hesitation peak value.

[0011] By extracting failure rate, cross-trial response delay trend coefficient, and intra-trial hesitation dwell peak, the interaction state is characterized from three independent dimensions: macro results, cross-trial time sequence, and intra-trial time sequence. The intra-trial time structure, which was originally compressed into a single response delay scalar, is expanded, allowing the system to characterize the continuity of operation from the perspective of intra-trial time distribution.

[0012] Preferably, the node stagnation depth index includes: when the cross-attempt response delay trend coefficient is greater than zero, multiplying the failure rate, the cross-attempt response delay trend coefficient, and the peak indecisiveness dwell time within the attempt to obtain the node stagnation depth index; when the cross-attempt response delay trend coefficient is not greater than zero, setting the node stagnation depth index directly to zero.

[0013] This ensures that the stagnation depth index is significantly positive only when multiple pieces of evidence are jointly established, which conforms to the rule of joint judgment of multiple evidence. It significantly reduces the probability of misjudging repeated trial-and-error exploration as a stagnation state and has the ability to accurately identify cognitive stagnation at the level of a single interaction node.

[0014] Preferably, the step of entering the scaffold injection process involves determining the scaffold strength level based on the specific range into which the node stagnation depth index falls, and performing the corresponding level of injection operation. This includes: when the node stagnation depth index falls between the first-level trigger threshold and the second-level trigger threshold, it is determined to be a first-level scaffold, and a text prompt is presented in the form of a spatial anchor point within the virtual reality scene; when the node stagnation depth index falls between the second-level trigger threshold and the third-level trigger threshold, it is determined to be a second-level scaffold, and a spatial highlight effect is superimposed on the three-dimensional contour of the target interactive object; when the node stagnation depth index exceeds the third-level trigger threshold, it is determined to be a third-level scaffold, and a guiding animation for the operation steps of that node is played within the virtual reality space.

[0015] By using a three-tiered spatial anchor-type micro-scaffold with progressive strength and an adaptive level selection based on the stagnation depth index, precise intervention is achieved for cognitive stagnation at a single interaction node without resetting the current virtual reality scene or disrupting the immersive experience. This effectively reduces the risk of actively leaving the learning scene after repeated frustration at the same node.

[0016] Preferably, after performing the injection operation at the corresponding level, the method further includes: listening for subsequent interaction events of the node in an event-driven manner; if the node operation is successfully completed during the scaffold presentation, the scaffold automatically disappears, and the scene state advances to the next node; if the node operation is not successfully completed after the scaffold presentation ends, the current scaffold level is increased by one level, the corresponding level injection operation is re-executed, and the subsequent listening state of the node is reset; if the node operation is still not successfully completed after the third-level scaffold is triggered, the node is recorded as an unreachable node in the current session, guiding the user to skip the node and continue advancing the subsequent scene.

[0017] Preferably, the step of updating the node-level mastery graph based on the complete interaction records of each node includes: extracting the first attempt success rate of the interactive node in the current session, the cumulative success rate of the interactive node in historical sessions, and whether the interactive node triggered the micro-scaffold and its trigger level in the current session; calculating the mastery score based on the first attempt success rate, the cumulative success rate, and the trigger level; and updating the mastery score to the corresponding node in the node mastery graph.

[0018] Preferably, the step of performing path replanning based on the updated node-level mastery graph and selecting the scene sequence for the next session includes: prioritizing several nodes with the lowest mastery scores in the node-level mastery graph as targets, and searching the scene library for scene modules containing the target nodes; selecting the scene sequence for the next session from the candidate scene module set according to the dual objective principle of maximizing target node coverage and minimizing total scene duration; when multiple candidate scene sequences have the same coverage, taking the sequence with the shortest total duration as the final planning result; when there is a conflict between coverage and duration, prioritizing coverage and using duration as a secondary sorting criterion.

[0019] In a second aspect, this application provides an interactive intelligent education system based on artificial intelligence and virtual reality, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned interactive intelligent education method based on artificial intelligence and virtual reality is implemented.

[0020] By adopting the above technical solution, an interactive intelligent education method combining artificial intelligence and virtual reality is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, facilitating its use.

[0021] This application achieves accurate identification of cognitive stagnation states at micro-interaction nodes by combining multimodal native signal acquisition and personal baseline calibration with dimensional transformation of time series matrices and extraction of multidimensional objective statistics. It effectively overcomes the problem of misjudgment due to individual differences caused by traditional fixed threshold judgment, provides reliable data support for subsequent precise intervention, and further enhances the system's ability to perceive the learner's true cognitive state.

[0022] This application utilizes spatial anchor-type micro-scaffolds for adaptive intensity intervention and performs dual-objective path replanning based on node-level mastery maps after the session ends. Without disrupting the learner's immersive experience, it achieves precise contextual guidance and also uses the micro-stagnation detection results as the basis for planning long-term learning paths, enabling learning resources to be concentrated on cognitive weaknesses and significantly enhancing the relevance and coherence of long-term learning. Attached Figure Description

[0023] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart illustrating an interactive intelligent education method and system based on artificial intelligence and virtual reality according to the present invention.

[0024] Figure 2 The diagram illustrates a comparison of response latency trends between an interactive intelligent education method based on artificial intelligence and virtual reality in this invention and existing technologies. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] This invention discloses an interactive intelligent education method and system combining artificial intelligence and virtual reality, referring to... Figure 1 This includes steps S1-S4: S1. Collect native interactive signals to establish a detection benchmark.

[0028] In an optional embodiment, when a VR learning session is started, the system will simultaneously open three parallel native signal acquisition channels, corresponding to gaze persistence signal, interactive event stream and node response delay, respectively. All three types of signals are directly from the native output interface of the head-mounted display hardware layer, without the need to deploy any external sensors on the learner.

[0029] The gaze persistence signal is continuously output by the eye-tracking module built into the head-mounted display at a fixed sampling frequency. The system determines the spatial affiliation of each interactive object by matching the object identifier with the gaze point coordinates of the current frame, and calculates the learner's continuous dwell time on the object. The interactive event stream records all actions performed by the learner at each atomic interaction node in an event-driven manner. Each event record includes a timestamp of the trigger time, an operation type label, and the system's judgment result for the operation. These three fields together constitute the original data unit for subsequent matrix construction. The node response delay is the time interval between the moment the system presents the task prompt when the learner enters a certain interaction node and the moment the learner first performs any operation. This time interval is recorded as the response delay value, which directly reflects the learner's initial cognitive reaction speed when faced with the task prompt.

[0030] Furthermore, the VR scene is pre-decomposed into a directed acyclic graph (DAG) structure at the design level. Each node in the graph corresponds to an atomic interaction task, and the directed edges between nodes represent the pre-dependencies of the tasks. Taking a virtual chemistry experiment scene as an example, its DAG may contain several sequential nodes such as sampling, heating, observation, and recording. Each node maintains an independent interaction record queue during system operation. After each attempt, the operation result label and response latency value of this attempt are appended to the end of the record queue of the corresponding node in real time. The queue saves all historical attempt records in a first-in-first-out order, continuously providing raw data streams for the construction of subsequent time-series matrices. This ensures that the node-level historical interaction information is completely retained throughout the entire session cycle and is not overwritten or cleared as the scene state progresses.

[0031] Before the formal session begins, the system executes a personal baseline calibration process, summarizing the baseline response latency values ​​and minimum observation window sizes for each difficulty level into a personal latency baseline profile to establish a detection benchmark. The calibration process presents learners with several standard interactive nodes of known difficulty levels in sequence. The difficulty level labels are marked according to a preset grading system, for example, they can be divided into three levels: beginner, intermediate, and advanced.

[0032] During the calibration process, the system collects all response latency values ​​for each learner at each difficulty level. For multiple collections at the same difficulty level, its distribution characteristics are calculated, and the median of this distribution is used as the learner's baseline response latency value at that corresponding difficulty level. The baseline values ​​for all difficulty levels are then aggregated into the learner's personal latency baseline profile. Each record in the baseline profile contains two fields: a difficulty level label and the corresponding baseline response latency value. The profile is written before the session begins and remains read-only throughout the session, only being updated after the learner completes a new round of calibration.

[0033] Next, the system writes the median number of average attempts by learners at each difficulty level collected during the calibration phase into the configuration field, which serves as the basis for the minimum observation window size in subsequent stagnation detection. The introduction of the minimum observation window ensures that the system has accumulated a sufficient number of attempt records before performing a stagnation determination on a certain node, avoiding statistical distortion due to insufficient sample size. For example, assuming that the calibration phase shows that the median number of average attempts by the learner on intermediate difficulty nodes is 4, the system will not perform stagnation depth calculation on that node until the learner's cumulative number of attempts on intermediate nodes reaches 4. Instead, it will continue to accumulate interaction records, and trigger the subsequent matrix construction process only after the window is satisfied.

[0034] In this way, through the coordinated execution of three parallel native signal acquisitions and personalized baseline calibration, the system completes the establishment of a detection benchmark with the learner's individual differences as the reference anchor point before the session officially begins. This effectively avoids the systematic bias of misjudging learners with slower reaction pace as being in a stagnant state, and provides a reliable personalized reference basis for subsequent node-level stagnation depth measurement.

[0035] S2, Reconstruct the two-dimensional time-series matrix to calculate the stagnation depth of nodes.

[0036] In an optional embodiment, once the cumulative number of attempts for a certain interaction node reaches the number specified by the minimum observation window, the system immediately performs a dimensionality transformation operation on the historical interaction record queue of that node, reconstructing the attempt records originally stored as a one-dimensional sequence into a two-dimensional time-series matrix. The row index of the matrix corresponds to the attempt round, and the column index corresponds to the time slice within each attempt. The sampling interval of the slice is taken as the median of the average single operation time of the learner at the current difficulty level during the calibration phase, to ensure that the slice level matches its own operation rhythm.

[0037] Each element in the matrix represents the total count of operation events that occurred within the corresponding time slice for the corresponding attempt round. The count range covers all operation event types, including triggering, confirmation, and cancellation. This dimensional transformation unfolds the internal time structure of the attempt, which was originally compressed into a single response delay scalar, into a two-dimensional data structure that can be analyzed column by column. This allows the system to characterize the learner's operational continuity from the perspective of the internal time distribution of the attempt.

[0038] Furthermore, the number of columns in each row is uniformly taken as the maximum number of valid slices across all trial rounds. For rows where the number of valid slices is less than the maximum, zeros are padded to the end of the row to a uniform number of columns, maintaining the matrix's regular two-dimensional structure. In the subsequent statistics extraction stage, only the range of valid slices within each row is counted, i.e., the continuous zero segment from column 1 to the column corresponding to the actual number of valid slices in that row. The zero-padding portion is not included in the statistics. The normalized denominator is the actual number of valid slices in that row, not the total number of columns in the matrix, to ensure that the statistical results of the gap length in each row are not affected by the padding operation.

[0039] Based on the constructed time series matrix, the system extracts three mutually orthogonal objective statistics to characterize the learner's interaction state at the node from three independent dimensions: macro results, cross-trial time series, and intra-trial time series.

[0040] The first statistic is the failure rate, which is the ratio of the number of failed attempts to the total number of attempts. It is a dimensionless ratio ranging from 0 to 1, directly reflecting the macro-level failure frequency at that point. The second statistic is the cross-attempt response delay trend coefficient, determined by the slope obtained from performing a least-squares linear regression on the response delay value sequence of each round of attempts. The unit is milliseconds per attempt. A positive value indicates that the response delay increases with the number of attempts, meaning the learner becomes more hesitant with each attempt; a non-positive value indicates that the delay does not increase, and the learner is in a self-correcting state. The third statistic is the peak in-trial hesitation dwell time, calculated using... The method is as follows: For each row of the time series matrix, count the number of the longest consecutive segments with a value of zero within the actual effective slice range of that row, and use this count as the gap length of that row. Take the arithmetic mean of the gap lengths of all rows, and then divide it by the actual effective slice count of the corresponding row to normalize it, resulting in a dimensionless ratio between 0 and 1. The closer this statistic is to 1, the more likely the learner has long-term operational interruptions within each attempt, such as staring at the target interactive object for a long time without taking any action; the closer it is to 0, the more likely the operational actions within the attempt are continuous and smooth.

[0041] Next, based on the three extracted statistics, the system calculates the node stagnation depth index. The failure rate characterizes the frequency of failure from a macro-result perspective; the cross-trial response delay trend coefficient characterizes the trend of increasing hesitation across trials from a temporal perspective; and the intra-trial hesitation peak characterizes the degree of cognitive gap within a single trial from a temporal perspective. These three metrics originate from three orthogonal observation dimensions, and the absence of any one dimension is insufficient to independently constitute a complete scale of cognitive stagnation. In this scheme, the three metrics are multiplied together, ensuring that the stagnation depth index is significantly positive only when all three dimensions are simultaneously positive, conforming to the judgment rule of multiple pieces of evidence being jointly valid. The specific relationship is as follows: When the cross-attempt response delay trend coefficient is greater than 0:

[0042] When the cross-attempt response delay trend coefficient is not greater than 0:

[0043] in, For failure rate, This is the trend coefficient for response latency across attempts, expressed in milliseconds per attempt. To attempt to reach the peak of internal hesitation. When When the value is not greater than 0, it indicates that the learner's response delay does not increase with the number of failures and may even decrease. The system judges the learner to be in an active exploration and self-correction state, does not trigger subsequent intervention processes, and sets the stagnation depth index directly to zero.

[0044] because and All are dimensionless ratios between 0 and 1. The actual magnitude is determined by the learner's response latency level. For example, suppose a learner... Approximately 50 milliseconds per cycle. It is 0.8. If it is 0.6, then Approximately 24 milliseconds per iteration. During the group calibration phase, the system performs calibration on a large number of learners. By statistically analyzing the historical distribution, the 75th, 90th, and 99th percentiles can be used as the three-level trigger thresholds, so that the specific values ​​of the three thresholds are automatically determined based on the population calibration data.

[0045] The system compares the calculated stagnation depth index with a preset trigger threshold to determine whether intervention scaffolding needs to be injected into the learner. This comparison is retried each time a new attempt record is added to the queue, ensuring that the identification of stagnation states has real-time responsiveness, rather than performing batch judgments only at fixed time points. If the stagnation depth index does not exceed the trigger threshold, the system continues to accumulate subsequent interaction records without making any modifications to the current scenario state; if it exceeds the trigger threshold, the scaffolding injection process is immediately initiated, and the specific execution logic is described in subsequent steps.

[0046] Thus, by reconstructing a one-dimensional attempt sequence into a two-dimensional time-series matrix and extracting objective statistics from three orthogonal dimensions, the system has for the first time the ability to accurately identify cognitive stagnation at the level of a single interaction node within a single session. This effectively overcomes the structural defect of existing session-level assessments that are unaware of stagnation nodes under micro-level conditions. At the same time, through the multiplicative relationship of joint judgment of multiple evidences, the probability of misjudging repeated trial-and-error exploration as a stagnation state is significantly reduced.

[0047] S3, matching node stagnation depth triggers the corresponding level of support.

[0048] In an optional embodiment, when the stagnation depth index exceeds the first-level trigger threshold, the system determines that the node has entered a stagnation state that requires intervention, and determines the stent strength level according to the specific range in which the stagnation depth index falls, and then performs the corresponding level of injection operation.

[0049] Furthermore, the micro-scaffolds are designed with three intensity levels, each progressively increasing in the degree of immersive intrusion. The first-level scaffold presents text prompts as spatial anchors within the VR scene, pointing to the key points of the current node's operation without revealing the specific answer, providing only directional guidance. This is suitable when the stagnation depth index falls between the first and second-level trigger thresholds. The second-level scaffold overlays a spatial highlight effect on the 3D outline of the target interactive object, guiding the learner's gaze to the correct object through increased visual salience. This is suitable when the stagnation depth index falls between the second and third-level trigger thresholds. The third-level scaffold plays a guided animation within the VR space, fully demonstrating the correct operation process. After the animation ends, the scene remains unchanged, and the learner still needs to complete the operation at that node independently. This is suitable when the stagnation depth index exceeds the third-level trigger threshold. All three levels of scaffolds are embedded into the scene as spatial anchors, their coordinates bound to the center of the 3D bounding box of the interactive object corresponding to the current stagnation node. This ensures that the scaffold content is precisely aligned with the problem object in space, rather than interrupting the learner's immersive experience with a full-screen pop-up.

[0050] Specifically, the maximum duration of the guided animation for the third-level scaffold is anchored to the learner's baseline response delay at the current difficulty level, and can be up to three times the baseline response delay. The single baseline response delay reflects the natural cognitive cycle required for the learner to complete one operational judgment. A length that is an integer multiple of three can cover the three consecutive cognitive segments of observation, understanding, and reproduction, without exceeding the learner's single-time attentional stability window, thus preventing attention loss due to excessively long animations. For example, assuming a learner's baseline response delay at the intermediate difficulty level is 2000 milliseconds, the maximum duration of the guided animation for the third-level scaffold is 6000 milliseconds. The system trims the animation content based on this maximum duration during the animation production stage to ensure that the animation fully conveys the key operational points within the time constraint.

[0051] Following scaffolding injection, the system continues to monitor subsequent interaction events at that node. If the learner successfully completes the node operation during scaffolding presentation, the scaffolding immediately and automatically disappears, and the scene state seamlessly advances to the next node in the directed acyclic graph. The entire transition process is transparent to the learner and does not cause any scene reset or loading interruption. If the learner still fails to complete the node operation after the scaffolding presentation ends, the system upgrades the current scaffolding level by one level, re-executes the injection operation at the corresponding level, and resets the subsequent monitoring state of the node. When the learner still fails to complete the node operation after the third-level scaffolding is triggered, the system records the node as an unreachable node in the current session, automatically guides the learner to skip the node and continue advancing to the subsequent scene, and marks the node as a key review object after the session ends, including it in the priority processing queue for subsequent path replanning.

[0052] In this way, through the three-level progressive spatial anchor-type micro-scaffolding and the adaptive level selection based on the stagnation depth index, the system can achieve precise intervention on cognitive stagnation at a single interaction node without resetting the current VR scene or disrupting the learner's immersive experience. This effectively reduces the risk of learners actively leaving the learning scene after being repeatedly frustrated at the same node.

[0053] S4. Update the node mastery graph and perform path replanning.

[0054] In an optional embodiment, after each VR learning session, the system updates the learner's node-level mastery graph based on the complete interaction records of each node in the session. The mastery graph uses nodes of a directed acyclic graph as basic units, with each node maintaining a mastery score. The score is calculated by comprehensively considering three dimensions: the first-attempt success rate of the node in the current session, reflecting the learner's immediate mastery of the node; the cumulative success rate of the node in historical sessions, reflecting the learner's long-term memory stability of the node; and whether the node triggered a micro-scaffolding in the current session and its trigger level. A higher trigger level results in a greater reduction in the mastery score. Let the first-attempt success rate of the current session be... The historical cumulative success rate is The trigger level reduction item for micro-stents is When not triggered Set to 1, triggering the first level. Take 0.8, when the second level is triggered Take 0.6, when the third level is triggered Taking 0.4, the normalized weights for the three dimensions are as follows: , , Node mastery score The relation is:

[0055] The weights of the three dimensions are determined by the ratio of the absolute values ​​of the Pearson correlation coefficients of each dimension in predicting learners' subsequent test scores: first, the absolute values ​​of the Pearson correlation coefficients of each dimension are taken to eliminate the sign interference introduced by negative correlations; then, the absolute values ​​of each dimension are normalized by dividing the sum of the absolute values ​​of the three dimensions to ensure... , , The sum of all values ​​is strictly equal to 1 and all are non-negative, which provides an objective basis for the predictive validity of the weight allocation.

[0056] Furthermore, based on the updated mastery graph, the system performs path replanning before the next session begins: the planning logic prioritizes the nodes with the lowest mastery scores in the graph, searches the VR scene library for scene modules containing these target nodes, and selects the scene sequence for the next session from the candidate scene module set according to the dual objective principle of maximizing target node coverage and minimizing total scene duration. Prioritizing target node coverage ensures that learning resources are concentrated on nodes where learners truly have cognitive weaknesses, while minimizing total scene duration controls the time cost of a single session while meeting coverage requirements, avoiding diluting effective practice density due to scene redundancy. When multiple candidate scene sequences have the same coverage, the system selects the sequence with the shortest total duration as the final planning result; when there is a Pareto conflict between coverage and duration, the system prioritizes coverage, using duration as a secondary ranking criterion.

[0057] Specifically, for nodes marked as unreachable in the current session, the system plans a separate, progressive preparatory path for them in subsequent sessions, rather than directly reintegrating them into the formal session sequence. The preparatory path is constructed based on the node's preceding dependencies in the directed acyclic graph. The system backtracks upwards along these dependencies, extracting all directly preceding nodes and arranging reinforcement practice scenarios for them in the scenario library. The system continuously monitors the mastery scores of each preceding node, using the 50th percentile of the historical mastery distribution of preceding nodes as the readiness threshold. When a learner's mastery score for a preceding node exceeds this threshold, it is determined that the learner's mastery of that preceding node has surpassed their historical average, and they possess the knowledge to challenge the previously stalled node again. Once the mastery scores of all directly preceding nodes have reached the readiness threshold, the system reintegrates the previously unreachable node into the formal session sequence and configures a pre-loaded state of the first-level micro-scaffold in the first session in which the node reappears. This ensures the scaffold is in an activated state when the learner first enters the node, shortening the response delay from entering the node to receiving guidance.

[0058] Next, the updated mastery graph results also serve as the data source for the system-level learning progress report. After each session, the system generates a node-level learning status summary based on the mastery scores of each node in the graph. The summary includes the number of newly added mastery nodes in this session, a list of nodes with significantly decreased mastery scores, and the cumulative number of unreachable nodes in the current session, which can be viewed by teachers or learners themselves. The summary data uses the directed acyclic graph topology of the graph nodes as the display framework, mapping the mastery scores to the corresponding nodes in the form of a heatmap. This allows learners or teachers to intuitively locate weak areas in the current knowledge graph, providing objective data support for manual intervention or curriculum adjustments.

[0059] Figure 2To illustrate the trend comparison of response delay, two broken lines and corresponding trend lines are used to represent the evolution of response delay in the existing method and the method of the present invention during multiple attempts. It can be observed that the response delay of the existing method shows a continuous upward trend with the number of attempts, while the slope of the response delay trend line of the method of the present invention is significantly lower and even tends to be stable. This indicates that the present invention effectively suppresses the cumulative increase in learners' hesitation after repeated setbacks through personalized baseline calibration and stagnation depth measurement.

[0060] In this way, learning resources are focused on the learners' actual cognitive weaknesses in each session, which significantly enhances the relevance of long-term learning. At the same time, the design of a progressive preparatory path breaks the stagnation cycle caused by the system repeatedly presenting the same high-difficulty node before the learners have acquired the prior knowledge.

[0061] This invention also discloses an interactive intelligent education system based on artificial intelligence and virtual reality, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an interactive intelligent education method based on artificial intelligence and virtual reality according to this invention.

[0062] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. An interactive intelligent education method combining artificial intelligence and virtual reality, characterized in that, include: Open parallel native signal acquisition channels to collect interactive event streams and node response delays respectively, execute the personal baseline calibration process, and summarize the baseline response delay values ​​and minimum observation window size under each difficulty level into a personal delay baseline file to establish a detection benchmark. When the cumulative number of attempts of a certain interaction node reaches the number specified by the minimum observation window, a dimension transformation operation is performed on the historical interaction record queue of the node to reconstruct it into a two-dimensional time series matrix, and the node stagnation depth index is calculated based on the two-dimensional time series matrix. The node stagnation depth index is compared with a preset trigger threshold. When the node stagnation depth index exceeds the trigger threshold, the stent injection process is initiated. The stent strength level is determined based on the specific range in which the node stagnation depth index falls, and the corresponding level of injection operation is performed. Based on the complete interaction records of each node, the node-level mastery graph is updated. Based on the updated node-level mastery graph, path replanning is performed to select the scene sequence for the next session.

2. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, The process involves opening parallel native signal acquisition channels to collect interactive event streams and node response delays, executing a personal baseline calibration process, and summarizing the baseline response delay values ​​and minimum observation window sizes at each difficulty level into a personal delay baseline profile to establish a detection benchmark, including: Spatial attribution is determined by the object identifier of each interactive object and the gaze landing point coordinates of the current frame, and the continuous dwell time on the object is accumulated as the gaze dwell signal. Record all operations issued on each atomic interaction node in an event-driven manner as an interaction event stream; The time interval between the moment the task prompt is displayed when entering an interactive node and the moment the first arbitrary action is issued is used as the node response delay. The system sequentially presents learners with several standard interactive nodes of known difficulty levels, collects all response latency values ​​for each difficulty level, calculates the distribution characteristics of multiple collection results for the same difficulty level, uses the median of the distribution as the baseline response latency value for the corresponding difficulty level, and summarizes the baseline response latency values ​​for all difficulty levels into a personal latency baseline profile.

3. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, The step of performing a dimensionality transformation operation on the historical interaction record queue of the node to reconstruct it into a two-dimensional time series matrix includes: The attempt records stored in the historical interaction record queue as a one-dimensional sequence are reconstructed into the two-dimensional time-series matrix; The row index of the two-dimensional time series matrix corresponds to the trial round, and the column index corresponds to the time slice within each trial. The sampling interval of the time slice is taken as the median of the average time taken for a single operation during the calibration phase at the current difficulty level. The value of each element in the two-dimensional time sequence matrix is ​​the total count of operation events that occurred in the corresponding time slice for the corresponding attempt round.

4. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, The extraction of objective statistics based on the two-dimensional time series matrix includes: Extract the failure rate, which is the ratio of the number of failed attempts to the total number of attempts; Extract the cross-trial response delay trend coefficient, which is determined by the slope obtained by performing least squares linear regression on the response delay value sequence of each round of trials; Extract the peak value of hesitation dwelling within the trial. For each row of the two-dimensional time series matrix, count the number of the longest segment among the continuous segments with an element value of zero in that row as the gap length of that row. Take the arithmetic mean of the gap lengths of all rows, and then divide it by the actual effective number of segments in that row to normalize and obtain the peak value of hesitation dwelling within the trial.

5. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 4, characterized in that, The stagnation depth metric for computing nodes includes: When the cross-attempt response delay trend coefficient is greater than zero, the failure rate, the cross-attempt response delay trend coefficient, and the peak in-attempt hesitation dwell time are multiplied to obtain the node stagnation depth index. When the cross-attempt response delay trend coefficient is not greater than zero, the node stagnation depth index is directly set to zero.

6. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, The stent injection process involves determining the stent strength level based on the specific range within which the node stagnation depth index falls, and then performing the corresponding level of injection operation, including: When the node stagnation depth index falls between the first-level trigger threshold and the second-level trigger threshold, it is determined to be a first-level support, and a text prompt is presented in the form of a spatial anchor point in the virtual reality scene. When the node stagnation depth index falls between the second-level trigger threshold and the third-level trigger threshold, it is determined to be a second-level support, and a spatial highlighting effect is superimposed on the three-dimensional outline of the target interactive object. When the node stagnation depth index exceeds the third-level trigger threshold, it is identified as a third-level support, and a guiding animation for the operation steps of that node is played in the virtual reality space.

7. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, After performing the injection operation at the corresponding level, the process also includes: Listen for subsequent interaction events of this node in an event-driven manner; If the node operation is successfully completed during the scaffold presentation, the scaffold will automatically disappear, and the scene state will advance to the next node; If the node operation is not successfully completed after the scaffold presentation is finished, the current scaffold level will be increased by one level, the injection operation at the corresponding level will be re-executed, and the subsequent listening state of the node will be reset. If the node operation fails to be completed after the third-level support is triggered, the node is recorded as an unreachable node in the current session, and the session is guided to skip the node and continue to the next scenario.

8. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, The step of updating the node-level mastery graph based on the complete interaction records of each node includes: Extract the first-attempt success rate of the interaction node in this session, the cumulative success rate of the interaction node in historical sessions, and whether the interaction node in this session triggered the micro-scaffold and its trigger level; A mastery score is calculated based on the first attempt success rate, cumulative success rate, and trigger level. Update the mastery score to the corresponding node in the mastery graph.

9. The interactive intelligent education method combining artificial intelligence and virtual reality according to claim 1, characterized in that, The process of performing path replanning based on the updated node-level mastery graph and selecting the scenario sequence for the next session includes: Prioritize the nodes with the lowest mastery scores in the node-level mastery graph, and search the scene library for scene modules containing the target nodes. Based on the dual objective principle of maximizing target node coverage and minimizing total scene duration, the scene sequence for the next session is selected from the candidate scene module set. When multiple candidate scene sequences have the same coverage, the sequence with the shortest total duration is taken as the final planning result. When there is a conflict between coverage and duration, coverage is given priority, and duration is used as a secondary sorting criterion.

10. An interactive intelligent education system combining artificial intelligence and virtual reality, characterized in that: include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement an interactive intelligent education method for artificial intelligence and virtual reality according to any one of claims 1-9.

Images