User health data monitoring method and system based on smart watch

Abstract

The invention discloses a user health data monitoring method and system based on a smart watch, and relates to the technical field of smart wearable health monitoring, and the method comprises the steps: carrying out the optimization through Hilbert transform and cross-spectral density, carrying out the iteration of a sparse coefficient through employing MCMC, carrying out the dimension reduction construction of a track matrix through employing Ortho-SSA, carrying out the screening through ReliefF and VC-SSA, and initializing a GBTN model. NAS is used for adjustment, integration is carried out through federated average and incremental learning, and the health state is monitored. According to the method, through the VIB model, cross-spectral density optimization, GMM noise modeling and MCMC sparse iteration, the dynamic coupling relation between modes is accurately captured, the robustness of data representation is improved, and through fusion of causal reasoning and federated learning, the accuracy of the health monitoring state is improved.

Description

Technical Field

[0001] The present invention relates to the field of smart wearable health monitoring technology, and in particular to a method and system for monitoring user health data based on a smart watch. Background Art

[0002] With the popularity of wearable devices, smart watches, as a portable health monitoring tool, have been widely used in the real-time collection and analysis of user health data. Early smart watches mainly relied on single-modal data for basic health monitoring and adopted simple statistical analysis methods. With the integration of multimodal sensors, smart watches can collect diversified data streams including physiological signals and environmental information. Machine learning-based technologies, such as support vector machines, random forests, and deep neural networks, have gradually been applied to health status monitoring and anomaly detection. In recent years, the introduction of federated learning and incremental learning technologies has enabled distributed devices to achieve model collaborative training while protecting user privacy.

[0003] Existing user health data monitoring methods based on smartwatches still have shortcomings. The time delay and noise interference of cross-modal data often lead to a decrease in the accuracy of feature extraction. Existing health monitoring technologies have weak modeling capabilities for causal relationships, making it difficult to accurately distinguish between causal factors and correlated factors of changes in health status, resulting in insufficient accuracy of monitoring results. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a user health data monitoring method and system based on a smart watch to solve the problem that time delay and noise interference in cross-modal data often lead to a decrease in the accuracy of feature extraction, and existing health monitoring technologies have weak modeling capabilities for causal relationships, making it difficult to accurately distinguish between causal factors and related factors of changes in health status, resulting in insufficient accuracy of monitoring results.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, the present invention provides a user health data monitoring method based on a smart watch, which includes collecting multimodal data, constructing VIB to generate an initial potential representation, optimizing through Hilbert transform and cross-spectral density, constructing an MLP model, obtaining the final potential representation through EAWD boundary optimization, introducing GMM noise modeling, using MCMC to iterate sparse coefficients, calculating cross-modal synchronization delay, and performing multimodal feature extraction, using Ortho-SSA dimensionality reduction to construct a trajectory matrix, and performing singular value decomposition, screening through ReliefF and VC-SSA, constructing a causal graph through K-Means grouping, DIM and Do-Calculus, using a CNN model to update edge weights, initializing a GBTN model, using NAS for adjustment, integrating through federated averaging and incremental learning, monitoring health status, converting monitoring results into instructions and executing them, and uploading instructions to the cloud.

[0007] As a preferred solution of the user health data monitoring method based on a smartwatch according to the present invention, wherein: the multimodal data is collected, the VIB is constructed to generate the initial potential representation, the Hilbert transform and cross-spectral density are optimized, the MLP model is constructed, and the final potential representation is obtained by EAWD boundary optimization, including: Deploy multimodal sensors in smartwatches, collect multimodal data and initialize sparse dictionaries; Use sliding windows to segment multimodal data, generate time series segments, and calculate the mean and variance to obtain the mean weight and variance weight; Construct a VIB model, input the mean and variance into the VIB model to obtain the initial potential representation; Based on the time series segments, mean, initial latent representation and variance, the instantaneous frequency distribution is obtained by fast Hilbert transform, and the cross-spectral density between modes is calculated by cross-power spectral density estimation; Sort the cross-spectral density in descending order, select the maximum cross-spectral density, and set it as the cross-spectral intensity; Construct an MLP model, input the initial latent representation into the MLP model, and obtain the mutual information estimation of health status; Based on the cross-spectral strength, health state mutual information estimation and time series segments, the EAWD decomposition boundary is initialized, optimized by Lagrange, and the optimized decomposition boundary is generated, and the final potential representation is generated.

[0008] As a preferred solution of the user health data monitoring method based on a smart watch described in the present invention, wherein: the introduction of GMM noise modeling, the use of MCMC to iterate sparse coefficients, the calculation of cross-modal synchronization delay, and the multimodal feature extraction include: The LDPC sparse sensing matrix is ​​constructed using a low-density parity check algorithm. The modal weights are calculated using statistical variance, and the modal sparsity is adjusted. Regularization is then used based on the adjustment results to generate a compressed matrix. Initialize the parameters of the GMM noise model based on the variance and modal weight; Based on the compression matrix and the adjusted LDPC matrix, combined with the initialized sparse dictionary, the MCMC is initialized to obtain the initialized sparse coefficients; Using the parameters of the GMM noise model, combined with the compression matrix, the adjustment of the LDPC matrix, the initialization of the sparse dictionary, and the initialization of the sparse coefficients, the GMM weighted likelihood is calculated, combined with the compression matrix, the adjustment of the LDPC matrix, and the initialization of the sparse dictionary, the MCMC posterior distribution is updated, and the final sparse coefficients are output; Based on the final sparse coefficients and the initialized sparse dictionary, the health state mutual information estimation is reconstructed and combined with the cross-spectral density to calculate the cross-modal synchronization delay, output the synchronized health state information estimation, and perform feature extraction to generate a multimodal feature vector.

[0009] As a preferred solution of the user health data monitoring method based on a smartwatch of the present invention, wherein: the Ortho-SSA dimension reduction is used to construct a trajectory matrix, the ReliefF and VC-SSA are used for screening, the causal graph is constructed, and the CNN model is used to update the edge weights, including: Ortho-SSA is used to reduce the dimension of the multimodal feature vector to construct a trajectory matrix, and singular value decomposition is performed to generate a standardized feature pool. The feature weight is calculated using the ReliefF algorithm. Initialize the VC-SSA particle swarm, define the particle fitness function, iterate using VC-SSA, update the particle position through velocity clamping, output the global optimal position, and set it as the screening weight threshold; Filter features whose weights are greater than the weight threshold to generate a filtered feature pool, which is then encrypted using the differential privacy Laplace mechanism to generate a private feature pool. Use K-Means clustering to group the features in the filtered feature pool and generate a matrix view; Based on DIM fusion Do-Calculus, grouping matrix and matrix view are used to build CIN, generate causal graph, input CNN model, generate updated causal graph, combine causal effects, generate dynamic causal graph embedding through Laplacian eigenmap, combine matrix view for weighted fusion, generate causal embedding matrix through 1D convolution and softmax function, and generate fused causal embedding through weighted fusion; The privacy protection pool and the fused causal embedding are feature spliced ​​and singular value decomposition is performed to generate a singular value matrix and a dimensionality reduction feature pool. The dimensionality reduction feature pool is then reconstructed and optimized through VAE to generate an optimized dimensionality reduction feature pool.

[0010] As a preferred solution of the user health data monitoring method based on a smartwatch according to the present invention, the initialization of the GBTN model, adjustment using NAS, integration through federated averaging and incremental learning, and monitoring of health status include: Based on the grouping results, a subtree is constructed, the GBTM model is initialized, the GBTN model parameters are updated using adaptive learning, the NAS optimization objective function is defined, and the subtree is optimized within the search space; The Wasserstein distance of the client distribution is calculated through optimal transmission, and aggregated through federated averaging to generate global model parameters. The user model parameters are output by combining the optimized model parameters. The user model parameters are used to monitor the health status based on the optimized dimensionality reduction feature pool.

[0011] As a preferred solution of the user health data monitoring method based on a smart watch according to the present invention, the converting of the monitoring results into instructions and executing the instructions includes: The monitoring results are converted into personalized feedback instructions and executed through smart watches.

[0012] As a preferred solution of the user health data monitoring method based on a smart watch according to the present invention, the uploading instruction to the cloud includes: The personalized feedback instructions generated by the analysis are uploaded to the cloud for storage via the MQTT protocol.

[0013] In a second aspect, the present invention provides a user health data monitoring system based on a smart watch, comprising: The collection and processing module is used to collect multimodal data, construct the VIB to generate the initial latent representation, optimize it through Hilbert transform and cross-spectral density, build the MLP model, obtain the final latent representation through EAWD boundary optimization, introduce GMM noise modeling, use MCMC to iterate sparse coefficients, calculate cross-modal synchronization delay, and perform multimodal feature extraction; The analysis and monitoring module is used to construct a trajectory matrix using Ortho-SSA dimensionality reduction and perform singular value decomposition, filter using ReliefF and VC-SSA, construct a causal graph using K-Means grouping, DIM, and Do-Calculus, update edge weights using a CNN model, initialize the GBTN model, adjust using NAS, integrate through federated averaging and incremental learning, and monitor health status; Execute the upload module, convert the monitoring results into instructions, execute them, and upload the instructions to the cloud.

[0014] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the user health data monitoring method based on a smart watch as described in the first aspect of the present invention is implemented.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the user health data monitoring method based on a smart watch as described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are as follows: the present invention optimizes through Hilbert transform and cross-spectral density, uses MCMC to iterate sparse coefficients, uses Ortho-SSA dimensionality reduction to construct a trajectory matrix, performs screening through ReliefF and VC-SSA, initializes the GBTN model, uses NAS for adjustment, integrates through federated averaging and incremental learning, and monitors the health status; accurately captures the dynamic coupling relationship between modalities, improves the robustness of data representation, and improves the accuracy of health monitoring status. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 This is a flowchart of the user health data monitoring method based on a smart watch in Example 1.

[0019] Figure 2 Schematic diagram of the user health data monitoring system based on a smart watch in Example 1.

[0020] Figure 3 This is a schematic diagram of the GBTN model optimization in the user health data monitoring method based on a smart watch in Example 1.

[0021] Figure 4 This is a schematic diagram of the construction and optimization of a causal graph in the user health data monitoring method based on a smart watch in Example 1. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Example 1, with reference to Figures 1 to 4 , which is the first embodiment of the present invention, provides a method for monitoring user health data based on a smart watch, comprising the following steps: S1. Collect multimodal data, construct VIB to generate the initial latent representation, optimize it through Hilbert transform and cross-spectral density, build an MLP model, obtain the final latent representation through EAWD boundary optimization, introduce GMM noise modeling, use MCMC to iterate sparse coefficients, calculate cross-modal synchronization delay, and perform multimodal feature extraction; Specifically, we collect multimodal data, construct VIB to generate the initial potential representation, optimize it through Hilbert transform and cross-spectral density, build the MLP model, and obtain the final potential representation through EAWD boundary optimization, including: Deploy multimodal sensors in smart watches to collect and preprocess multimodal data; The multimodal sensor includes a biosensor and an environmental sensor; The multimodal data includes heart rate, blood pressure, cortisol concentration, body surface temperature, galvanic skin response (indicating emotions and stress), and altitude data, and is normalized; Initialize sparse dictionary based on multimodal data; Use sliding windows to segment multimodal data and generate time series segments; Calculate the mean and variance based on the time series segments, and obtain the mean weight and variance weight by the variance weighting method and the cross-spectral density peak weighting method; Construct a VIB model, including input layer, encoder layer, information bottleneck constraint layer, decoder layer, and output layer; Collect the MMIMDb dataset and train the VIB model; Input the mean and variance into the VIB model to obtain the initial latent representation; Based on the time series segments, mean, initial latent representation and variance, the instantaneous frequency distribution is obtained by fast Hilbert transform; The cross-spectral density between modes is calculated by cross-power spectral density estimation based on the instantaneous frequency distribution; Sort the cross-spectral density in descending order, select the maximum cross-spectral density, and set it as the cross-spectral intensity; Build an MLP model, including input layer, hidden layer and output layer; The MLP model was trained using the UCI Wine dataset. The initial latent representation is input into the MLP model to obtain the mutual information estimation of health status; Based on the cross-spectral intensity, health state mutual information estimation and time series fragments, the EAWD decomposition boundary is initialized and optimized by Lagrange to generate the optimized decomposition boundary. The formula is: ; in, For modal The decomposition level ( represents multimodal data), is the mutual information estimation of health status, For modal In time time series, Preserve weights for dynamic information (based on cross-spectral strength), For modal In time The initial latent representation of is the conditional distribution of health status (obtained based on the MLP model), is the regularization weight (based on the variance weight), For modal The initial latent representation of Variance of (calculated based on a heuristic decomposition)

[0026] The multimodal regularization term is calculated based on the optimized decomposition boundary to generate the final potential representation. The formula is: ; in, is the final potential representation, For modal The variance weight of For modal and modal The regularization weights between (based on the cross-spectral density obtained by fast Fourier transform), is the covariance.

[0027] Multimodal data is collected through multimodal sensors, which improves the comprehensiveness of information. Normalization processing is used to reduce the influence of modal heterogeneity. The VIB model is used to compress redundant information through information bottleneck constraints, reducing computing delays. Sparse dictionary initialization is combined with sliding window segmentation to capture the dynamic characteristics of multimodal time series. Fast Hilbert transform is used to extract instantaneous frequency distribution, capture non-stationary signal characteristics, and improve frequency resolution. The cross-spectral density is used to screen the maximum interactive features and improve the correlation between modalities. Descending sorting of cross-spectral strength reduces computational complexity and reduces feature screening time. The MLP model maps the initial potential representation to the mutual information of health status. Nonlinear mapping captures complex health patterns and improves diagnostic accuracy. EAWD optimizes the decomposition boundary through Lagrange optimization, balances information retention and regularization, and the combination of regularization weights and dynamic information retention weights improves the robustness of features. The multimodal regularization term optimizes inter-modal consistency. Furthermore, GMM noise modeling is introduced, MCMC is used to iterate sparse coefficients, calculate cross-modal synchronization delay, and perform multimodal feature extraction, including: Based on the final potential representation, cross-spectral density, variance weight, and mean weight, the LDPC sparse sensing matrix is ​​constructed and compressed using the low-density parity check algorithm. The formula is: ; in, For time The compressed LDPC sparse sensing matrix, For time LDPC sparse sensing matrix; The modal weight is calculated by statistical variance and the modal sparsity is adjusted. The formula is: ; in, For modal The modal weight of For modal The adjusted modal sparsity of is a constant scaling factor (set based on experimental experience); The adjusted LDPC matrix is ​​generated based on the adjusted modal sparsity, and the formula is: ; in, For time Adjusted LDPC matrix; Regularization is applied to the adjusted LDPC matrix, and the compressed matrix is ​​generated by matrix multiplication, as follows: ; in, Adjust the LDPC matrix for regularization, For modal In time The final potential representation of For modal In time Adjusted LDPC matrix; Initialize the parameters of the GMM noise model based on the variance and modal weight; Based on the compression matrix and the adjusted LDPC matrix, combined with the initialized sparse dictionary, the MCMC is initialized to obtain the initialized sparse coefficients; Using the parameters of the GMM noise model, combined with the compression matrix, adjustment of the LDPC matrix, initialization of the sparse dictionary, and initialization of the sparse coefficients, the GMM weighted likelihood is calculated as follows: ; in, is the GMM weighted likelihood factor, For time Iteration The sparse coefficients of (based on the sparse coefficient dimension setting), To initialize the sparse dictionary, 、 as well as is the first parameters (the parameters are weight, mean and variance) is the compression matrix; Based on the CMM weighted likelihood and the initialized sparse coefficients, combined with the compression matrix, adjustment of the LDPC matrix, and the initialization of the sparse dictionary, the MCMC posterior distribution is updated. When the maximum number of iterations is reached, it stops (based on the empirical rule) and outputs the final sparse coefficients. The formula is: ; in, For time Iteration The sparse coefficient of is the sparse regularization weight (based on cross-validation settings); Based on the final sparse coefficients and the initialized sparse dictionary, the health state mutual information estimation is reconstructed through multiplication operation; Based on the reconstructed health status information estimate and cross-spectral density, the cross-modal synchronization delay is calculated and the synchronized health status information estimate is output. The formula is: ; in, is the cross-modal synchronization delay, is the initial cross-modal synchronization delay (obtained by variational approximation based on the cross-spectral density); Kalman filtering is used to extract time series features of synchronized health status information estimation, and PCA is used for dimensionality reduction to generate multimodal feature vectors.

[0028] Through GMM noise modeling, complex noise is fitted by multi-Gaussian distribution, which improves the robustness of multimodal data processing. The iterative update of MCMC approximates the posterior distribution of sparse coefficients through random sampling, overcomes the local extreme value problem in high-dimensional data optimization, and significantly improves the accuracy and stability of sparse solutions. The LDPC matrix is ​​constructed by the low-density parity check algorithm, which can achieve efficient data compression with lower computational complexity and storage requirements, optimize the sparse representation of multimodal data, reduce data redundancy, and retain key feature information. The modal weight is calculated by statistical variance, and the modal sparsity is adjusted according to the formula to ensure that The contributions of homomodal data in the compression and reconstruction processes are balanced. The GMM weighted likelihood formula, combined with the compression matrix and sparse dictionary, provides high-precision likelihood estimation and enhances the stability of sparse coefficient updates. The iterative update of the MCMC posterior distribution balances computational efficiency and the global optimality of the solution through the empirical setting of the maximum number of iterations. The cross-modal synchronization delay is calculated based on the cross-spectral density, which solves the problem of time asynchrony of multimodal data and improves the accuracy of health status information estimation. Kalman filtering combined with PCA dimensionality reduction further extracts time series features, reduces the dimension of the feature vector, and improves computational efficiency and model generalization ability.

[0029] S2. Use Ortho-SSA to reduce dimensionality and construct the trajectory matrix, perform singular value decomposition, filter it with ReliefF and VC-SSA, construct a causal graph with K-Means grouping, DIM, and Do-Calculus, use the CNN model to update edge weights, initialize the GBTN model, use NAS for adjustment, integrate it with federated averaging and incremental learning, and monitor health status; Specifically, Ortho-SSA is used to reduce the dimension and construct the trajectory matrix, and singular value decomposition is performed. ReliefF and VC-SSA are used for screening. K-Means grouping, DIM and Do-Calculus are used to construct a causal graph. The CNN model is used to update the edge weights, including: The multimodal feature vectors are standardized, and the trajectory matrix is ​​constructed using Ortho-SSA dimensionality reduction. Singular value decomposition is performed, and the singular values ​​are sorted in descending order. The top K values ​​(based on the inflection point method) are selected, where K is the number of singular values, to generate a standardized feature pool. Based on the standardized feature pool, the ReliefF algorithm is used to calculate the feature weights; Initialize the VC-SSA particle swarm (each particle represents an initial candidate weight threshold (based on random generation combined with variational constraint setting)), define the particle fitness function based on the weighted sum of the information entropy of the standardized feature pool and the dimensionality constraint, and the formula is: ; in, For particles Initial candidate weight threshold The fitness function value of Information entropy The constant coefficient of (optimal feature selection based on information theory), is the information entropy (calculated based on the Shannon entropy formula), is the standardized feature pool, Index feature set A constant coefficient (based on grid search settings) for filtering feature indices whose feature weights are greater than the initial candidate weight threshold. Use VC-SSA to iterate, update the particle position through velocity clamping, stop when the maximum number of iterations is reached (based on the early stopping mechanism), output the global optimal position, and set it as the screening weight threshold; Filter features whose weights are greater than the weight threshold (based on the VC-SSA algorithm setting) to generate a screening feature pool; The screening feature pool is encrypted through the differential privacy Laplace mechanism to generate a private feature pool; Use K-Means clustering to group the features in the filtered feature pool, generate a grouping matrix, and convert it into a matrix view through matrix recombination technology; Based on DIM (Dynamic Interaction Module), the matrix view is passed through the RBF kernel to generate the interaction matrix; Do-Calculus (causal inference method) derives causal effects based on the statistical distribution of the interaction matrix and the grouping matrix, defines them as edge weights, defines features as nodes, and defines causal relationships (based on the causal relationship matrix) as edges. It constructs CIN and aggregates them to generate a causal relationship matrix. Build a CNN model, including input layer, convolution layer, pooling layer, fully connected layer and output layer; Collect the ImageNet image dataset and train the CNN model; Input CIN into the CNN model, update the edge weights, and generate an updated causal graph; Based on the updated causal graph and combined with the causal effects, a dynamic causal graph embedding is generated through Laplace eigenmap; Perform weighted fusion of dynamic causal embedding and matrix view to generate fused matrix view; Extract the fusion matrix view features through 1D convolution, calculate the sparse causal attention through the softmax function, and generate the causal embedding matrix; Based on the causal embedding matrix, it is transformed by linear change, the attention weight is calculated by Softmax normalization, and the fused causal embedding is generated by weighted fusion; The privacy protection pool and the fused causal embedding are concatenated and subjected to singular value decomposition to generate a singular value matrix and a reduced-dimensional feature pool. Add orthogonal constraints to the dimensionality reduction feature pool, sort the singular values ​​in descending order, filter the first V singular values ​​(based on the cumulative variance contribution rate), where V is the number of singular values, and reconstruct the dimensionality reduction feature pool. The formula is: ; in, To reconstruct the dimensionality reduction feature pool, is the front of the left singular value matrix List, is the front of the right singular value matrix Transpose of rows ; Use VAE to optimize the reconstructed dimensionality reduction feature pool to generate an optimized dimensionality reduction feature pool.

[0030] Through orthogonal constraints and singular value decomposition, Ortho-SSA can effectively reduce the high dimensionality of multimodal feature vectors, retain the main information of the data, reduce redundancy and noise, and screen the first K singular values ​​based on the inflection point method to ensure the representativeness and robustness of the feature pool after dimensionality reduction. The ReliefF algorithm screens out the features that contribute most to the task through feature weight calculation, reduces the dimension and improves the model performance at the same time. VC-SSA dynamically adjusts the feature screening threshold through particle swarm optimization and variational constraints to enhance the stability and global optimality of feature selection. The weighted sum of information entropy and dimensional constraints further optimizes the feature selection process to ensure that the screening results take into account both the amount of information and dimensional constraints. The screened feature pool is encrypted through the Laplace mechanism to generate a privacy feature pool, which effectively protects data privacy and prevents the leakage of sensitive information while maintaining the characteristics. The statistical characteristics of K-Means clustering are used to group features into matrix views. DIM and Do-Calculus are combined to construct CIN, generate a causal relationship matrix, accurately capture the causal effects between features, and update the edge weights based on the causal effect. The CNN model is used to further optimize the CIN structure, improve the explanatory power and monitoring accuracy of the causal relationship, and dynamically update the edge weights through ImageNet pre-training and CIN input to generate an updated causal graph, thereby improving the dynamic adaptability of the causal relationship. Laplace eigenmaps and sparse causality are used to reconstruct the dimensionality reduction feature pool through orthogonal constraints and singular value decomposition to ensure the orthogonality and information integrity of the feature pool. VAE further optimizes the feature pool and enhances the feature generation ability and robustness. The attention mechanism further enhances the expression ability of causal embedding and reduces computational complexity.

[0031] Furthermore, the GBTN model is initialized, adjusted using NAS, integrated through federated averaging and incremental learning, and the health status is monitored, including: Based on the grouping results, a subtree is constructed using a greedy splitting algorithm. Nodes are split based on features and the algorithm stops when the maximum depth is reached. The final node is output and set as a leaf node to initialize the GBTM model. Collect samples (the samples refer to collecting historical multimodal data to generate a historical optimized dimensionality reduction feature pool). Define the GBTM optimization loss function based on the subtree GBTM and the optimized dimensionality reduction feature pool. The formula is: ; in, To optimize the total loss value of the loss function, is the number of samples, To optimize the loss function value, is the sample feature The real health status of the smartwatch (the real health status is obtained from the history of the smartwatch through the API interface), is the sample feature Initial monitoring health status (referring to the sample obtained through GBTM monitoring), is the regularization coefficient (based on L2 regularization setting), is the number of subtrees, is the subtree complexity (based on the depth limit of the tree), 、 as well as Subtrees , sample characteristics And sample characteristics The prediction function in the CBTM model is is the causal regularization coefficient (based on causal inference settings), For causal effect, is the cross regularization term (based on L1 regularization setting); The prediction function in the CBTM model is as follows: ; ; in, is an exponential function, For a single sample feature pair The sum of the predicted contributions of For subtree For samples The predicted contribution value of is the index of the leaf node, is a leaf node, Leaf nodes In the subtree The weight of (calculated based on Newton's method); Use adaptive learning to update GBTN model parameters; Define the search space as the tree depth and the number of split points, define the NAS optimization objective function, and optimize the subtree in the search space. The formula is: ; in, Optimize the objective function value for NAS, For samples The accuracy rate (the ratio of the number of samples to the number of true health states), is the complexity weight (based on Pareto optimization setting), is the model complexity (based on MAC inference delay); The Wasserstein distance of client distribution is calculated through optimal transmission, and aggregated through federated averaging. The aggregated results and optimized model parameters are weighted and fused to generate global model parameters. Based on the global model parameters and the optimized model parameters, they are integrated through weighted calculation and updated through incremental learning. When the maximum number of updates is reached (based on the computing resource limit setting), the user model parameters are output; The user model parameters are used to monitor the health status based on the optimized dimensionality reduction feature pool (the model output values ​​are set to 1, 0, and -1, corresponding to health, sub-health, and abnormality, respectively).

[0032] Subtree construction enables independent modeling of the heterogeneity of different features, enhancing the model's adaptability to complex health data. Initializing the GBTM model provides a stable infrastructure for subsequent optimization and reduces the risk of overfitting. Optimizing the dimensionality reduction feature pool reduces data redundancy and improves feature expression efficiency through dimensionality reduction techniques. The GBTM optimization loss function comprehensively considers monitoring error, regularization terms, causal effects, and cross-regularization terms, ensuring monitoring accuracy while controlling model complexity and enhancing causal interpretation capabilities. The regularization coefficient and causal regularization term effectively balance the model's generalization and causal reasoning capabilities. Adaptive learning accelerates model convergence and improves adaptability to dynamic data by dynamically adjusting the learning rate and parameter update strategy. NAS optimization generates an efficient subtree structure by balancing monitoring accuracy and model complexity (based on MAC inference latency), enabling high-performance monitoring on resource-constrained devices while reducing energy consumption and latency. Wasserstein distance quantifies differences in client data distribution, optimizes the federated average aggregation process, and generates a more robust global model. Federated averaging protects user privacy and avoids the risks of centralized data storage, while improving the model's adaptability to heterogeneous data. Weighted fusion combines global and local model parameters to enhance the model's personalized monitoring capabilities. Incremental learning reduces dependence on historical data through dynamic updates and adapts to changes in user health status. By setting output values, the model can intuitively reflect the user's health status, facilitating user understanding and medical intervention.

[0033] S3, convert the monitoring results into instructions and execute them, uploading the instructions to the cloud; Specifically, the monitoring results are converted into instructions and executed, including: The monitoring results are converted into personalized feedback instructions through natural language generation (for example, if healthy, it will display "keep in good condition"; if sub-healthy, it will display "moderate relaxation and meditation recommended"; if abnormal, it will display "abnormal condition, please pay attention"). The results are transmitted to the smart watch through the BLE protocol, and the smart watch receives the personalized feedback instructions and executes them.

[0034] The monitoring results are converted into personalized feedback instructions, and the algorithm accurately matches the user's health status to generate differentiated instructions, effectively improving the pertinence and practicality of health management. The execution on the smart watch screen significantly improves the user's perception and compliance with health information, reduces health risks, and promotes the user's physical and mental health.

[0035] Furthermore, the instructions are uploaded to the cloud, including: The personalized feedback instructions generated by the analysis are uploaded to the cloud for storage via the MQTT protocol.

[0036] Efficient and real-time health data upload is achieved through the MQTT protocol, ensuring stable and low-latency data transmission. Visual charts enhance data readability, helping users to intuitively grasp health trends.

[0037] This embodiment also provides a user health data monitoring system based on a smartwatch, comprising: The collection and processing module is used to collect multimodal data, construct the VIB to generate the initial latent representation, optimize it through Hilbert transform and cross-spectral density, build the MLP model, obtain the final latent representation through EAWD boundary optimization, introduce GMM noise modeling, use MCMC to iterate sparse coefficients, calculate cross-modal synchronization delay, and perform multimodal feature extraction; The analysis and monitoring module is used to construct a trajectory matrix using Ortho-SSA dimensionality reduction and perform singular value decomposition, filter using ReliefF and VC-SSA, construct a causal graph using K-Means grouping, DIM, and Do-Calculus, update edge weights using a CNN model, initialize the GBTN model, adjust using NAS, integrate through federated averaging and incremental learning, and monitor health status; Execute the upload module, convert the monitoring results into instructions, execute them, and upload the instructions to the cloud.

[0038] This embodiment also provides a computer device, which is suitable for the user health data monitoring method based on a smart watch, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the user health data monitoring method based on a smart watch proposed in the above embodiment.

[0039] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0040] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the user health data monitoring method based on a smart watch as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0041] In summary, the present invention optimizes through Hilbert transform and cross-spectral density, iterates sparse coefficients using MCMC, constructs trajectory matrices through dimensionality reduction using Ortho-SSA, performs screening through ReliefF and VC-SSA, initializes the GBTN model, uses NAS for adjustment, and integrates through federated averaging and incremental learning to monitor health status; accurately captures the dynamic coupling relationship between modalities, improves the robustness of data representation, and enhances the accuracy of health monitoring status.

[0042] 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 the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for monitoring user health data based on a smartwatch, characterized by: include, Collect multimodal data, construct VIB to generate the initial latent representation, optimize it through Hilbert transform and cross-spectral density, build an MLP model, obtain the final latent representation through EAWD boundary optimization, introduce GMM noise modeling, use MCMC to iterate sparse coefficients, calculate cross-modal synchronization delay, and perform multimodal feature extraction; Ortho-SSA is used for dimensionality reduction to construct a trajectory matrix, followed by singular value decomposition. ReliefF and VC-SSA are used for filtering. A causal graph is constructed using K-Means grouping, DIM, and Do-Calculus. A CNN model is used to update edge weights, the GBTN model is initialized, and NAS is used for adjustment. Federated averaging and incremental learning are used for integration to monitor health status. Convert monitoring results into instructions, execute them, and upload them to the cloud.

2. The method for monitoring user health data based on a smartwatch according to claim 1, wherein: The method of collecting multimodal data, constructing VIB to generate initial potential representation, optimizing through Hilbert transform and cross-spectral density, building MLP model, and obtaining final potential representation through EAWD boundary optimization includes: Deploy multimodal sensors in smartwatches, collect multimodal data and initialize sparse dictionaries; Use sliding windows to segment multimodal data, generate time series segments, and calculate the mean and variance to obtain mean weight and variance weight; Construct a VIB model, input the mean and variance into the VIB model to obtain the initial potential representation; Based on the time series segments, mean, initial latent representation and variance, the instantaneous frequency distribution is obtained by fast Hilbert transform, and the cross-spectral density between modes is calculated by cross-power spectral density estimation; Sort the cross-spectral density in descending order, select the maximum cross-spectral density, and set it as the cross-spectral intensity; Construct an MLP model, input the initial latent representation into the MLP model, and obtain the mutual information estimation of health status; Based on the cross-spectral strength, health state mutual information estimation and time series segments, the EAWD decomposition boundary is initialized, optimized by Lagrange, and the optimized decomposition boundary is generated, and the final potential representation is generated.

3. The method for monitoring user health data based on a smartwatch according to claim 2, wherein: The method introduces GMM noise modeling, uses MCMC to iterate sparse coefficients, calculates cross-modal synchronization delay, and performs multimodal feature extraction, including: The LDPC sparse sensing matrix is ​​constructed using a low-density parity check algorithm. The modal weights are calculated using statistical variance, and the modal sparsity is adjusted. Regularization is then used based on the adjustment results to generate a compressed matrix. Initialize the parameters of the GMM noise model based on the variance and modal weight; Based on the compression matrix and the adjusted LDPC matrix, combined with the initialized sparse dictionary, the MCMC is initialized to obtain the initialized sparse coefficients; Using the parameters of the GMM noise model, combined with the compression matrix, the adjustment of the LDPC matrix, the initialization of the sparse dictionary, and the initialization of the sparse coefficients, the GMM weighted likelihood is calculated, combined with the compression matrix, the adjustment of the LDPC matrix, and the initialization of the sparse dictionary, the MCMC posterior distribution is updated, and the final sparse coefficients are output; Based on the final sparse coefficients and the initialized sparse dictionary, the health state mutual information estimation is reconstructed and combined with the cross-spectral density to calculate the cross-modal synchronization delay, output the synchronized health state information estimation, and perform feature extraction to generate a multimodal feature vector.

4. The method for monitoring user health data based on a smartwatch according to claim 3, wherein: The method uses Ortho-SSA to reduce the dimension and construct the trajectory matrix, filters it through ReliefF and VC-SSA, builds the causal graph, and uses the CNN model to update the edge weights, including: Ortho-SSA is used to reduce the dimension of the multimodal feature vector to construct a trajectory matrix, and singular value decomposition is performed to generate a standardized feature pool. The feature weight is calculated using the ReliefF algorithm. Initialize the VC-SSA particle swarm, define the particle fitness function, iterate using VC-SSA, update the particle position through velocity clamping, output the global optimal position, and set it as the screening weight threshold; Filter features whose weights are greater than the weight threshold to generate a filtered feature pool, which is then encrypted using the differential privacy Laplace mechanism to generate a private feature pool. Use K-Means clustering to group the features in the filtered feature pool and generate a matrix view; Based on DIM fusion Do-Calculus, grouping matrix and matrix view are used to build CIN, generate causal graph, input CNN model, generate updated causal graph, combine causal effects, generate dynamic causal graph embedding through Laplacian eigenmap, combine matrix view for weighted fusion, generate causal embedding matrix through 1D convolution and softmax function, and generate fused causal embedding through weighted fusion; The privacy protection pool and the fused causal embedding are feature spliced ​​and singular value decomposition is performed to generate a singular value matrix and a dimensionality reduction feature pool. The dimensionality reduction feature pool is then reconstructed and optimized through VAE to generate an optimized dimensionality reduction feature pool.

5. The method for monitoring user health data based on a smart watch according to claim 4, wherein: The GBTN model is initialized, adjusted using NAS, integrated through federated averaging and incremental learning, and health status is monitored, including: Based on the grouping results, a subtree is constructed, the GBTM model is initialized, the GBTN model parameters are updated using adaptive learning, the NAS optimization objective function is defined, and the subtree is optimized within the search space; The Wasserstein distance of the client distribution is calculated through optimal transmission, and aggregated through federated averaging to generate global model parameters. The user model parameters are output by combining the optimized model parameters. The user model parameters are used to monitor the health status based on the optimized dimensionality reduction feature pool.

6. The method for monitoring user health data based on a smart watch according to claim 5, wherein: The converting the monitoring results into instructions and executing them includes: The monitoring results are converted into personalized feedback instructions and executed through smart watches.

7. The method for monitoring user health data based on a smart watch according to claim 6, wherein: The uploading instruction to the cloud includes: The personalized feedback instructions generated by the analysis are uploaded to the cloud for storage via the MQTT protocol.

8. A user health data monitoring system based on a smart watch, based on the user health data monitoring method based on a smart watch according to any one of claims 1 to 7, characterized in that: include, The collection and processing module is used to collect multimodal data, construct the VIB to generate the initial latent representation, optimize it through Hilbert transform and cross-spectral density, build the MLP model, obtain the final latent representation through EAWD boundary optimization, introduce GMM noise modeling, use MCMC to iterate sparse coefficients, calculate cross-modal synchronization delay, and perform multimodal feature extraction; The analysis and monitoring module is used to construct a trajectory matrix using Ortho-SSA dimensionality reduction and perform singular value decomposition, filter using ReliefF and VC-SSA, construct a causal graph using K-Means grouping, DIM, and Do-Calculus, update edge weights using a CNN model, initialize the GBTN model, adjust using NAS, integrate through federated averaging and incremental learning, and monitor health status; Execute the upload module, convert the monitoring results into instructions, execute them, and upload the instructions to the cloud.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the user health data monitoring method based on a smart watch are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the user health data monitoring method based on a smart watch according to any one of claims 1 to 7 are implemented.

Images