A health data fusion and syndrome differentiation processing method of a multi-modal intelligent ring

Abstract

The application discloses a health data fusion and syndrome differentiation processing method of a multi-modal intelligent ring, and belongs to the technical field of data processing based on machine learning. The method comprises the following steps: deploying and synchronously collecting multi-modal physiological data through built-in devices; extracting original pulse wave data and implementing enhancement, determining pressure parameters and collecting after secondary enhancement; constructing a graph network with a condition bias of a latent factor, and fusing a temperature distillation and a competitive sparse attention mechanism, so as to deeply fuse a pulse wave curve family and multi-modal physiological data; modeling a physical sign differentiation result as a multi-objective Markov decision process carrying a constraint and solving the problem; and summarizing an individualized conditioning execution sequence, a physiotherapy scheme and multi-modal physiological data, constructing an individualized health report and uploading the report. The application aims to construct a deep learning hybrid architecture to fuse and differentiate human health data, and realizes intelligent individualized health monitoring.

Description

Technical Field

[0001] This invention relates to the field of data processing technology based on machine learning, and in particular to a method for health data fusion and dialectical processing of a multimodal smart ring. Background Technology

[0002] With the rapid iteration of smart healthcare and home health monitoring technologies, lightweight, non-intrusive, and all-weather human health status monitoring has become the core trend in wearable device technology development. Compared with traditional smartwatches, wristbands, and other wrist-worn devices, smart rings, relying on the structural advantages of dense blood vessels in the fingertips, thin subcutaneous tissue, and short physiological signal acquisition distance, can acquire human physiological signals with a higher signal-to-noise ratio. They have a natural advantage in the monitoring of core health indicators such as static heart rate, heart rate variability, blood oxygen, and body surface temperature, and are gradually becoming the core development direction of the next generation of personal health monitoring terminals. In recent years, domestic and foreign smart ring hardware sensing, signal acquisition, and data transmission technologies have continued to break through. The industry as a whole has gone through three stages of technological development: single-parameter monitoring, multi-parameter independent acquisition, and preliminary data integration. However, there are still significant technical shortcomings in multimodal data fusion and health status diagnosis and judgment, making it difficult to meet the needs of refined, intelligent, and professional health assessment.

[0003] The existing technology still has significant technical shortcomings, including: 1. Wearable devices suffer from distortion in HRV and pulse feature extraction due to signal noise; 2. Fixed pulse-taking pressure cannot adapt to individual differences, resulting in insufficient pulse identification ability; 3. Lack of incorporation of traditional Chinese medicine theory for deep integration and dialectical analysis, resulting in weak interpretability and limited accuracy; 4. Single-objective recommendation cannot dynamically balance the contradiction between effectiveness and compliance. Summary of the Invention

[0004] This invention discloses a method for health data fusion and dialectical processing of a multimodal smart ring. By constructing an enhanced and optimized hybrid method chain, the method deeply optimizes the human health data monitored by the multimodal smart ring, thereby solving the shortcomings of the existing technology. The plan is as follows: This invention discloses a method for health data fusion and dialectical processing in a multimodal smart ring. The method includes: synchronously collecting multimodal physiological data through a built-in device; constructing a pulse wave quality enhancement mechanism and an optimal pulse pressure parameter search mechanism; extracting raw pulse wave data from the multimodal physiological data; enhancing the pulse wave quality through the pulse wave quality enhancement mechanism; and then enhancing the pulse wave data again by determining the pressure parameter based on the enhanced signal and applying pressure after collection, thereby obtaining a family of pulse wave curves; and decomposing the constitution identification into potential factors based on the dialectical knowledge of Qi, Blood, and Body Fluids in Traditional Chinese Medicine, wherein the potential factors include Qi factors, Blood factors, and Body Fluids. Factors and body fluid factors; a graph network is constructed using conditional biases of qi factors, blood factors, and body fluid factors, and temperature distillation and competitive sparsification attention mechanisms are integrated to deeply fuse pulse wave curve families and multimodal physiological data to obtain the results of physical sign differentiation; the results of physical sign differentiation are modeled as a constrained multi-objective Markov decision process, and solved using an evolutionary strategy guided by adaptive reference vectors to obtain personalized conditioning execution sequences and physiotherapy plans; the personalized conditioning execution sequences and physiotherapy plans, along with multimodal physiological data, are summarized to create a personalized health report and uploaded to a database for subscription.

[0005] Furthermore, the multimodal physiological data includes fingertip pulse wave sequences, biomimetic pulse pressure sequences, tongue and facial image data, and sleep physiological parameters.

[0006] Furthermore, the pulse wave quality enhancement mechanism includes: A progressive cooperative adversarial network is constructed using raw pulse wave data, pre-trained weights to initialize the spectrum normalization attenuation coefficient and wavelet gating factor. A progressive cooperative adversarial network is used to generate artifacts, perform wavelet-gated denoising and reconstruction, and evaluate the discriminator quality of the original pulse wave data, resulting in high-fidelity reconstructed pulse waves and point-by-point quality confidence. A joint gradient backpropagation model is constructed using adversarial loss, reconstruction fidelity loss, and consistency loss. Based on high-fidelity reconstructed pulse waves and point-by-point quality confidence, the weights and dynamic parameters of the progressive cooperative adversarial network are iteratively updated to obtain denoised pulse waves and optimized enhancement models.

[0007] Furthermore, the optimal pulse pressure parameter search mechanism includes: Using high-fidelity reconstructed pulse wave as the reference signal, a Gaussian process surrogate model is constructed and the variational posterior of the combined kernel hyperparameter is initialized to obtain the initial search model of adaptive Bayesian optimization. The posterior mean and variance of the initial search model are obtained, and the confidence parameter is calculated by incorporating the constitution differentiation confidence level. An α-conditional entropy acquisition function is constructed. By maximizing this function, the candidate pulse pressure combination for the next round is obtained. The system acquires and collects pulse wave curves from candidate pulse pressure combinations, calculates information scores by combining them with reference signals, updates kernel parameters and GP posteriors through variational inference, and outputs personalized optimal pressure and high-fidelity pulse wave curve families.

[0008] Furthermore, the decomposition of constitution identification into potential factors based on traditional Chinese medicine theory includes: The pulse wave family was embedded and weighted with quality confidence using 1D-CNN and positional encoding. Tongue and face image data were encoded using ViT and sleep physiological parameters were encoded using TCN. The modal tokens were concatenated into a unified multimodal token sequence. Based on the predefined prior associations of Qi, Blood, Body Fluids and physiological characteristics in Traditional Chinese Medicine theory, the association feature statistics are extracted from the multimodal token sequence and mapped as initial factor nodes through linear projection to construct Qi factor, Blood factor and Body Fluid factor.

[0009] Furthermore, the conditional bias construction graph network based on Qi factor, Blood factor, and Body Fluid factor includes: By using graph attention message passing and GRU iterative updates to the node embeddings of Qi, Blood, and Body Fluid factors, corresponding layer-normalized conditional biases are generated.

[0010] Furthermore, the deep fusion of pulse wave curve families and multimodal physiological data includes... By employing a cross-modal attention mechanism with learnable temperature parameter scaling and multi-head competitive orthogonal constraints, multimodal token sequences are fused to generate a deep fused CLS token representation with injected factor bias for body type classification. Construct a total loss function, which includes constitution classification cross-entropy loss, consistency loss, competition sparsification loss and factor graph auxiliary loss, and perform offline training to obtain a converged constitution dialectical model; The constitution differentiation model is used to perform online fusion inference on multimodal token sequences. The model outputs constitution type, confidence level, and quantitative values ​​of Qi factor, blood factor and body fluid factor as the constitution differentiation result.

[0011] Furthermore, the process of modeling the results of physical examination as a constrained multi-objective Markov decision process includes: The results of physical symptom identification are mapped to continuous states, and a constrained multi-objective Markov decision process is constructed, including action masking, probability transition, and health-burden dual-objective reward. The policy network is then parameterized by mask softmax to obtain the initial policy parameters.

[0012] Furthermore, the solution using the evolutionary strategy guided by the adaptive reference vector includes: Population generation and reference vector construction and processing are performed on the initial policy parameter space and the target space respectively to obtain the initial policy population, the uniform reference vector set and their initial association relationships; For individuals in the initial policy population, the discounted rewards for the two objectives of health improvement and burden are accumulated according to the action sequence with the highest probability, to obtain the multi-objective fitness vector of all individuals; The parent strategy population is subjected to targeted perturbation and gene recombination to obtain the offspring strategy population that incorporates user compliance preferences. The offspring strategy population and the parent strategy population are then merged into a merged population.

[0013] Furthermore, the evolutionary strategy guided by the adaptive reference vector for solving the problem also includes: By performing dominance stratification, density-driven parameter updates, and diversity-oriented screening on the multi-objective fitness vectors of individuals in the merged population, a new generation of strategic populations with uniform distribution and preservation of frontier diversity is obtained. The new generation of strategy populations is subjected to strategy selection and trajectory sampling to generate personalized conditioning execution sequences and physiotherapy plans.

[0014] Compared with the prior art, the present invention achieves at least one of the following beneficial effects: 1. By using a progressive collaborative adversarial generative network, the original pulse wave signal is enhanced through spectral normalization attenuation coefficient and multi-scale wavelet attention gating dynamic scaling. This effectively suppresses motion artifact interference and solves the problem of HRV and pulse feature extraction distortion caused by signal noise in existing wearable devices.

[0015] 2. Adaptive kernelized Bayesian optimization is used to perform online search by combining kernel functions and α-conditional entropy acquisition functions to personalize the MEMS pressure parameters, thereby obtaining the best matching combination of floating, middle and sinking pressures and the corresponding family of pulse wave curves that best match individual vascular characteristics. This solves the problem that the fixed pulse-taking pressure of the existing technology cannot adapt to individual differences, resulting in insufficient pulse identification ability.

[0016] 3. A deep factor graph attention network is constructed, and the factors of Qi, Blood and Body Fluid are injected into the normalized layer as conditional biases. The temperature distillation and competitive sparsification attention mechanisms are integrated to deeply fuse multimodal physical signs data, thereby achieving high-precision constitution differentiation. This solves the problem that the existing technology lacks the introduction of TCM theory constraints for deep integration and differentiation, resulting in weak interpretability and limited accuracy.

[0017] 4. The conditioning decision is modeled as a constrained multi-objective Markov decision process, and solved by an evolutionary strategy guided by an adaptive reference vector. This optimizes the balance between health improvement and user burden, generating a personalized conditioning execution sequence. This solves the problem that existing single-objective recommendations cannot dynamically balance the contradiction between effectiveness and compliance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the health data fusion and dialectical processing method for a multimodal smart ring according to the present invention. Figure 2 Peak curves of measured curves for fingertip pulse wave sequence, biomimetic pulse pressure sequence, tongue and facial image data, and sleep physiological parameters; Figure 3 This is a display image of the current tongue appearance; Figure 4 This is a diagram showing the appearance of a healthy tongue. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0024] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0026] Example 1: Please see the appendix Figure 1 A method for health data fusion and dialectical processing of a multimodal smart ring, the method comprising: S100: Synchronously collects multimodal physiological data through built-in devices; S200. Extract raw pulse wave data from multimodal physiological data, construct a collaborative iterative framework of pulse wave quality enhancement mechanism and optimal pulse pressure parameter search mechanism, use the raw pulse wave after preliminary quality enhancement as the reference signal to drive the search for optimal pulse pressure parameters, re-acquire and perform secondary enhancement under candidate pressure, feed the enhanced signal back to the optimal pulse pressure parameter search mechanism, perform alternating iterations until convergence, and obtain a family of high-fidelity pulse wave curves under optimal pressure. S300. Based on traditional Chinese medicine theory, the constitution identification is decomposed into potential factors, which include Qi factors, blood factors and body fluid factors. S400, using conditional bias construction graph networks of Qi, Blood and Body Fluid factors, integrates temperature distillation and competitive sparsification attention mechanisms to deeply fuse pulse wave curve families and multimodal physiological data to obtain the results of physical signs differentiation. S500: The results of the physical signs diagnosis are modeled as a multi-objective Markov decision process with constraints, and an evolutionary strategy guided by an adaptive reference vector is used to solve the problem, resulting in a personalized conditioning execution sequence and physiotherapy plan. S600: The personalized conditioning sequence and physiotherapy plan, as well as multimodal physiological data, are compiled into a personalized health report and uploaded to the database for subscription use.

[0027] Its core components include: 1- progressive collaborative adversarial generative network signal quality enhancement; 2- adaptive kernelized Bayesian optimization for personalized pulse pressure search; 3- multimodal constitution differentiation using deep factor graph attention networks that integrate traditional Chinese medicine theory; 4- construction of constrained multi-objective Markov decision processes and conditioning and optimization of evolutionary strategies.

[0028] In specific implementation, the multimodal physiological data includes fingertip pulse wave sequences, biomimetic pulse pressure sequences, tongue and facial image data, and sleep physiological parameters.

[0029] The acquisition of the fingertip pulse wave sequence utilizes a PPG photoelectric sensor inside the smart ring, employing four-wavelength LEDs (green 530nm, red 660nm, infrared 880nm, and near-infrared 940nm) and a high-sensitivity photodiode, with a sampling rate of 256Hz (heart rate, blood oxygen, and HRV data are all acquired using this sensor). The biomimetic pulse pressure sequence uses a MEMS pressure sensor array arranged along the digital artery, with a sensor spacing of 2mm and a measurement range of 0-300mmHg, dynamically applying pressure using a piezoelectric ceramic microactuator. Tongue and facial image data are obtained by accessing photos and facial videos uploaded by the user from the cloud. Sleep physiological parameters are synchronized to the smart ring by accessing data recorded in the user's personal information database. Simultaneously, the smart ring is equipped with a body temperature sensor (fitted against the inner side of the fingertip), a 6-axis IMU, and a GSR skin conductance sensor, providing skin temperature data, GSR data, and three-axis acceleration + angular velocity data, respectively, which are sent to the APP for monitoring via data packets. (See appendix for details.) Figure 2 Peak values ​​of measured curves for fingertip pulse wave sequence, biomimetic pulse pressure sequence, tongue and facial image data, and sleep physiological parameters.

[0030] In specific implementation, the pulse wave quality enhancement mechanism includes: A progressive cooperative adversarial network is constructed using raw pulse wave data, pre-trained weights to initialize the spectrum normalization attenuation coefficient and wavelet gating factor. A progressive cooperative adversarial network is used to generate artifacts, perform wavelet-gated denoising and reconstruction, and evaluate the discriminator quality of the original pulse wave data, resulting in high-fidelity reconstructed pulse waves and point-by-point quality confidence. A joint gradient backpropagation model is constructed using adversarial loss, reconstruction fidelity loss, and consistency loss. Based on high-fidelity reconstructed pulse waves and point-by-point quality confidence, the weights and dynamic parameters of the progressive cooperative adversarial network are iteratively updated to obtain denoised pulse waves and optimized enhancement models.

[0031] Its construction of a progressive cooperative adversarial network includes: Step 1) Construct a three-party game network including an artifact generator, a signal reconstructor, and a discriminator, and initialize two core improvement parameters, including the pre-trained weight initialization spectrum normalization attenuation coefficient and the wavelet gating factor; Step 2) Based on Step 1) the spectral normalization attenuation coefficient for initializing the pre-trained weights, when applying spectral normalization to the weights of each layer of the discriminator, the multiplication by the attenuation coefficient is calculated using the following formula: In the formula The discriminant weights of the l-th layer of discriminator D that actually participate in the forward computation; ρ is the original weight matrix of the l-th layer of the discriminator D; t ρ is the decay coefficient for the t-th training round; ρ0 is the initial value of the decay coefficient (set to 1.0); The maximum singular value of the original weight matrix of the l-th layer of the discriminator D (used for spectral normalization); Step 3) Based on Step 2), a learnable gating factor is introduced into the signal reconstructor using the wavelet gating factor. The initial value of this gating factor is dynamically updated using the following formula: In the formula The estimated signal-to-noise ratio for the current input raw pulse wave data segment; This is a fragment of the raw pulse wave data currently being processed; This is the reconstructed signal output by the signal reconstructor from the previous round; The total energy of the original signal; This represents the energy difference between the original signal and the signal reconstructed in the previous round.

[0032] Its process for artifact generation, wavelet-gated denoising and reconstruction, and discriminator quality evaluation of raw pulse wave data includes: Step 1) For artifact generation, a random noise vector is used, and a clean template signal (i.e., a pre-stored resting high-quality PPG) is used as a conditional input to the artifact generator. The artifact generator outputs the artifact signal. Step 2) For wavelet-gated denoising and reconstruction, a signal reconstructor is used to perform discrete wavelet transform on the noisy original pulse wave data and calculate the gating value to reconstruct the gating application, outputting a high-fidelity reconstructed pulse wave. Step 3) Based on Step 1) and Step 2), use a discriminator to process the raw pulse wave data, artifact signals and high-fidelity reconstructed pulse waves, and output the classification probability and point-by-point quality confidence.

[0033] The calculation of its adversarial loss, reconstruction fidelity loss, and consistency loss includes: Step 1) The formula for calculating adversarial losses is: L 对抗 =-E[log(p 真实 )]-E[log(1-p 伪影 )]-E[log(p 重建 )], where L 对抗 To combat losses; p 真实 p 伪影 and p重建 These represent the output probabilities of the discriminator for the real clean signal, the output probabilities of the discriminator for the artifact signal, and the output probabilities of the discriminator for the reconstructed signal, respectively; E is the expected value. Step 2) The formula for calculating the reconstruction fidelity loss is: In the formula L 重建 This refers to the reconstruction fidelity loss of the signal reconstructor; and λ represents the amplitude of the reconstructed pulse wave output by the signal reconstructor at time t and the amplitude of the real clean pulse wave at time t, respectively; 相关 is the weighting coefficient of the correlation coefficient penalty term (set to 0.1); Pearson is the Pearson correlation coefficient; and The reconstructed pulse wave and the real clean pulse wave are output by the signal reconstructor; Step 3) The formula for calculating consistency loss is: In the formula L 一致性 For consistency loss; γ t γ0 is the wavelet attention gating scaling factor of the actual output of the network in the current batch; γ0 is the initial scaling factor (set to 0.1); β is the decay rate constant (set to 0.5). This represents the theoretically optimal gating strength based on the signal-to-noise ratio. Step 4) Based on steps 1) to 3), the total loss function formula is: L 总 =L 对抗 +10×L 重建 +L 一致性 In the formula L 总 The total loss is used to update network weights and dynamic parameters. Adam is used to update the artifact generator, signal reconstructor, and discriminator in sequence to obtain the denoised pulse wave and the optimized enhancement model.

[0034] In specific implementation, the optimal pulse pressure parameter search mechanism includes: Using high-fidelity reconstructed pulse wave as the reference signal, a Gaussian process surrogate model is constructed and the variational posterior of the combined kernel hyperparameter is initialized to obtain the initial search model of adaptive Bayesian optimization. The posterior mean and variance of the initial search model are obtained, and the confidence parameter is calculated by incorporating the constitution differentiation confidence level. An α-conditional entropy acquisition function is constructed. By maximizing this function, the candidate pulse pressure combination for the next round is obtained. The system acquires and collects pulse wave curves from candidate pulse pressure combinations, calculates information scores by combining them with reference signals, updates kernel parameters and GP posteriors through variational inference, and outputs personalized optimal pressure and high-fidelity pulse wave curve families.

[0035] Its construction of the Gaussian process surrogate model and initialization of the combined kernel hyperparameter variational posterior includes: Step 1), Construct the formula for the combined kernel function as follows: In the formula, K(pp) 1 ) is a combined kernel function (used to measure the similarity between two pressure parameter points); p and p 1 There are two three-dimensional pressure vectors; θ1 and θ2 are the RBF kernel weight hyperparameter and the periodic kernel weight hyperparameter, respectively (initially set to 1.0 and 0.5, respectively). and These are the length scales of the RBF nucleus and the periodic nucleus, respectively; P is the periodic parameter (set to 10 mmHg). The Euclidean distance between the two pressure points; The first term, RBF kernel, captures the smooth trend; the second term, periodic kernel, depicts the periodic changes of the human pulse with pressure. Step 2) Initialize the set of variational posterior formulas for the combined kernel hyperparameters as follows: In the formula q(θ) 组合 ) is the variational posterior distribution (used to approximate the true posterior of θ); θ 组合 The kernel hyperparameter vector is a combination of θ1 and θ2; LogNormal is the log-normal distribution function. and , respectively, represent the initial mean and initial covariance of the variational posterior; I2 is the 2×2 identity matrix; ln is the natural logarithm (mapping positive real numbers to an unconstrained space for optimization); Step 3), based on Step 1) and Step 2), construct the Gaussian process surrogate model formula as follows: y(p)~GP(m(p),K(pp 1 In the formula, y(p) is the pulse wave information content score at the pressure point p (a black box function that needs to be modeled and predicted); GP is a Gaussian process; m(p) is the mean function (set as the mean of the historical information content scores to provide a reasonable global baseline); the Gaussian process surrogate model is used as the initial search model (its search domain is [20,60]×[40,100]×[80,200]mmHg).

[0036] It incorporates the confidence parameter for calculating the confidence level of constitution differentiation and constructs an α-conditional entropy acquisition function. The solution involves maximizing this function to obtain the following: Step 1) The formula for calculating the confidence parameter is: α 信心 =α min +(α max -α min )·(1-c 置信), where α 信心 α is the confidence parameter (used to determine the balance weight between exploration and exploitation in the acquisition function). max and α min These are the minimum exploration weight and the maximum exploration weight (set to 0.05 and 0.45 respectively); c 置信 The confidence level of the current physical constitution diagnosis (i.e., the point-by-point quality confidence level); Step 2), based on Step 1), construct the α-conditional entropy acquisition function formula set as follows: In the formula α ACE (p) is the value of the α-conditional entropy acquisition function at pressure point p (i.e., the comprehensive score, which determines the selection of the next pulse pressure point); σ(p) is the posterior standard deviation of GP (i.e., the posterior standard deviation of the initial search model); α CE (p) represents the standard entropy search acquisition function value; (1-α) 信心 )·α CE (p) represents the information gain utilization term; H[...] represents the entropy function; p(y|D) represents the value of y given the observed data D. 优 D represents the global optimum; E represents the observed dataset; p(y|D) To calculate the expected value of y that may be observed at pressure point p; DU{(p,y 优 To add hypothetical sampling points (p, y) to the existing observed data. 优 The extended dataset follows; the first term in the formula group is the standard entropy search, and the second term is the function obtained by improving the standard entropy search in this invention; Step 3), based on steps 1) to 2), solve for the candidate pressure formula as follows: In the formula p 候选 The candidate pulse pressure combinations for the next round are sent to the MEMS controller of the smart ring for pressure acquisition; argmax is the independent variable function that maximizes the function; S is the search space (i.e., the search domain); the ring applies three levels of pressure sequentially, each lasting 2 seconds, and records the pressure-pulse wave curves simultaneously.

[0037] It combines the reference signal to calculate the information content score, and the variational inference updates the kernel parameters and GP posterior, including: Step 1), the formula for calculating the information content score is: In the formula y 信息 The information content score for the current stress combination (as the objective function value of Yeats optimization; a higher score indicates that more dialectical information has been obtained under this stress); λ 压力 Pressure weight (values ​​[0.3, 0.35, 0.35]); η is the multi-scale sample entropy of the pulse wave under pressure level j; η is the weight of the difference term (constant 0.2). This refers to the pulse wave signal collected under pressure level j. For high-fidelity reconstruction of the pulse wave at rest; The distribution difference between the pressure pulse wave and the reference resting pulse wave (measuring the information increment of the waveform relative to the resting state at this pressure); for the multi-scale sample entropy of the pulse wave at pressure level j and the distribution difference between the pressure pulse wave and the reference resting pulse wave, the joint calculation formula set is as follows: In the formula h represents the probability of the k-th template vector appearing in the embedding dimension m; w (b) is the normalized probability of the pressure pulse wave in the b-th amplitude interval; The normalized probability of the baseline resting pulse wave in the b-th amplitude interval; For box-by-box logarithmic comparison; Step 2) Based on Step 1), calculate the variational inference update kernel hyperparameters and maximize ELBO to update q(θ). 组合 Its update formula is: In the formula L 证据 This serves as the lower bound for evidence (i.e., the optimization objective of variational inference, maximizing which is equivalent to making the variational posterior distribution approximate the true posterior). To calculate the expectation of the variational posterior distribution; The KL divergence between the variational posterior and prior; p(θ) is the log-marginal likelihood of the observed data. 组合 ) represents the prior distribution of the kernel hyperparameter; Step 3) Based on step 2), perform GP posterior update, and the update formula set is as follows: In the formula μ(p) ) is the new point p The posterior predicted mean at the point (the best estimate of the information content score for that pressure point, serving as the core basis for the data acquisition function); k This is the kernel vector between the new point and all observed points; K1 is the transpose of the kernel vector; K2 is the n×n kernel matrix between observed points. Let I be the observation noise variance; I be the n×n identity matrix; y be the observation noise variance. 信息 The information content score for the current pressure combination (obtained from the above calculation); σ 2 (p ) is the new point p The posterior prediction variance at that point (a measure of the uncertainty in the estimation of that point, serving as the core basis for the exploration term of the acquisition function); k(p ,p () represents the kernel function value of the new point itself; Step 4) Repeat the construction, solution, and variational inference of the α-conditional entropy acquisition function, updating the kernel parameters and the GP posterior stage, until the value of the α-conditional entropy acquisition function at the pressure point p is α. ACE (p) (i.e., the overall score) < 0.01, or after 15 iterations, the optimal pressure is output.

[0038] In specific implementation, the decomposition of constitution identification into potential factors based on traditional Chinese medicine theory includes: The pulse wave family was embedded and weighted with quality confidence using 1D-CNN and positional encoding. Tongue and face image data were encoded using ViT and sleep physiological parameters were encoded using TCN. The modal tokens were concatenated into a unified multimodal token sequence. Based on the predefined prior associations of Qi, Blood, Body Fluids and physiological characteristics in Traditional Chinese Medicine theory, the association feature statistics are extracted from the multimodal token sequence and mapped as initial factor nodes through linear projection to construct Qi factor, Blood factor and Body Fluid factor.

[0039] Traditional Chinese medicine theory acquires data through a data interface connected to a domain-pre-trained large language model (an "AI Cyber ​​TCM Large Model Diagnosis System"), which contains over 5 million tongue images precisely annotated by experts, integrates a multi-dimensional health profile of symptoms, signs, treatments, and plans, and can identify 75+ negative feature points and 80+ TCM constitutions. See Table 1 for some of the TCM theory knowledge. Table 1 Four encoders are constructed, including a pulse wave encoder, a tongue image encoder, a face image encoder, and a sleep sequence encoder. The pulse wave encoder uses 1D-CNN+ positional encoding to convert the family of pulse wave curves into a token sequence. The tongue image encoder uses a pre-trained ViT-B / 16 to extract 75+ tongue image features and outputs CLS tokens and patch tokens. The face image encoder uses the same ViT to extract facial color and gloss features for keyframes. The sleep sequence encoder uses a temporal convolutional network (TCN) to encode the sleep sequence.

[0040] The construction of Qi factor, Blood factor, and Body Fluid factor includes: For the initial features of Qi factor, the average amplitude of the dicrotic wave region in the pulse wave token and the weighted sum of the respiratory rate and deep sleep percentage components in the sleep token are obtained through linear projection. The feature extraction formula is as follows: In the formula is the initial eigenvector of the gas factor; Mean (diphtheria wave amplitude) is the mean amplitude of the diphtheria wave region in the pulse wave token; W 气因子 b is the linear projection weight matrix of the gas factor; 气因子Let be the bias vector for the Qi factor; for the initial features of the blood factor, the SpO2 variation coefficient, the mean facial blood color saturation, and the mean tongue color are taken and obtained through linear projection. The feature extraction formula is as follows: In the formula W represents the initial feature vector of blood factors. 血因子 b is the linear projection weight matrix of blood factors; 血因子 This is the blood factor bias vector; The average saturation of facial blood color; The average value of tongue color brightness; The mean value of tongue color along the red-green axis; The initial features of the body fluid factor are: tongue coating moisture (i.e., texture statistics), mean skin conductivity, and mean snoring energy, obtained through linear projection. The feature extraction formula is as follows: In the formula W represents the initial eigenvector of the body fluid factor. 津液 b is the linear projection weight matrix of the body fluid factor; 津液 The bias vector for the body fluid factor; This represents the average skin conductivity. The average snoring energy is used; the three factor vectors are used as graph nodes, the edges of the factor graph are fully connected, the edge type is interaction, the edge weight matrix is ​​initialized, and the factor nodes are mapped to the same dimension as the token through a linear layer.

[0041] In specific implementation, the conditional bias construction graph network based on Qi factor, Blood factor, and Body Fluid factor includes: By using graph attention message passing and GRU iterative updates to the node embeddings of Qi, Blood, and Body Fluid factors, corresponding layer-normalized conditional biases are generated.

[0042] Its conditional bias generation process includes: Step 1) For each factor node, the aggregation formula for neighbor node information is: In the formula Let N(i) be the aggregated message vector received by factor node i in the t-th iteration; N(i) is the set of neighboring nodes of node i. W represents the attention weight for node j to pass a message to node i after the (t-1)th iteration. 变换 The message transformation matrix; Let be the embedding vector of neighbor node j after the (t-1)th iteration; where the attention weight is calculated using the following formula: In the formula, softmax j To perform softmax normalization on the neighbor dimension j; W q To query the projection matrix; W kThe key projection matrix; The square root of the dimension of the attention head (as a scaling factor). Step 2) Update the node based on Step 1), and the update formula is: In the formula This is the updated embedding vector of node i after the t-th iteration (used for the final conditional bias generation). The embedding vector of neighbor node i after the (t-1)th iteration; c 全局 The global context vector is used; after KQ rounds (set to 3), the final factor node embedding is obtained. Step 3) Based on Step 1) and Step 2), the three factor embeddings are used to generate affine transformation parameters for LayerNorm through a small network, which serve as conditional biases for LayerNorm.

[0043] In specific implementation, the deep fusion of pulse wave curve families and multimodal physiological data includes By employing a cross-modal attention mechanism with learnable temperature parameter scaling and multi-head competitive orthogonal constraints, multimodal token sequences are fused to generate a deep fused CLS token representation with injected factor bias for body type classification. Construct a total loss function, which includes constitution classification cross-entropy loss, consistency loss, competition sparsification loss and factor graph auxiliary loss, and perform offline training to obtain a converged constitution dialectical model; The constitution differentiation model is used to perform online fusion inference on multimodal token sequences. The model outputs constitution type, confidence level, and quantitative values ​​of Qi factor, blood factor and body fluid factor as the constitution differentiation result.

[0044] The process of fusing multimodal token sequences includes: Step 1) Construct an L-layer improved Transformer encoder, each layer containing multi-head cross attention (MCA) and a feedforward network (FFN), and inject factor bias into the normalization layer; Step 2), Improvement 1, involves introducing a learnable temperature parameter for head-by-head scaling when calculating attention weights. The scaling formula is as follows: In the formula, Attention h Q is the output of the h-th attention head; h K h and V h These are the h-th query matrix, key matrix, and value matrix, respectively; To query the dot product similarity matrix with the key; τ h The learnable temperature parameter for the h-th attention head (set to 1.0); This is the scaling factor after temperature scaling; This is the attention score matrix after temperature distillation; Step 3) Introduce a competition loss into the weight matrix of each attention head, the formula of which is: In the formula L 竞争 For competitive sparsity loss; Let h be the query projection matrix of the h-th attention head; H is the transpose of the query projection matrix of the h-th attention head; 注意力 I represents the total number of attention heads; I is the identity matrix. Let h be the key projection matrix of the h-th attention head; Let be the transpose of the key projection matrix of the h-th attention head; The projection matrix is ​​the value of the h-th attention head; The transpose of the projection matrix of the h-th attention head; To query the projection orthogonality penalty term; The key projection orthogonality penalty term; The value is a penalty term for projection orthogonality; Step 4) For the normalization operation of each layer, select the corresponding factor bias according to the processing mode. Specifically, the pulse token uses the Qi factor bias, the face / tongue token uses the blood factor bias, the sleep / tongue coating moisture related token uses the body fluid factor bias, and the CLS token is fused with the average of the three. After passing through the L=6 layer encoder, the CLS token is output.

[0045] Its construction of the total loss function includes: Step 1) The calculation formula for the cross-entropy loss of physical constitution classification is as follows: In the formula L 体质 Cross-entropy loss (measures the difference between the model-predicted constitution probability distribution and the true label); C is the total number of constitution categories (80 types based on traditional Chinese medicine theory); y c The true label for class c; is the CLS token vector after deep fusion; log is the natural logarithm function (applied to the prediction probability); Step 2) The formula for calculating consistency loss is as follows: In the formula L 一致性 For consistency loss; KL divergence (used to measure the difference between two probability distributions); P 网络 and P 因子图 These are the probability distributions of body constitution prediction for the factor graph branch and the probability distributions of body constitution prediction for the deep network branch, respectively. , and These are the embeddings of Qi factor, Blood factor, and Body Fluid factor after the KQ round of iterations; b偏置 b is the bias vector for the factor graph classifier head; cls Here, represents the classifier head bias vector; softmax is the normalization function; and concat is the vector concatenation function. Step 3) The formula for calculating the factor plot-assisted loss is: In the formula L 辅助 The factor plot auxiliary loss is used; f is the factor type identifier; v f This is the embedding vector of the f-th class factor in the current sample; and These represent the positive and negative sample factor embeddings, respectively; cos is the cosine calculation function; ∑ f∈{气,血,津液} To calculate and sum the three factors of Qi, Blood, and Body Fluids separately; Step 4), based on steps 1) to 3), calculate the total loss function formula as follows: L 总 =L 体质 +λ1×L 一致性 +λ2×L 竞争 +λ3×L 辅助 In the formula L 总 λ1 represents the total loss function (which serves as the objective function for joint optimization and the source of backpropagation); λ2 and λ3 represent the consistency loss weight, competitive sparsification loss weight, and factor graph auxiliary loss weight, respectively; the remaining loss terms are obtained through the above calculations.

[0046] In specific implementation, the process of modeling the results of physical examination as a constrained multi-objective Markov decision process includes: The results of physical symptom identification are mapped to continuous states, and a constrained multi-objective Markov decision process is constructed, including action masking, probability transition, and health-burden dual-objective reward. The policy network is then parameterized by mask softmax to obtain the initial policy parameters.

[0047] The construction of its constrained multi-objective Markov decision process includes: Step 1) Extract latent variables of Qi factor, blood factor and body fluid factor from the physical signs diagnosis results, and concatenate them into state basis factors; encode the current time as sine and cosine components (to introduce time context); embed the action sequence executed in the last 24 hours and take the average to obtain the historical behavior vector; input all information into a fully connected network for fusion and output a continuous state vector; Step 2) Extract 50 executable actions from the conditioning knowledge base (obtained through the AI ​​Cyber ​​TCM Big Data Model Diagnosis System), including dietary recommendations, acupressure, guided exercises, and herbal conditioning categories. Each action has inherent attributes (including estimated time, execution difficulty score, applicable time window, and contraindications based on TCM rules). During decision-making, dynamically generate an action mask m based on the current state. 掩码 If the action is within the allowed time window and does not conflict with the user's allergy information or symptoms, the mask is 1, indicating that the action is optional; otherwise, the mask is 0, which forces the action to be excluded from the candidate set, thus completing the constraint construction. Step 3) The state transition is modeled using a pre-trained probabilistic neural network. Given the current state s and the selected action a, the probabilistic neural network outputs the change of the basic factor at the next moment. By superimposing the current factor and the change, a new factor estimate is obtained. The remaining state components are updated according to deterministic rules (i.e., time advances by a fixed step size, and the historical vector window slides to include the new action). Step 4) The reduction in Euclidean distance between the current factor vector and the individual's ideal factor (the standard state determined by the dialectical constitution) is used as the health improvement reward; the negative value of the action cost is defined as the user burden reward. Maximizing the user burden reward is equivalent to minimizing the time occupation and operation difficulty, thus completing the definition of the health-burden dual-goal reward. Step 5) Based on the strategies in Steps 1) to 4), parameterize the system using a stochastic policy network. After inputting the state s, the original preference value of each action is output through two layers of MLP and then combined with the action mask. Through the above construction process, a complete multi-objective Markov decision process with constraints is obtained.

[0048] In specific implementation, the evolutionary strategy guided by adaptive reference vectors for solving the problem includes: Population generation and reference vector construction and processing are performed on the initial policy parameter space and the target space respectively to obtain the initial policy population, the uniform reference vector set and their initial association relationships; For individuals in the initial policy population, the discounted rewards for the two objectives of health improvement and burden are accumulated according to the action sequence with the highest probability, to obtain the multi-objective fitness vector of all individuals; The parent strategy population is subjected to targeted perturbation and gene recombination to obtain the offspring strategy population that incorporates user compliance preferences. The offspring strategy population and the parent strategy population are then merged into a merged population.

[0049] Its population generation and reference vector construction and processing include: Step 1) Set the population size to N=100 and the target number to 2 (i.e., health improvement 1 and user burden 2). Each individual in the population represents a complete set of policy network parameter vectors (randomly sampled from an isotropic Gaussian distribution) to complete the population initialization. Step 2) In the two-dimensional target space, n reference vectors are generated using the simplex uniform design method. Each reference vector satisfies the condition that the sum is 1 and each component is non-negative. Step 3) Assign a selected pressure coefficient to each reference vector and initialize it to 1; Step 4) Randomly assign each individual to a reference vector to establish an initial association, so that each reference vector has associated individuals in the initial state.

[0050] Its maximum probability action sequence unfolding process and fitness calculation include: Step 1) Calculate the current set of valid actions and the probability distribution of the strategy, use the maximum value function to find the action with the highest probability and calculate the real-time reward update status; Step 2) Calculate the cumulative reward of multiple objectives (using two examples) as the fitness vector. The fitness vector calculation formula is as follows: In the formula, F1(θ) i ) and F2(θ i ) represent the cumulative rewards (i.e., fitness vectors) for the first and third objectives, respectively; θ i Let γ be the strategy parameter vector for the i-th individual; H is the total number of planning steps (H=48, simulating a full day of conditioning execution, 48 steps covering 24 hours); t r is the discount factor raised to the power of t; 1,t and r 2,t Let t represent the immediate health improvement reward and the immediate user burden reward, respectively; combine the fitness vectors of all individuals in the population into a single vector.

[0051] Its targeted perturbation and gene recombination treatments include: Step 1) Train a shallow network based on the user's historical compliance data (including state s, action a, and whether the action was actually performed by the user) to predict the probability that the user will perform the recommended action (the network takes the embedding of state s and action a as input and outputs the probability that the user accepts the recommendation). Step 2) Based on Step 1), a tournament selection is performed from the current population. Two individuals are randomly selected each time, and their non-dominance level and crowding degree are compared. The winner is selected as the parent (this process is repeated to select n parent individuals, each parent carrying its policy parameters and corresponding preference gradient). Step 3) Based on Step 2), simulated binary crossover (SBX) is used to generate basic parameters for offspring for each pair of parents (the crossover operation is performed dimension-wise on the k-th dimension, generating an expansion factor according to a specific distribution; the offspring's dimension is the weighted average of the two parents). This process produces n offspring individuals, inheriting the gene combinations from their parents. Step 4) Based on Step 3), perform targeted mutation on each offspring to inject user compliance preferences. Traditional mutation adds zero-mean Gaussian noise to the parameters. The improvement of this invention lies in adding noise with mean shift. The improved mutation formula set is as follows: In the formula Let be the final strategy parameter vector of the i-th offspring individual after targeted mutation; The basic parameter vector of the i-th offspring (generated by simulated binary crossover SBX); Let be the random perturbation vector of the i-th offspring; η is the preference injection strength coefficient (taken as a constant 3); Let be the preference gradient of the i-th offspring corresponding to the parent; I is the identity matrix; σ 2 The basic variance (as a covariance matrix); To conform to the distribution symbol, the mean of the mutated noise is no longer zero, but rather biased towards improving user acceptance. This allows the offspring to maintain diversity while naturally recommending conditioning actions that are easier for users to follow. Finally, the offspring strategy population is output and merged with the parent strategy population to form a merged population.

[0052] In specific implementation, the evolutionary strategy guided by adaptive reference vectors for solving the problem also includes: By performing dominance stratification, density-driven parameter updates, and diversity-oriented screening on the multi-objective fitness vectors of individuals in the merged population, a new generation of strategic populations with uniform distribution and preservation of frontier diversity is obtained. The new generation of strategy populations is subjected to strategy selection and trajectory sampling to generate personalized conditioning execution sequences and physiotherapy plans.

[0053] Its dominant stratification, density-driven parameter updates, and diversity-oriented screening include: Step 1) Perform Pareto dominance sorting on the individuals in the merged population. For any two individuals A and B, if the fitness vectors F1(A)≥F1(B) and F2(A)≥F2(B), then A is said to dominate B. Based on the dominance relationship, all individuals are stratified. The first non-dominated layer L1 includes all individuals that are not dominated by any individual. After removing L1, the second non-dominated layer L2 includes the non-dominated individuals among the remaining individuals. This process continues until all individuals are assigned. Step 2) Calculate the ideal point (i.e. the optimal value that the current population can achieve on each of the two objectives), subtract the ideal point from the fitness vector of all individuals to complete the translation, and then normalize each objective dimension to eliminate the difference in dimensions; for each normalized individual, calculate the acute angle between it and each reference vector (establish the correspondence that each reference vector is responsible for a target space region) to complete the individual-reference vector association. Step 3) Calculate the density deficit by counting the number of individuals associated with each reference vector in the current merged population. The formula is: def j =max(0,n 期望 -n j ), where def j The density deficit of the j-th reference vector (quantifying the sparsity of the target space region covered by this reference vector); max is the maximum value function; n 期望 n represents the desired number of individuals (i.e., the number of individuals ideally associated with each reference vector). j n is the number of actual individuals currently associated with the j-th reference vector; 期望 -n j The difference between the expected value and the actual value (a positive value indicates that there are no individuals in that direction; a non-positive difference is truncated to zero after maxing out). Step 4) Based on steps 1) to 3), starting from the first non-dominated layer L1, the entire layer of individuals is incorporated into the next generation population layer by layer (until the layer cannot accommodate all individuals). After the iteration is completed, a new generation of strategy population is obtained.

[0054] Its strategy selection and trajectory sampling for the new generation of strategy populations include: Step 1) The formula for calculating the user's current health urgency index is: In the formula δ 健康 For health urgency indicators; f 基础 f is the fundamental factor vector in the current state; 理想 For the individual's ideal factor (the standard state determined by dialectical constitution); set the preference weights for the two goals, including the weight for health improvement (set as w1=0.5+0.5·δ). 健康 The weight of user burden (set as w2=1-w1) is determined by the following: when the health status is good, w1 is close to 0.5 to balance the therapeutic effect and burden; when the health status is seriously deviated, w2 is close to 1.0 to prioritize the conditioning effect. Step 2) For each individual θ in the non-dominated solution set, evaluate its comprehensive value by calculating a linear weighted value function, the formula of which is: V(θ) = w1·F1(θ) + w2·F2(θ), where V(θ) is the comprehensive value score of strategy θ; F1(θ) and F2(θ) are the cumulative fitness of strategy θ on the health improvement goal and the cumulative fitness of strategy θ on the user burden goal, respectively; select the optimal individual; Step 3) Using the selected optimal strategy, perform a forward simulation starting from the current state with a step size of half an hour, for a total of 48 steps. Each step involves: calculating the probability distribution of each action based on the current state, selecting the action with the highest probability, and performing a state transition to obtain the next state. Record the action, immediate reward, and state characteristics of that step. After the simulation is completed, the complete optimal trajectory is obtained. Step 4) Compress the optimal trajectory by merging adjacent steps with the same continuous action and removing redundancy to obtain the compressed execution sequence; Step 5) Format the compressed execution sequence into a daily schedule. Each record includes: 1- Execution time window, action name (e.g., "drink chrysanthemum tea", "massage Taichong acupoint", "15 minutes of midday guidance"); 2- Triggering condition; 3- Expected effect description. At the same time, generate a summary of the physiotherapy plan based on traditional Chinese medicine theory, suggestions and precautions, and finally output a structured personalized conditioning execution sequence and physiotherapy plan.

[0055] Using rule templates, structured indicators are converted into text descriptions (e.g., transforming the HRV downward trend into "Autonomic nervous system regulation ability weakened this week compared to last week; it is recommended to increase soothing guided exercises"). All charts are rendered using mobile-optimized visualization components. Tongue image comparison charts use a reference template that displays the current tongue image and a healthy tongue image side-by-side; see the attached template for an example. Figure 3 and attached Figure 4 .

[0056] The generated physiotherapy plans are saved in PDF format and encrypted with AES-256 before being uploaded to the cloud server via HTTPS. The cloud server pushes the report to the terminal app of the linked family members according to the user's subscription relationship. Family members can only view the desensitized summary information (such as comprehensive score, risk warning or abnormal event). Complete physiological data can only be shared with separate authorization from the user.

[0057] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A method for health data fusion and dialectical processing of a multimodal smart ring, characterized in that, The methods include: Multimodal physiological data is collected synchronously through built-in devices; A pulse wave quality enhancement mechanism and an optimal pulse pressure parameter search mechanism are constructed. Raw pulse wave data are extracted from multimodal physiological data, enhanced by the pulse wave quality enhancement mechanism, and then enhanced again by the optimal pulse pressure parameter search mechanism based on the enhanced signal to determine the pressure parameter and apply pressure to collect data, resulting in a family of pulse wave curves. Based on the knowledge of Qi, Blood and Body Fluid differentiation in traditional Chinese medicine theory, the constitution identification is decomposed into potential factors, which include Qi factor, Blood factor and Body Fluid factor. By constructing a network of conditional biases for Qi, Blood, and Body Fluid factors, and integrating temperature distillation and competitive sparsification attention mechanisms, a deep fusion of pulse wave curve families and multimodal physiological data is performed to obtain the results of syndrome differentiation. The results of the physical signs diagnosis are modeled as a constrained multi-objective Markov decision process, and an evolutionary strategy guided by an adaptive reference vector is used to solve the problem, resulting in a personalized conditioning execution sequence and physiotherapy plan. Personalized treatment sequences and physiotherapy plans, along with multimodal physiological data, are compiled into personalized health reports and uploaded to a database for subscription.

2. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The multimodal physiological data includes fingertip pulse wave sequences, biomimetic pulse pressure sequences, tongue and facial image data, and sleep physiological parameters.

3. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The pulse wave quality enhancement mechanism includes: A progressive cooperative adversarial network is constructed using raw pulse wave data, pre-trained weights to initialize the spectrum normalization attenuation coefficient and wavelet gating factor. A progressive cooperative adversarial network is used to generate artifacts, perform wavelet-gated denoising and reconstruction, and evaluate the discriminator quality of the original pulse wave data, resulting in high-fidelity reconstructed pulse waves and point-by-point quality confidence. A joint gradient backpropagation model is constructed using adversarial loss, reconstruction fidelity loss, and consistency loss. Based on high-fidelity reconstructed pulse waves and point-by-point quality confidence, the weights and dynamic parameters of the progressive cooperative adversarial network are iteratively updated to obtain denoised pulse waves and optimized enhancement models.

4. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 3, characterized in that, The optimal pulse pressure parameter search mechanism includes: Using high-fidelity reconstructed pulse wave as the reference signal, a Gaussian process surrogate model is constructed and the variational posterior of the combined kernel hyperparameter is initialized to obtain the initial search model of adaptive Bayesian optimization. The posterior mean and variance of the initial search model are obtained, and the confidence parameter is calculated by incorporating the constitution differentiation confidence level. An α-conditional entropy acquisition function is constructed. By maximizing this function, the candidate pulse pressure combination for the next round is obtained. The system acquires and collects pulse wave curves from candidate pulse pressure combinations, calculates information scores by combining them with reference signals, updates kernel parameters and GP posteriors through variational inference, and outputs personalized optimal pressure and high-fidelity pulse wave curve families.

5. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The decomposition of constitution identification into potential factors based on traditional Chinese medicine theory includes: The pulse wave family was embedded and weighted with quality confidence using 1D-CNN and positional encoding. Tongue and face image data were encoded using ViT and sleep physiological parameters were encoded using TCN. The modal tokens were concatenated into a unified multimodal token sequence. Based on the predefined prior associations of Qi, Blood, Body Fluids and physiological characteristics in Traditional Chinese Medicine theory, the association feature statistics are extracted from the multimodal token sequence and mapped as initial factor nodes through linear projection to construct Qi factor, Blood factor and Body Fluid factor.

6. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The conditionally biased graph network constructed using Qi factor, blood factor, and body fluid factor includes: By using graph attention message passing and GRU iterative updates to the node embeddings of Qi, Blood, and Body Fluid factors, corresponding layer-normalized conditional biases are generated.

7. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The deep fusion of pulse wave curve families and multimodal physiological data includes: By employing a cross-modal attention mechanism with learnable temperature parameter scaling and multi-head competitive orthogonal constraints, multimodal token sequences are fused to generate a deep fused CLS token representation with injected factor bias for body type classification. Construct a total loss function, which includes constitution classification cross-entropy loss, consistency loss, competition sparsification loss and factor graph auxiliary loss, and perform offline training to obtain a converged constitution dialectical model; The constitution differentiation model is used to perform online fusion inference on multimodal token sequences. The model outputs constitution type, confidence level, and quantitative values ​​of Qi factor, blood factor and body fluid factor as the constitution differentiation result.

8. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The process of modeling the results of physical examination as a constrained multi-objective Markov decision process includes: The results of physical symptom identification are mapped to continuous states, and a constrained multi-objective Markov decision process is constructed, including action masking, probability transition, and health-burden dual-objective reward. The policy network is then parameterized by mask softmax to obtain the initial policy parameters.

9. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 1, characterized in that, The evolutionary strategy guided by adaptive reference vectors for solving the problem includes: Population generation and reference vector construction and processing are performed on the initial policy parameter space and the target space respectively to obtain the initial policy population, the uniform reference vector set and their initial association relationships; For individuals in the initial policy population, the discounted rewards for the two objectives of health improvement and burden are accumulated according to the action sequence with the highest probability, to obtain the multi-objective fitness vector of all individuals; The parent strategy population is subjected to targeted perturbation and gene recombination to obtain the offspring strategy population that incorporates user compliance preferences. The offspring strategy population and the parent strategy population are then merged into a merged population.

10. The method for health data fusion and dialectical processing of a multimodal smart ring according to claim 9, characterized in that, The evolutionary strategy guided by adaptive reference vectors for solving the problem also includes: By performing dominance stratification, density-driven parameter updates, and diversity-oriented screening on the multi-objective fitness vectors of individuals in the merged population, a new generation of strategic populations with uniform distribution and preservation of frontier diversity is obtained. The new generation of strategy populations is subjected to strategy selection and trajectory sampling to generate personalized conditioning execution sequences and physiotherapy plans.

Images