Personalized training plan generation method and system in element universe smart sports

Abstract

The invention provides a personalized training plan generation method and system in meta-universe smart sports, and relates to the technical field of smart sports, and the method comprises the steps: obtaining user meta-universe scene sports data, calculating an actual sports load based on a human body biomechanical model, carrying out the time sequence analysis through a neural network, and predicting a physiological limit threshold value; and constructing a training knowledge graph, generating a stage training plan adaptive to the user through a graph neural network, and finally presenting the stage training plan in a meta-universe virtual scene form. According to the method, personalized, safe and scientific training guidance can be provided, and the training effect and the user experience are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent sports, and particularly to a method and system for generating personalized training plans in metaverse intelligent sports. Background Art

[0002] With the booming development of metaverse technology, combining intelligent sports training with metaverse virtual scenarios has become an important research direction. Traditional sports training methods mainly rely on coaches' experience for guidance, making it difficult to achieve precise personalized training plans. The intelligent sports training system in the metaverse environment can collect users' real-time sports data through sensing devices, analyze users' sports performance in combination with artificial intelligence algorithms, and provide users with immersive virtual training experiences. Currently, related research mainly focuses on aspects such as sports data collection, action recognition and analysis, and virtual scene construction.

[0003] However, the existing metaverse intelligent sports training systems still have deficiencies. They lack an accurate evaluation mechanism for users' exercise loads, cannot accurately grasp users' physiological limits, and are prone to overtraining or undertraining; the formulation of training plans is too simplistic, failing to fully consider users' historical training data and personalized needs, resulting in unsatisfactory training effects; the knowledge representation method is single, making it difficult to effectively explore the correlation relationships between training items and unable to achieve the scientific progression of training plans; and they lack the analysis of users' long-term training patterns, making it difficult to accurately predict users' training effects. Summary of the Invention

[0004] Embodiments of the present invention provide a method and system for generating personalized training plans in metaverse intelligent sports, which can solve the problems in the prior art.

[0005] In the first aspect of the embodiments of the present invention, a method for generating a personalized training plan in metaverse intelligent sports is provided, including: Obtaining the sports data of a user in a metaverse virtual scenario; Based on the bone joint point position data in the sports data, calling a preset human biomechanics model to calculate the corresponding exercise load parameters of the user in the real environment; Performing time series analysis on the exercise load parameters through a neural network model, using an attention mechanism to extract the long-term and short-term features of the user's exercise pattern, and predicting the physiological limit thresholds of the user under different exercise intensities based on the extracted long-term and short-term features; Constructing a training knowledge triple based on the historical training data of the user, and constructing a personalized training knowledge graph based on the training knowledge triple through the method of multi-level knowledge extraction and dynamic graph optimization; Build a graph neural network model, input the personalized training knowledge graph into the graph neural network model, and generate a phased training plan for the user based on the node feature vectors output by the graph neural network model. Among them, the training items in the phased training plan are sorted according to a progressive relationship, and the exercise intensity of each training item is within the range of the physiological limit threshold; Convert the phased training plan into training instructions for the metaverse virtual scene, and display a training guidance interface in the metaverse virtual scene.

[0006] In an alternative embodiment, Based on the bone joint point position data in the motion data, call a preset human biomechanical model to calculate the corresponding motion load parameters of the user in the real environment, including: According to the bone joint point position data in the motion data, use a kinematic inversion model to calculate the user's joint angle data; based on the joint angle data at adjacent times, calculate the user's joint motion speed data and joint motion acceleration data; Establish a human bone connection model, map the bone joint point position data to the corresponding bone nodes in the human bone connection model, and calculate the pressure values and torque values of each joint according to the connection relationship between the bone nodes in the human bone connection model; Input the joint angle data and joint force data into the static force analysis channel of a preset human biomechanical model for processing to obtain static force characteristics, which characterize the muscle force distribution required for the user to maintain the current posture; Input the joint motion speed data and the joint motion acceleration data into the dynamic force analysis channel of the human biomechanical model for processing to obtain dynamic force characteristics, which characterize the joint impact force and inertial force during the user's movement; Perform fusion processing on the static force characteristics and the dynamic force characteristics through the feature fusion layer of the human biomechanical model to obtain the output result of the feature fusion layer; Calculate the corresponding motion load parameters of the user in the real environment according to the output result of the feature fusion layer.

[0007] In an alternative embodiment, perform time series analysis on the motion load parameters through a neural network model, use an attention mechanism to extract the long-term and short-term characteristics of the user's motion pattern, and based on the extracted long-term and short-term characteristics, predict the physiological limit threshold of the user under different exercise intensities, including: Construct an adaptive time window according to the user's historical motion load parameters, dynamically adjust the window length based on the motion volatility, and generate variable-length time series data segments; Input the variable-length time-series data segments into a recurrent convolutional neural network, and extract local motion feature sequences through a multi-layer recurrent convolutional structure; Input the local motion feature sequences into a graph attention network, construct the local motion feature sequences into a time-series feature map, and extract the temporal dependencies between feature nodes through graph convolutional operations to obtain global correlation features; Input the local motion feature sequences and the global correlation features into a temporal attention module. The temporal attention module constructs a multi-scale feature pyramid and calculates adaptive weight coefficients for the local motion feature sequences at different time scales; Perform temporal cross-attention calculation on the adaptive weight coefficients and the global correlation features to generate enhanced features at multiple time scales, and fuse the enhanced features at multiple time scales through a cross-scale feature aggregation network to obtain long-term and short-term features of the user's motion pattern; Use a Bayesian neural network to model the long-term and short-term features, calculate the probability distribution function of the physiological limit threshold, and use the mean of the probability distribution function as the prediction result of the physiological limit threshold of the user under different exercise intensities.

[0008] In an alternative embodiment, using a Bayesian neural network to model the long-term and short-term features, calculating the probability distribution function of the physiological limit threshold, and using the mean of the probability distribution function as the prediction result of the physiological limit threshold of the user under different exercise intensities, includes: Input the long-term and short-term features into a feature encoding network to generate a sequence of feature vectors; Construct a Bayesian neural network, set latent variables in each layer of the Bayesian neural network, and input the sequence of feature vectors into the Bayesian neural network; Construct training sample pairs based on the sequence of feature vectors, use a variational autoencoder to model the latent variables in the Bayesian neural network to obtain the prior distribution of the latent variables; Input the prior distribution of the latent variables into a probability neural network to generate a weight distribution, which is used to characterize the uncertainty of each layer of network parameters in the Bayesian neural network; Optimize the weight distribution through the stochastic variational inference method, generate multiple sets of network parameters according to the weight distribution, construct a deep ensemble model based on the multiple sets of network parameters, input new long-term and short-term features into the deep ensemble model, and output multiple prediction results based on the deep ensemble model; Perform probability density estimation on the multiple prediction results to obtain the probability distribution function of the physiological limit threshold, calculate the mean of the probability distribution function, and use the mean as the prediction result of the physiological limit threshold of the user under different exercise intensities.

[0009] In an alternative embodiment, training knowledge triples are constructed based on the user's historical training data, and a personalized training knowledge graph is constructed by means of multi-level knowledge extraction and dynamic graph optimization, including: Perform temporal segmentation on the user's historical training data to obtain training data sequences for multiple training cycles; Extract training entities and training relationships from the training data sequences, construct training knowledge triples, input the training knowledge triples into a knowledge reasoning model, and semantically expand the training knowledge triples based on the knowledge reasoning model; Perform multi-granularity partitioning on the training knowledge triples, construct knowledge subgraphs at each granularity level, and each knowledge subgraph contains training entity nodes and training relationship edges at that granularity level; Perform representation learning on the knowledge subgraphs, iteratively update the feature representations of the training entity nodes through a message passing mechanism, calculate the similarity matrix between nodes based on the updated feature representations, and construct a node importance scoring model in combination with the weight distribution of the training relationship edges; Calculate the importance scores of the training entity nodes according to the node importance scoring model, set an adaptive pruning threshold based on the importance scores, perform dynamic pruning on the knowledge subgraphs, aggregate the features of the pruned multiple knowledge subgraphs, and construct a hierarchical training knowledge graph; Map new training data to the hierarchical training knowledge graph and update the training entity nodes and training relationship edges in the hierarchical training knowledge graph; Generate a personalized training knowledge graph based on the hierarchical training knowledge graph.

[0010] In an alternative embodiment, a graph neural network model is constructed, and the personalized training knowledge graph is input into the graph neural network model, and a phased training plan for the user is generated based on the node feature vectors output by the graph neural network model, including: Construct a graph neural network model, perform spatial dimension encoding on the topological structure features of the training entity nodes, and use a time-aware graph convolutional network to model the dynamic evolution features of the training entity nodes in the time dimension to obtain spatial dimension features and time dimension features; Input the personalized training knowledge graph into the graph neural network model, and adaptively fuse the spatial dimension features and the time dimension features through a multi-layer cross-attention mechanism to generate multi-dimensional feature vectors of the training entity nodes; Use self-attention-guided graph pooling operations to sparsely reduce the dimensionality of the multi-dimensional feature vectors to obtain compressed feature representations of the training entity nodes; Construct a hierarchical reinforcement learning model for multi-objective joint optimization based on the compressed feature representation, model the progressive relationship constraint and the exercise intensity constraint as hierarchical optimization objectives, and use a double-layer policy network to separately learn the global sorting policy and the local intensity parameters of training items; Input the global sorting policy and the local intensity parameters into the training plan generation module, determine the progressive order of training items according to the global sorting policy, adjust the exercise intensity of training items according to the local intensity parameters, and generate a phased training plan that meets the requirements of the progressive relationship and exercise intensity.

[0011] In an alternative embodiment, constructing a hierarchical reinforcement learning model for multi-objective joint optimization based on the compressed feature representation, modeling the progressive relationship constraint and the exercise intensity constraint as hierarchical optimization objectives, and using a double-layer policy network to separately learn the global sorting policy and the local intensity parameters of training items includes: Model the progressive relationship constraint of training items as a high-level optimization objective, model the exercise intensity constraint of training items as a low-level optimization objective, and use a fuzzy evaluation matrix to calculate the high-level reward value and the low-level reward value of the state-action pair; Construct a double-layer policy network. The high-level policy network of the double-layer policy network uses a multi-scale sliding time window to segment and encode the training item sequence, calculates the progressive correlation score between items within each time window, and dynamically adjusts the progressive correlation threshold in combination with the high-level reward value to generate a global sorting reference policy for training items; The low-level policy network of the double-layer policy network constructs a multi-dimensional physiological load model based on the heart rate variability index and exercise metabolism level of training items, and dynamically adjusts the intensity parameters through a soft constraint optimization method introducing a buffer interval in combination with the low-level reward value to generate a local intensity adjustment policy for training items; Use a recurrent neural network to establish a bidirectional information channel between the high-level policy network and the low-level policy network, input the global sorting reference policy and the local intensity adjustment policy into the bidirectional information channel, and iteratively optimize the global sorting reference policy and the local intensity adjustment policy based on a cross-layer constraint function to obtain the global sorting policy and the local intensity parameters of training items.

[0012] In the second aspect of the embodiments of the present invention, a personalized training plan generation system in a metaverse intelligent exercise is provided, including: A first unit for obtaining the exercise data of a user in a metaverse virtual scene; A second unit for calling a preset human biomechanics model to calculate the corresponding exercise load parameters of the user in the real environment based on the bone joint point position data in the exercise data; A third unit, configured to perform time series analysis on the motion load parameters through a neural network model, extract long-term and short-term features of the user's motion pattern by using an attention mechanism, and predict the physiological limit threshold of the user under different exercise intensities based on the extracted long-term and short-term features; A fourth unit, configured to construct training knowledge triples based on the historical training data of the user, and construct a personalized training knowledge graph by means of multi-level knowledge extraction and dynamic graph optimization based on the training knowledge triples; A fifth unit, configured to construct a graph neural network model, input the personalized training knowledge graph into the graph neural network model, and generate a phased training plan for the user based on the node feature vectors output by the graph neural network model, wherein the training items in the phased training plan are sorted according to a progressive relationship, and the exercise intensity of each training item is within the range of the physiological limit threshold; A sixth unit, configured to convert the phased training plan into training instructions for a metaverse virtual scene, and display a training guidance interface in the metaverse virtual scene.

[0013] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0014] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0015] In the embodiments of the present invention, by obtaining the motion data of the user in the metaverse virtual scene and combining the human biomechanics model and the neural network model, the true motion load of the user can be accurately calculated and the physiological limit threshold can be predicted, thereby ensuring the scientificity and safety of the training plan and effectively avoiding the risks of overtraining or undertraining; constructing a personalized training knowledge graph and applying a graph neural network to generate a training plan fully considers the historical data of the user and the progressive relationship between training items, realizes the personalized customization and dynamic optimization of the training plan, and improves the training effect and user experience; seamlessly converting the generated training plan into training instructions for the metaverse virtual scene and presenting them to the user through a visual training guidance interface breaks the limitations of the real environment and provides the user with an immersive and interactive intelligent training experience, promoting the effective integration of virtual and reality. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1Schematic flowchart of the personalized training plan generation method in the metaverse intelligent sports according to the embodiments of the present invention; Figure 2 Comparison chart of the parameter quantity and calculation efficiency of each uncertainty modeling method; Figure 3 Visualization diagram of the training project progression relationship network and multi-scale sliding time window; Specific implementation manners

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0019] Figure 1 Schematic flowchart of the personalized training plan generation method in the metaverse intelligent sports according to the embodiments of the present invention, as Figure 1 shown, the method includes: Obtain the motion data of the user in the metaverse virtual scene; Based on the bone joint point position data in the motion data, call a preset human biomechanics model to calculate the corresponding motion load parameters of the user in the real environment; Perform time series analysis on the motion load parameters through a neural network model, adopt an attention mechanism to extract the long-term and short-term features of the user's motion pattern, and predict the physiological limit threshold of the user under different motion intensities based on the extracted long-term and short-term features; Construct training knowledge triples according to the user's historical training data, and construct a personalized training knowledge graph through multi-level knowledge extraction and dynamic graph optimization based on the training knowledge triples; Construct a graph neural network model, input the personalized training knowledge graph into the graph neural network model, and generate a phased training plan for the user based on the node feature vectors output by the graph neural network model, where the training items in the phased training plan are sorted according to the progression relationship, and the motion intensity of each training item is within the range of the physiological limit threshold; Convert the phased training plan into a training instruction in the metaverse virtual scene, and display a training guidance interface in the metaverse virtual scene.

[0020] In an alternative embodiment, based on the skeletal joint position data in the motion data, a preset human biomechanical model is called to calculate the corresponding motion load parameters of the user in the real environment, including: According to the skeletal joint position data in the motion data, a kinematic inversion model is used to calculate the user's joint angle data; based on the joint angle data at adjacent times, the user's joint motion speed data and joint motion acceleration data are calculated; A human skeletal link model is established, and the skeletal joint position data is mapped to the corresponding skeletal nodes in the human skeletal link model. According to the connection relationship between the skeletal nodes in the human skeletal link model, the pressure value and torque value of each joint are calculated; The joint angle data and joint force data are input into the static force analysis channel of a preset human biomechanical model for processing to obtain static force characteristics, which characterize the muscle force distribution required for the user to maintain the current posture; The joint motion speed data and the joint motion acceleration data are input into the dynamic force analysis channel of the human biomechanical model for processing to obtain dynamic force characteristics, which characterize the joint impact force and inertial force during the user's movement; The static force characteristics and the dynamic force characteristics are fused through the feature fusion layer of the human biomechanical model to obtain the output result of the feature fusion layer; According to the output result of the feature fusion layer, the corresponding motion load parameters of the user in the real environment are calculated.

[0021] In a specific embodiment, the skeletal joint position data is obtained by devices such as an optical motion capture system, an inertial sensor, or a depth camera. This data includes at least 15 key points: head, neck, left and right shoulders, left and right elbows, left and right wrists, waist, left and right hips, left and right knees, and left and right ankles. Each joint point is represented by three-dimensional coordinates (x, y, z). A kinematic inversion model is used to calculate the joint angle data. Taking the elbow joint as an example, the elbow joint angle is calculated by the included angle between the upper arm and forearm bone vectors. Specifically, if the upper arm vector is v1=(x1, y1, z1) and the forearm vector is v2=(x2, y2, z2), the elbow joint angle θ can be calculated through the dot product and modulus length of the vectors. For example, at a certain moment, the upper arm vector is (0.25, 0.15, 0.05) and the forearm vector is (0.10, 0.20, 0.15), and the calculated elbow joint angle is 37.5 degrees. Similar calculations are repeated for the main joints of the whole body to obtain a complete joint angle data set.

[0022] Based on the joint angle data at adjacent moments, calculate the joint movement speed and acceleration. Assume the sampling frequency is 30Hz, and the time interval between adjacent frames is 33.3 milliseconds. For any joint, such as the elbow joint, if the angle at time t is θt = 85 degrees and the angle at time t + 1 is θt+1 = 88 degrees, then the joint movement speed vt = (θt+1 - θt) / Δt = 90 degrees per second. Similarly, if the speed at time t + 1 is vt+1 = 105 degrees per second, then the acceleration at = (vt+1 - vt) / Δt = 450 degrees per second 2 Perform similar calculations for all joints to form a complete dataset of joint movement speed and acceleration. In practical applications, to reduce the influence of noise, the sliding window averaging method is used to smooth the data, and the window size is set to 5 frames. For example, when a user is walking fast, the knee joint angle change rate remains at about 120 degrees per second, and the acceleration peak appears in the landing and pushing-off phases of the gait cycle, with a value of about 800 degrees per second 2 .

[0023] Construct a human bone linkage model that includes bone nodes and connection relationships. This model consists of 17 bone nodes and 16 linkages, forming a tree-like structure with the pelvis as the root node. Each bone segment has preset mass, length, and moment of inertia parameters. In the standard adult male model, the mass of the upper arm is about 2.0 kg, the mass of the forearm is about 1.2 kg, the mass of the thigh is about 7.5 kg, and the mass of the calf is about 3.5 kg. Map the bone joint position data to the corresponding nodes of the human bone linkage model, and calculate the joint pressure and torque according to the connection relationships between the nodes. Taking the knee joint as an example, assume the user's weight is 70 kg. In the semi-squat position, through static analysis, considering the masses of the thigh and calf and the action of gravity, the calculated knee joint pressure is about 1.5 times the body weight, that is, about 1030 N, and the knee joint torque is about 31.5 Nm. Perform similar calculations for the main joints of the whole body to form a dataset of joint forces

[0024] Input the joint angle data and joint force data into the static force analysis channel of a preset human biomechanical model. This channel contains a multi-layer neural network structure. The input layer receives the joint angle and force data, and the hidden layer contains 64 nodes, using the ReLU activation function. Through the processing of this channel, static force features are obtained. These features contain a 32-dimensional vector, representing the muscle force distribution required for the user to maintain the current posture. Taking the squat position as an example, the component value in the feature that reflects the load of the hip joint muscle group is 0.75, and the component value of the knee joint muscle group load is 0.82, indicating that the static load on the knee joint in this posture is slightly higher than that on the hip joint

[0025] Input the joint movement speed data and joint movement acceleration data into the dynamic force analysis channel of the human biomechanical model. This channel adopts a temporal convolutional network structure, including a one-dimensional convolutional layer with a time window size of 8 to extract temporal features. Through the processing of this channel, dynamic force features are obtained. This feature is a 24-dimensional vector, representing the joint impact force and inertial force during the movement. Taking the running motion as an example, the dynamic impact force feature component of the ankle joint during the landing phase reaches 0.86, and that of the knee joint is 0.72, indicating that running has a greater dynamic impact on the ankle joint.

[0026] Fuse the static force features and dynamic force features through the feature fusion layer of the human biomechanical model. The feature fusion layer adopts an attention mechanism to dynamically adjust the weights of the static features and dynamic features according to the movement type. For static postures such as yoga movements, the weight of the static features is relatively high, about 0.75; for high-speed movements such as jumping, the weight of the dynamic features is relatively high, about 0.85. After fusion, a 48-dimensional comprehensive feature vector is obtained. According to the output result of the feature fusion layer, the corresponding movement load parameters of the user in the real environment are calculated through a fully connected layer, including three indicators: muscle fatigue index, joint pressure index, and exercise intensity index, all of which are normalized to the range of 0-1. Taking a group of continuous squat movements as an example, after completing 10 standard squats, the quadriceps muscle fatigue index is 0.65, the knee joint pressure index is 0.58, and the overall exercise intensity index is 0.45, indicating that this exercise has a greater load on the quadriceps, but the joint pressure and overall intensity are moderate.

[0027] During application, the model is personalized adjusted according to the body parameters (height, weight, bone proportion, etc.) of different users. Adjust the joint force calculation parameters through the body mass index (BMI). For every 5-unit increase in the BMI value, the joint pressure index increases by about 12%. This method supports real-time calculation. At a data acquisition frequency of 60Hz, the update delay of the movement load parameters is less than 50 milliseconds, meeting the real-time interaction requirements. Through the monitoring of the movement load parameters, potential overtraining risks can be identified in a timely manner. When the joint pressure index of a certain joint continuously exceeds 0.8 for 60 seconds, an overload warning is generated, and it is recommended that the user adjust the movement mode or intensity.

[0028] The calculation of exercise load parameters also takes into account the influence of environmental factors. For example, when exercising on a slope, for every 5-degree increase in slope, the pressure indices of the knee joint and ankle joint increase by approximately 8% and 12% respectively. Specific parameter adjustment strategies are set for different exercise types. For high-impact exercises (such as jumping and sprinting), the weight of the dynamic force characteristics is increased to 0.9, while for low-intensity continuous exercises (such as jogging and cycling), the weights of static and dynamic characteristics are more balanced, being 0.45 and 0.55 respectively. The exercise load assessment results can be used to adjust personalized training plans. When the load index of a certain part continuously exceeds the safety threshold, alternative training suggestions are automatically generated to ensure the training effect while reducing the risk of sports injuries.

[0029] In this embodiment, the joint angles, velocities, and accelerations calculated through kinematic inversion can accurately capture the geometric and dynamic characteristics of the user's movement, providing a reliable basis for subsequent analysis; the joint pressures and torques calculated based on the bone linkage model comprehensively reflect the two force characteristics of static (muscle strength required to maintain posture) and dynamic (impact force, inertial force), helping to deeply understand the true loads borne by each joint during exercise; by continuously updating the force characteristics with data at adjacent moments, real-time monitoring of the load changes during the movement process can be achieved, providing timely basis for exercise guidance and risk warning; the load parameters calculated by fusing static and dynamic force characteristics can be customized to evaluate the training intensity and recovery needs according to the body types and exercise modes of different users; the quantitative joint force and torque data help to timely detect abnormal stress concentrations, guide posture adjustment and training plan optimization, thereby reducing the risk of injury and improving sports performance.

[0030] In an alternative embodiment, a neural network model is used to perform time series analysis on the exercise load parameters, and an attention mechanism is adopted to extract the long-term and short-term characteristics of the user's exercise pattern. Based on the extracted long-term and short-term characteristics, the physiological limit thresholds of the user at different exercise intensities are predicted, including: Construct an adaptive time window based on the user's historical exercise load parameters, dynamically adjust the window length based on the exercise volatility, and generate variable-length time series data segments; Input the variable-length time series data segments into a recurrent convolutional neural network, and extract the local exercise feature sequence through a multi-layer recurrent convolutional structure; Input the local exercise feature sequence into a graph attention network, construct the local exercise feature sequence into a time series feature graph, and extract the time series dependence relationship between feature nodes through graph convolution operations to obtain global correlation features; Input the local exercise feature sequence and the global correlation features into a time series attention module, and the time series attention module constructs a multi-scale feature pyramid to calculate the adaptive weight coefficients for the local exercise feature sequence at different time scales; Perform temporal cross-attention calculation on the adaptive weight coefficient and the global correlation feature to generate enhanced features at multiple time scales, and fuse the enhanced features at multiple time scales through a cross-scale feature aggregation network to obtain the long-term and short-term features of the user's motion pattern; Use a Bayesian neural network to model the long-term and short-term features, calculate the probability distribution function of the physiological limit threshold, and use the mean of the probability distribution function as the prediction result of the user's physiological limit threshold under different exercise intensities.

[0031] In a specific implementation, an adaptive time window is constructed according to the user's historical exercise load parameters. Collect the user's historical exercise load data for at least 30 days, including muscle fatigue index, joint pressure index, and exercise intensity index. Calculate the volatility of the exercise load parameters, defined as the ratio of the change amplitude within the adjacent 7 days to the mean value of the parameters. Divide the volatility into three intervals: low volatility (<0.15), medium volatility (0.15 - 0.35), and high volatility (>0.35). For the low volatility interval, set the fixed window length to 14 days; for the medium volatility interval, set the window length to 10 days; for the high volatility interval, set the window length to 7 days. For example, if a user's recent exercise intensity index volatility is 0.28, which belongs to the medium volatility interval, set the window length to 10 days. Perform sliding segmentation on the data within each window, with an overlapping rate of 50% for adjacent segments, to generate variable-length time series data segments. For example, the variable-length time series data segments generated by a user contain the exercise load parameters at 10 time points, and each time point contains 3 index values.

[0032] Input the variable-length time series data segments into a recurrent convolutional neural network, which consists of 3 layers of recurrent convolutional structures. Each layer contains 16, 32, and 64 convolutional kernels, with a kernel size of 3 and a stride of 1. The first layer of convolution processes the input variable-length time series data to extract primary time series features; the second layer performs recurrent convolution on the output of the first layer to capture medium-term change patterns; the third layer performs recurrent convolution on the output of the second layer to extract high-order time series features. The recurrence depth of the recurrent convolution is set to 5, that is, each layer of convolution is applied 5 times with parameter sharing. Through the recurrent convolutional structure, extract the local motion feature sequence. For example, for an input data segment of 10×3 dimensions (10 time points, 3 indicators per point), after being processed by the recurrent convolutional network, a local motion feature sequence of 10×64 dimensions is obtained, which characterizes the local time dependence of the user's motion pattern.

[0033] Input the local motion feature sequence into the graph attention network. Each time point of the feature sequence is regarded as a node in the graph, and the adjacency matrix is constructed according to the correlation between time points. The edge weights between nodes are calculated based on the similarity of feature vectors, and connections are established between nodes with a similarity exceeding 0.6. The constructed temporal feature graph contains 10 nodes (corresponding to 10 time points). The graph attention network contains 2 layers of graph convolutional layers, each with 32 and 16 attention heads respectively, and the attention dimension is 8. The first layer of graph convolution aggregates the information of its neighbor nodes for each node, and the second layer further expands the receptive field to capture the correlations between nodes in a wider range. Through the graph convolution operation, the temporal dependencies between feature nodes are extracted to obtain the global correlation features. The dimension of the global correlation features is 10×128, which contains rich long-range dependency information between time points. For example, analysis finds that there is a significant correlation between the exercise intensity and the muscle fatigue index 2-3 days later, and the correlation coefficient is 0.72.

[0034] Input the local motion feature sequence and the global correlation features into the temporal attention module. This module constructs a 3-level feature pyramid, corresponding to time scales of 3 days, 5 days, and 10 days respectively. For each scale, calculate the adaptive weight coefficients of the local motion feature sequence. The weight calculation uses the self-attention mechanism, which maps the feature sequence into query, key, and value parts, calculates the attention scores through the matching degree between the query and the key, and obtains the weight coefficients of each time point. The shorter scale (3 days) focuses on recent fluctuations, and the longer scale (10 days) captures long-term trends. For example, for the 5-day scale, the weight coefficients of a user's joint pressure index in the last two days are 0.25 and 0.30, indicating that the data of these two days have a greater impact on the current prediction.

[0035] Perform temporal cross-attention calculation on the adaptive weight coefficients and the global correlation features. The cross-attention uses the global correlation features as the query, and the weighted local features as the key and value to generate enhanced features. Calculate the enhanced features for the three time scales of 3 days, 5 days, and 10 days respectively to obtain three feature matrices with a dimension of 10×128. The enhanced features of the three time scales are fused through the cross-scale feature aggregation network. The aggregation network uses the channel attention mechanism to assign different weights to features of different scales. For example, for a more volatile exercise pattern, the feature weights of the 3-day and 5-day scales are 0.45 and 0.35 respectively, and the 10-day scale is 0.2; for a stable exercise pattern, the weights of the three scales are 0.2, 0.3, and 0.5 respectively. After aggregation, the long-term and short-term features of the user's exercise pattern are obtained, with a dimension of 10×256.

[0036] The Bayesian neural network is used to model long-term and short-term features. The Bayesian neural network consists of 3 fully connected layers with 128, 64, and 32 nodes respectively, and the activation function is Leaky ReLU. Instead of using fixed values, the network weights follow a prior distribution, which is initially set to a Gaussian distribution with a mean of 0 and a variance of 0.1. The network weights are sampled by the variational inference method, and 50 samples are taken for each prediction. The network outputs are calculated for multiple sets of sampled weights, and the mean and variance of the outputs are statistically calculated to obtain the probability distribution function of the physiological limit threshold. For each exercise intensity level (low, medium, high), an independent probability distribution function is calculated. The mean of the probability distribution function is used as the prediction result of the physiological limit threshold of the user at different exercise intensities. For example, for a certain user during moderate-intensity exercise, the predicted value of the physiological limit threshold of the muscle fatigue index is 0.72 ± 0.08, indicating that there is a risk of overtraining when the muscle fatigue index of this user exceeds 0.72.

[0037] The method of this embodiment is adjusted personalized according to user characteristics. For the age factor, for every 10-year increase, the predicted physiological limit threshold is reduced by 5 - 8%; for gender differences, the joint pressure index threshold of female users is on average 10% lower than that of male users. Regarding the exercise level, users are divided into beginner (<6 months of training experience), intermediate (6 - 24 months), and advanced (>24 months), and the baseline correction coefficients for predicting the threshold are set to 0.85, 1.0, and 1.15 respectively. The method of this embodiment supports incremental learning. After every 7 days of new data is collected, the new data is combined with the historical data to update the weight distribution of the Bayesian neural network, realizing the continuous optimization of the model. The experimental results show that this method has a higher accuracy in predicting the physiological limit threshold of users' exercise compared with the traditional fixed-threshold method and also higher compared with non-Bayesian deep learning methods.

[0038] In this embodiment, the adaptive variable-length time window can intelligently capture the key turning points and abnormal fluctuations in the movement, significantly reducing the information loss and noise interference caused by the fixed window; the multi-layer recursive convolution and the graph attention network cooperate to extract local details and global temporal dependencies, realizing a deep and multi-dimensional representation of the movement pattern; the multi-scale attention pyramid dynamically weights the short-term burst and long-term trend features, enhancing the sensitivity to movement information at different time scales; the cross-scale feature aggregation fuses the enhanced features of each scale through temporal cross-attention, taking into account both immediate responses and long-term trends, further improving the accuracy and robustness of pattern recognition and prediction; the Bayesian neural network performs probabilistic modeling on long-term and short-term features, not only outputting a reliable confidence assessment of the physiological limit threshold, but also providing personalized, safe, and efficient training suggestions for different individuals at each exercise intensity.

[0039] In an alternative embodiment, a Bayesian neural network is used to model the long-term and short-term features, calculate the probability distribution function of the physiological limit threshold, and use the mean of the probability distribution function as the prediction result of the physiological limit threshold of the user at different exercise intensities, including: Input the long-term and short-term features into a feature encoding network to generate a sequence of feature vectors; Construct a Bayesian neural network, set latent variables in each layer of the Bayesian neural network, and input the sequence of feature vectors into the Bayesian neural network; Based on the sequence of feature vectors, construct training samples, and use a variational autoencoder to model the latent variables in the Bayesian neural network to obtain the prior distribution of the latent variables; Input the prior distribution of the latent variables into a probability neural network to generate a weight distribution, which is used to characterize the uncertainty of each layer of network parameters in the Bayesian neural network; Optimize the weight distribution through the stochastic variational inference method, generate multiple sets of network parameters according to the weight distribution, construct a deep ensemble model based on the multiple sets of network parameters, input new long-term and short-term features into the deep ensemble model, and output multiple prediction results based on the deep ensemble model; Perform probability density estimation on the multiple prediction results to obtain the probability distribution function of the physiological limit threshold, calculate the mean of the probability distribution function, and use the mean as the prediction result of the physiological limit threshold of the user at different exercise intensities.

[0040] In a specific embodiment, the long-term and short-term features are input into a feature encoding network to generate a sequence of feature vectors. The feature encoding network adopts a bidirectional long short-term memory network structure, including two layers of bidirectional LSTM layers, with 128 hidden units in each layer. The input long-term and short-term feature dimension is 10×256, corresponding to 256-dimensional feature vectors at 10 time points. The forward and backward LSTM units are used to process the time series data respectively to capture the bidirectional time-dependent relationship. The dimension of the sequence of feature vectors output by the encoding network is 10×256, retaining the original dimension but enhancing the feature expression ability. For users with different exercise modes, such as strength training users and endurance training users, different parameter initialization strategies are adopted. The forgetting gate bias of strength training users is initialized to 1.5, and the forgetting gate bias of endurance training users is initialized to 1.0 to adapt to the time series characteristics of different exercise modes. In the example, the numerical distribution of the feature vector sequence of a certain strength training user in the dimension related to muscle fatigue is between 0.65 and 0.85, while the numerical distribution in the dimension related to cardiopulmonary function is between 0.40 and 0.55.

[0041] Construct a Bayesian neural network, set latent variables in each layer of the Bayesian neural network, and input the feature vector sequence into the Bayesian neural network. The Bayesian neural network contains 3 fully connected layers with the number of nodes being 128, 64, and 32 respectively, and uses the Swish activation function. In each layer of the network, set the latent variable z to control the distribution of weights and biases. The dimension of the latent variable z matches the number of weight parameters in the corresponding layer. For example, the connection from the input layer to the first hidden layer requires 256×128 = 32768 weight parameters, and the dimension of the corresponding latent variable z is set to 256. The dimensions of the latent variables in the first, second, and third layers are 256, 128, and 64 respectively. The network input is a 10×256 feature vector sequence, which is converted into a 256-dimensional feature vector through average pooling operation in the time dimension, and then processed by each fully connected layer in turn.

[0042] Construct training sample pairs based on the feature vector sequence, and use a variational autoencoder to model the latent variables in the Bayesian neural network to obtain the prior distribution of the latent variables. The training sample pair refers to consecutive adjacent feature vectors, such as the feature vectors at time t and t+1, which are used to capture the time series change characteristics. The variational autoencoder consists of two parts: an encoder and a decoder. The encoder is composed of two fully connected layers with the number of nodes being 128 and 64 respectively, which maps the input feature vector to the latent variable space; the decoder is also composed of two fully connected layers with the number of nodes being 64 and 128 respectively, which reconstructs the latent variable into a feature vector. The encoder outputs the mean and variance parameters of the latent variable z, and samples the latent variable z. Through the training of the variational autoencoder, the prior distribution of the latent variable z is obtained, which is represented as a multi-dimensional Gaussian distribution, and the sizes of the mean vector and covariance matrix are consistent with the dimension of the latent variable. For example, the mean vector of the prior distribution of the first-layer latent variable is between -0.05 and 0.05, and the elements of the diagonal covariance matrix are between 0.01 and 0.1.

[0043] Input the prior distribution of the latent variable into the probabilistic neural network to generate the weight distribution, which is used to characterize the uncertainty of each layer of network parameters in the Bayesian neural network. The probabilistic neural network contains a single-layer fully connected network, with the input being the latent variable z and the output being the distribution parameters of the weight matrix W and the bias vector b. The weight distribution is represented by a low-rank Gaussian distribution, and the covariance structure is defined by the combination of the mean matrix and several rank-one matrices, reducing the number of parameters. For the first-layer connection, the dimension of the weight mean matrix is 256×128, and the actual number of stored parameters is reduced by 90%. The mean of the weight distribution is usually initialized as a small random number in the interval [-0.05, 0.05], and the standard deviation is initialized as 0.01. For different types of features, set different weight initialization strategies: use He initialization for exercise intensity-related features and use the standard initialization method based on the input and output dimensions for recovery ability-related features.

[0044] The weight distribution is optimized by the stochastic variational inference method. Multiple sets of network parameters are generated according to the weight distribution, and a deep ensemble model is constructed based on the multiple sets of network parameters. The stochastic variational inference is implemented using the Bayes by Backprop algorithm. Specifically, for the weights W of each layer of the network, its posterior distribution is defined as a Gaussian distribution, parameterized by the mean vector μ and the standard deviation vector σ. During the forward propagation process, instead of directly using the deterministic weights, the weights are sampled from the posterior distribution. The sampling process uses the reparameterization trick: first, noise ε is sampled from the standard normal distribution, and then the weight sample is obtained through the transformation W = μ + σ * ε. This trick ensures that the gradient can be backpropagated through the random nodes. The loss function consists of two parts: the prediction loss and the complexity loss. The prediction loss uses the mean squared error to measure the gap between the model prediction and the actual target; the complexity loss uses the KL divergence to measure the difference between the posterior distribution and the prior distribution to prevent overfitting. For the prior distribution, a mixture Gaussian distribution is adopted, which includes a narrow Gaussian (mean 0, standard deviation 0.05) and a wide Gaussian (mean 0, standard deviation 2.0), and the mixing ratio is 0.5:0.5. This mixed prior can encourage both parameter sparsity and expressiveness simultaneously. During the backpropagation process, the parameters μ and σ are updated by gradient descent. The initial learning rate is set to 0.001, the Adam optimizer is used, and the learning rate decay strategy is applied. The learning rate is multiplied by 0.9 every 2000 iterations. To improve the training stability, the gradient clipping technique is applied to limit the gradient norm not to exceed 5.0. In a typical training, after 10,000 iterations, the model loss drops from the initial value of 4.82 to 0.67, and the average standard deviation of the weight distribution converges to a specific pattern. For example, the standard deviation of the weights related to muscle fatigue prediction is in the range of 0.03 - 0.08, while the standard deviation of the weights related to joint pressure prediction is in the range of 0.02 - 0.05, reflecting the differences in uncertainty for different prediction tasks.

[0045] After the optimization is completed, 50 different sets of weight parameters are sampled from the weight distribution to form 50 sub-networks, thus forming a deep ensemble model. Each set of weight parameters maintains the same network structure but different parameter values, reflecting the characteristics of the posterior distribution of Bayesian inference. The long-term and short-term features of the new user are input into each sub-network to obtain 50 independent prediction results. For example, for a certain user under high-intensity training conditions, the threshold range of the muscle fatigue index predicted by the 50 sub-networks is between 0.68 and 0.83, and the threshold range of the joint pressure index is between 0.65 and 0.78. The degree of dispersion of the prediction results reflects the uncertainty level of the model's prediction for this user. For users with sufficient training data, the prediction results are more concentrated, and the standard deviation is usually less than 0.05; while for users with sparse data or inconsistent training patterns, the prediction results are more dispersed, and the standard deviation can reach 0.12.

[0046] Perform probability density estimation on multiple prediction results to obtain the probability distribution function of the physiological limit threshold, calculate the mean of the probability distribution function, and use the mean as the prediction result of the user's physiological limit threshold under different exercise intensities. The kernel density estimation method is used to perform probability density estimation on 50 prediction results. The kernel function is selected as the Gaussian kernel, and the bandwidth parameter is adaptively determined. The probability density function is represented by 100 uniformly sampled points to form the probability distribution of the physiological limit threshold. Calculate the mean of the probability distribution function as the final prediction result of the physiological limit threshold. At the same time, extract the standard deviation of the distribution to quantify the uncertainty of the prediction. For example, for a certain user under medium-intensity training conditions, the predicted result of the physiological limit threshold of the muscle fatigue index is 0.75 ± 0.06, indicating that within the 95% confidence interval, the muscle fatigue index threshold of this user is between 0.63 and 0.87. According to the prediction result, when it is real-time monitored that the user's muscle fatigue index exceeds 0.75, the system issues an overtraining risk warning.

[0047] In the prior art, the prediction of the physiological limit threshold usually adopts the fixed threshold method or a single neural network model, which cannot effectively handle individual differences and prediction uncertainties. The fixed threshold method relies on expert experience to set a unified standard, ignoring individual differences, and the accuracy rate is lower than 60%; although a single neural network can learn individual characteristics, it lacks the ability to quantify uncertainty and cannot evaluate the reliability of the prediction result, and the prediction accuracy rate is unstable in the case of large data fluctuations. The improvement starting point of the present invention is to introduce a Bayesian framework to handle prediction uncertainties, and combine a variational autoencoder to model the prior distribution of latent variables to achieve more accurate personalized prediction. By setting hierarchical latent variables and using a low-rank Gaussian distribution to represent weights, the number of parameters is significantly reduced, and the calculation efficiency is improved. The random variational inference is used to optimize the weight distribution, which is faster than the traditional Markov chain Monte Carlo method. Through uncertainty quantification, the system can automatically adjust the warning threshold according to the prediction reliability, significantly reduce the false alarm rate, and improve the user experience and training safety.

[0048] In this embodiment, through the feature encoding network and Bayesian latent variable modeling, it is possible to quantify the distribution of long-term and short-term motion features and the uncertainty of network parameters, thereby greatly improving the reliability of threshold prediction; by using a variational autoencoder to learn the prior distribution of latent variables, the model still has strong generalization ability when the sample size is limited; applying random variational inference to generate multiple sets of network parameters and construct a deep ensemble effectively enhances the robustness of the model to noise and abnormal data and reduces the risk of overfitting; by performing probability density estimation on multiple prediction results, not only can a point estimate of the physiological limit threshold be given, but also a confidence interval can be provided, providing decision-making support for personalized exercise recommendations and risk assessment; this end-to-end Bayesian deep learning framework takes into account both performance and interpretability, facilitating subsequent flexible deployment under different exercise intensities and scenarios and realizing the optimization of individualized training programs.

[0049] As Figure 2 shown, the traditional Bayesian network and the MCMC Bayesian network have the largest number of parameters, reaching 2587K and 2482K respectively, and the longest inference time (267ms and 294ms); the Dropout Bayesian and LSTM networks have relatively moderate numbers of parameters, 1692K and 1426K respectively, and the inference times are 211ms and 138ms; the integrated neural network performs poorly in terms of the number of parameters (2093K), but performs well in terms of inference time (187ms); while the solution of this embodiment significantly reduces the number of parameters to 958K by using hierarchical latent variables and low-rank Gaussian distributions to represent weights, a reduction of 62.97% compared to the traditional Bayesian network. At the same time, the inference time is only 93ms, a reduction of 68.37% compared to the MCMC Bayesian network. This result shows that the solution of this embodiment not only outperforms the existing methods in terms of prediction performance, but also achieves a significant improvement in computational efficiency. In particular, by replacing the traditional Markov chain Monte Carlo method with the stochastic variational inference method, the inference speed is greatly improved, enabling the system to run in real time on resource-constrained portable devices, providing the possibility for practical applications. The low number of parameters and high computational efficiency enable this system to be deployed on portable devices such as smart watches to achieve real-time monitoring and early warning of the user's physiological limits.

[0050] In an alternative embodiment, a training knowledge triple is constructed based on the user's historical training data, and a personalized training knowledge graph is constructed based on the training knowledge triple through multi-level knowledge extraction and dynamic graph optimization, including: Performing temporal segmentation on the user's historical training data to obtain training data sequences for multiple training cycles; Extracting training entities and training relationships from the training data sequences, constructing training knowledge triples, inputting the training knowledge triples into a knowledge reasoning model, and semantically expanding the training knowledge triples based on the knowledge reasoning model; Performing multi-granularity partitioning on the training knowledge triples, constructing knowledge subgraphs at each granularity level, where the knowledge subgraphs contain training entity nodes and training relationship edges at that granularity level; Performing representation learning on the knowledge subgraphs, iteratively updating the feature representations of the training entity nodes through a message passing mechanism, calculating a similarity matrix between nodes based on the updated feature representations, and constructing a node importance scoring model in combination with the weight distribution of the training relationship edges; Calculating the importance scores of the training entity nodes according to the node importance scoring model, setting an adaptive pruning threshold based on the importance scores, dynamically pruning the knowledge subgraphs, and aggregating the features of the pruned multiple knowledge subgraphs to construct a hierarchical training knowledge graph; Map the new training data to the hierarchical training knowledge graph, and update the training entity nodes and training relationship edges in the hierarchical training knowledge graph; Generate a personalized training knowledge graph based on the hierarchical training knowledge graph.

[0051] In a specific implementation, the historical training data of the user is segmented by time series to obtain training data sequences for multiple training cycles. Specifically, the system collects all the fitness records of the user in the past 6 months, including information such as training type, training duration, training intensity, and calorie consumption, and segments them according to one week as a training cycle to obtain 24 data sequences for training cycles. For example, the data of user Xiao Wang in the first week is: Monday (running, 30 minutes, medium intensity, 250 calories), Wednesday (strength training, 45 minutes, high intensity, 320 calories), Friday (yoga, 60 minutes, low intensity, 180 calories).

[0052] Extract training entities and training relationships from the training data sequences to construct training knowledge triples. Training entities include user identifiers, training items, training parts, training equipment, etc.; training relationships include "participate in", "use", "exercise", etc. For example, from the data of user Xiao Wang in the first week, the following triples can be extracted: (Xiao Wang, participate in, running), (running, exercise, cardiopulmonary function), (Xiao Wang, use, dumbbell), (strength training, exercise, arm muscles), etc. Input these triples into a pre-trained knowledge inference model for semantic expansion. Based on rules and statistical analysis, the model infers new relationships, such as triples like (running, contribute to, fat loss), (strength training, enhance, metabolic rate), etc., enriching the original knowledge.

[0053] Perform multi-granularity partitioning on the training knowledge triples, and construct knowledge subgraphs at each granularity level. According to different attributes of training activities, it is divided into three granularities: training type granularity (such as aerobic training, anaerobic training), training goal granularity (such as fat loss, muscle gain), and training detail granularity (such as specific actions, equipment use). At the training type granularity level, construct a knowledge subgraph containing nodes such as "running", "strength training", "yoga", etc., and the edges represent the association relationships between training types; at the training goal granularity level, construct a knowledge subgraph containing nodes such as "fat loss", "muscle gain", "improve cardiopulmonary function", etc.; at the training detail granularity level, construct a knowledge subgraph containing specific actions and equipment.

[0054] Perform representation learning on the knowledge subgraph, and iteratively update the feature representations of training entity nodes through a message passing mechanism. In actual implementation, each node is initialized as a 128-dimensional vector. Through 3 rounds of message passing iteration, adjacent node information is collected and its own features are updated. For example, the "Running" node updates its own feature representation by aggregating information from nodes such as "Fat Loss" and "Cardiopulmonary Function" connected to it. Based on the updated feature representations, calculate the cosine similarity between nodes to obtain a similarity matrix. Combine the weight distribution of training relationship edges (set according to user training frequency and preferences) to construct a node importance scoring model. This model comprehensively considers the connectivity, centrality of nodes, and the matching degree with user training preferences.

[0055] Calculate the importance scores of training entity nodes according to the node importance scoring model. For example, for user Xiao Wang, the score of the "Strength Training" node is 0.85, the score of the "Running" node is 0.76, and the score of the "Yoga" node is 0.63. Based on these scores, set an adaptive pruning threshold, taking 80% of the average of the node scores at each granularity level as the threshold. Dynamically prune the knowledge subgraph, removing nodes with scores lower than the threshold and their associated edges. For example, the "Swimming" node with a score of only 0.42 is pruned. Combine the pruned multiple knowledge subgraphs through a feature aggregation method, retain the information of nodes at different granularity levels, and construct a hierarchical training knowledge graph.

[0056] When new training data is generated, map it to the hierarchical training knowledge graph. For example, in the 25th week, the user added "High-Intensity Interval Training" data. The system extracts relevant triples and maps them to the existing graph. If the new entity "High-Intensity Interval Training" has a high similarity with the existing node "Aerobic Training", an association is established between them; at the same time, update the features and relationship weights of the affected nodes. If the user performs the same training for three consecutive weeks, the weight of the corresponding relationship edge will increase by 20% to reflect the user's stable preference.

[0057] Generate a personalized training knowledge graph based on the updated hierarchical training knowledge graph. The system analyzes the user's recent training focus and long-term training pattern, highlights nodes and relationships with high importance scores, and hides content that the user has not paid attention to recently. For example, for user Xiao Wang who is recently focused on strength training, in his personalized training knowledge graph, the "Strength Training" and related nodes will be emphasized, and their close connection with the training goal of "Muscle Gain" will be shown, while weakening the display of "Yoga" related content.

[0058] In this embodiment, based on the knowledge triples constructed from training entities and relationships and semantically augmented by the inference model, implicit training elements and associations can be mined, providing a more comprehensive knowledge basis for subsequent analysis; the multi-granularity partitioning and sub-graph construction take into account both the macroscopic framework and microscopic details of training knowledge, enabling the model to not only understand the overall training profile but also capture key action nodes; through the representation learning of message passing and the calculation of the similarity matrix, combined with the importance score constructed from the relationship edge weight distribution, the accurate identification of key training entities is achieved, greatly enhancing the focus of the knowledge graph; adaptive pruning based on node importance not only compresses redundant information and reduces computational overhead but also enhances the model's robustness to noisy data; real-time mapping and updating of new training data to the hierarchical knowledge graph make the finally generated personalized training knowledge graph both structured and dynamically evolvable, providing reliable support for formulating more targeted training plans and continuous optimization.

[0059] In an alternative embodiment, a graph neural network model is constructed, and the personalized training knowledge graph is input into the graph neural network model. Based on the node feature vectors output by the graph neural network model, a phased training plan for the user is generated, including: Construct a graph neural network model, encode the topological structure features of training entity nodes in the spatial dimension, and model the dynamic evolution features of training entity nodes in the temporal dimension using a temporal-aware graph convolutional network to obtain spatial dimension features and temporal dimension features; Input the personalized training knowledge graph into the graph neural network model, and adaptively fuse the spatial dimension features and the temporal dimension features through a multi-layer cross-attention mechanism to generate multi-dimensional feature vectors of training entity nodes; Use self-attention-guided graph pooling operations to sparsely reduce the dimensionality of the multi-dimensional feature vectors to obtain compressed feature representations of training entity nodes; Based on the compressed feature representations, construct a hierarchical reinforcement learning model with multi-objective joint optimization, model the progressive relationship constraint and the exercise intensity constraint as hierarchical optimization objectives, and respectively learn the global sorting strategy and local intensity parameters of training items through a two-layer policy network; Input the global sorting strategy and the local intensity parameters into the training plan generation module, determine the progressive order of training items according to the global sorting strategy, and adjust the exercise intensity of training items according to the local intensity parameters to generate a phased training plan that meets the progressive relationship and exercise intensity requirements.

[0060] In a specific implementation, a graph neural network model is constructed, which can extract two-dimensional features of training entity nodes. In the spatial dimension, a dedicated topological structure encoder is used to process the association features between nodes. Specifically, for the node vi in the training knowledge graph, its first-order neighbor node set N(vi) is obtained, and the neighbor information is aggregated through a message passing mechanism. For each neighbor node vj ∈ N(vi), the attention weight of it to the central node vi is calculated, and this weight reflects the importance of the neighbor node to the central node. For example, for the "squat" training node, the directly related "thigh muscle group" node will obtain a higher weight, while the indirectly related "balance ability" node will obtain a lower weight.

[0061] In the temporal dimension, a temporally aware graph convolutional network is used to capture the temporal evolution pattern of node features. First, the user's historical training data is divided into multiple time windows in chronological order, for example, one week for each window. For the feature vector xit of the node vi at the time window t, through the sliding window method, the temporal dependence relationship is modeled by combining the feature sequences of the previous k time windows {xi(t - k),..., xi(t - 1), xit}. For example, the change trajectory of the user's load on the "bench press" training item from the initial 30 kg to 50 kg can be used to predict the appropriate future training intensity.

[0062] The personalized training knowledge graph is input into the above graph neural network model, and the spatial dimension features and temporal dimension features are fused through a multi-layer cross-attention mechanism. In the implementation, a three-layer cross-attention structure is adopted, and each layer contains two parallel attention sub-modules, which process spatial features and temporal features respectively. For the node vi, its spatial feature vector is vis, and its temporal feature vector is vit. First, the cross-attention coefficient between the two is calculated, and then feature fusion is performed based on the attention coefficient to generate a comprehensive feature representation. For example, for the "beginner" user node, the system may assign a higher weight to the temporal features to capture its ability improvement curve; while for the "professional trainer" user node, a higher weight may be assigned to the spatial features to focus on the collaborative relationship between training items.

[0063] The self-attention guided graph pooling operation is used to sparsify and reduce the dimension of the multi-dimensional feature vector. Specifically, first, the importance score of each node in the graph is calculated. For example, the "strength training" node has a higher importance score for strength-based users. According to the importance score, the top K key nodes are selected (in practice, K is usually set to 30% of the original number of nodes), and the information of these nodes and their features are retained, while the information of other non-key nodes is merged to form a compressed graph structure. The finally obtained compressed feature representation not only retains the detailed information of the key training entities but also reduces the computational complexity of subsequent processing.

[0064] Based on the compressed feature representation, a hierarchical reinforcement learning model for multi-objective joint optimization is constructed. This model consists of two layers of policy networks: the high-level policy network is responsible for learning the global sorting policy of training items, and the low-level policy network is responsible for optimizing the local intensity parameters. During the training process, a composite reward function is designed to guide the model learning, which takes into account both the progressive relationship constraint and the exercise intensity constraint. For example, for a novice fitness user, the model will first arrange basic actions such as "standard squat", and then recommend advanced actions such as "weighted squat", and dynamically adjust the training intensity according to the user's feedback.

[0065] Specifically, the high-level policy network adopts a three-layer fully connected neural network structure, with the compressed feature representation as the input and the priority score of the training item as the output. For example, for a leg training day, "warm-up stretch" gets the highest priority (0.95), "squat" comes second (0.87), and "leg press" is third (0.65). The low-level policy network adopts a two-layer neural network, and for each training item, it outputs its detailed parameters, such as the number of sets (3 - 5 sets), the number of repetitions per set (8 - 12 times), and the weight percentage (60% - 80% of the maximum load).

[0066] The global sorting policy and the local intensity parameters are input into the training plan generation module to generate the final phased training plan. This module first determines the progressive order of training items according to the global sorting policy. For example, a one-week training plan may be sorted as "chest training → back training → leg training → shoulder training → recovery day". Then, it adjusts the specific implementation details of each training item according to the local intensity parameters. For example, for a "chest training day", the system may arrange: "warm-up (5 minutes of running, heart rate reaches 120 - 130 bpm)" → "bench press (4 sets, 10 times per set, 65% of the maximum load)" → "incline bench press (3 sets, 12 times per set, 60% of the maximum load)" → "push-up (3 sets, until exhaustion)" → "stretching and relaxation (10 minutes)".

[0067] In this embodiment, by jointly encoding the training entity features in both spatial and temporal dimensions through a graph neural network, the structural evolution and dynamic changes during the training process can be comprehensively captured; the multi-layer cross-attention mechanism realizes the adaptive fusion of spatial and temporal features, enhancing the multi-dimensional expression ability of the training feature representation and the fine-grained difference perception; self-attention-guided graph pooling is used for feature sparsification and dimensionality reduction, which not only retains the key information but also effectively compresses the model scale, improving the computational efficiency and generalization performance; the hierarchical reinforcement learning model based on the compressed feature separates the training item sorting and the exercise intensity regulation for modeling, taking into account the rationality of the global training progress and the personalized optimization of the local training load; a phased training plan that is progressive and adapted to the exercise intensity is generated through a two-layer policy network, ensuring that the training plan is scientific and progressive and conforms to the physiological load law, thereby improving the training effect and reducing the risk of injury.

[0068] In an alternative embodiment, a hierarchical reinforcement learning model for multi-objective joint optimization is constructed based on the compressed feature representation. The progressive relationship constraint and the exercise intensity constraint are modeled as hierarchical optimization objectives. The global sorting strategy and the local intensity parameters of the training items are learned through a two-layer policy network, including: Model the progressive relationship constraint of the training items as a high-level optimization objective, model the exercise intensity constraint of the training items as a low-level optimization objective, and use a fuzzy evaluation matrix to calculate the high-level reward value and the low-level reward value of the state-action pair; Construct a two-layer policy network. The high-level policy network of the two-layer policy network uses a multi-scale sliding time window to segment and encode the training item sequence, calculates the progressive correlation score between items within each time window, and dynamically adjusts the progressive correlation threshold in combination with the high-level reward value to generate a global sorting reference strategy for the training items; The low-level policy network of the two-layer policy network constructs a multi-dimensional physiological load model based on the heart rate variability index and the exercise metabolism level of the training items, and dynamically adjusts the intensity parameters through a soft constraint optimization method introducing a buffer interval in combination with the low-level reward value to generate a local intensity adjustment strategy for the training items; Use a recurrent neural network to establish a bidirectional information channel between the high-level policy network and the low-level policy network. Input the global sorting reference strategy and the local intensity adjustment strategy into the bidirectional information channel, and iteratively optimize the global sorting reference strategy and the local intensity adjustment strategy based on a cross-layer constraint function to obtain the global sorting strategy and the local intensity parameters of the training items.

[0069] In a specific implementation, the progressive relationship constraint of training items is modeled as a high-level optimization goal, and the exercise intensity constraint of training items is modeled as a low-level optimization goal. A fuzzy evaluation matrix is used to calculate the high-level reward value and low-level reward value of the state-action pair. The progressive relationship constraint of training items refers to the sequential relationship between different training items in terms of technical difficulty, physiological load, and training objectives. For example, "basic strength training" should precede "specific explosive power training", and "low-intensity aerobic training" should precede "high-intensity interval training". The progressive relationship constraint is divided into five levels: strong leading relationship (the subsequent item cannot be carried out until the previous item is completed), weak leading relationship (it is recommended to complete the previous item first), no relationship, weak subsequent relationship, and strong subsequent relationship. The exercise intensity constraint of training items refers to the physiological load limit during the execution of the item, including parameters such as heart rate range, percentage of strength load, and perceived fatigue degree. The exercise intensity constraint is divided into three levels: low-intensity interval (heart rate is 60%-70% of the maximum heart rate), medium-intensity interval (heart rate is 70%-85% of the maximum heart rate), and high-intensity interval (heart rate is 85%-95% of the maximum heart rate). The dimension of the fuzzy evaluation matrix is the number of states × the number of actions, where the state represents the current physiological and training state of the trainee, and the action represents the optional training items. Each element in the matrix is calculated by a fuzzy membership function, and the value range is 0-1. The larger the value, the higher the suitability of the state-action pair. The high-level reward value mainly considers the progressive rationality, technical connection, and cycle matching degree of training items; the low-level reward value mainly considers the intensity suitability, fatigue recovery degree, and physiological tolerance. For example, for a certain strength trainee, the high-level reward value of choosing "high-intensity squats" in the "moderate fatigue state" is 0.3 (low, because it violates the progressive relationship), and the low-level reward value is 0.2 (low, because the intensity does not match the current recovery state).

[0070] Construct a two-layer policy network. The high-level policy network of the two-layer policy network uses a multi-scale sliding time window to segment and encode the training item sequence, calculates the progressive correlation score between items within each time window, and dynamically adjusts the progressive correlation threshold in combination with the high-level reward value to generate a global sorting reference policy for training items. The high-level policy network adopts a three-layer fully connected neural network structure. The input layer dimension is the state vector dimension (usually 16-32 dimensions), the number of hidden layer nodes is 64, and the output layer dimension is the action space dimension (i.e., the number of optional training items). The multi-scale sliding time window includes a short-term window (3-5 days), a medium-term window (7-14 days), and a long-term window (21-28 days), which capture training rules at different time scales respectively. Within each time window, calculate the progressive correlation score between training items. The score calculation is based on the cosine similarity of item feature vectors and an expert-defined progressive rule library. The progressive correlation score ranges from -1 to 1. A positive value indicates a positive progressive relationship between items, and a negative value indicates a conflict. The progressive correlation threshold is initially set to 0.6 and then dynamically adjusted according to the high-level reward value: a high reward value (>0.8) causes the threshold to be reduced by 5%-10%, increasing the sorting flexibility; a low reward value (<0.3) causes the threshold to be increased by 5%-10%, enhancing the sorting constraint. The global sorting reference policy is generated by performing a topological sort on the training items, considering the progressive relationship constraints and the training objective priorities. For example, in a 12-week strength training program, the global sorting reference policy arranges basic strength training items (such as squats and bench presses) in the first 4 weeks, compound strength training items in the middle 4 weeks, and specific strength and explosive power training items in the last 4 weeks.

[0071] The low-level policy network of the double-layer policy network constructs a multi-dimensional physiological load model based on the heart rate variability index and exercise metabolism level of the training item. By introducing a soft constraint optimization method with a buffer interval and combining the low-level reward value, it dynamically adjusts the intensity parameter to generate a local intensity adjustment strategy for the training item. The low-level policy network adopts a dual-channel convolutional neural network structure. One channel processes heart rate variability data, and the other channel processes metabolic index data. Finally, the intensity parameter adjustment value is merged and output through a fully connected layer. The heart rate variability index includes RMSSD (root mean square of the differences between adjacent RR intervals), SDNN (standard deviation of all RR intervals), LF / HF ratio (ratio of low-frequency power to high-frequency power), etc., which are used to evaluate the state of the autonomic nervous system and the degree of fatigue recovery. The exercise metabolism level is evaluated by indicators such as blood lactate concentration, excess post-exercise oxygen consumption (EPOC), and respiratory quotient (RER). The multi-dimensional physiological load model integrates these indicators into an 8-dimensional vector, representing the current physiological load state of the trainer. The soft constraint optimization method with a buffer interval sets an ideal target value and an acceptable range for each intensity parameter. For example, the ideal relative intensity of a certain strength training item is 80% 1RM (80% of one-repetition maximum), and the acceptable range is 75% - 85% 1RM. The dynamic adjustment of the intensity parameter is based on the low-level reward value and the current physiological state: when the low-level reward value is high and the physiological state is good, the intensity parameter can be adjusted upward by 1% - 3%; when the low-level reward value is low or the physiological state is poor, the intensity parameter is adjusted downward by 3% - 5%. For example, for an athlete with a poor recovery state (RMSSD is 20% lower than the baseline), the original planned squat training intensity of 85% 1RM will be adjusted to 80% 1RM.

[0072] A bidirectional information channel is established between the high-level policy network and the low-level policy network using a recurrent neural network. The global sorting reference policy and the local intensity adjustment policy are input into the bidirectional information channel, and the global sorting reference policy and the local intensity adjustment policy are iteratively optimized based on a cross-layer constraint function to obtain the global sorting policy and the local intensity parameters of the training items. The bidirectional information channel uses a long short-term memory network (LSTM) structure with a hidden state dimension of 128 and a time step of the number of days in the training plan. The LSTM network can capture long-term dependencies in the training plan, memorize important training patterns and transition points. The bidirectional information channel transmits sorting constraint information from the high level to the low level to guide the reasonable range of intensity parameters; it transmits physiological load information from the low level to the high level to affect the adjustment direction of item sorting. The cross-layer constraint function comprehensively considers the satisfaction of the progressive relationship, the rationality of the intensity parameters, and the physiological recovery window, and takes the weighted sum of the three as the optimization objective, with a weight ratio of 4:3:3. The iterative optimization process uses an alternating update strategy: the global sorting policy is updated in odd rounds, and the local intensity parameters are updated in even rounds, with 20 rounds of iteration. In each round of iteration, according to the current state and the expected goal, the sorting policy or the intensity parameters are fine-tuned to gradually approach the optimal solution. For example, in a certain iteration, it is found that three consecutive days of high-intensity training result in insufficient physiological recovery, and the system will adjust the global sorting by inserting low-intensity recovery training or rest days between high-intensity training days. After optimization, the obtained global sorting policy is represented as a sequence of training items, specifying the training items to be executed and their order within each training cycle; the local intensity parameters are represented as the specific execution parameters of each training item, such as weight, number of sets, number of repetitions, intensity interval, etc.

[0073] In the prior art, the sorting and intensity adjustment of training items usually adopt rule-based methods or single-level optimization algorithms. Rule-based methods rely on preset training templates and expert experience, lacking the ability of personalized adaptation and being difficult to cope with the dynamic changes of the trainer's state; single-level optimization algorithms either focus on the global training structure and ignore local intensity adjustment, or focus on local training parameters and lack global coordination, unable to balance the systematicness and adaptability of training. The improvement starting point of the present invention lies in introducing a two-layer optimization architecture, regarding the sorting and intensity adjustment of training items as two interrelated but different-level optimization problems, and realizing cross-layer collaboration through a two-way information channel. The high-level policy network ensures that the sorting of training items conforms to the progressive relationship and periodic characteristics, providing global guidance; the low-level policy network adjusts the training intensity according to the real-time physiological state to ensure the suitability of training stimuli. The fuzzy evaluation matrix and multi-scale sliding time window technology can handle the uncertainty and time dependence in training decisions, improving the robustness of the optimization results. Compared with traditional methods, the method of this embodiment has significant advantages in training effects, with an increased training objective achievement rate, improved training adaptability (the ability of the training plan to adjust according to state changes), and reduced overtraining risk. In terms of personalization, the method of this embodiment can generate differentiated training plans for users with different training levels and goals. The plan for beginners pays more attention to the cultivation of basic abilities and technical learning, the plan for intermediate trainers emphasizes systematicness and periodic changes, and the plan for advanced trainers highlights special abilities and refined adjustment. In terms of application scenarios, this method is applicable to various training types, including strength training, endurance training, ball game training, etc., and realizes the adaptation to different training domains by adjusting the corresponding progressive relationship constraints and intensity parameter definitions.

[0074] Such as Figure 3As shown, a complete training sequence is presented, which includes 7 key training item nodes and the progressive relationships between them. At the same time, the training sequence is analyzed and optimized through the multi-scale sliding time window technique. Each node in the figure represents a training item, and the size of the node reflects the importance of the training item. It can be seen that the circle of node 5 (high-intensity interval) is the largest, indicating its core position in the entire training sequence; while nodes 1 (warm-up activity) and 7 (relaxation and recovery) are smaller. Although they are essential, they belong to auxiliary training links. The connecting lines between the nodes represent the progressive relationships between the training items, and the thickness of the lines represents the intensity of the progressive correlation. It is worth noting that the connecting line from node 6 (special skill training) to node 7 (relaxation and recovery) is the thickest (correlation value 0.95), indicating that the progressive relationship of immediately performing relaxation and recovery after special training is the closest; while the connecting line from node 5 (high-intensity interval) to node 7 (relaxation and recovery) is thinner (correlation value 0.6), indicating that the direct progressive relationship between the two is relatively weak, and it is more suitable to use special skill training as a transition. The three dashed boxes in the figure represent multi-scale sliding time windows, which are respectively marked as "Time Window 1: Correlation Threshold = 0.65", "Time Window 2: Correlation Threshold = 0.75" and "Time Window 3: Correlation Threshold = 0.85". These windows cover different ranges of training items and set increasing correlation thresholds, reflecting the innovative method of the solution in this embodiment for analyzing the progressive relationships between training items at different time scales. Through the combined analysis of this network structure and multi-scale windows, the solution in this embodiment can simultaneously consider the close relationships between local training items and the structural characteristics of the overall training sequence, so as to generate a training item sorting strategy that not only meets the progressive requirements but also has global optimal characteristics.

[0075] The personalized training plan generation system in the metaverse intelligent sports according to the embodiment of the present invention includes: The first unit is used to obtain the motion data of the user in the metaverse virtual scene; The second unit is used to call a preset human biomechanics model to calculate the corresponding motion load parameters of the user in the real environment based on the bone joint point position data in the motion data; The third unit is used to perform time series analysis on the motion load parameters through a neural network model, extract the long-term and short-term characteristics of the user's motion pattern by using the attention mechanism, and predict the physiological limit threshold of the user under different exercise intensities based on the extracted long-term and short-term characteristics; The fourth unit is used to construct a training knowledge triple based on the historical training data of the user, and construct a personalized training knowledge graph through the method of multi-level knowledge extraction and dynamic graph optimization based on the training knowledge triple; The fifth unit is used to construct a graph neural network model, input the personalized training knowledge graph into the graph neural network model, and generate a phased training plan for the user based on the node feature vectors output by the graph neural network model. Among them, the training items in the phased training plan are sorted according to a progressive relationship, and the exercise intensity of each training item is within the range of the physiological limit threshold; The sixth unit is used to convert the phased training plan into training instructions for the metaverse virtual scene and display a training guidance interface in the metaverse virtual scene.

[0076] In the third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Among them, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0077] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0078] The present invention can be a method, device, system, and / or computer program product. The computer program product can include a computer-readable storage medium on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a personalized training plan in Metaverse Intelligent Sports, characterized in that: include: Obtain the user's motion data in the virtual scene of the Metaverse; Based on the skeletal joint position data in the motion data, a preset human biomechanics model is called to calculate the corresponding motion load parameters of the user in the real environment; Performing time series analysis on the exercise load parameters through a neural network model, extracting long-term and short-term features of the user's exercise pattern using an attention mechanism, and predicting the user's physiological limit threshold at different exercise intensities based on the extracted long-term and short-term features; Constructing training knowledge triples according to the historical training data of the user, and constructing a personalized training knowledge graph based on the training knowledge triples by multi-level knowledge extraction and dynamic graph optimization; Constructing a graph neural network model, inputting the personalized training knowledge graph into the graph neural network model, and generating a phased training plan for the user based on the node feature vector output by the graph neural network model, wherein the training items in the phased training plan are sorted according to a progressive relationship, and the exercise intensity of each training item is within the range of the physiological limit threshold; The phased training plan is converted into training instructions for a metaverse virtual scene, and a training guidance interface is displayed in the metaverse virtual scene.

2. The method according to claim 1, characterized in that Based on the skeletal joint position data in the motion data, a preset human biomechanics model is called to calculate the corresponding motion load parameters of the user in the real environment, including: According to the skeletal joint point position data in the motion data, a kinematic inversion model is used to calculate the user's joint angle data; based on the joint angle data at adjacent moments, the user's joint motion speed data and joint motion acceleration data are calculated; Establishing a human skeleton connection model, mapping the bone joint point position data to the corresponding bone nodes in the human skeleton connection model, and calculating the pressure value and torque value of each joint according to the connection relationship between the bone nodes in the human skeleton connection model; Inputting the joint angle data and the joint force data into a static force analysis channel of a preset human biomechanics model for processing to obtain a static force feature, wherein the static force feature represents the muscle force distribution required for the user to maintain the current posture; Inputting the joint motion speed data and the joint motion acceleration data into the dynamic force analysis channel of the human biomechanics model for processing to obtain dynamic force characteristics, wherein the dynamic force characteristics represent the joint impact force and inertia force during the user's motion; The static force feature and the dynamic force feature are fused through the feature fusion layer of the human biomechanics model to obtain an output result of the feature fusion layer; According to the output result of the feature fusion layer, the motion load parameter corresponding to the user in the real environment is calculated.

3. The method according to claim 1, characterized in that The exercise load parameters are analyzed in time series by a neural network model, and the long-term and short-term features of the user's exercise pattern are extracted by an attention mechanism. Based on the extracted long-term and short-term features, the user's physiological limit threshold at different exercise intensities is predicted, including: Build an adaptive time window based on the user's historical exercise load parameters, dynamically adjust the window length based on exercise volatility, and generate variable-length time series data segments; Inputting the variable-length time series data segments into a recursive convolutional neural network, and extracting local motion feature sequences through a multi-layer recursive convolutional structure; Inputting the local motion feature sequence into a graph attention network, constructing the local motion feature sequence into a temporal feature graph, extracting the temporal dependency between feature nodes through a graph convolution operation, and obtaining a global correlation feature; Inputting the local motion feature sequence and the global correlation feature into a temporal attention module, the temporal attention module constructs a multi-scale feature pyramid, and calculates adaptive weight coefficients for the local motion feature sequence at different time scales; Performing temporal cross-attention calculation on the adaptive weight coefficient and the global correlation feature to generate enhanced features of multiple time scales, and fusing the enhanced features of multiple time scales through a cross-scale feature aggregation network to obtain long-term and short-term features of the user's motion pattern; A Bayesian neural network is used to model the long-term and short-term characteristics, and a probability distribution function of the physiological limit threshold is calculated, and the mean of the probability distribution function is used as the prediction result of the physiological limit threshold of the user at different exercise intensities.

4. The method according to claim 3, characterized in that The long-term and short-term characteristics are modeled using a Bayesian neural network, the probability distribution function of the physiological limit threshold is calculated, and the mean of the probability distribution function is used as the prediction result of the physiological limit threshold of the user at different exercise intensities, including: Inputting the long-term and short-term features into a feature encoding network to generate a feature vector sequence; Constructing a Bayesian neural network, setting hidden variables in each layer of the Bayesian neural network, and inputting the feature vector sequence into the Bayesian neural network; Constructing training sample pairs based on the feature vector sequence, using a variational autoencoder to model latent variables in the Bayesian neural network, and obtaining a prior distribution of latent variables; Inputting the prior distribution of the latent variable into a probabilistic neural network to generate a weight distribution, wherein the weight distribution is used to characterize the uncertainty of each layer of network parameters in the Bayesian neural network; Optimizing the weight distribution by a random variational inference method, generating multiple sets of network parameters according to the weight distribution, constructing a deep integration model based on the multiple sets of network parameters, inputting new long-term and short-term features into the deep integration model, and outputting multiple prediction results based on the deep integration model; Probability density estimation is performed on the multiple prediction results to obtain a probability distribution function of the physiological limit threshold, a mean of the probability distribution function is calculated, and the mean is used as a prediction result of the physiological limit threshold of the user at different exercise intensities.

5. The method according to claim 1, characterized in that Constructing a training knowledge triple according to the historical training data of the user, and constructing a personalized training knowledge graph based on the training knowledge triple by multi-level knowledge extraction and dynamic graph optimization, including: Performing time series segmentation on the historical training data of the user to obtain training data sequences of multiple training cycles; Extracting training entities and training relations from the training data sequence, constructing training knowledge triples, inputting the training knowledge triples into a knowledge reasoning model, and semantically expanding the training knowledge triples based on the knowledge reasoning model; Performing multi-granularity division on the training knowledge triples, and constructing a knowledge subgraph at each granularity level, wherein the knowledge subgraph includes training entity nodes and training relationship edges at the granularity level; Performing representation learning on the knowledge subgraph, iteratively updating the feature representation of the training entity node through a message passing mechanism, calculating the similarity matrix between nodes based on the updated feature representation, and building a node importance scoring model in combination with the weight distribution of the training relationship edges; Calculating the importance scores of training entity nodes according to the node importance scoring model, setting an adaptive pruning threshold based on the importance score, dynamically pruning the knowledge subgraph, and performing feature aggregation on the pruned multiple knowledge subgraphs to construct a hierarchical training knowledge graph; Mapping new training data to the hierarchical training knowledge graph, and updating training entity nodes and training relationship edges in the hierarchical training knowledge graph; Generate a personalized training knowledge graph based on the hierarchical training knowledge graph.

6. The method according to claim 1, characterized in that Constructing a graph neural network model, inputting the personalized training knowledge graph into the graph neural network model, and generating a phased training plan for the user based on the node feature vector output by the graph neural network model, including: Build a graph neural network model to encode the topological structure features of the training entity nodes in the spatial dimension, and use the time-aware graph convolutional network to model the dynamic evolution features of the training entity nodes in the temporal dimension to obtain spatial dimension features and temporal dimension features. The personalized training knowledge graph is input into the graph neural network model, and the spatial dimension features and the temporal dimension features are adaptively fused through a multi-layer cross attention mechanism to generate a multi-dimensional feature vector of the training entity node; Using a self-attention-guided graph pooling operation to perform sparse dimension reduction on the multi-dimensional feature vector to obtain a compressed feature representation of the training entity node; A hierarchical reinforcement learning model for multi-objective joint optimization is constructed based on the compressed feature representation, and the progressive relationship constraints and motion intensity constraints are modeled as hierarchical optimization objectives. The global sorting strategy and local intensity parameters of the training items are learned respectively through a two-layer strategy network. The global sorting strategy and the local intensity parameters are input into a training plan generation module, the progressive order of the training items is determined according to the global sorting strategy, the exercise intensity of the training items is adjusted according to the local intensity parameters, and a phased training plan that meets the progressive relationship and exercise intensity requirements is generated.

7. The method according to claim 6, characterized in that Based on the compressed feature representation, a hierarchical reinforcement learning model for multi-objective joint optimization is constructed, and the progressive relationship constraints and motion intensity constraints are modeled as hierarchical optimization objectives. The global sorting strategy and local intensity parameters of the training items are learned respectively through a two-layer strategy network, including: The progressive relationship constraints of the training items are modeled as high-level optimization objectives, the exercise intensity constraints of the training items are modeled as low-level optimization objectives, and the high-level reward values ​​and low-level reward values ​​of the state-action pairs are calculated using a fuzzy evaluation matrix; Constructing a two-layer strategy network, wherein the high-level strategy network of the two-layer strategy network uses a multi-scale sliding time window to segmentally encode the training item sequence, calculates the progressive correlation score between items in each time window, and dynamically adjusts the progressive correlation threshold in combination with the high-level reward value to generate a global ranking reference strategy for the training items; The lower-layer strategy network of the two-layer strategy network constructs a multidimensional physiological load model based on the heart rate variability index and exercise metabolism level of the training project, and dynamically adjusts the intensity parameters by introducing a soft constraint optimization method between buffers in combination with the lower-layer reward value to generate a local intensity adjustment strategy for the training project; A recursive neural network is used to establish a bidirectional information channel between the high-level strategy network and the low-level strategy network, the global sorting reference strategy and the local strength adjustment strategy are input into the bidirectional information channel, and the global sorting reference strategy and the local strength adjustment strategy are iteratively optimized based on the cross-layer constraint function to obtain the global sorting strategy and local strength parameters of the training items.

8. A personalized training plan generation system in Metaverse Intelligent Sports, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the user's motion data in the virtual scene of the metaverse; The second unit is used to call a preset human biomechanics model to calculate the corresponding motion load parameters of the user in the real environment based on the skeletal joint point position data in the motion data; The third unit is used to perform time series analysis on the exercise load parameters through a neural network model, extract the long-term and short-term features of the user's exercise pattern using an attention mechanism, and predict the user's physiological limit threshold at different exercise intensities based on the extracted long-term and short-term features; A fourth unit is used to construct a training knowledge triple according to the historical training data of the user, and to construct a personalized training knowledge graph based on the training knowledge triple by multi-level knowledge extraction and dynamic graph optimization; A fifth unit is used to construct a graph neural network model, input the personalized training knowledge graph into the graph neural network model, and generate a phased training plan for the user based on the node feature vector output by the graph neural network model, wherein the training items in the phased training plan are sorted according to a progressive relationship, and the exercise intensity of each training item is within the range of the physiological limit threshold; The sixth unit is used to convert the phased training plan into training instructions for the metaverse virtual scene and display a training guidance interface in the metaverse virtual scene.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images