Data processing method and system for virtual physical education

Abstract

The invention relates to the technical field of virtual teaching, and discloses a data processing method and system for virtual physical education teaching, and the method comprises the steps: collecting and preprocessing the exercise data of a learner; analyzing the difference between the action posture of the learner and the standard action; predicting a skill forgetting rule of the learner; generating a personalized review strategy; executing a self-adaptive learning forgetting review closed-loop process; the motion recognition accuracy is improved through multi-source data fusion, skill attenuation is predicted through a personalized forgetting curve, the optimal review strategy is generated through reinforcement learning, the technical problems that in virtual physical education, the motion capture accuracy is insufficient, skill forgetting is difficult to predict, and the review strategy lacks personalization are solved, and the virtual physical education teaching experience is improved. The learning efficiency and the skill retention rate are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of virtual teaching, more particularly, it relates to a data processing method and system for virtual sports teaching. BACKGROUND

[0002] With the development of information technology, virtual sports teaching as a new teaching mode is gradually being widely applied. Virtual sports teaching provides an immersive sports skill learning environment for learners through computer simulation, motion capture, virtual reality and other technologies, breaking through the limitations of time, space and resources in traditional sports teaching.

[0003] Currently, virtual sports teaching mainly uses a single camera or sensor for motion capture, uses a unified learning schedule, and provides skill guidance and evaluation according to the preset teaching process.

[0004] However, these existing technologies have three main problems: the accuracy of motion capture by a single data source is insufficient, especially in complex lighting, high-speed motion or occlusion conditions, it is difficult to accurately capture the details of the learner's motion, resulting in delayed feedback and inability to effectively correct incorrect movements; the skill forgetting rate of different learners is different, and the existing technology lacks accurate modeling of the skill forgetting law, and cannot predict the degree of skill decline caused by interval learning; review strategies lack personalization and fail to provide optimal review plans based on the characteristics and skill mastery of learners, resulting in waste of teaching resources and poor learning effectiveness. In addition, the various data sources (such as posture data, learning history, physiological state, etc.) in the existing technology are not effectively integrated, resulting in the teaching system being unable to comprehensively assess the learning state and being difficult to provide targeted teaching guidance.

[0005] Therefore, there is a need for a virtual sports teaching data processing method that can accurately capture motion, accurately predict skill forgetting patterns and provide personalized review strategies to improve learning efficiency and skill retention rate. SUMMARY

[0006] The present application provides a data processing method and system for virtual sports teaching, which solves the technical problems of insufficient motion capture accuracy, difficulty in predicting skill forgetting patterns and lack of personalized review strategies in related technologies.

[0007] The present application provides a data processing method for virtual sports teaching, comprising the following steps:

[0008] Collecting learner motion data and preprocessing;

[0009] Based on the preprocessed data, analyzing the differences between the learner's action posture and the standard action;

[0010] According to the action difference analysis result, the skill forgetting law of the learner is predicted;

[0011] Based on the skill forgetting prediction result, an individualized review strategy is generated;

[0012] Based on the individualized review strategy, an adaptive learning forgetting review closed loop process is executed.

[0013] In a preferred embodiment, the step of collecting learner motion data and preprocessing includes:

[0014] Collecting multi-source heterogeneous data including video stream data, inertial sensor data and pressure sensor data;

[0015] Standardizing the collected data;

[0016] Calculating adaptive weights of each data source;

[0017] Based on the adaptive weights, multi-source data fusion is performed.

[0018] In a preferred embodiment, the step of analyzing the difference between the learner's action posture and the standard action includes:

[0019] Using a three-level linked posture feature extraction algorithm to analyze the action;

[0020] Comparing the learner and the standard action sequence by dynamic time warping algorithm;

[0021] Generating an action difference heat map;

[0022] Based on the difference analysis result, automatically generating improvement suggestions.

[0023] In a preferred embodiment, the three-level linked posture feature extraction algorithm includes:

[0024] Key point layer: extracting three-dimensional coordinate points of human joints;

[0025] Relationship layer: analyzing the relative position and motion relationship between key points through graph convolution network;

[0026] Semantic layer: identifying the semantic information of the action through bidirectional long short-term memory network.

[0027] In a preferred embodiment, the step of predicting the skill forgetting law of the learner includes:

[0028] Decomposing the entire sports skill into multiple atomic action units, and constructing a skill dependency graph;

[0029] Based on the extended Ebbinghaus forgetting curve, a basic forgetting model is established;

[0030] Bayesian knowledge tracing algorithm is applied to estimate the mastery probability of each sub-skill for a learner;

[0031] A recurrent neural network based on long short-term memory or gated recurrent unit is constructed to predict the skill retention level of a learner.

[0032] In a preferred embodiment, the extended Ebbinghaus forgetting curve establishes a base forgetting model as:

[0033]

[0034] where R mem (t) represents the memory retention rate, t represents the time interval, S stab represents the relative memory stability, f(P, D diff , E) is a stability adjustment function, P represents the practice frequency, D diff represents the skill difficulty, E represents the experience level of the learner, and e represents the base of the natural logarithm.

[0035] In a preferred embodiment, the step of generating a personalized review strategy comprises:

[0036] Defining the state space of reinforcement learning, including the skill mastery vector of the learner, the historical review record, and the personal characteristics;

[0037] Defining the action space of reinforcement learning, including review content selection, review difficulty setting, and review form determination;

[0038] Designing a multi-step reward function to balance learning effectiveness, time efficiency, and learner engagement;

[0039] Training the optimal review strategy through a deep Q network.

[0040] In a preferred embodiment, the multi-step reward function is:

[0041] R reward = α reward · ΔSkill + β reward · Time -1 + γ reward · Engagement;

[0042] where R reward represents the multi-step reward function, ΔSkill represents the improvement of skill mastery after review, Time -1 represents the inverse of time efficiency, Engagement represents the engagement score of the learner, α reward , β reward , γ rewardrespectively represent the weight coefficients of the improvement of the mastery of the reviewed skills, the reciprocal of the time efficiency and the importance of the participation score of the learners.

[0043] In a preferred embodiment, the step of performing the adaptive learning-forgetting-review closed loop process comprises:

[0044] generating a personalized learning plan;

[0045] monitoring and evaluating the learning process in real time;

[0046] predicting and updating the skill forgetting curve;

[0047] scheduling optimal review timing and content;

[0048] establishing a long-term maintenance and closed loop optimization mechanism.

[0049] In a preferred embodiment, a data processing system for virtual sports teaching is used to perform a data processing method for virtual sports teaching, comprising:

[0050] a data acquisition and preprocessing module for collecting learner motion data and preprocessing;

[0051] an action analysis module for analyzing the difference between the learner's action posture and the standard action;

[0052] a forgetting prediction module for predicting the skill forgetting law of the learner;

[0053] a review strategy generation module for generating personalized review strategies;

[0054] a closed loop execution module for performing an adaptive learning-forgetting-review closed loop process.

[0055] The beneficial effects of the present application are:

[0056] Multi-source heterogeneous data fusion and adaptive weight calculation are adopted to improve the accuracy of motion capture, especially in complex lighting conditions and high-speed action scenes to maintain stable performance;

[0057] The personalized skill forgetting prediction model reduces the prediction error, making the review timing arrangement more accurate;

[0058] The review strategy generated based on reinforcement learning improves the skill mastery level of the learners in the same learning time, and reduces the total time required for learning to proficiency;

[0059] The adaptive closed loop mechanism enables the system to dynamically adjust the teaching strategy according to the actual performance of the learners, adapt to the characteristics of different learners, and meet the needs of personalized teaching. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flow chart of a data processing method for virtual sports teaching according to the present application;

[0061] Figure 2 is a detailed flow chart of collecting and pre-processing learner motion data according to the present application;

[0062] Figure 3 is a detailed flow chart of analyzing the difference between the learner's action posture and the standard action according to the present application;

[0063] Figure 4 is a detailed flow chart of predicting the skill forgetting rule of the learner according to the present application;

[0064] Figure 5 is a detailed flow chart of generating personalized review strategy according to the present application;

[0065] Figure 6 is a detailed flow chart of performing adaptive learning forgetting review closed loop process according to the present application. DETAILED DESCRIPTION

[0066] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand so as to be able to implement the subject matter described herein, and variations of elements discussed can be made by one skilled in the art, without departing from the scope of the present specification. Various examples can omit, substitute, or add various procedures or components as appropriate, and the methods described can not require the particular order of steps described, where appropriate. Also, features described in relation to one example can be combined in other examples.

[0067] In at least one embodiment of the present application, a data processing method for virtual sports teaching is disclosed, as shown in Figures 1 to 6 comprising the following steps:

[0068] Step 1, collecting and pre-processing learner motion data;

[0069] Specifically comprising the following sub-steps:

[0070] Step 1.1, multi-source data acquisition;

[0071] The present application collects multi-source heterogeneous information including camera video stream, inertial sensor data and pressure sensor data, etc.

[0072] The camera video stream provides visual representation of the learner's action, the inertial sensor (such as accelerometer, gyroscope) provides motion parameters and posture information, and the pressure sensor provides measurement data of the ground reaction force.

[0073] These sensors can be placed at different locations in the learning environment or directly on the learner's key joints, providing multi-dimensional observations of the learner's movements.

[0074] Step 1.2, data preprocessing;

[0075] The collected raw data is denoised, time-synchronized, and standardized;

[0076] For camera video streams, Gaussian filtering is applied to remove noise and extract key frames;

[0077] For inertial sensor data, Kalman filtering algorithm is used to eliminate random errors;

[0078] For pressure sensor data, median filtering is used to eliminate spike interference.

[0079] Through timestamp calibration, the data streams of different sensors are aligned to a unified time axis, ensuring the timing consistency of the data.

[0080] Standardize different types of data to the same scale range to prepare for subsequent fusion.

[0081] Step 1.3, adaptive weight calculation;

[0082] According to the signal-to-noise ratio, delay and reliability of each sensor data, the application adaptively calculates the fusion weight.

[0083] The fusion weight calculation formula is:

[0084] W i =f(SNR i ,Latency i ,Reliability i );

[0085] Where W i represents the fusion weight of the i-th sensor, used to determine the importance of the sensor data in the final fusion result; SNR i represents the signal-to-noise ratio of the sensor data, i.e. the ratio of effective signal to noise, the higher the value, the better the data quality; Latency i represents the data delay, i.e. the time difference from data generation to system reception and processing, the smaller the value, the better the real-time performance; Reliability i represents the reliability score based on historical performance evaluation, reflecting the stability and accuracy of the sensor in the past use. The function f adopts the weighted sum form to consider these three parameters comprehensively, and dynamically calculates the optimal fusion weight:

[0086]

[0087] where, represents the optimal fusion weight of the ith sensor; SNR i represents the signal-to-noise ratio of the ith sensor data, the higher the value, the better the data quality; Latency i represents the data latency of the ith sensor, the lower the value, the better the real-time performance of the data; Reliability i represents the reliability score of the ith sensor based on historical performance evaluation, the higher the value, the more reliable the sensor; α sensor , β sensor , γ sensor respectively represent the weight coefficients of controlling the relative importance of signal-to-noise ratio, latency and reliability, satisfying α sensor + β sensor + γ sensor = 1; n sensor is the total number of sensors participating in fusion; represents the sum of the corresponding parameters of all sensors for normalization. The system dynamically adjusts these coefficients according to different motion types and environmental conditions to optimize the fusion effect.

[0088] Step 1.4, data fusion;

[0089] The multi-source data is weighted and fused by the calculated weights to generate a unified motion posture data stream.

[0090] The fusion process adopts a hierarchical strategy:

[0091] Fusion at the sensor level, weighted combination of data from the same type of sensor;

[0092] Fusion at the feature level, integrating features extracted by different types of sensors;

[0093] Fusion at the decision level, integrating various types of information to form the final motion posture representation.

[0094] Therefore, the fused data contains the three-dimensional coordinates of the key points of the human skeleton, joint angles and motion parameters (such as speed and acceleration), providing comprehensive and accurate input data for subsequent posture analysis.

[0095] Step 2, based on the preprocessed data, analyze the differences between the learner's motion posture and the standard motion;

[0096] Specifically, the following sub-steps are included:

[0097] Step 2.1, three-layer cascaded motion posture feature extraction;

[0098] According to the embodiments of the present application, a three-layer cascading pose feature extraction algorithm is used to process the fusion data and extract multi-level features of the action. The specific implementation structure of the algorithm is as follows:

[0099] The key point layer is responsible for extracting the key point coordinates of the human skeleton, including the three-dimensional position information of the main joint points such as the head, torso, limbs, etc.

[0100] In the specific implementation, the present application adopts a multi-branch convolutional network structure to process visual data, and each branch is responsible for detecting a specific body area, and then locates the key points through the confidence heat map method. For the forward bending action, the system pays special attention to the key point coordinates of the spine, hip joint and knee joint to capture the bending angle and body balance state.

[0101] The relationship layer analyzes the relative position relationship between the key points, calculates the joint angle, the distance between the body parts and the relative motion trajectory.

[0102] In the present application, the relationship layer adopts a graph convolutional network (GCN) structure, in which the human joint points are nodes of the graph and the joint connections are edges of the graph. The GCN includes multiple graph convolutional layers, and the calculation of each layer is represented as:

[0103]

[0104] wherein H (l) represents the node feature matrix of the lth layer, which includes the feature representation of each joint point of the human body at the layer; H (l +1) represents the node feature matrix of the (l+1)th layer, i.e. the output feature after the graph convolution operation; represents the adjacency matrix with self-connection, which describes the connection relationship between the joint points of the human skeleton, and the self-connection ensures that each node considers its own features; represents the corresponding degree matrix, which is used to normalize the adjacency matrix; W (l) is a learnable weight matrix, which is used to transform the feature space and is optimized through training; σ act is a sigmoid activation function; is a normalized Laplacian matrix, which ensures that the feature aggregation process will not be unstable due to the difference in the number of node connections.

[0105] In the sports action analysis scene, this layer pays special attention to the coordination between the limbs and the transfer trajectory of the body center of gravity, for example, in the basketball shooting action, the system will analyze the coordination relationship between the arms and the torso and the forward and backward movement pattern of the center of gravity.

[0106] Semantic layer: Based on the features extracted from the previous two layers, this layer identifies the semantic information of the action, including action type, action stage, and action intention.

[0107] In this application, the semantic layer adopts a bidirectional long short-term memory network (Bi-LSTM) structure, which contains a forward and a backward LSTM, can consider past and future context information at the same time, and enhance the understanding of action sequences.

[0108] The formula of Bi-LSTM is represented as:

[0109]

[0110]

[0111] where, represents the hidden state of the forward LSTM at time step t, which is used to capture the sequence information from the past to the present; represents the hidden state of the backward LSTM at time step t, which is used to capture the sequence information from the future to the present; LSTM f represents the calculation function of the forward LSTM unit, which processes the forward timing information; LSTM b represents the calculation function of the backward LSTM unit, which processes the reverse timing information; x t represents the input feature at time step t, which contains the action features extracted by the key point layer and the relationship layer; represents the hidden state of the forward LSTM at time step t-1, which is used as the historical information for the current time step calculation; represents the hidden state of the backward LSTM at time step t+1, which is used as the future information for the current time step calculation; h t represents the final output hidden state, which is the concatenation of the forward and backward hidden states, and contains the complete context information.

[0112] In practical applications, such as football kicking action analysis, this layer can identify the preparation period, core force period, and follow-up period of the action, and understand the technical points of each stage.

[0113] The three levels of information are linked through a residual connection mechanism to ensure that the bottom layer features can directly affect the extraction process of the high-level features, and the high-level features can also guide the focus of the bottom layer features in the opposite direction, forming a closed-loop feedback structure.

[0114] The formula of residual connection is represented as:

[0115] F l = H l + F l-1 ;

[0116] where, lHl represents the fused feature of the l-th layer, i.e., the final output feature of the l-th layer after residual connection; H l F represents the feature directly output by the l-th layer of the neural network, containing new information extracted by the current layer; F l-1 Hl-1 represents the fused feature of the previous layer, i.e., the final output of the previous layer after residual connection; l represents the index of the current layer, used to identify different levels in the network. The residual connection passes information from the shallow layer to the deep layer without loss through a direct addition operation, effectively alleviating the gradient vanishing problem of the deep network, while retaining useful features extracted by each layer.

[0117] In actual application scenarios, for example, when a learner is learning the "cloud hand" action of Tai Chi, the key point layer first captures the position data of the hand, shoulder and waist; the relationship layer calculates the relative position and movement trajectory of the hand and shoulder, analyzes the smoothness and radius change of the circular arc; the semantic layer identifies the action intention in different stages, such as the coherent process of "lifting-rotating-sinking", and evaluates the fluency of force transmission. The system then integrates the three layers of information to generate a complete action understanding, and compares it with the standard action to find the problem of insufficient coordination between the waist and the hand in the "rotating" stage.

[0118] Step 2.2, the dynamic time warping algorithm calculates the difference;

[0119] The action sequence of the learner is compared with the standard action sequence to calculate the difference between the two. Since the speed and rhythm of different learners performing actions may be different, the present application uses the dynamic time warping (DTW) algorithm to time-align the two sequences to ensure the accuracy of the comparison.

[0120] The core of the DTW algorithm is to construct a distance matrix D dtw , where D dtw (i, j) represents the distance between the i-th time point in the learner's action sequence and the j-th time point in the standard action sequence. The distance calculation formula is:

[0121] D dtw (i,j)=d(x i ,y j )+min{D dtw (i-1,j),D dtw (i,j-1),D dtw (i-1,j-1)};

[0122] where D dtw (i, j) represents the cumulative distance between the i-th time point in the learner's action sequence and the j-th time point in the standard action sequence; d(x i , y j(x) is a distance metric that measures the difference in actions at two points in time; it can be expressed as Euclidean distance, cosine similarity, etc. i y j D represents the action feature vectors at the i-th and j-th time points in the learner's action sequence, respectively; dtw (i-1, j) represents the cumulative distance between the (i-1)th time point in the learner's sequence and the jth time point in the standard sequence, corresponding to vertical path selection; D dtw (i, j-1) represents the cumulative distance between the i-th time point in the learner's sequence and the (j-1)-th time point in the standard sequence, corresponding to the horizontal path selection; D dtw (i-1, j-1) represents the cumulative distance between the (i-1)th time point in the learner sequence and the (j-1)th time point in the standard sequence, corresponding to the diagonal path selection; min represents selecting the path with the smallest cumulative distance among the three possible paths to achieve optimal path search.

[0123] By backtracking the path of minimum distance, the best correspondence between the two sequences is found, and the overall difference score is calculated.

[0124] Step 2.3: Generate the deviation heatmap;

[0125] Based on the degree of difference calculated, this application generates a heatmap representing the distribution of motion deviations.

[0126] Heat maps are based on the human skeletal structure and use different shades of color to represent the degree of deviation of various joints and body parts, with red indicating larger deviations and green indicating smaller deviations.

[0127] Therefore, heat maps visually display the parts of a learner's movements that need to be corrected, making it easier for learners to understand the shortcomings of their own movements.

[0128] Step 2.4: Automatically generate correction guidelines;

[0129] Based on the deviation heatmap and the difference calculation results, this application automatically generates targeted corrective guidance. The corrective guidance includes the following aspects:

[0130] Deviation description: Clearly indicate the specific location and type of deviation in the learner's movements (such as excessive angle, positional deviation, timing error, etc.).

[0131] Correction suggestions: Specific correction methods are provided for each deviation, including correct posture requirements, key points of movement, and practice tips.

[0132] Prioritization: Based on the severity of the deviation and its impact on the overall quality of the movement, the corrective actions are prioritized to ensure that learners can start with the most critical issues for improvement.

[0133] Visualization demonstration: combined with the comparison of standard actions and learner actions, generate visual correction demonstration, intuitive display of the difference between correct and wrong.

[0134] Step 3, according to the action difference analysis result, predict the skill forgetting law of the learner;

[0135] Specifically, the following sub-steps are included:

[0136] Step 3.1, skill decomposition and structured representation;

[0137] First, the whole sports skill is decomposed into multiple atomic action units, and a skill dependency graph is constructed.

[0138] In application scenarios, for example, for basketball shooting skills, the system decomposes it into "standing posture preparation-holding posture-knee bending-pulling arm-stretching wrist-following action" and other sub-skill units.

[0139] This application uses a directed acyclic graph (DAG) to represent the dependency relationship between skills, where nodes represent sub-skills and edges represent dependency relationships. For complex skills such as volleyball spiking, basic actions such as approach, jump, and arm swing need to be mastered first, and then coordinated skills are learned, so the constructed skill graph will present a clear hierarchical structure and pre-dependence relationship.

[0140] Skill decomposition uses a combination of expert knowledge and data-driven methods:

[0141] Based on sports teaching experience, the skill is systematically decomposed;

[0142] Through techniques such as principal component analysis (PCA) and independent component analysis (ICA), key action components are automatically extracted from a large amount of standard action data.

[0143] Taking swimming breaststroke technology as an example, the system will not only decompose it into leg kicking, arm stroking, and breathing coordination according to the guidance of professional coaches, but also identify subtle but important technical details such as "elbow high" through analysis of the action data of excellent athletes.

[0144] Step 3.2, based on the decay model of Ebbinghaus forgetting curve;

[0145] Based on the Ebbinghaus forgetting curve, a basic forgetting model is established to represent the change law of skill retention rate over time.

[0146] The classic Ebbinghaus forgetting formula is:

[0147]

[0148] where R mem represents the retention rate, i.e., the percentage of skill level that a learner can maintain after a certain time interval t; S stab represents the relative memory stability, which is a parameter reflecting the degree of stability of skill in memory, and a larger value indicates a more stable memory and a slower forgetting rate; e represents the base of natural logarithm.

[0149] In this application, the basic formula is extended to consider the particularity of sports skill learning, and various factors are introduced to establish a more accurate forgetting model:

[0150]

[0151] where R mem (t) represents the retention rate, t represents the time interval, S stab represents the relative memory stability, f(P, D diff , E) is the stability adjustment function, P represents the practice frequency, D diff represents the skill difficulty coefficient, E represents the experience level of the learner, and e represents the base of natural logarithm.

[0152] The implementation of the adjustment function adopts the form of weighted product:

[0153]

[0154] where f(P, D diff , E) represents the stability adjustment function, P represents the practice frequency, i.e., the number of practice times per unit time, and a larger value indicates more frequent practice; D diff represents the skill difficulty coefficient, and a larger value indicates that the skill is more difficult to master; E represents the experience level of the learner, reflecting the prior knowledge and skill base in the relevant field, and a larger value indicates more experience; α stab is the weight coefficient of practice frequency, controlling the degree of influence of practice frequency on stability, and a larger value indicates that the influence of practice frequency is more significant; β stab is the weight coefficient of skill difficulty, and since a negative exponent is used, a larger value indicates that the negative influence of difficulty on stability is stronger; γ stab is the weight coefficient of experience level, controlling the degree of influence of prior experience on stability, and a larger value indicates that the positive influence of experience is more significant.

[0155] For different types of skills and different learning stages, the system will adaptively adjust these parameters to reflect the forgetting characteristics in different situations. For example, for high-difficulty skills such as track and field high jump, the system will set a higher β stabThe value reflects that high-difficulty actions are more likely to be forgotten without continuous practice; and for learners with relevant motor experience, the value of γ stab increases, indicating the positive effect of prior experience on skill stability.

[0156] Step 3.3, Bayesian knowledge tracing algorithm estimates the mastery probability;

[0157] Apply the Bayesian knowledge tracing (BKT) algorithm to estimate the mastery probability of each sub-skill for the learner.

[0158] The core parameters of the BKT model include:

[0159] p(L0): Initial mastery probability, indicating the probability that the learner has mastered the skill at first contact;

[0160] p(T learn ): Learning probability, indicating the probability of the learner transitioning from the unmastered state to the mastered state;

[0161] p(G): Guessing probability, indicating the probability of the learner performing correctly without mastering the skill;

[0162] p(S fail ): Error probability, indicating the probability of the learner performing incorrectly with mastered skill;

[0163] The Bayesian update formula is as follows:

[0164] p(L n+1 |Evidence n) = p(L n |Evidence n ) + (1-p(L n |Evidence n ))·p(T learn );

[0165] Where p(L n |Evidence n ) represents the posterior probability of skill mastery after observing the nth practice evidence; p(L n+1 |Evidence n ) represents the predicted probability of skill mastery in the n+1th practice after observing the nth practice evidence; (1-p(L n |Evidence n )) represents the probability that the learner does not currently master the skill; p(T learn ) represents the learning conversion rate, i.e., the probability of the learner transitioning from the unmastered state to the mastered state.

[0166]

[0167] Among them, p(L n Evidence n p(Evidence) represents the posterior probability that the learner has mastered the skill after observing evidence from the nth practice session; n |L n p(L) represents the conditional probability of observing evidence of the current practice, assuming the learner has already mastered the skill; n p(Evidence) represents the prior probability that the learner has mastered the skill; n () represents the marginal probability of observing the current practice evidence, which can be calculated using the law of total probability:

[0168]

[0169] Wherein, p(Evidence) n p(Evidence) represents the marginal probability of observing the current practice evidence; n |L n ), Let p(L) represent the conditional probabilities of observing current practice evidence, assuming the learner has already mastered the skill and assuming the learner has not mastered the skill, respectively; n ), These represent the prior probabilities that the learner has mastered the skill and those that have not, respectively. This indicates that the learner has not yet mastered the skill.

[0170] In sports teaching scenarios, the system continuously collects learners' practice results data (such as success rate, accuracy, and fluency) and updates the mastery probability estimate. For example, for the "jab attack" technique in fencing, the system records the learner's hit rate and posture accuracy at different distances and speeds, comprehensively assesses their mastery level, and continuously adjusts the model parameters as practice accumulates. This dynamic estimation allows the system to adapt to changes in the learner's learning state, improving the accuracy of predictions.

[0171] Compared to the traditional BKT model, this application makes the following improvements to the algorithm:

[0172] By introducing continuous scoring, it is no longer limited to binary "correct / incorrect" judgments, but uses a continuous scoring system of 0-1 to more precisely characterize the quality level of sports movements.

[0173] Considering the adjustment of task difficulty, the parameters are adaptively adjusted according to the difficulty of the practice, giving higher weight to success in high-difficulty situations, and more accurately reflecting the learner's ability.

[0174] By integrating the time factor and combining it with the forgetting curve model, when the time interval between two practice sessions is long, the mastery probability estimate should be appropriately reduced to reflect the natural decline of skills.

[0175] Step 3.4: Recurrent neural network predicts skill retention level;

[0176] A recurrent neural network based on Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) is constructed to predict learners' skill retention levels. This network takes learner characteristics, historical learning data, and temporal information as input and outputs the probability of skill retention at different future time points.

[0177] The input characteristics of the network include:

[0178] Learner personal characteristics: age, gender, motor ability score, learning style, etc.;

[0179] Skill characteristics: difficulty level, classification, and degree of skill mastery required;

[0180] Learning history: past practice records, learning curves, error patterns, etc.;

[0181] Time characteristics: the time interval since the last practice, total study time, etc.

[0182] The core structure of an LSTM network includes an input gate, a forget gate, and an output gate, enabling it to effectively handle long-term dependencies. Its main computational process is as follows:

[0183] i t =σ act (W xi x t +W hi h t-1 +W ci c t-1 +b i );

[0184] f t =σ act (W xf x t +W hf h t-1 +W cf c t-1 +b f );

[0185] c t =f t ⊙c t-1 +i t ⊙tanh(W xc xt +W hc h t-1 +b c );

[0186] o t =σ act (W xo x t +W ho h t-1 +W co c t +b o );

[0187] h t =o t ⊙tanh(c t );

[0188] Among them, i t f t and o t These are the activation values ​​of the input gate, forget gate, and output gate, respectively, c. t It is the unit state, h t It is in a hidden state, σ act It is the sigmoid activation function, ⊙ represents element-wise multiplication; x t The input vector at the current time step contains learner features, skill features, and temporal features; W xi W xf W xc W xo W represents the weight matrix connecting the input vector to the input gate, forget gate, cell state, and output gate, respectively; hi W hf W hc W ho W represents the weight matrix connecting the hidden state of the previous time step to the input gate, forget gate, cell state, and output gate, respectively; ci W cf W co These represent the weight matrices connecting the unit states to the input gate, forget gate, and output gate, respectively; b i b f b c b o These represent the bias terms for the input gate, forget gate, cell state, and output gate, respectively; h t-1 Indicates the hidden state at the previous time step; c t-1 This represents the cell state at the previous time step; tanh is the hyperbolic tangent activation function, used to adjust the range of cell states and output values.

[0189] The output layer of the network uses a sigmoid activation function to map the output to between 0 and 1, representing the skill retention probability:

[0190] p(t) = σ act (W p h t +b p );

[0191] where p(t) represents the skill retention probability at time point t, with a value range of 0-1; σ act represents the sigmoid activation function, mapping the output value to between 0 and 1; W p represents the weight matrix of the output layer, used to convert the hidden state to the output value; h t represents the hidden state of the LSTM network at time step t, containing the encoding of historical information; b p represents the bias term of the output layer, used to adjust the output value.

[0192] Step 4, based on the skill forgetting prediction results, generate personalized review strategy;

[0193] Specifically, the following sub-steps are included:

[0194] Step 4.1, state space definition;

[0195] Define the state space of reinforcement learning, including the skill mastery vector of the learner, historical review records and personal characteristics.

[0196] The state is represented as:

[0197] S state = [Mastery, History, PersonalFeatures];

[0198] where S state represents the state space of reinforcement learning, Mastery represents a vector of length n, representing the current mastery level of the learner on n sub-skills, with a value range of [0, 1], and the larger the value, the higher the mastery level; History represents the time, content and effect of the past k reviews, forming a time sequence feature, where time represents the specific date and duration of each review, content represents the specific sub-skills and practice forms of the review, and effect represents the change in skill mastery after the review; n represents the total number of sub-skills to be learned, which will be different for different sports; PersonalFeatures represents the personal characteristics of the learner, such as learning style, athletic ability baseline, etc.

[0199] In practical applications, such as learning the skill of passing a football, the Mastery vector may include scores of the mastery level of "inside of the foot touch", "passing force control", "direction accuracy", etc. The History records the review activities in the past two weeks, including the specific content of each review, the duration and the effect of improvement. The Personal Features records the characteristics of the learner, such as preferring practical learning and being good at lower limb control.

[0200] Step 4.2, action space definition;

[0201] The action space of reinforcement learning is defined, including review content selection, review difficulty setting and review form determination.

[0202] The action is represented as:

[0203] A action = [Content, Difficulty, Form];

[0204] Wherein, A action represents the action space, Content represents the selection of the combination of sub-skills that need to be reviewed, and the skills with lower mastery level or faster decay rate are preferred; Difficulty represents the difficulty level of setting the review content, ranging from 1 to 5, and the larger the value, the higher the difficulty; Form represents the determination of the form of review, which is a discrete variable, and the selected values include video demonstration (value 1), interactive game (value 2), real-time feedback practice (value 3), peer collaboration learning (value 4), competition mode (value 5) and the like. Different review forms are suitable for learners with different learning styles, and the system will select the most suitable form according to the characteristics and historical performance of the learner.

[0205] In the review scene of basketball dribbling skill, the system may generate the action of selecting "cross dribbling" and "change direction and decelerate" two sub-skills with faster decay, setting the difficulty level to 3 (medium difficulty), and adopting the "interactive game" form, requiring the learner to complete the dribbling route among virtual obstacles.

[0206] Step 4.3, reward function design;

[0207] A multi-step reward function is designed to balance learning effectiveness, time efficiency and learner engagement.

[0208] The reward function is defined as:

[0209] R reward = α reward · ΔSkill + β reward · Time -1 + γ reward · Engagement;

[0210] wherein R reward represents the reward function, Time -1 represents the inverse of time efficiency, the larger the value, the higher the time efficiency; Engagement represents the engagement score of the learner, with a value range of 0-1, which is the weighted average value based on attention duration, interaction frequency, effort level and emotional state, the larger the value, the higher the engagement; α reward , β reward , γ reward represent the weight coefficients of skill improvement, time efficiency and engagement respectively, and satisfy α reward + β reward + γ reward = 1; ΔSkill represents the improvement of skill mastery after review, and the calculation formula is:

[0211]

[0212] wherein w i is the importance weight of the i-th sub-skill, represents the mastery of the i-th sub-skill after review, represents the mastery of the i-th sub-skill before review, and n represents the total number of sub-skills.

[0213] In practical applications, such as the review of tennis serving skills, different weights are assigned to force accuracy, ball point height, and racket angle in the calculation of ΔSkill; Time -1 encourages the system to design efficient review content and avoid redundancy; Engagement assesses engagement by capturing eye movement data, facial recognition, and interaction frequency, ensuring that the review process is both effective and interesting.

[0214] Step 4.4, Deep Q Network Training;

[0215] A Deep Q Network (DQN) model is constructed to learn the optimal review strategy through training with the interaction data of the learner. The DQN network consists of the following components:

[0216] Input layer: receives state vector S state ;

[0217] Feature extraction layer: composed of multiple fully connected layers, extracts key features of the state;

[0218] Q value prediction layer: outputs the Q value of all possible actions, i.e. the long-term expected return of taking action A state in state S action ;

[0219] The core formula for updating Q value is:

[0220] Q(S state,t ,A action,t )←Q(S state,t ,A action,t )+α lean [R t +γ discount max a Q(S state,t+1 ,a)-Q(S state,t ,A action,t )];

[0221] where Q(S state,t ,A action,t ) denotes the Q-value of taking action A state,t in state S action，t , i.e., the expected long-term return; α learn denotes the learning rate, controlling the step size of Q-value update, usually ranging from (0, 1], the larger the value, the more aggressive the update; R t denotes the immediate reward, representing the reward value obtained after executing action A state,t in state S action,t ; γ discount is the discount factor, used to balance the importance of immediate rewards and future rewards, taking a value range of [0, 1], the larger the value, the more emphasis on long-term returns; max a Q(S state，t+1 ,a) denotes the maximum Q-value of all possible actions in the next state S state，t+1 , representing the estimate of the future optimal return; [R t + γ discount max a Q(S state，t+1 ,a) - Q(S state，t ,A action，t ) denotes the time difference error, representing the gap between the actual reward obtained and the expected reward.

[0222] To improve training stability, the following improvements are adopted in the present application:

[0223] Experience Replay: Store past <state, action, reward, next state> quadruples, randomly sample batches for training, break the temporal correlation between samples.

[0224] Target Network: Maintain two networks with the same structure, where the target network has a lower parameter update frequency, used to calculate the target Q-value, reduce the shock in the training process.

[0225] Double Q-learning: separate action selection and evaluation, avoid overestimation of Q value. Update formula is modified as:

[0226] Q(S state,t ,A action,t )←Q(S state,t ,A action,t )+α learn [R t +γ discount Q′(S state,t+1 ,argmax a Q(S state,t+1 ,a))-Q(S state,t ,A action,t )];

[0227] Where Q(S state，t ,A action，t ) represents the Q value of taking action A action，t in state S state，t , i.e. the expected long-term return predicted by the current network; α learn represents the learning rate, controlling the step size of Q value update, determining the influence of new information on existing Q value; R t represents the immediate reward, representing the reward value obtained after executing action A action，t in state S state，t ; γ disount represents the discount factor, used to balance the importance of immediate reward and future reward; Q′ represents the Q function of the target network, used to calculate the target Q value, which has a lower parameter update frequency than the main network, improving the stability of training; S state，t+1 represents the new state transferred after executing action A action，t ; argmax a Q(S state，t+1 ,a) represents the action with the maximum Q value in the next state S state，t+1 selected by the main network; Q′(S state，t+1 ,argmax a Q(S state，t+1 ,a)) represents the evaluation of the action selected by the main network by the target network, avoiding overestimation of Q value; [R t +γ discount Q′(S state,t+1 ,argmax a Q(S state,t+1 ,a))-Q(S state,t ,A action,t )] represents the temporal difference (TD) error, representing the difference between the actual reward obtained and the expected reward.

[0228] Prioritized Experience Replay: Assign different priorities to samples based on TD error, increase the sampling probability of important samples, and accelerate the learning process.

[0229] In actual training process, the system will continuously collect review data of learners, including skill assessment before and after review, review time and participation index, etc., to update the reinforcement learning model. Through the accumulation of a large amount of learner data, the system gradually optimizes the review strategy generation model, and can tailor the most effective review scheme for different types of learners.

[0230] Step 5, based on the personalized review strategy, execute the adaptive learning forgetting review closed-loop process;

[0231] Specifically, the following sub-steps are included:

[0232] Step 5.1, personalized learning plan generation;

[0233] Based on the characteristics and skill mastery of learners, automatically generate a personalized learning plan. The learning plan includes the following elements:

[0234] Learning objectives: Clearly define the skills to be mastered and the expected proficiency level, quantified as measurable indicators.

[0235] Learning path: Determine the optimal learning order based on the skill dependency graph to avoid learning obstacles.

[0236] Time arrangement: According to the available time of learners and learning efficiency, reasonably allocate the learning time of each skill.

[0237] Adaptive adjustment mechanism: Set trigger conditions to automatically adjust the plan when the learning progress deviates significantly from the expected progress.

[0238] Step 5.2, real-time learning monitoring and evaluation;

[0239] Continuously monitor the performance of learners during the learning process and evaluate the learning effect. The specific evaluation indicators include:

[0240] Action accuracy: The degree of coincidence between the learner's action and the standard action.

[0241] Learning speed: The time and number of exercises required to master new skills.

[0242] Error mode: The type and frequency of common errors, reflecting the difficulties of learning.

[0243] Attention state: Assess the degree of learning concentration through eye tracking, facial expression analysis, etc.

[0244] These information are fed back to learners and teachers in real-time and stored in the database, providing the basis for subsequent personalized guidance. In the context of yoga posture learning, the system will capture the learner's key joint angles and body balance data in real-time, analyze the differences with the standard posture, identify common problems such as "shoulder sinking" or "forward leaning of the center of gravity", and provide timely visual and audio prompts when the posture deviation exceeds the threshold, helping learners adjust immediately.

[0245] Step 5.3, Skill Forgetting Curve Prediction and Update;

[0246] Based on real-time observation data, continuously predict and update the skill forgetting curve. The update mechanism includes:

[0247] Periodic testing: Schedule skill tests after different time intervals to collect actual retention data.

[0248] Model calibration: Adjust the forgetting curve parameters based on test data to improve prediction accuracy.

[0249] Dynamic update: Adjust the forgetting model parameters accordingly as the learner's skill level improves.

[0250] Critical point identification: Determine the time point when the skill retention rate drops to a certain threshold as the trigger condition for review.

[0251] Step 5.4, Optimal Review Timing and Content Arrangement;

[0252] Based on the skill forgetting curve and the generated review strategy, arrange the optimal review timing and content. Specifically including:

[0253] Review timing determination: Review before the skill retention rate drops to the critical value but has not yet seriously declined.

[0254] Review content combination: Combine multiple sub-skills that need review to optimize learning efficiency.

[0255] Review effect evaluation: Record the effect of each review and update the corresponding skill mastery.

[0256] Dynamic adjustment of review plan: Adjust the subsequent review plan according to the review effect and learner feedback.

[0257] In fencing skill training, the system will monitor the forgetting curve of multiple technical actions (such as thrust, block, and dodge), and arrange review when the predicted retention rate of a skill drops to 75%. According to the learner's time availability and energy state, the system may combine multiple skills close to the critical point (such as thrust and block) into the same review session, and design a confrontation exercise that includes the alternating use of these two skills, not only improving review efficiency, but also enhancing the coordination between skills. After review, the system will immediately evaluate the improvement effect and adjust the forgetting curve parameters of these skills accordingly.

[0258] Step 5.5, closed-loop optimization and long-term maintenance;

[0259] Establish a complete learning-forgotten-review closed-loop system to achieve long-term maintenance of skills. The closed-loop optimization mechanism includes:

[0260] Model periodic retraining: Use accumulated learner data to periodically retrain models to improve system performance.

[0261] User feedback integration: Collect learners' feedback on system guidance to evaluate system performance.

[0262] Long-term learning path optimization: Analyze a large number of learners' learning path data to find the optimal learning mode.

[0263] Individual adaptability improvement: With increasing interaction with a specific learner, the system's adaptability to that learner improves continuously.

[0264] In the application of group dance teaching, the system will track the learner's entire process from beginner to master (which may last for several months), record the learning efficiency and forgetting characteristics of each link, and gradually adjust the model to better adapt to the characteristics of the learner. At the same time, the system will analyze all learners' data, identify common difficulties and optimal teaching order in dance movements, and continuously optimize teaching content and methods. Learners can also directly influence the system's decision-making process through rating and feedback functions, such as pointing out that certain review arrangements are inconvenient or the content is not interesting enough, and the system will adjust its strategy accordingly to achieve human-machine collaborative teaching optimization.

[0265] Specific application examples:

[0266] The application process of the present application is illustrated by specific examples as follows:

[0267] Example 1: Basketball shooting teaching;

[0268] A sports teaching application uses the method and system of the present application to teach basketball shooting skills. The specific process is as follows:

[0269] Data collection and preprocessing:

[0270] The system collects video stream data of the learner's shooting process through the camera of the smart phone;

[0271] The smart watch worn by the learner provides inertial sensor data;

[0272] The sensors in the specially designed basketball provide pressure distribution data;

[0273] The system pre-processes and fuses these three heterogeneous data to obtain accurate motion data;

[0274] Action posture difference analysis:

[0275] The system uses a three-level link posture feature extraction algorithm to analyze the learner's shooting action;

[0276] The comparative analysis results show that the learner has problems in the following aspects:

[0277] Insufficient wrist bending angle (15° difference from standard action);

[0278] Insufficient elbow extension (12° difference from standard action);

[0279] Knee bending degree is too low (18° difference from standard action);

[0280] Early shooting opportunity (timing difference 0.3 seconds);

[0281] The system generates an action difference heat map to visually display the problem area;

[0282] Automatically generate improvement suggestions: "Increase wrist bending amplitude, fully extend elbow, increase knee bending depth, delay shooting opportunity";

[0283] Skill forgetting prediction:

[0284] The system decomposes the shooting skill into the following atomic action units:

[0285] Preparation posture (lower center of gravity);

[0286] Lifting action (arm lifting);

[0287] Wrist bending (positioning ball);

[0288] Stretch release (leg straightening + arm stretching + wrist shooting);

[0289] Follow-up action (maintain posture);

[0290] Based on the learner's exercise frequency (2 times a week), skill difficulty (medium), and experience level (beginner), the system predicts the forgetting curve:

[0291] Wrist bending: retention rate drops to 65% after 7 days;

[0292] Stretch release: retention rate drops to 50% after 5 days;

[0293] Other action units: retention rate drops to 70% after 10 days;

[0294] Review strategy generation:

[0295] The system defines the state space, including the learner's current mastery probability of each sub-skill, historical review records, and personal characteristics;

[0296] Define action space, including review content (focus on wrist flexion and extension release), review difficulty and review form;

[0297] The optimal review strategy generated by the deep Q network is:

[0298] Day 3: Special training for wrist flexion, moderate difficulty, video guidance + practice;

[0299] Day 4: Special training for extension release, slightly lower difficulty, game-based practice;

[0300] Day 7: Complete shooting action training, increasing difficulty, real-time feedback practice;

[0301] Closed loop execution:

[0302] The system generates a personalized learning plan, including initial learning, periodic assessment and regular review;

[0303] In the review on the 3rd day, the system monitors the improvement of the learner's wrist flexion angle in real time;

[0304] Find that the learner has made significant progress (the angle difference has decreased from 15° to 5°), and the system updates the forgetting curve of wrist flexion skills, predicting that the retention rate can be increased to 75% after 10 days;

[0305] Adjust the subsequent review plan accordingly, reduce the frequency of special training for wrist flexion, and increase the special training for extension release.

[0306] Example two: swimming posture correction;

[0307] A swimming training center uses the method and system of the invention to teach freestyle swimming technology. The specific process is as follows:

[0308] Data acquisition and preprocessing:

[0309] The system collects multi-angle video data of the learner's swimming process through an underwater camera array;

[0310] The waterproof inertial sensor worn by the learner provides body posture data;

[0311] The system calculates the adaptive weight of each data source. Since underwater video data is easily affected by water waves, the system automatically gives higher weight to inertial sensor data;

[0312] Action posture difference analysis:

[0313] The system analysis found that the learner has problems in the following aspects:

[0314] The arm entry angle is too steep (25° difference from the standard action);

[0315] Inadequate body roll (30° difference from standard motion);

[0316] Leg whip motion is uncoordinated (0.5 seconds timing difference);

[0317] System generates targeted improvement suggestions: "Lower arm entry angle, increase body roll amplitude, adjust leg whip rhythm";

[0318] Skill forgetting prediction:

[0319] System predicts the forgetting pattern of the learner's freestyle skill:

[0320] Arm stroke: retention rate drops to 40% after 14 days;

[0321] Body roll: retention rate drops to 35% after 10 days;

[0322] Leg whip: retention rate drops to 30% after 7 days;

[0323] Review strategy generation:

[0324] The optimal review strategy generated by the system is:

[0325] Day 5: Conduct leg whip specialized training, with rhythm machine providing timing guidance;

[0326] Day 8: Conduct body roll training, using AR augmented reality technology to provide real-time posture guidance;

[0327] Day 10: Conduct arm stroke training, combined with underwater resistance feedback equipment;

[0328] Day 12: Conduct complete freestyle motion training, with the system providing real-time comprehensive scoring;

[0329] Closed loop execution:

[0330] The system assesses learning effectiveness after each review training, updates skill mastery level and forgetting prediction model;

[0331] Find that the learner's improvement on leg whip motion is not obvious, the system increases the training frequency and feedback detail of this part;

[0332] At the same time, it is found that the arm stroke technology has made significant progress, and the corresponding specialized training of this part is reduced;

[0333] It can be seen from the above application examples that the virtual sports teaching experience provided by the application is highly personalized and adaptive according to individual characteristics, skill mastery and forgetting rules of learners. The system continuously optimizes the learning plan and review strategy by continuously monitoring the progress of the learners, thereby effectively improving the learning efficiency and long-term retention rate of sports skills.

[0334] The above describes the embodiments of the application, but the embodiments are not limited to the specific implementation described above, which is only illustrative but not restrictive. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A data processing method for virtual sports teaching, characterized by, The method comprises the following steps: Collecting and preprocessing learner motion data; Analyzing the differences between the learner's action posture and the standard action based on the preprocessed data; Predicting the skill forgetting pattern of the learner according to the action difference analysis result; Generating personalized review strategies based on the skill forgetting prediction result; Executing an adaptive learning forgetting review closed-loop process based on the personalized review strategies.

2. The data processing method for virtual sports teaching according to claim 1, wherein, The step of collecting and preprocessing learner motion data comprises: Collecting multi-source heterogeneous data including video stream data, inertial sensor data and pressure sensor data; Standardizing the collected data; Calculating adaptive weights of each data source; Fusing multi-source data based on the adaptive weights.

3. The data processing method for virtual sports teaching according to claim 1, wherein, The step of analyzing the differences between the learner's action posture and the standard action comprises: Using a three-level linked posture feature extraction algorithm to analyze the action; Comparing the learner's action sequence with the standard action sequence through a dynamic time warping algorithm; Generating an action difference heat map; Automatically generating improvement suggestions based on the difference analysis result.

4. The data processing method for virtual sports teaching according to claim 3, wherein, The three-level linked posture feature extraction algorithm comprises: Key point layer: extracting three-dimensional coordinate points of human joints; Relationship layer: analyzing the relative position and motion relationship between key points through a graph convolution network; Semantic layer: identifying semantic information of the action through a bidirectional long short-term memory network.

5. The data processing method for virtual sports teaching according to claim 1, wherein, The step of predicting the skill forgetting pattern of the learner comprises: Decomposing the entire sports skill into multiple atomic action units and constructing a skill dependency graph; Establishing a basic forgetting model based on an extended Ebbinghaus forgetting curve; Applying a Bayesian knowledge tracking algorithm to estimate the learner's mastery probability of each sub-skill; Constructing a recurrent neural network based on a long short-term memory network or a gated recurrent unit to predict the learner's skill retention level.

6. The data processing method for virtual sports teaching according to claim 5, wherein, The extended Ebbinghaus forgetting curve establishes a basic forgetting model. where R mem (t) represents the memory retention rate, t represents the time interval, S stab represents the relative memory stability, f(P, D diff ,E) is the stability adjustment function, P represents the practice frequency, D diff represents the skill difficulty, E represents the experience level of the learner, and e represents the base of the natural logarithm.

7. The data processing method for virtual sports teaching according to claim 1, wherein, The step of generating personalized review strategies comprises: Defining the state space of reinforcement learning, including the learner's skill mastery vector, historical review records and personal characteristics; Defining the action space of reinforcement learning, including review content selection, review difficulty setting and review form determination; Designing a multi-step reward function to balance learning effectiveness, time efficiency and learner engagement; Training the optimal review strategy through a deep Q network.

8. The data processing method for virtual sports teaching according to claim 7, wherein, The multi-step reward function is: R reward = a reward · ΔSkill + β reward · Time -1 γ reward · Engagement; where R reward represents a multi-step reward function, ΔSkill represents the improvement of skill mastery after review, Time -1 represents the inverse of time efficiency, Engagement represents the engagement score of the learner, and α reward , β reward , and γ reward represent the weight coefficients of the importance of controlling the improvement of skill mastery after review, the inverse of time efficiency, and the engagement score of the learner, respectively.

9. The data processing method for virtual sports teaching according to claim 1, wherein, The step of executing an adaptive learning forgetting review closed-loop process comprises: Generating a personalized learning plan; Monitoring and evaluating the learning process in real time; Predicting and updating the skill forgetting curve; Arranging the optimal review time and content; Establishing a long-term maintenance and closed-loop optimization mechanism.

10. A data processing system for virtual sports teaching, configured to perform a data processing method for virtual sports teaching according to any one of claims 1-9. It comprises: A data collection and preprocessing module for collecting and preprocessing learner motion data; An action analysis module for analyzing the differences between the learner's action posture and the standard action; A forgetting prediction module for predicting the skill forgetting pattern of the learner; A review strategy generation module for generating personalized review strategies; A closed-loop execution module for executing an adaptive learning forgetting review closed-loop process.

Images