Body building guidance method and system based on knowledge graph

Abstract

The invention discloses a fitness guidance method based on a knowledge graph, and further discloses a system used by the fitness guidance method based on the knowledge graph. According to the method, the multi-modal heterogeneous features are determined, so that the features for describing the relevance among physiological signals, body-building actions and subjective feedback of the user during body-building of the target user can be obtained; the determination of the multi-modal heterogeneous features is beneficial to comprehensively perceiving the physiological status, exercise performance and subjective feeling of the user by fusing the time-frequency change features of the biological signals, the spatial-temporal features of the key skeleton points and the semantic features fed back by the user, so that the comprehensive understanding of the fitness state of the user is formed.

Description

Technical Field

[0001] The present invention belongs to the field of fitness guidance technology, and specifically relates to a fitness guidance method and system based on knowledge graph. Background Art

[0002] With the rapid development of intelligent fitness technology, users have an increasing demand for personalized, safe and efficient fitness guidance. Traditional fitness guidance methods mainly rely on general training plans and do not fully consider individual differences. In recent years, knowledge graph technology has been widely used in the field of artificial intelligence. It provides new possibilities for realizing personalized fitness guidance by representing entities and their relationships in a structured way.

[0003] Existing technologies provide scientific exercise guidance by constructing a fitness knowledge graph, using fitness videos and explanatory text as input, and using deep neural networks to extract action semantics to generate standard knowledge graph triples, which enriches the content of fitness guidance to a certain extent.

[0004] However, existing knowledge graph-based fitness guidance systems primarily provide fitness guidance by building static relationships between fitness movements and related information. These systems are unable to obtain real-time physiological signals and motion data from users, and they fail to effectively integrate heterogeneous data from multiple sources, including biosignals, motion movements, and user feedback. This limits understanding of the user's overall status, making it impossible to adjust fitness plans based on their status, impacting training effectiveness and safety. Therefore, effectively integrating multimodal data from users during fitness and building a dynamic, personalized fitness knowledge graph has become a challenge facing the industry. Summary of the Invention

[0005] The purpose of the present invention is to provide a fitness guidance method based on a knowledge graph to solve the problems raised in the above background technology. The fitness guidance method provided by the present invention is characterized by being able to effectively integrate the multimodal data of users during fitness and construct a dynamic and personalized fitness knowledge graph.

[0006] Another object of the present invention is to provide a system, a computer device and a computer-readable storage medium used in a fitness guidance method based on a knowledge graph.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a fitness guidance method based on knowledge graph, comprising the following steps:

[0008] S1. Obtain bioelectric signals, exercise videos, and feedback voice when the target user is exercising;

[0009] S2. Extract the time-frequency variation features of the target user's bioelectrical signals during exercise, extract the spatiotemporal features of the target user's key skeletal points from the exercise video, and extract the target user's feedback semantic features from the feedback voice;

[0010] S3. Perform cross-dimensional heterogeneous feature fusion on time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user;

[0011] S4. Based on multimodal heterogeneous features, construct all entity triplets between user feedback, fitness movements, and motion biosignals during the target user's fitness process, and then determine the confidence level of the relationship between the feedback semantic entity, fitness movement entity, and biosignal entity within each entity triplet;

[0012] S5. Based on all entity triples whose relationship confidence is higher than a preset threshold, construct multi-dimensional dynamic constraints for the target user's fitness, and then generate a training knowledge graph for personalized fitness guidance for the target user based on the multi-dimensional dynamic constraints.

[0013] In the present invention, further in S2, the method for extracting the spatiotemporal features of the key skeleton points of the target user from the motion video includes the following steps:

[0014] S21, detecting the key points of the human skeleton in each frame of the motion video to obtain a joint coordinate dataset of the key skeleton points of the target user;

[0015] S22. Determine, based on the joint coordinate data set, a joint position change sequence of each key skeletal point of the target user between consecutive frames;

[0016] S23. Perform spatiotemporal fitting on all joint position change sequences to obtain the spatiotemporal features of the target user's key skeletal points.

[0017] Furthermore, in S3, the method of performing cross-dimensional heterogeneous feature fusion on the time-frequency variation features, the spatiotemporal features, and the feedback semantic features to obtain multimodal heterogeneous features of the target user includes the following steps:

[0018] S31. Feature mapping of time-frequency variation features based on convolutional neural networks to obtain low-dimensional embedding representation under physiological modalities;

[0019] S32, feature encoding the spatiotemporal features to obtain a low-dimensional embedding representation under the spatiotemporal modality;

[0020] S33, determine the low-dimensional embedding representation under semantic modality through fully connected neural network and feedback semantic features;

[0021] S34. Through the cross-modal attention mechanism, the low-dimensional embedding representations under all modalities are fused to obtain the multimodal heterogeneous features of the target user.

[0022] In S4, based on multimodal heterogeneous features, the method for constructing all entity triplets between user feedback, fitness movements, and motion biosignals during the fitness process of the target user includes the following steps:

[0023] S411. Determine the feedback semantic entity of the target user feedback through multimodal heterogeneous features;

[0024] S412: Determine the action entity of the target user's fitness action;

[0025] S413. Determine a biosignal entity of the target user's motion biosignal based on the multimodal heterogeneous features;

[0026] S414: Perform spatiotemporal alignment on the feedback semantic entity, the fitness action entity, and the biosignal entity to obtain all entity triplets between the user feedback, fitness action, and motion biosignal during the fitness process of the target user.

[0027] In the present invention, further, in S4, the method for determining the confidence level of the relationship between the feedback semantic entity, the fitness action entity, and the biosignal entity in each entity triplet includes the following steps:

[0028] S421. For each entity triple, determine the temporal dependency features between entities in the entity triple by using an attention mechanism;

[0029] S422. Perform sequence modeling on the temporal dependency features to obtain the confidence distribution of the relationship between entities;

[0030] S423. Determine the relationship confidence between entities in the entity triplet through the confidence distribution, and then obtain the relationship confidence between the feedback semantic entity, the fitness action entity and the biological signal entity in each entity triplet.

[0031] Furthermore, in the present invention, in S5, the method for constructing multi-dimensional dynamic constraints for the target user's fitness based on all entity triples whose relationship confidence is higher than a preset threshold includes the following steps:

[0032] S51, filtering out all entity triples whose relationship confidence is higher than a preset threshold;

[0033] S52: Generate an incremental knowledge set for fitness guidance for the target user based on all entity triples whose relationship confidence is higher than a preset threshold;

[0034] S53, integrating the incremental knowledge set into the global knowledge graph for updating, and obtaining the entity relationship subgraph in the federated temporal knowledge graph updating process;

[0035] S54. Construct multi-dimensional dynamic constraints for target users during fitness based on the entity relationship subgraph.

[0036] In the present invention, further, in S54, the method for constructing multi-dimensional dynamic constraints for target users during fitness based on the entity relationship subgraph includes the following steps:

[0037] S541. Determine the security constraints of the target user during the fitness process based on the entity relationship subgraph;

[0038] S542, performing performance prediction on the entity relationship subgraph, thereby generating enhanced constraints on the target user's fitness performance;

[0039] S543. Conflicts between the safety constraint and the enhancement constraint are resolved through a multi-objective optimization algorithm to obtain multi-dimensional dynamic constraints for the target user during fitness.

[0040] Furthermore, in the present invention, a system used in a fitness guidance method based on a knowledge graph includes a multimodal dynamic graph construction unit, and the multimodal dynamic graph construction unit includes:

[0041] An acquisition module is used to obtain bioelectric signals, exercise videos, and feedback voice when the target user is doing fitness exercises;

[0042] A processing module is used to extract the time-frequency variation features of the target user's bioelectric signals during fitness, extract the spatiotemporal features of the target user's key skeletal points from the fitness video, and extract the target user's feedback semantic features from the feedback voice; to perform cross-dimensional heterogeneous feature fusion on the time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user; and to construct all entity triplets between user feedback, fitness movements, and exercise biosignals during the target user's fitness process based on the multimodal heterogeneous features, and then determine the relationship confidence between the feedback semantic entity, fitness movement entity, and biosignal entity in each entity triplet;

[0043] The execution module is used to construct multi-dimensional dynamic constraints for the target user's fitness based on all entity triples whose relationship confidence is higher than a preset threshold, and then generate a training knowledge graph for personalized fitness guidance for the target user based on the multi-dimensional dynamic constraints.

[0044] Further in the present invention, a computer device is provided, which includes a memory and a processor, wherein the memory stores code, and the processor is configured to obtain the code and execute the above-mentioned fitness guidance method based on knowledge graph.

[0045] Furthermore, in the present invention, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the above-mentioned fitness guidance method based on the knowledge graph is implemented.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. The present invention determines multimodal heterogeneous features to obtain features that describe the correlation between the target user's physiological signals, fitness movements, and subjective feedback during fitness. The determination of multimodal heterogeneous features helps to fully perceive the user's physiological condition, exercise performance, and subjective feelings by fusing the time-frequency variation characteristics of biological signals, the spatiotemporal characteristics of key skeletal points, and the semantic characteristics of user feedback, thereby forming a comprehensive understanding of the user's fitness status.

[0048] 2. The present invention determines the relationship confidence to obtain a probability indicator for measuring the credibility of the relationship between entities. The determination of the relationship confidence helps to effectively evaluate the reliability of the relationship between entities, thereby giving priority to retaining high-confidence entity triples in the knowledge graph, ensuring the accuracy and robustness of the knowledge graph construction.

[0049] 3. The present invention determines multi-dimensional dynamic constraints to obtain guidance measures that comprehensively consider safety and efficiency improvement during the fitness process to ensure that the target user obtains the best training effect within a safe range. The determination of multi-dimensional dynamic constraints helps the system to monitor the user's status in real time, and dynamically adjust the fitness plan according to the target user's real-time physiological signals, movement movements and feedback information, and generate a personalized training knowledge map, thereby providing fitness guidance that is more in line with individual needs and ensuring the safety and effectiveness of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Flow chart of the method of the present invention.

[0051] Figure 2 Flowchart for determining multimodal heterogeneous features for the present invention.

[0052] Figure 3 Flowchart for determining multi-dimensional dynamic constraints for the present invention.

[0053] Figure 4 This is a system block diagram of the present invention.

[0054] Figure 5 This is a diagram of the internal structure of a computer device. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] Example 1

[0057] See also Figure 1-Figure 5 , this embodiment provides the following technical solution: a fitness guidance method based on knowledge graph, comprising the following steps:

[0058] S1. Obtain bioelectric signals, exercise videos, and feedback voice when the target user is exercising;

[0059] It should be noted that bioelectric signals are electrical signals generated by the potential difference of human tissue that reflect the physiological functional status of nerves, muscles or organs. They can be used to monitor and analyze the user's physiological responses during fitness in real time; sports video is a sequence of continuous image frames arranged in chronological order during the target user's fitness exercise; feedback voice refers to the spoken feedback audio signal of the target user when performing fitness exercise.

[0060] In specific implementation, obtaining the target user's bioelectric signals, motion videos and feedback voice when performing fitness exercises can be achieved in the following ways, namely: first, the target user's electromyographic signals and electrocardiographic signals can be collected through wearable devices (such as surface electromyographic sensors and heart rate monitoring wristbands) when the target user performs fitness exercises, and the collection of the collected electromyographic signals and electrocardiographic signals is used as the target user's biosignals; then, a high-definition camera can be used to collect motion videos that record the target user's limb changes when the target user performs fitness exercises; finally, the target user's feedback voice can be collected through a microphone with noise reduction function when the target user performs fitness exercises.

[0061] S2. Extract the time-frequency variation features of the target user's bioelectrical signals during exercise, extract the spatiotemporal features of the target user's key skeletal points from the exercise video, and extract the target user's feedback semantic features from the feedback voice;

[0062] Specifically, the following steps can be used to extract the time-frequency variation characteristics of the target user's bioelectrical signal during fitness:

[0063] De-noise the target user's bioelectrical signals during fitness to obtain noise-free bioelectrical signals;

[0064] The feature extraction of the noise-free biological signal is performed through wavelet transformation to obtain the time-frequency variation feature of the bioelectrical signal of the target user when exercising.

[0065] In specific implementation, the bioelectric signals of the target user during fitness are denoised to obtain noise-free biosignals, which can be achieved in the following way: the bioelectric signals of the target user during fitness can be denoised by existing denoising methods (such as Kalman filtering) to obtain noise-free biosignals; wherein the noise-free biosignals are high signal-to-noise ratio time series data obtained after removing background noise and power frequency interference in the bioelectric signals, which can provide purer and more reliable physiological feature inputs, avoid noise interference in time-frequency analysis and feature extraction, and thus improve the accuracy and robustness of downstream multimodal fusion and relationship extraction.

[0066] It should be noted that the time-frequency variation characteristics are characteristics that describe the changes in the target user's biological signals in the two dimensions of time and frequency when exercising. The time-frequency variation characteristics can be used to deeply analyze the user's physiological reactions during fitness and provide data support for personalized and efficient fitness guidance. In specific implementation, the noise-free biological signal is extracted through wavelet transform to obtain the time-frequency variation characteristics of the target user's bioelectric signal when exercising. This can be achieved in the following way, namely: the denoised biological signal can be subjected to multi-scale analysis using wavelet transform and a time-frequency graph reflecting its changes in time-frequency can be extracted, and the time-frequency graph can be used as the time-frequency variation characteristics of the target user's biological signal when exercising. Among them, the wavelet transform can be performed by convolving the biological signal with a wavelet basis function with finite support, and then analyzing the local characteristics of the biological signal at different scales and positions to obtain the representation of the signal on the time-frequency plane.

[0067] Specifically, in S2, the method for extracting the spatiotemporal features of the key skeleton points of the target user from the motion video includes the following steps:

[0068] S21, detecting the key points of the human skeleton in each frame of the motion video to obtain a joint coordinate dataset of the key skeleton points of the target user;

[0069] S22. Determine, based on the joint coordinate data set, a joint position change sequence of each key skeletal point of the target user between consecutive frames;

[0070] S23. Perform spatiotemporal fitting on all joint position change sequences to obtain the spatiotemporal features of the target user's key skeletal points.

[0071] In a specific implementation, the human skeleton key points are detected frame by frame on the motion video, and the joint coordinate dataset of the key skeleton points of the target user is obtained in the following manner, namely: the existing key skeleton point detection algorithm (such as OpenPose) can be used to decode the motion video into image frames frame by frame and perform key skeleton point detection on each frame of the image to obtain the three-dimensional coordinates of each key skeleton point of the target user's body in each frame, and the set of all three-dimensional coordinates is used as the joint coordinate dataset of the key skeleton points of the target user; wherein, OpenPose is a real-time multi-person 2D pose estimation algorithm based on convolutional neural networks, which can use a multi-stage convolutional neural network to generate a key point confidence map for each frame of motion video to represent the joint position and bone connection information, thereby detecting the key skeleton points of the human body.

[0072] It should be noted that, in this application, the joint coordinate dataset is a dataset composed of the three-dimensional coordinate information of the key skeletal points in the target user's body during the fitness process. The key skeletal points are usually the main joints of the human body (such as the head, shoulders, elbows, wrists, hips, knees and ankles, etc.). This dataset reflects the target user's motion change trajectory during exercise and is the basis for subsequent analysis of the user's motion stability, coordination and rhythm; the joint position change sequence is a time series that describes the motion changes of the key skeletal points during the target user's fitness process. The position change sequence reflects the dynamic changes of the target user's joints during exercise, and provides basic data support for subsequent spatiotemporal feature extraction and personalized fitness guidance.

[0073] In specific implementation, the following method can be used to determine the joint position change sequence of each key skeletal point of the target user between consecutive frames based on the joint coordinate data set, namely: for each key skeletal point of the target user, all three-dimensional coordinates of the key skeletal point are obtained in the joint coordinate data set, and all the three-dimensional coordinates are arranged and combined in chronological order to form a sequence as the joint position change sequence between consecutive frames of the key skeletal point. The above steps can be used to obtain the joint position change sequence of each key skeletal point of the target user between consecutive frames.

[0074] It should be noted that the spatiotemporal features are features that describe the changes in the target user's key skeletal points in time and space when exercising. The spatiotemporal features can be used to deeply analyze the user's movement patterns during fitness and provide data support for personalized and efficient fitness guidance. In specific implementation, all joint position change sequences are spatiotemporally fitted to obtain the spatiotemporal features of the target user's key skeletal points. This can be achieved in the following way, namely: the target user's continuous motion trajectory in time and space can be generated according to the joint position change sequence of each key skeletal point through an existing trajectory planning algorithm (such as the quintic polynomial interpolation method or the non-uniform B-spline curve method), and the continuous motion trajectory is used as the spatiotemporal feature of the target user's key skeletal point. Among them, the trajectory planning algorithm can smoothly fit the continuous motion trajectory of the key skeletal points in time and space according to the joint position change sequence. Through the continuous motion trajectory, features reflecting the stability, coordination and rhythm of the user's movements can be extracted for personalized fitness guidance and provide support for sports performance evaluation and action recognition.

[0075] Specifically, extracting the target user's feedback semantic features from the feedback speech can be achieved by using the following steps:

[0076] Performing speech preprocessing on the feedback speech to generate a clean speech stream;

[0077] Converting the speech stream into feedback text using speech recognition technology;

[0078] Extract semantic features of the target user's feedback on the fitness guidance from the feedback text.

[0079] In a specific implementation, the feedback voice is subjected to voice preprocessing to generate a clean voice stream, which can be achieved in the following manner: the feedback voice can be preprocessed using existing voice preprocessing technologies (such as endpoint detection, denoising, and voice enhancement, etc.) to obtain a clean voice stream; wherein the voice stream refers to a continuous voice signal collected and preprocessed by a microphone during the target user's fitness process.

[0080] In specific implementation, the voice stream can be converted into feedback text through voice recognition technology in the following manner, namely: the voice stream can be converted into feedback text using existing voice recognition technology (such as an automatic speech recognition system); wherein, the feedback text refers to the text data provided by the target user during the fitness process. By transcribing the voice signal into text, the user's subjective experience and needs can be further analyzed, thereby providing a basis for personalized fitness guidance.

[0081] In a specific implementation, the semantic features of the target user's feedback on fitness guidance can be extracted from the feedback text in the following manner: natural language processing technology (such as the BERT model) can be applied to perform sentiment analysis and semantic understanding on the feedback text and identify the sentiment tendency vector of the target user's feedback, and then use the sentiment tendency vector as the semantic feature of the target user's feedback on fitness guidance; wherein, the feedback semantic feature is a feature that describes the emotions and intentions expressed by the target user when interacting with the fitness guidance system during the fitness process. The feedback semantic feature can be used to deeply understand user needs, thereby providing personalized and accurate fitness guidance for the user, and improving user experience and satisfaction.

[0082] S3. Perform cross-dimensional heterogeneous feature fusion on time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user;

[0083] Specifically, in S3, the method of performing cross-dimensional heterogeneous feature fusion on time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user includes the following steps:

[0084] S31. Feature mapping of time-frequency variation features based on convolutional neural networks to obtain low-dimensional embedding representation under physiological modalities;

[0085] S32, feature encoding the spatiotemporal features to obtain a low-dimensional embedding representation under the spatiotemporal modality;

[0086] S33, determine the low-dimensional embedding representation under semantic modality through fully connected neural network and feedback semantic features;

[0087] S34. Through the cross-modal attention mechanism, the low-dimensional embedding representations under all modalities are fused to obtain the multimodal heterogeneous features of the target user.

[0088] In specific implementation, feature mapping of the time-frequency change features is performed based on a convolutional neural network to obtain a low-dimensional embedding representation under the physiological modality. This can be achieved in the following manner: an existing convolutional neural network (such as a two-dimensional convolutional neural network) can be loaded, and the time-frequency change features are input into the convolutional neural network. The local frequency band features are extracted through the convolution layer, and the convolution output is reduced in dimensionality using a global average pooling layer to obtain a fixed-length 64-dimensional time-frequency embedding vector, which is then used as the low-dimensional embedding representation under the physiological modality. The convolution kernel size in the convolutional neural network can be set to 3×3, and the number of output channels can be set to 64 to extract local frequency band information. A global average pooling layer is then added after the convolution layer to globally summarize each channel of the entire time-frequency graph to generate a single feature value for each channel, thereby generating a 64-dimensional time-frequency embedding representation. This method can effectively retain the key information in the time-frequency change features and provide a unified low-dimensional representation for subsequent multimodal feature fusion.

[0089] It should be noted that the low-dimensional embedding representation is a dense vector representation of a high-dimensional feature vector converted into a low-dimensional space. The low-dimensional embedding representation can be used to match other modal features while retaining key feature information, thereby improving the processing efficiency and performance of the model.

[0090] In specific implementation, feature encoding of the spatiotemporal features to obtain a low-dimensional embedding representation under the spatiotemporal modality can be achieved in the following manner, namely: the spatiotemporal features can be input into an existing encoder (such as a Transformer encoder), and the spatiotemporal features can be encoded in combination with position encoding and a multi-head self-attention mechanism to output a fixed-length 64-dimensional spatiotemporal embedding vector, and the 64-dimensional spatiotemporal embedding vector is used as a low-dimensional embedding representation under the spatiotemporal modality.

[0091] In specific implementation, the low-dimensional embedding representation under the semantic modality is determined by a fully connected neural network and the feedback semantic features. This can be achieved in the following manner: an existing fully connected neural network (such as a multilayer perceptron) can be loaded, and the feedback semantic features are input into the fully connected neural network model for dimensionality reduction processing to obtain a 64-dimensional semantic embedding vector of a fixed length, and the 64-dimensional semantic embedding vector is used as the low-dimensional embedding representation under the semantic modality; wherein, the fully connected neural network model can perform nonlinear transformation through the ReLU activation function to achieve compression and conversion of semantic features, and apply the fully connected layer to reduce the dimensionality of the semantic features to 64 dimensions. Through this process, a concise and representative 64-dimensional semantic embedding can be extracted from the high-dimensional semantic features, ensuring that the semantic features can be effectively integrated with the time-frequency change features and the spatiotemporal features in the same low-dimensional space.

[0092] It should be noted that the multimodal heterogeneous features are features that describe the correlation between the target user's physiological signals, fitness movements and user feedback during fitness. The multimodal heterogeneous features can accurately correlate the target user's physiological signal abnormalities, movement deviations and user subjective feedback, thereby providing a decision-making basis for dynamically adjusting the fitness plan that takes into account both safety and training effectiveness. In specific implementation, the low-dimensional embedding representations under all modalities are fused through the cross-modal attention mechanism to obtain the multimodal heterogeneous features of the target user. This can be achieved in the following way, namely: first, the low-dimensional embedding representations under these three modalities are processed as Query, Key and Value respectively through the multi-head attention mechanism module to calculate the interactive attention weights between each modality, and the weighted features are combined. The fusion generates a joint embedding representation, and then feature alignment constraints can be imposed on the joint embedding representation through existing feature alignment methods (such as contrastive loss or covariance constraints) to obtain multi-dimensional features of spatiotemporal-physiological-semantic collaborative fusion, and the multi-dimensional features are used as multimodal heterogeneous features of the target user. Among them, the low-dimensional embedding representations under all modalities are fused through the cross-modal attention mechanism, which can dynamically allocate information weights of different modalities, fully capture the complementary characteristics between the modalities, and make up for the limitations of a single modality, thereby improving the discriminability and robustness of the overall features. At the same time, the mechanism can automatically suppress noise and irrelevant information, achieve efficient alignment and collaborative fusion between modalities, and provide more accurate and reliable data support for subsequent personalized fitness guidance.

[0093] S4. Based on multimodal heterogeneous features, construct all entity triplets between user feedback, fitness movements, and motion biosignals during the target user's fitness process, and then determine the confidence level of the relationship between the feedback semantic entity, fitness movement entity, and biosignal entity within each entity triplet;

[0094] Specifically, in S4, based on multimodal heterogeneous features, the method for constructing all entity triplets between user feedback, fitness movements, and motion biosignals during the fitness process of the target user includes the following steps:

[0095] S411. Determine the feedback semantic entity of the target user feedback through multimodal heterogeneous features;

[0096] S412: Determine the action entity of the target user's fitness action;

[0097] S413. Determine a biosignal entity of the target user's motion biosignal based on the multimodal heterogeneous features;

[0098] S414: Perform spatiotemporal alignment on the feedback semantic entity, the fitness action entity, and the biosignal entity to obtain all entity triplets between the user feedback, fitness action, and motion biosignal during the fitness process of the target user.

[0099] In a specific implementation, determining the feedback semantic entity of the target user feedback through the multimodal heterogeneous features can be achieved in the following manner, namely: the user feedback semantics can be extracted from the multimodal heterogeneous features and the existing entity recognition method (such as the BERT+CRF model) is used to perform named entity recognition on the user feedback semantics to obtain a feedback semantic entity containing a timestamp in the user feedback semantics (such as "knee pain_10:05:22").

[0100] In specific implementation, the action entity of the target user's fitness action can be determined in the following way: extracting a skeleton key point sequence from the multimodal heterogeneous features and using an existing entity recognition method (such as a CNN-LSTM action classification model) to perform dynamic feature analysis on the skeleton key point sequence, mapping the original action data into a standardized action label, and thus obtaining an action entity containing a timestamp (such as "squat_10:05:22").

[0101] In specific implementation, the biosignal entity of the target user's motion biosignal based on the multimodal heterogeneous features can be determined in the following manner, namely: the motion biosignal in the multimodal heterogeneous features is monitored through the existing anomaly detection algorithm (such as Isolation Forest) and the biosignal entity containing the timestamp is extracted (such as heart rate>150bpm_10:05:22).

[0102] It should be noted that entities are the basic units for constructing triples in knowledge graphs, and usually represent specific things or concepts in the real world (such as people, places, things, etc.). In this application, the feedback semantic entity refers to words or phrases that express specific emotions and evaluation opinions in user feedback. Accurately identifying feedback semantic entities can deeply understand user needs and experiences, thereby providing valuable references for personalized fitness guidance; the action entity refers to a specific fitness action instance of the target user during fitness. Accurately identifying the action entity can analyze the user's movement behavior and posture, thereby providing data support for personalized fitness guidance and effect evaluation; the biosignal entity refers to an electrical or physical signal that reflects the target user's physiological activities during fitness. Accurately identifying the biosignal entity can monitor the user's physiological state in real time, providing a scientific basis for personalized fitness guidance and health management.

[0103] In specific implementation, the feedback semantic entity, the fitness action entity and the bio-signal entity are temporally and spatially aligned to obtain all entity triplets between user feedback, fitness actions and motion bio-signals during the fitness process of the target user. This can be achieved in the following manner: first, a sliding time window method (for example, a 5-second window) can be used to temporally and spatially align the feedback semantic entity, action entity and bio-signal entity, and aggregate entities within the same time period into multiple candidate sets. Then, according to the time sequence and semantic relationship (such as action appears before feedback), the preset combination rules (such as causal relationship verification or lagged correlation analysis) are used to combine the entities in each candidate set and construct all entity triplets reflecting the relationship between user feedback, fitness actions and bio-signals.

[0104] It should be noted that the entity triple is a triple that describes the relationship between the target user's feedback and fitness movements and physiological reactions, for example: (squat, knee pain, heart rate > 150bpm). In this solution, by constructing the entity triple, the relationship between user feedback, fitness movements and biological signals can be clearly expressed, which can be used to deeply analyze user behavior patterns, thereby providing data support for personalized fitness guidance.

[0105] Specifically, in S4, the method for determining the confidence level of the relationship between the feedback semantic entity, the fitness action entity, and the biological signal entity in each entity triplet includes the following steps:

[0106] S421. For each entity triple, determine the temporal dependency features between entities in the entity triple by using an attention mechanism;

[0107] S422. Perform sequence modeling on the temporal dependency features to obtain the confidence distribution of the relationship between entities;

[0108] S423. Determine the relationship confidence between entities in the entity triplet through the confidence distribution, and then obtain the relationship confidence between the feedback semantic entity, the fitness action entity and the biological signal entity in each entity triplet.

[0109] In specific implementation, the following method can be used to determine the temporal dependency features between entities in an entity triple based on the attention mechanism, namely: first, a relation extraction model based on the attention mechanism (such as a Transformer model) can be loaded, and the triplet can be input into the relation extraction model. The input triplet is converted into a vector representation through an embedding layer and processed by a multi-layer Transformer encoder to generate a temporal context vector, and then the temporal context vector is used as the temporal dependency feature between entities in the entity triple; wherein the temporal dependency feature is a feature representation that describes the contextual semantic relationship between entities in the entity triple, and the temporal dependency feature can be used to capture the potential correlation between entities, providing important basic information for subsequent relationship analysis; the Transformer model uses the self-attention mechanism to capture the dependency relationship between entity words in the sequence, thereby effectively identifying the potential connection between entities.

[0110] In specific implementation, sequence modeling is performed on the temporal dependency features to obtain the confidence distribution of the relationship between entities, which can be achieved in the following way: the hidden state of the Transformer encoder in the temporal dependency features can be mapped to the relationship score space through multiple layers of fully connected layers and the scores are normalized using the Softmax activation function to obtain the confidence distribution of the relationship between each entity; wherein the confidence distribution is a probability distribution that describes the strength of the relationship between entities in the triple, and the confidence distribution can be used to evaluate the reliability of the relationship between entities, thereby helping to screen out high-credibility relationships and improving the accuracy and robustness of knowledge graph construction.

[0111] It should be noted that in this application, relationship confidence is a probabilistic indicator for measuring the degree of credibility of relationships between entities. This relationship confidence can be used to screen out high-quality relationships in a triple sequence, thereby ensuring the accuracy and robustness of knowledge graph construction. In specific implementation, the relationship confidence between entities in an entity triple can be determined by the confidence distribution in the following manner, namely: the average value of all values ​​in the confidence distribution can be used as the relationship confidence between entities in the entity triple.

[0112] S5. Based on all entity triples whose relationship confidence is higher than a preset threshold, construct multi-dimensional dynamic constraints for the target user's fitness, and then generate a training knowledge graph for personalized fitness guidance for the target user based on the multi-dimensional dynamic constraints.

[0113] In S5, the method for constructing multi-dimensional dynamic constraints for the target user's fitness based on all entity triples whose relationship confidence is higher than a preset threshold includes the following steps:

[0114] S51, filtering out all entity triples whose relationship confidence is higher than a preset threshold;

[0115] S52: Generate an incremental knowledge set for fitness guidance for the target user based on all entity triples whose relationship confidence is higher than a preset threshold;

[0116] S53, integrating the incremental knowledge set into the global knowledge graph for updating, and obtaining the entity relationship subgraph in the federated temporal knowledge graph updating process;

[0117] S54. Construct multi-dimensional dynamic constraints for target users during fitness based on the entity relationship subgraph.

[0118] In specific implementation, the following method can be used to filter out all entity triples whose relationship confidence is higher than a preset threshold, that is: the entity triples whose relationship confidence is greater than a preset confidence threshold (such as 0.8) among all entity triples can be regarded as entity triples whose relationship confidence is higher than the preset threshold, thereby obtaining all entity triples whose relationship confidence is higher than the preset threshold.

[0119] In specific implementation, the generation of an incremental knowledge set for fitness guidance for the target user based on all entity triples with a relationship confidence higher than a preset threshold can be achieved in the following manner, namely: an existing rule engine (such as Drools) can be used to define conflict resolution rules and detect whether there is semantic repetition or conflict (for example, "squat → knee pain" and "squat → knee discomfort") in all entity triples with a relationship confidence higher than a preset threshold to eliminate redundant information. Then, an exponential decay function is used to assign a timeliness weight to each corresponding triple according to the generation time of each entity triple after removing the redundancy, thereby constructing a deduplicated and timeliness-weighted incremental knowledge set, thereby obtaining an incremental knowledge set for fitness guidance for the target user; wherein, the incremental knowledge set is a knowledge set that knows that the target user performs fitness, and the incremental knowledge set can be used to dynamically enrich the global knowledge graph and improve the accuracy and effectiveness of personalized fitness guidance.

[0120] It should be noted that the entity-relationship subgraph is a subset graph that describes the entities and their mutual relationships in the target user's personal information. The entity-relationship subgraph can be used to focus on the target user's fitness behavior and feedback information, thereby providing personalized and accurate fitness guidance services. In specific implementation, the incremental knowledge set is integrated into the global knowledge graph for updating, and the entity-relationship subgraph obtained in the updating process of the federated temporal knowledge graph can be implemented in the following way, namely: through a federated learning framework (such as FedAvg or FedProx), the triple embedded data in the incremental knowledge set can be encrypted (for example, using Paillier homomorphic encryption or differential privacy mechanism) and securely uploaded to the central server of the fitness guidance system, and the updated data of the target user is integrated through a secure aggregation algorithm, and the merged incremental knowledge is structurally optimized in combination with a graph convolutional network (such as T-GCN) to update the global knowledge graph, and the federated temporal knowledge graph subgraph containing the latest fitness information of the target user is extracted from it. As the entity relationship subgraph in the updating process of the federated temporal knowledge graph; among them, the federated temporal knowledge graph is a distributed knowledge representation framework that combines the temporal knowledge graph with federated learning. It collaboratively constructs and updates the knowledge graph containing time information (such as timestamps or time intervals) under the premise of ensuring the data privacy of each participating node. The temporal knowledge graph adds a time dimension to the relationship on the basis of the entities and relationships of the traditional knowledge graph, and records the temporal evolution of facts in the form of a quadruple (head entity, relationship, tail entity, timestamp), which can express the timeliness and dynamic changes of events. The federated knowledge graph is a knowledge graph owned by different clients locally. The model parameters or embedded vectors are shared through the federated learning algorithm without directly exchanging the original data, so as to realize the collaborative training and fusion of the distributed knowledge graph, protect data privacy and autonomy, and realize the real-time update and privacy protection of the knowledge graph under multi-user collaboration through federated security aggregation and graph structure optimization.

[0121] Example 2

[0122] This embodiment differs from the first embodiment in that, specifically, in S54, the method for constructing multi-dimensional dynamic constraints for the target user's fitness based on the entity relationship subgraph includes the following steps:

[0123] S541. Determine the security constraints of the target user during the fitness process based on the entity relationship subgraph;

[0124] S542, performing performance prediction on the entity relationship subgraph, thereby generating enhanced constraints on the target user's fitness performance;

[0125] S543. Conflicts between the safety constraint and the enhancement constraint are resolved through a multi-objective optimization algorithm to obtain multi-dimensional dynamic constraints for the target user during fitness.

[0126] In specific implementation, the safety constraints of the target user during fitness training can be determined based on the entity relationship subgraph in the following manner: first, the target user's biometric characteristics (such as weight, joint health score and historical sports injury record) can be collected through the user information acquisition module of the fitness guidance system. Then, the "action→risk" relationship in the entity relationship subgraph and the target user's biometric characteristics are input into an existing biomechanical simulation tool (such as OpenSim) to perform force analysis on the specific action (such as squat) in the entity relationship subgraph, calculate the force conditions of the target user's motion joints and generate personalized safety thresholds (such as knee joints). The risk confidence level is updated using the biosignals in the entity relationship subgraph and the dynamic adjustment function based on time decay, and personalized safety constraints are output (such as "when the knee joint force exceeds the safety threshold, it is recommended to reduce the squat depth or reduce the training intensity"). This results in safety constraints that avoid fitness exercise risks. The safety constraints are restrictive measures set during fitness to prevent target users from overtraining or improper posture leading to exercise risks. The purpose of the safety constraints is to ensure that users train within a safe range, prevent potential injuries, and thus protect their health.

[0127] In specific implementation, the performance prediction of the entity relationship subgraph and the generation of enhanced constraints for the target user's fitness performance can be achieved in the following manner, namely: first, the performance constraints set by the target user (such as the movement speed or number of sets of squat training) can be collected through the user information acquisition module of the fitness guidance system, and then the entity relationship subgraph and the performance constraints are input into a pre-trained performance prediction model (for example, based on XGBoost) to extract the target user's fitness movements and corresponding biological signal prediction training effect indicators (such as the percentage of muscle activation increase) and generate enhanced constraints for the target user's fitness movements based on the prediction results and the performance constraints set by the user's goals; wherein, the enhanced constraints are guiding measures set during the fitness process to optimize the training effect and improve the target user's physiological indicators and sports performance. The role of the enhanced constraints is to formulate personalized training strategies to promote users to achieve the best training effect by analyzing the user's biological signals and training feedback.

[0128] It should be noted that the multidimensional dynamic constraint is a guiding measure that comprehensively considers safety and efficiency improvement during the fitness process to ensure that the target user obtains the best training effect within a safe range. In specific implementation, the multidimensional dynamic constraint of the target user during fitness is obtained by resolving conflicts between the safety constraint and the enhancement constraint through a multi-objective optimization algorithm. The following method can be used to obtain the multidimensional dynamic constraint of the target user during fitness, namely: the existing multi-objective optimization algorithm (such as NSGA-II) can be used to solve the Pareto optimal solution between the safety constraint and the enhancement constraint, thereby resolving the conflict between the two and outputting the safety-efficacy-oriented multidimensional dynamic constraint of the target user during fitness. Among them, by resolving conflicts between the safety constraint and the enhancement constraint, the training efficiency can be optimized under the premise of ensuring user safety, avoiding insufficient training intensity due to overly strict safety restrictions and ignoring potential risks due to the pursuit of efficiency, thereby providing users with a personalized, dynamically balanced fitness guidance plan, reducing the risk of sports injuries and improving training effects and overall health levels.

[0129] Specifically, generating a training knowledge graph for providing personalized fitness guidance to target users based on the multi-dimensional dynamic constraints can be achieved by using the following steps:

[0130] Generate a constraint-driven graph structure for the target user through the multi-dimensional dynamic constraints;

[0131] The constraint-driven graph structure is graph-embedded to obtain a training knowledge graph for providing personalized fitness guidance to target users.

[0132] In specific implementation, the constraint-driven graph structure for the target user generated by the multi-dimensional dynamic constraints can be achieved in the following manner, namely: first, the multi-dimensional dynamic constraints can be converted into logical rules in the knowledge graph based on the network ontology language, SPARQL protocol and RDF query language through an existing rule engine (such as Drools), and a constraint-driven graph structure suitable for the target user is generated based on the logical rules; wherein the constraint-driven graph structure is a graph structure for guiding the target user to perform personalized fitness. The constraint-driven graph structure can ensure that only user-related information is contained in the graph for providing personalized fitness guidance, thereby improving the accuracy and personalization of the recommendation.

[0133] It should be noted that the training knowledge graph is a knowledge structure that contains multi-dimensional dynamic constraint transformation logic rules and is specifically used to guide the target user's personalized fitness. It can be used to provide users with customized fitness guidance plans. In specific implementation, the constraint-driven graph structure is embedded in the graph to obtain a training knowledge graph for personalized fitness guidance for the target user. This can be achieved in the following way: first, the entity relationship subgraph updated by the federated temporal knowledge graph can be embedded into the constraint-driven graph structure through existing graph embedding methods (such as knowledge graph embedding technology) and the entities and relationships in the entity relationship subgraph are mapped to a low-dimensional vector space to obtain an initial training graph. Then, the existing pruning technology (such as a pruning method based on graph partitioning) can be used to trim the nodes in the initial training graph according to the entity heat to obtain a training knowledge graph for personalized fitness guidance for the target user. Among them, the pruning technology can divide a large graph into multiple subgraphs for parallel processing in a distributed environment to improve processing efficiency.

[0134] Example 3

[0135] This embodiment provides the following technical solution: a system used in a fitness guidance method based on a knowledge graph, including a multimodal dynamic graph construction unit, which includes:

[0136] An acquisition module is used to obtain bioelectric signals, exercise videos, and feedback voice when the target user is doing fitness exercises;

[0137] A processing module is used to extract the time-frequency variation features of the target user's bioelectric signals during fitness, extract the spatiotemporal features of the target user's key skeletal points from the fitness video, and extract the target user's feedback semantic features from the feedback voice; to perform cross-dimensional heterogeneous feature fusion on the time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user; and to construct all entity triplets between user feedback, fitness movements, and exercise biosignals during the target user's fitness process based on the multimodal heterogeneous features, and then determine the relationship confidence between the feedback semantic entity, fitness movement entity, and biosignal entity in each entity triplet;

[0138] The execution module is used to construct multi-dimensional dynamic constraints for the target user's fitness based on all entity triples whose relationship confidence is higher than a preset threshold, and then generate a training knowledge graph for personalized fitness guidance for the target user based on the multi-dimensional dynamic constraints.

[0139] Each module in the aforementioned knowledge graph-based fitness guidance system can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0140] Example 4

[0141] This embodiment provides the following technical solutions: a computer device, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data on a multimodal dynamic graph construction method. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a multimodal dynamic graph construction method is implemented.

[0142] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0143] This embodiment also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned multimodal dynamic graph construction method embodiment are implemented.

[0144] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-mentioned multimodal dynamic graph construction method embodiment.

[0145] This embodiment provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the above-mentioned multimodal dynamic graph construction method embodiment.

[0146] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0147] In summary, the present invention determines multimodal heterogeneous features to obtain features that describe the correlation between the target user's physiological signals, exercise movements, and subjective feedback during fitness. Determining multimodal heterogeneous features helps comprehensively perceive the user's physiological condition, exercise performance, and subjective feelings by integrating the time-frequency variation characteristics of biological signals, the spatiotemporal characteristics of key skeletal points, and the semantic characteristics of user feedback, thereby forming a comprehensive understanding of the user's fitness status. Determining relationship confidence in the present invention can obtain a probabilistic indicator that measures the credibility of relationships between entities. Determining relationship confidence helps effectively assess the reliability of relationships between entities, thereby prioritizing the retention of high-confidence entity triples in the knowledge graph, ensuring the accuracy and robustness of knowledge graph construction. Determining multidimensional dynamic constraints in the present invention can obtain guidance measures that comprehensively consider safety and efficiency improvements during fitness to ensure that the target user achieves the best training results within a safe range. Determining multidimensional dynamic constraints helps enable the system to monitor the user's status in real time and dynamically adjust the fitness plan based on the target user's real-time physiological signals, exercise movements, and feedback information, generating a personalized training knowledge graph, thereby providing fitness guidance that better meets individual needs and ensures the safety and effectiveness of training.

[0148] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A fitness guidance method based on knowledge graph, characterized in that: The following steps are involved: S1. Obtain bioelectric signals, exercise videos, and feedback voice when the target user is exercising; S2. Extract the time-frequency variation features of the target user's bioelectrical signals during exercise, extract the spatiotemporal features of the target user's key skeletal points from the exercise video, and extract the target user's feedback semantic features from the feedback voice; S3. Perform cross-dimensional heterogeneous feature fusion on time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user; S4. Based on multimodal heterogeneous features, construct all entity triplets between user feedback, fitness movements, and motion biosignals during the target user's fitness process, and then determine the confidence level of the relationship between the feedback semantic entity, fitness movement entity, and biosignal entity within each entity triplet; S5. Based on all entity triples whose relationship confidence is higher than a preset threshold, construct multi-dimensional dynamic constraints for the target user's fitness, and then generate a training knowledge graph for personalized fitness guidance for the target user based on the multi-dimensional dynamic constraints.

2. The fitness guidance method based on knowledge graph according to claim 1, characterized in that: In S2, the method for extracting the spatiotemporal features of the target user's key skeleton points from the motion video includes the following steps: S21, detecting the key points of the human skeleton in each frame of the motion video to obtain a joint coordinate dataset of the key skeleton points of the target user; S22. Determine, based on the joint coordinate data set, a joint position change sequence of each key skeletal point of the target user between consecutive frames; S23. Perform spatiotemporal fitting on all joint position change sequences to obtain the spatiotemporal features of the target user's key skeletal points.

3. The fitness guidance method based on knowledge graph according to claim 1, characterized in that: In S3, the method of performing cross-dimensional heterogeneous feature fusion on the time-frequency variation features, the spatiotemporal features, and the feedback semantic features to obtain the multimodal heterogeneous features of the target user includes the following steps: S31. Feature mapping of time-frequency variation features based on convolutional neural networks to obtain low-dimensional embedding representation under physiological modalities; S32, feature encoding the spatiotemporal features to obtain a low-dimensional embedding representation under the spatiotemporal modality; S33, determine the low-dimensional embedding representation under semantic modality through fully connected neural network and feedback semantic features; S34. Through the cross-modal attention mechanism, the low-dimensional embedding representations under all modalities are fused to obtain the multimodal heterogeneous features of the target user.

4. The fitness guidance method based on knowledge graph according to claim 1, characterized in that: In S4, the method for constructing all entity triplets between user feedback, fitness movements, and motion biosignals during the fitness process of the target user based on multimodal heterogeneous features includes the following steps: S411. Determine the feedback semantic entity of the target user feedback through multimodal heterogeneous features; S412: Determine the action entity of the target user's fitness action; S413. Determine a biosignal entity of the target user's motion biosignal based on the multimodal heterogeneous features; S414: Perform spatiotemporal alignment on the feedback semantic entity, the fitness action entity, and the biosignal entity to obtain all entity triplets between the user feedback, fitness action, and motion biosignal during the fitness process of the target user.

5. The fitness guidance method based on knowledge graph according to claim 1, characterized in that: In S4, the method for determining the confidence level of the relationship between the feedback semantic entity, the fitness action entity, and the biological signal entity in each entity triplet comprises the following steps: S421. For each entity triple, determine the temporal dependency features between entities in the entity triple by using an attention mechanism; S422. Perform sequence modeling on the temporal dependency features to obtain the confidence distribution of the relationship between entities; S423. Determine the relationship confidence between entities in the entity triplet through the confidence distribution, and then obtain the relationship confidence between the feedback semantic entity, the fitness action entity and the biological signal entity in each entity triplet.

6. The fitness guidance method based on knowledge graph according to claim 1, characterized in that: In S5, the method for constructing multi-dimensional dynamic constraints for the target user's fitness based on all entity triples whose relationship confidence is higher than a preset threshold includes the following steps: S51, filtering out all entity triples whose relationship confidence is higher than a preset threshold; S52: Generate an incremental knowledge set for fitness guidance for the target user based on all entity triples whose relationship confidence is higher than a preset threshold; S53, integrating the incremental knowledge set into the global knowledge graph for updating, and obtaining the entity relationship subgraph in the federated temporal knowledge graph updating process; S54. Construct multi-dimensional dynamic constraints for target users during fitness based on the entity relationship subgraph.

7. The fitness guidance method based on knowledge graph according to claim 6, characterized in that: In S54, the method for constructing multi-dimensional dynamic constraints for target users' fitness based on the entity relationship subgraph includes the following steps: S541. Determine the security constraints of the target user during the fitness process based on the entity relationship subgraph; S542, performing performance prediction on the entity relationship subgraph, thereby generating enhanced constraints on the target user's fitness performance; S543. Conflicts between the safety constraint and the enhancement constraint are resolved through a multi-objective optimization algorithm to obtain multi-dimensional dynamic constraints for the target user during fitness.

8. A system used in a fitness guidance method based on a knowledge graph according to any one of claims 1 to 7, characterized in that: Including a multimodal dynamic graph construction unit, the multimodal dynamic graph construction unit includes: An acquisition module is used to obtain bioelectric signals, exercise videos, and feedback voice when the target user is doing fitness exercises; A processing module is used to extract the time-frequency variation features of the target user's bioelectric signals during fitness, extract the spatiotemporal features of the target user's key skeletal points from the fitness video, and extract the target user's feedback semantic features from the feedback voice; to perform cross-dimensional heterogeneous feature fusion on the time-frequency variation features, spatiotemporal features, and feedback semantic features to obtain multimodal heterogeneous features of the target user; and to construct all entity triplets between user feedback, fitness movements, and exercise biosignals during the target user's fitness process based on the multimodal heterogeneous features, and then determine the relationship confidence between the feedback semantic entity, fitness movement entity, and biosignal entity in each entity triplet; The execution module is used to construct multi-dimensional dynamic constraints for the target user's fitness based on all entity triples whose relationship confidence is higher than a preset threshold, and then generate a training knowledge graph for personalized fitness guidance for the target user based on the multi-dimensional dynamic constraints.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores code, and the processor is configured to obtain the code and execute the fitness guidance method based on knowledge graph described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a computer-readable storage medium, and when the computer program is executed by a processor, the fitness guidance method based on the knowledge graph described in any one of claims 1 to 7 is implemented.

Images