A health monitoring method integrating multimodal bioinformation

Abstract

The present application provides a health monitoring method integrating multimodal bio-information, comprising: when a user wears a wearable device for the first time, deploying a multimodal sensor to collect the user's actual bio-signals, including heart rate, blood pressure and body temperature, and recording the tightness level and position information of the wearable device; collecting the user's tightness perception of the wearable device at different position information and corresponding to different tightness levels, and associating the tightness perception with the tightness level and position information; calculating the monitoring accuracy of the wearable device according to the actual bio-signal and wearable device data, and establishing a regression analysis model in combination with the current user's tightness perception, tightness level and position information to describe the correlation between different tightness levels and position information and the monitoring accuracy of the wearable device; determining a recommended wearing position and a recommended tightness level through a regression model in combination with the minimum monitoring accuracy corresponding to the target indicator, and dynamically adjusting the recommended wearing position and the recommended tightness by real-time monitoring of user feedback information.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a health monitoring method integrating multi-modal biological information. Background Art

[0002] In the biosignal monitoring of wearable devices, the wearing position and tightness of the device have a crucial impact on the accuracy of biosignals. However, in practical applications, there is often a contradiction between the user's personal tolerance and the minimum accuracy required for the disease. In order to solve this technical problem, it is first necessary to collect a large number of users' personal tolerance data, and obtain the user's subjective feelings and physiological reactions under different wearing positions and tightness through questionnaires, physiological tests, etc. According to the characteristics of different diseases, the minimum accuracy requirements for biosignal monitoring are determined. In actual use, it may happen that the user's personal tolerance cannot meet the minimum accuracy requirements of the disease. At this time, it is necessary to improve the accuracy of biosignal collection by optimizing the wearing position, improving sensor design and other technical means while ensuring user comfort. At the same time, it is also necessary to consider the physiological differences and wearing habits of different users to adapt to the needs of different users. In addition, how to improve the accuracy of biosignals while taking into account the comfort, portability and sustainability of the device is also a problem worthy of in-depth discussion. Summary of the invention

[0003] The present invention provides a health monitoring method integrating multimodal bioinformation, which mainly includes:

[0004] When a user wears a wearable device for the first time, deploy multimodal sensors to collect the user's actual biological signals, including heart rate, blood pressure, and body temperature, and record the tightness level and location information of the wearable device;

[0005] Collecting users' tightness perception of wearable devices at different locations and at different tightness levels, and associating the tightness perception with the tightness level and location information;

[0006] The monitoring accuracy of the wearable device is calculated based on the actual biological signals and wearable device data. A regression analysis model is established based on the tightness perception, tightness level, and location information of the current user to describe the correlation between different tightness levels and location information and the monitoring accuracy of the wearable device.

[0007] Receive the target index and personal tightness threshold corresponding to the wearable device input by the user, determine the personal tightness level range by combining the personal tightness threshold and the optimal tightness level corresponding to the user's optimal tightness feeling, and evaluate the matching degree of the wearable device to the user based on the preset adjustable range of the wearable device;

[0008] Determine the minimum monitoring accuracy according to the target indicator, and determine the highest tightness level corresponding to the minimum monitoring accuracy through the regression model. If the highest tightness level is within the personal tightness level range, the user is identified as a tolerant user; if the highest tightness level is outside the personal tightness level range, the user is identified as a sensitive user;

[0009] Combined with the minimum monitoring accuracy corresponding to the target indicator, the recommended wearing position and recommended tightness level are determined through the regression model, and the recommended wearing position and recommended tightness are dynamically adjusted through real-time monitoring of user feedback information;

[0010] Apply long short-term memory neural network to analyze user personalized data, build a model to predict the accuracy of biological signals under different usage conditions, compare the model prediction results with the actual collected data, calculate the mean square error, and dynamically adjust the wearing strategy according to the error size until the error converges to an acceptable range;

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a health monitoring method integrating multimodal bio-information, which involves how to achieve accurate monitoring of user bio-signals and optimization of personalized wearing experience through integrated multimodal sensors and intelligent data processing technology. The present invention pays special attention to the influence of the contact tightness and position between the wearable device and the user's body on the monitoring accuracy, and how to adjust the device according to the user's personal feelings to achieve the best wearing effect and data collection quality.

[0013] In general, the present invention significantly improves the applicability and monitoring accuracy of wearable devices through highly personalized data processing and intelligent adjustment mechanisms, improves the user experience, and has important practical value and broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flowchart of a health monitoring method integrating multimodal bio-information. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1 In this embodiment, a health monitoring method integrating multimodal bio-information may specifically include:

[0017] Step S101, when a user wears the wearable device for the first time, deploy a multimodal sensor to collect the user's actual biological signals, including heart rate, blood pressure and body temperature, and record the tightness level and location information of the wearable device.

[0018] Obtain the user's basic personal information, including gender, age, height, weight, etc., and preliminarily estimate the user's physiological parameter range based on this information to provide a reference benchmark for the actual biological signals collected later. According to the user's body size and the structural characteristics of the wearable device, the optimal tightness level and wearing position are calculated through simulation, and the two are combined to obtain the initial sensor wearing parameters to guide the user to wear it correctly, ensure that the sensor fits tightly to the skin, and improve the accuracy of the collected signals. While the user is wearing the device, the user's movement state is detected in real time, and the relatively static time period is identified and screened. The heart rate, blood pressure, body temperature and other biological signals in this time period are collected. The characteristic parameters of the signal are extracted using time-frequency domain analysis methods such as wavelet analysis. The clustering algorithm is used for analysis to obtain the user's basic physiological indicators in a static state and establish a personal health baseline. Continuously record the changes in biological signals when users wear the device at different tightness levels and positions, use time domain analysis, frequency domain analysis and other methods to extract the statistical characteristics and spectral characteristics of the signal, calculate the mean, variance, correlation coefficient and other indicators of the characteristic parameters, and set reasonable thresholds to judge the validity and reliability of the collected data. Based on this, dynamically adjust the working parameters of the sensor, detect the tightness and offset of the wearable device in real time, and remind the user to readjust in time if the monitoring data exceeds the threshold range to ensure the quality of biological signal collection. According to the physiological characteristics of different user groups, such as the elderly, children, and patients with chronic diseases, adopt personalized multimodal sensor combination solutions, reasonably select the type, quantity and layout of sensors, take into account the comprehensiveness and comfort of monitoring, and meet the monitoring needs of different groups of people. The collected biological signal data of the user when wearing it for the first time is bound to the basic personal information, and after data desensitization, encryption and other processing, it is uploaded to the cloud platform through a secure channel, using blockchain technology to ensure the integrity and immutability of the data, strictly limit data access rights, and protect user privacy. The cloud platform compares and analyzes the received data with the health indicators in the medical big data database, evaluates the overall health level of the user, and forms a personalized health report. The user's biosignal baseline data obtained during the initial wearing phase is used as a reference standard for subsequent monitoring. The difference between the user's current biosignal and the baseline is compared in real time. Through the abnormal detection algorithm, abnormal changes in the user's physiological state are discovered in time, triggering an early warning mechanism to prompt the user to seek medical treatment or adjust their lifestyle in a timely manner, thereby achieving dynamic management of the user's health status and risk assessment.

[0019] Specifically, when obtaining basic user information, the frontal photo uploaded by the user can be analyzed through an image recognition algorithm to estimate the user's height, weight and other parameters. For example, a convolutional neural network is used to segment the image, extract the human body contour features, and then combined with a pre-trained regression model, taking height as an example, the estimated error can be controlled within ±2cm. According to the user's gender, age and estimated height and weight, the normal range of the user's physiological parameters can be preliminarily determined, such as the normal heart rate of an adult male is 60-100 times / minute, and the systolic blood pressure is 90-140mmHg. At the same time, using finite element simulation software, a three-dimensional human body model is established according to the user's body size, different tightness levels and sensor layout schemes are set, and the optimal sensor pressure distribution and wearing position are obtained through multiple iterative calculations, such as wearing the photoelectric volumetric pulse wave sensor on the radial artery of the wrist, and setting the pressure to 20-40mmHg. When the user wears the device, the user's motion data is collected through the accelerometer and gyroscope, and the user's motion state is classified using the support vector machine (SVM) algorithm. When the user is detected to be in a static state (such as sitting or lying) for at least 5 minutes, the biological signal collection program is triggered. The collected ECG, pulse wave and other signals are first pre-processed, including removing power frequency interference, baseline drift, etc., and then extracting time domain features such as heart rate, pulse wave peak interval, etc., and frequency domain features such as low frequency / high frequency power ratio (LF / HF). Through density clustering algorithm (DBSCAN) analysis, the normal physiological index range of the user in a static state is obtained. During the continuous monitoring process, the multi-scale energy distribution characteristics of the biological signal are extracted through time-frequency analysis methods such as wavelet transform, the energy ratio of different frequency bands is calculated, and the abnormal threshold (such as energy mutation exceeding 20%) is set to judge the reliability of the signal in real time. For example, 100 heart rate data samples of the user in a static state in a day are analyzed, and after removing the highest and lowest 5% abnormal values, the average heart rate is calculated to be 65 beats / minute, with a standard deviation of 3 beats / minute, which is used as the healthy baseline of the user's heart rate. The changes in the user's biological signals when wearing the device at different tightness levels and positions are continuously recorded, and the statistical characteristics and spectral characteristics of the signal are extracted by time domain analysis, frequency domain analysis and other methods, and the mean, variance, correlation coefficient and other indicators of the characteristic parameters are calculated. By setting a reasonable threshold, the validity and reliability of the collected data are judged, and the working parameters of the sensor are dynamically adjusted accordingly. The tightness and offset of the wearable device are detected in real time. If the monitoring data exceeds the threshold range, the user is reminded to readjust in time to ensure the quality of biological signal collection. For example, when the signal-to-noise ratio of the user's ECG signal is detected to be lower than 20dB, or the waveform distortion exceeds 10%, the sampling frequency and gain of the sensor are adjusted in time, and the user is prompted to check the firmness of the electrode patch.According to the physiological characteristics of different user groups, such as the elderly, children, and patients with chronic diseases, a personalized multimodal sensor combination solution is adopted, and the type, quantity, and layout of sensors are reasonably selected to take into account the comprehensiveness and comfort of monitoring and meet the monitoring needs of different groups of people. For example, for the elderly who need to focus on heart health, a chest strap sensor with integrated ECG, blood pressure, blood oxygen, and other multi-parameters can be selected to take into account measurement accuracy and wearing comfort. The collected biosignal data of the user when wearing it for the first time is bound to the basic personal information, and after data desensitization, encryption, etc., it is uploaded to the cloud platform through a secure channel. Blockchain technology is used to ensure the integrity and immutability of the data, strictly limit data access rights, and protect user privacy. For example, the user's identity information is hashed, the uploaded physiological data is encrypted using an asymmetric encryption algorithm, and a whitelist and multiple identity authentication mechanism for data access are set. The cloud platform compares and analyzes the received data with the health indicators in the medical big data database, evaluates the user's overall health level, and forms a personalized health report. For example, the user's heart rate, blood pressure, blood sugar and other data are compared with the statistical indicators of people of the same gender and age group, and the user's percentile on the corresponding indicators is given. Based on the medical expert knowledge base, the user's health risks are evaluated and warned. The baseline data of the user's biological signals obtained during the initial wearing stage is used as a reference standard for subsequent monitoring. The difference between the user's current biological signals and the baseline is compared in real time. Through the abnormal detection algorithm, abnormal changes in the user's physiological state are discovered in time, and the early warning mechanism is triggered to prompt the user to seek medical treatment or adjust his lifestyle in time, so as to achieve dynamic management of the user's health status and risk estimation. For example, when the user's average heart rate at night is monitored to be more than 20% higher than the baseline value for 3 consecutive days, the early warning information is pushed in time, and the user is advised to pay attention to rest and undergo medical examinations when necessary.

[0020] Step S102, collecting the user's tightness perception of the wearable device at different position information and different tightness levels, and associating the tightness perception with the tightness level and position information.

[0021] Through online questionnaire survey, users were subjectively evaluated for wearing comfort using the Likert scale, and the tightness was divided into five levels, corresponding to very comfortable, relatively comfortable, general, slightly uncomfortable and very uncomfortable, to obtain the subjective tightness data of users at different tightness levels. At the same time, the pressure sensor array was used to collect the pressure distribution data of the contact part between the wearable device and the skin of users at different tightness levels. Through preprocessing methods such as data normalization and feature mapping, the subjective tightness score was aligned and calibrated with the objective pressure distribution measurement results, and a quantitative relationship between the two was established. The Gaussian process regression algorithm was used to establish a quantitative relationship model between the tightness level and the pressure distribution, and the objective measurement of the tightness level was achieved. The body surface contour data of the user was obtained by using human body three-dimensional scanning technology, and combined with the three-dimensional structure model of the wearable device, the contact area and positive pressure distribution of different wearing positions were calculated through finite element simulation analysis, and the theoretical tightness level of different positions was obtained. Collect tightness feedback from users of different ages, genders, and body shapes, use feature selection algorithms to select the user attributes with the greatest correlation with tightness, build a decision tree model, select the optimal splitting attributes through indicators such as information gain, generate a tree structure, and use the post-pruning algorithm to optimize the decision tree to improve generalization ability. At the leaf nodes of the decision tree, use the Apriori algorithm to mine the association rules between different user attribute combinations and tightness feelings, and establish personalized tightness-feeling correspondence models for different user groups. Develop a tightness feedback APP. After users wear the device at different tightness levels and positions, they enter real-time subjective tightness feelings through the APP, and use the crowdsourcing mechanism to accumulate a large amount of user feedback data. Design data quality assessment and incentive feedback mechanisms, perform outlier detection, noise filtering and other pre-processing on the feedback data uploaded by users, eliminate false or invalid feedback, and improve data reliability. Give certain points to users who provide timely and accurate feedback to increase user participation. Aggregate multi-dimensional data such as user subjective tightness feelings, objective pressure distribution, theoretical tightness level, and user attributes, and use one-hot encoding, numerical normalization and other methods to represent and standardize input features. Design the LSTM network structure, including the input layer, embedding layer, LSTM layer and fully connected layer, and optimize the network's hyperparameters, such as the number of hidden layers, number of neurons, dropout rate, etc., through grid search and other methods. Use cross entropy as the loss function, use the Adam optimization algorithm for model training, and perform performance evaluation and model tuning on the validation set, and finally form an end-to-end mapping relationship from tightness level, position information to tightness perception. Based on the established tightness perception prediction model, the forward propagation algorithm is used to perform model inference for the personal attributes, expected wearing position and other information input by the new user, and the expected tightness perception distribution under different tightness levels is calculated.Combining the human body 3D model and wearable device virtual assembly technology, rendering engines such as Unity 3D are used to generate preview images of wearing effects at different tightness levels, which are intuitively displayed to users to assist them in selecting the optimal tightness level and wearing position, and guide users to wear comfortably and effectively.

[0022] Specifically, in the online questionnaire survey, a 5-point Likert scale was used to divide the tightness feeling into 1 to 5 points, corresponding to very uncomfortable to very comfortable. At the same time, the pressure sensor array was used to collect pressure distribution data at different tightness levels, and the pressure value was normalized to between 0 and 1. The Pearson correlation analysis was performed with the subjective score, and the correlation coefficient was 0.8, indicating that the two had a strong linear relationship. Using the Gaussian process regression algorithm, the tightness level was used as the independent variable and the pressure distribution characteristics were used as the covariate to establish a regression model. The hyperparameters were optimized by maximum likelihood estimation, and the regression mean square error was 0.1, with a high goodness of fit. The user's three-dimensional body surface contour data was gridded to generate a triangular patch model containing 10,000 nodes. The contact pressure distribution at different positions was calculated by finite element simulation and mapped with the tightness level to obtain the optimal tightness interval at different positions. In the process of building the decision tree model, the optimal splitting attribute, such as age and gender, was selected by the information gain ratio, and the tree structure was recursively generated, and the minimum number of samples of leaf nodes was set to 10 to prevent overfitting. The model was evaluated using 10-fold cross validation, with an average accuracy of 85%. On the leaf nodes, the Apriori algorithm was used to mine the association rules between user attribute combinations and tightness perception, with the minimum support set to 0.05 and the minimum confidence set to 0.8, to obtain the top 5 association rules, such as "female and BMI<18.5=>tightness perception is slightly uncomfortable". In the preprocessing of user feedback data, the Z-score method was used to detect outliers, and data exceeding 3 times the standard deviation was removed. The Kalman filter algorithm was used to smooth and denoise the data, and the forgetting factor was set to 0.9, and the noise estimate was dynamically updated. In the design of the LSTM network structure, the input layer dimension is 10, corresponding to features such as tightness level and position coordinates, the embedding layer dimension is 32, the LSTM layer contains 2 hidden layers, each with 64 neurons, the dropout rate is set to 0.5, and the fully connected layer outputs the tightness perception prediction score. The Adam optimizer was used, the initial learning rate was 0.001, the momentum factor was 0.9, and L2 regularization was used to prevent overfitting. The model's average absolute error on the validation set is 0.3, and the accuracy rate reaches 90%. Based on the user's personal attributes and wearing position, the model inference predicts the tightness distribution under different tightness levels, and combines the virtual wearing effect generated by Unity 3D engine rendering to select the tightness level with the highest expected tightness score and that meets the user's comfort requirements to generate a personalized wearing plan.

[0023] Step S103, calculating the monitoring accuracy of the wearable device based on the actual biological signal and the wearable device data, and establishing a regression analysis model based on the tightness perception, tightness level, and location information of the current user to describe the correlation between different tightness levels and location information and the monitoring accuracy of the wearable device.

[0024] Multimodal sensors are used to collect users' physiological signals such as ECG, pulse, and blood oxygen in real time, and to simultaneously record the acceleration, angular velocity and other motion data of wearable devices. Through data preprocessing, heterogeneous data from different sources are time-aligned and sample rate converted, and key features such as time domain and frequency domain are extracted to achieve multimodal data fusion and feature representation. Using deep learning models such as convolutional neural networks, physiological signals are evaluated for quality and anomaly detection, and noise, distortion and invalid fragments in the signals are automatically identified to improve the reliability and validity of the data. According to the signal quality evaluation results, high-quality physiological signal fragments are dynamically screened and compared with the monitoring results of reference medical equipment. The measurement errors and deviations of wearable devices on different physiological indicators are calculated to obtain a quantitative evaluation of monitoring accuracy. Questionnaires, human-computer interactions and other methods are used to collect users' subjective comfort scores at different tightness levels and wearing positions, and establish a corresponding relationship between tightness perception and user physiological parameters. Through feature selection algorithms, such as mutual information-based filtering and L1 regularization-based embedding, the correlation and importance between factors such as tightness level, location information, and user tightness perception and monitoring accuracy are evaluated, key influencing factors are screened out, and the dimension of feature space is reduced. For the selected continuous features, methods such as minimum-maximum normalization and Z-score standardization are used to scale data and transform distribution; for discrete features, methods such as unique hot encoding and serial number encoding are used to achieve quantitative representation of qualitative indicators. With monitoring accuracy as the dependent variable and key factors such as tightness level, location information, and user tightness perception as independent variables, a multivariate regression analysis model is constructed using machine learning algorithms. Before modeling, multicollinearity problems are diagnosed by calculating indicators such as the Pearson correlation coefficient matrix and variance inflation factor between features, and highly correlated redundant features are merged or eliminated. The least squares method is used to estimate the regression coefficient, and the weight and contribution of each independent variable are evaluated through hypothesis testing and significance analysis. Consider introducing second-order and third-order terms to characterize the interaction effects and nonlinear relationships between independent variables. For complex nonlinear patterns, kernel function methods such as polynomial kernel and Gaussian kernel are used to map the original features to high-dimensional space and enhance the expressive power of the model. Through cross-validation and grid search, the hyperparameters of the model, such as regularization coefficient and kernel function parameters, are optimized to improve the generalization performance and robustness of the model. Based on the trained regression model, the monitoring accuracy change trend and sensitivity under different tightness levels, location information, and user tightness perception are analyzed to predict the optimal wearing parameter combination to maximize the monitoring quality of physiological signals while ensuring user comfort. Resampling methods such as Bootstrap are used to evaluate the confidence interval and uncertainty of the model prediction, providing a statistical basis for the reliability estimation of monitoring accuracy. Through dynamic updating and online learning of the model, the parameters and structure of the regression model are continuously optimized to adapt to changes in user status and wearing environment, and to achieve adaptive adjustment and dynamic prediction of monitoring accuracy.

[0025] Specifically, when fusing multimodal sensor data, the dynamic time warping (DTW) algorithm is used to time-align physiological signals with different sampling rates such as electrocardiogram and pulse. By optimizing the path search, the mapping function with the minimum cumulative distance is found to achieve signal synchronization and registration. Time domain features such as mean, standard deviation, peak-to-peak value, etc., and frequency domain features such as power spectrum density and band energy ratio are extracted to construct a 20-dimensional feature vector. A convolutional neural network is used for signal quality assessment, and 1,000 labeled samples are used for training, including 500 high-quality signals and 500 noisy signals. Through 5-fold cross validation, the average accuracy of the model reaches 95%. When calculating the monitoring accuracy, a continuous 5-minute high-quality signal segment is selected, and the root mean square error (RMSE) analysis is performed with the measurement results of the medical reference equipment. The RMSE of indicators such as heart rate and blood oxygen saturation are 2 times / minute and 1% respectively. For the quantitative expression of tightness perception, the fuzzy comprehensive evaluation method is used to convert the five language levels from "very loose" to "very tight" into membership values ​​in the interval [0,1]. The comprehensive score of tightness perception is obtained by weighted average. In feature selection, the maximum relevance minimum redundancy (mRMR) algorithm is used. The correlation between features and monitoring accuracy, as well as the redundancy between features, are calculated through mutual information, and 10 optimal feature subsets are selected, reducing the feature dimension by 90%. When constructing the regression model, the second-order polynomial kernel function is used. Through maximum likelihood estimation, the regression coefficients of tightness level, position information, and tightness perception are 0.5, 0.3, and -0.2, respectively. Through hypothesis testing, the P values ​​are all less than 0.05, which is significant. Using 10-fold cross validation, the determination coefficient R of the model is 2 The mean square error is 0.85 and 0.1, indicating that the model has high goodness of fit and strong generalization ability.

[0026] Step S104, receiving the target indicator and personal tightness threshold corresponding to the wearable device input by the user, combining the personal tightness threshold and the optimal tightness level corresponding to the user's optimal tightness feeling, determining the personal tightness level range, and evaluating the matching degree of the wearable device to the user based on the preset adjustable range of the wearable device.

[0027] Through the human-computer interaction interface, the user's functional requirements for wearable devices and expected physiological monitoring indicators, such as heart rate, blood oxygen, respiratory rate, etc., are obtained. At the same time, the user is allowed to customize the comfort range and input the acceptable minimum and maximum tightness values ​​to form a personal tightness threshold. The fuzzy comprehensive evaluation method is used to construct a multi-level fuzzy evaluation matrix with tightness perception as the target layer, physiological parameters, physical pressure, wearing position, etc. as the criterion layer, and different tightness levels as the scheme layer. The weight coefficient and membership function of each evaluation indicator are determined through the expert knowledge base and user survey data. The weighted average method is used to perform multi-level fuzzy synthesis operations to obtain the quantitative mapping relationship between the user's subjective tightness perception and the objective tightness level. The k-means clustering algorithm is used to analyze the tightness perception feedback data of existing users. The optimal cluster number k is determined by the elbow rule and silhouette coefficient evaluation. The Euclidean distance is used to measure the similarity between the tightness perception feature vectors, and the cluster center is iteratively optimized to obtain the optimal tightness perception mode for different groups of people. The Apriori association rule mining algorithm is adopted, with the crowd clustering label as the consequent and individual attributes such as age, gender, and body shape as the antecedent. The minimum support and confidence thresholds are set to extract significant association rules and form a recommended mapping between user portraits and the optimal tightness level. When determining the personalized tightness level, a weighted voting decision fusion strategy is adopted. For the user-defined tightness threshold, a higher weight coefficient is assigned to reflect the priority of the user's subjective needs; for the optimal tightness level recommended based on the crowd portrait, the second highest weight coefficient is assigned to take into account the optimization of objective monitoring performance. By setting the weighted weight vector and threshold, the two tightness levels are weightedly fused to obtain the user's personalized tightness level range. By consulting patent literature and consulting material suppliers, the mechanical performance parameters of materials such as TPU and silicone rubber used in wearable devices, such as Young's modulus and Poisson's ratio, are obtained. Combined with the structural parameters in the product design drawings, such as bandwidth and thickness, a tightness level prediction model based on material mechanics is established. Finite element simulation method is used to numerically simulate the stress and deformation of wearable devices under different tightness levels, analyze the stress and strain distribution characteristics, and evaluate whether the mechanical properties of the material meet the user's tightness level requirements. A matching evaluation function is designed, and the user's personal tightness threshold, optimal tightness level, and material performance and adjustable range of the wearable device are comprehensively considered to construct a multi-criteria attribute decision matrix. The weighted normalized TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution) method is used to calculate the Euclidean distance between each alternative wearable device and the ideal optimal solution and the negative ideal solution, and the closeness coefficient is obtained as a quantitative indicator of matching. The higher the matching degree, the higher the fit between the wearable device and the user's needs.At the same time, a user feedback mechanism is introduced to collect users' satisfaction scores on the recommendation results, and a collaborative filtering algorithm is used to update and optimize the matching evaluation model online.

[0028] Specifically, in the human-computer interaction interface, through controls such as sliders and digital input boxes, users are allowed to input target indicators such as the expected heart rate monitoring accuracy (such as ±5 beats / minute), blood oxygen saturation error (such as ±2%), and acceptable tightness thresholds (such as 0.1-0.5 Newtons). In the fuzzy comprehensive evaluation process, through expert scoring and user questionnaires, a fuzzy evaluation matrix containing 10 criterion layer indicators and 5 tightness levels is constructed. The triangular membership function is used to calculate the comprehensive evaluation value of tightness perception through the weighted average operator, and the Pearson correlation analysis is performed with the measured physiological parameters, and the correlation coefficient r=0.8 is obtained, indicating that the subjective and objective evaluation results have good consistency. In cluster analysis, the tightness perception data of 1,000 users were normalized by min-max, and the optimal number of clusters k=4 was determined by the average silhouette coefficient and Calinski-Harabaz index. The k-means++ algorithm was used to select the initial cluster center, and 4 user groups were obtained after 20 iterations. The Apriori algorithm was used for association rule mining, and the minimum support was set to 0.1 and the minimum confidence was set to 0.8. Ten significant association rules were obtained, such as "age>60 years old and BMI<18.5→the best tightness level is 2". When determining the personalized tightness level, the user-defined threshold weight was set to 0.6, and the crowd-recommended tightness weight was set to 0.4. The user's tightness level range was obtained by weighted average [2,4]. In material mechanics modeling, the strain distribution cloud map of TPU material under 0.1-0.5 Newton pressure was obtained through finite element simulation analysis. The strain peak was less than 5%, which met the user's tightness level requirements. When evaluating the matching degree, we selected five alternative sports bracelets, constructed a weighted normalized matrix from six dimensions, including performance, comfort, and appearance, and calculated the Euclidean distance between each candidate bracelet and the ideal solution through the TOPSIS algorithm to obtain the matching coefficient. The highest matching degree of bracelet A was 0.85, and it was recommended to users for trial wearing and purchase. At the same time, we collected users' 5-star satisfaction ratings on the recommended results as explicit feedback data for the collaborative filtering algorithm, and updated the matching evaluation model online to improve the accuracy and personalization of subsequent recommendations.

[0029] Step S105, determine the minimum monitoring accuracy according to the target indicator, and determine the maximum tightness level corresponding to the minimum monitoring accuracy through the regression model. If the maximum tightness level is within the personal tightness level range, the user is identified as a tolerant user; if the maximum tightness level is outside the personal tightness level range, the user is identified as a sensitive user.

[0030] According to the application scenarios and functional positioning of wearable devices, such as sports health monitoring, sleep quality assessment, etc., the target indicators of various physiological parameter monitoring are determined by comprehensively considering the medical significance of physiological parameters and the technical characteristics of wearable sensors. Through experiments, the physiological parameter monitoring data of different subjects under different tightness levels, materials, wearing positions and other conditions are collected, and the data are preprocessed by cleaning, outlier processing and normalization. A multivariate linear regression model is constructed with the tightness level, material characteristics, wearing position and other conditions of wearable devices as independent variables and the monitoring error of physiological parameters as the dependent variable. The stepwise regression method is used for feature selection, and the model parameters are optimized by cross-validation and regularization techniques. The regression coefficients are analyzed for significance using F test and t test to obtain the quantitative relationship between tightness level, material characteristics, wearing position and monitoring accuracy. The monitoring target indicators are substituted into the regression equation to calculate the minimum tightness level required to achieve the target accuracy. Then, the adjustment accuracy of the tightness level and the individual differences of physiological parameters are comprehensively considered. Through sensitivity analysis and Monte Carlo simulation experiments, a reasonable tightness level margin is determined, and finally the highest tightness level threshold to ensure monitoring accuracy is obtained. The user's personal attribute information such as age, gender, height and weight filled in during registration, as well as the subjective tightness score reported during the first wearing, are obtained, and the user data is subjected to feature engineering and standardization. The k-means clustering algorithm is used to group users, and the optimal number of clusters is determined by the elbow rule and the silhouette coefficient to obtain user subgroups with different tightness perception characteristics. The C4.5 decision tree algorithm is used to establish classification rules between user attributes and clustering results, and the association patterns between different combinations of attributes such as age, gender, BMI and the optimal tightness level interval are mined to form a personalized user tightness perception portrait. The highest tightness level threshold is compared with the user's personal tightness level interval. If the highest threshold falls within the personal interval, the user is marked as tolerant, indicating that the user can adapt to a higher wearing tightness to achieve better monitoring accuracy; if the highest threshold exceeds the upper limit of the personal interval, the user is identified as sensitive, indicating that the user is more sensitive to wearing pressure and needs to prioritize wearing comfort. Based on the domain expert knowledge and historical user feedback data, the tightness level deviation threshold is empirically set and iteratively optimized to dynamically adjust the boundary conditions of user classification. For tolerant users, based on their personal optimal tightness level range and actual wearing habit data, fuzzy reasoning and reinforcement learning algorithms are used to give targeted tightness adjustment suggestions, guiding users to appropriately increase the tightness while ensuring wearing comfort to meet the monitoring accuracy requirements. For sensitive users, a lower tightness wearing solution corresponding to the lower limit of their personal tightness level range is recommended, and suggestions such as wearing position selection and wearing time control to relieve pressure discomfort are given. At the same time, users are reminded that the monitoring effect may be reduced at too low a tightness, helping users to strike a balance between comfort and functionality.

[0031] Specifically, the physiological parameter monitoring data of different subjects under different tightness levels, materials, wearing positions, etc. were collected through experiments, and preprocessing operations such as data cleaning, outlier processing and normalization were performed to construct a multivariate linear regression model with the tightness level, material characteristics, wearing position, etc. of the wearable device as independent variables and the monitoring error of physiological parameters as the dependent variable. For example, 30 subjects were selected, and each of them was monitored for heart rate and blood oxygen saturation under 3 different tightness levels (loose, moderate, tight), 2 materials (nylon, silicone), and 3 wearing positions (wrist, upper arm, ankle). The measurement was repeated 3 times under each combination condition, and a total of 30×3×2×3×3=1620 sets of data were obtained. The stepwise regression method was used for feature selection, and the model parameters were optimized through 5-fold cross validation and L1 regularization technology. The F test and t test were used to perform significance analysis on the regression coefficient (significance level α=0.5), and the quantitative relationship between tightness level, material characteristics, wearing position and monitoring accuracy was obtained, such as heart rate monitoring error = 5×tightness level-2×material hardness+8×wearing position+6 blood oxygen saturation monitoring error = 1×tightness level-9×material breathability+1×wearing position+5. The monitoring target indicators were substituted into the regression equation to calculate the minimum tightness level required to achieve the target accuracy. Then, considering the adjustment accuracy of the tightness level (such as 5 intervals per level), individual differences in physiological parameters (such as standard deviation 2 times / minute) and other factors, the reasonable tightness level margin was determined to be 5 through sensitivity analysis and 1000 Monte Carlo simulation experiments. Finally, the highest tightness level threshold to ensure monitoring accuracy was obtained, such as heart rate monitoring not exceeding level 3 and blood oxygen monitoring not exceeding level 4. The personal attribute information such as age, gender, height and weight filled in by 1000 users during registration, as well as the subjective tightness score (1-5 points) fed back during the first wearing, were obtained, and the user data was subjected to feature engineering and z-score standardization. The k-means clustering algorithm was used to group the users, and the optimal number of clusters was determined to be 4 by the elbow rule and silhouette coefficient, and four user subgroups with different tightness perception characteristics were obtained, namely, comfortable type (45%), general type (30%), sensitive type (20%), and extremely sensitive type (5%). The C5 decision tree algorithm was used to establish classification rules between user attributes and clustering results, such as If age < 30 and BMI < 24 Then comfortable type (confidence 86%) If age ≥ 60 and gender = female Then sensitive type (confidence 78%). The association patterns between different combinations of attributes such as age, gender, BMI and the optimal tightness level range were mined to form a personalized user tightness perception portrait, such as the optimal range for comfortable users is 2-4 levels, and the optimal range for sensitive users is 1-2 levels.The highest tightness level threshold is compared with the user's personal tightness level interval. If the highest threshold falls within the personal interval, the user is marked as tolerant, indicating that the user can adapt to a higher wearing tightness to achieve better monitoring accuracy; if the highest threshold exceeds the upper limit of the personal interval, the user is identified as sensitive, indicating that the user is sensitive to wearing pressure and needs to prioritize wearing comfort. Based on the domain expert knowledge and historical user feedback data, the tightness level deviation threshold is empirically set and iteratively optimized, such as dynamically adjusting the deviation from the initial 5 levels to 8 levels, so that the boundary conditions of user classification are more accurate. For tolerant users, based on their personal optimal tightness level interval (such as 2-4 levels) and actual wearing habit data (such as an average of 5 levels), fuzzy reasoning and Q-learning reinforcement learning algorithms are used to give targeted tightness adjustment suggestions, such as "It is recommended that you appropriately increase the tightness level from 5 to 3 levels to achieve better heart rate monitoring effects, while the wearing experience is still within the comfortable range", to guide users to appropriately increase the tightness while ensuring wearing comfort to meet the monitoring accuracy requirements. For sensitive users, we recommend a lower tightness wearing solution corresponding to the lower limit of their personal tightness level range (such as level 1), and provide suggestions on wearing position selection (such as the inside of the wrist) and wearing time control (such as taking a 10-minute break every 2 hours) to relieve pressure discomfort. At the same time, we remind users that "at the lower tightness level 1, the heart rate monitoring error may exceed 8 times / minute, and the blood oxygen monitoring error may exceed 5%. Please weigh it.

[0032] For tolerant users, adjustment instructions for the wearable device are generated to adjust the tightness level of the wearable device to a level higher than the maximum tightness level and closest to the optimal tightness level. If the minimum monitoring accuracy is lower than the threshold, the adjustment instructions include reducing the data collection frequency of the wearable device to extend its service life; if the minimum monitoring accuracy is higher than the threshold, the adjustment instructions include increasing the data collection frequency of the wearable device.

[0033] Through user surveys and physiological experiments, the comfort ratings and experience feedback of tolerant users on different tightness levels are obtained. Combined with the physiological signal monitoring accuracy of wearable devices at different tightness, the optimal tightness level range and monitoring accuracy requirements of users are determined and stored in the user information database to provide a basis for subsequent tightness adjustment. The optimal tightness level range and monitoring accuracy requirements of users are read from the database, and the upper limit of the interval is compared with the preset maximum tightness level threshold. If the upper limit is greater than or equal to the threshold, the threshold is used as the recommended tightness level; if the upper limit is less than the threshold, the upper limit value is taken as the recommended tightness level. According to the recommended tightness level, a fuzzy control algorithm is designed to generate tightness adjustment instructions for wearable devices. With the tightness level as the input variable and the control parameters of the actuator (such as voltage, current, etc.) as the output variable, Gaussian and triangular membership functions are defined, and a fuzzy rule base is constructed, such as "IF the tightness level is high THEN the voltage is large". Through fuzzy reasoning and defuzzification, the tightness level is mapped to a specific control parameter value. The adjustment command is sent to the embedded control unit of the wearable device. The control unit drives the actuator, such as piezoelectric material, memory alloy, etc., according to the command to change the tightness of the wearable device in real time. The PID control algorithm based on pressure sensor feedback is introduced to monitor the contact pressure between the wearable device and the skin in real time, calculate the pressure error and adjust the output of the actuator until the target tightness level is reached and the control accuracy is within the preset range. According to the minimum monitoring accuracy threshold, the data acquisition frequency of the wearable device is adaptively adjusted to minimize energy consumption and extend battery life while meeting the monitoring requirements. The current monitoring accuracy is calculated in real time. If it is lower than the threshold, the frequency reduction control logic is triggered. According to the ratio of the accuracy difference to the threshold, the data sampling rate is reduced by 1-5 gears, and each gear corresponds to a sampling frequency change of 10Hz; if it is higher than the threshold, the frequency increase control logic is triggered to increase the data sampling rate by 1-3 gears until the maximum allowable sampling frequency is reached. Wearable devices usually include multiple types of sensors such as accelerometers, gyroscopes, PPG, ECG, etc., which are distributed in different body positions such as wrists, chests, ankles, etc. The acquisition frequency ranges from 1Hz to 1kHz, and the data characteristics are different. Multi-sensor data fusion algorithms, such as Kalman filtering and Bayesian estimation, are used to comprehensively analyze and suppress noise in physiological signal data collected by different sensors to improve data quality and monitoring accuracy. Adaptive weight allocation strategies are used to dynamically adjust the data contribution of each sensor. When monitoring accuracy decreases, the fusion weight of high-confidence sensors is increased, such as increasing the weight of PPG by 20% and the weight of ECG by 10%. When monitoring accuracy increases, the fusion weight of each sensor is balanced to optimize data redundancy and energy consumption.

[0034] Specifically, a one-month follow-up survey was conducted on 100 tolerant users, who wore wearable devices of different tightness levels (1-5) every day, recorded the users' comfort scores for each tightness level (1-10 points), and tested the heart rate monitoring error of the device at each tightness level. Comprehensive analysis found that 80% of users had the highest comfort score (9.2 points on average) and the lowest heart rate monitoring error (<5%) at tightness levels 3-4, so the optimal tightness level range for this user group was determined to be 3-4, and the minimum monitoring accuracy requirement was 95%. For a certain target user, by querying the user information database, the upper limit of its optimal tightness level was obtained to be 4, which was compared with the preset maximum tightness threshold of 5. Since 4<5, 4 was selected as the recommended tightness level for this user. According to the recommended tightness level, a control algorithm containing 25 fuzzy rules was designed. The input variables are the tightness level (level 1-5), and the output variables are the driving voltage (0-5V) and current (0-100mA) of the actuator, such as the rule "IF tightness level is4THEN voltage is4Vand current is80mA". The fuzzy reasoning results were defuzzified by the centroid method, and the control instructions of driving voltage 3.8V and current 76mA were obtained and sent to the control unit of the wearable device. After receiving the instructions, the control unit gradually increased the driving voltage from 0V to 3.8V, while monitoring the output of the pressure sensor, calculating the pressure error in real time through the PID algorithm and adjusting the voltage, and finally stabilizing at 3.85V. At this time, the tightness level of the wearable device is level 4, and the contact pressure with the skin is (30±1)mmHg, which meets the requirements of the recommended tightness level. During the monitoring process, if the current monitoring accuracy is lower than 95%, the frequency reduction control logic is triggered, and the data sampling rate is reduced by 2 gears from 100Hz to 60Hz according to the accuracy difference ratio (such as 2%); if the monitoring accuracy is higher than 98%, the frequency increase control logic is triggered, and the sampling rate is increased by 1 gear from 60Hz to 80Hz. At the same time, the multi-source heterogeneous data of the accelerometer (50Hz), gyroscope (100Hz), PPG (500Hz) and ECG (1kHz) are fused through the Kalman filter algorithm, and the fusion weight of each sensor is adaptively adjusted. When the monitoring accuracy is lower than 95%, the weight of PPG is increased from 0.3 to 0.5, and the weight of ECG is increased from 0.2 to 0.3; when the monitoring accuracy is higher than 98%, the weights of PPG and ECG are reduced to 0.4 and 0.25 respectively, and the weights of the accelerometer and gyroscope are set to 0.175, which can reduce energy consumption while ensuring monitoring accuracy. Finally, the user wore the tuned wearable device for continuous monitoring for 1 week, and the average monitoring accuracy stabilized at (97±1)%, an increase of 5% compared with before wearing. The battery life was extended from 2 days to 3 days, meeting the user's comfort and monitoring needs.

[0035] For sensitive users, control instructions for the wearable device are generated to reduce the data collection frequency of the wearable device, and the tightness level of the wearable device is adjusted to the highest tightness level when the wearable device is collecting data, and the tightness level of the wearable device is adjusted to be closest to the optimal tightness level when the wearable device is not collecting data.

[0036] The optimal tightness level and data acquisition frequency requirements of the individual identified as sensitive users are read from the user information database. According to the user's sensitivity characteristics and physiological parameters, a quantitative relationship model between sensitivity, optimal tightness level, and data acquisition frequency is established. By constructing a ternary quadratic polynomial regression equation, the nonlinear mapping relationship between sensitivity, tightness level, and acquisition frequency is fitted, and the model parameters are estimated using the least squares method to obtain a quantitative prediction function of sensitivity, optimal tightness level, and data acquisition frequency. Based on this function, the highest tightness level that the user can accept for a long time and the amplitude of the data acquisition frequency that needs to be reduced are calculated. According to the frequency reduction amplitude requirements, data acquisition instructions are generated. The instruction content includes parameters such as the sampling period of the sensor, the size of the data buffer, and the time interval for data packaging. The fractional calculus model is used to describe the nonlinear dynamic relationship between the frequency reduction amplitude and the acquisition parameters. The mathematical expression of the model is D a x f(t)=t n-a / Γ(n-α)·(d n f(t) / dt n), where α is the fractional differential order, reflecting the rate and smoothness of the frequency reduction process, n is an integer order, and Γ is the Gamma function. By solving the fractional differential equation, a nonlinear mapping relationship between the frequency reduction amplitude and the acquisition parameters is established, and a parameter adaptive optimization algorithm, such as an adaptive weighted PSO algorithm, is used to search for the optimal model parameter combination, and dynamically adjust the various parameter values ​​in the acquisition instruction to ensure the integrity and real-time performance of the data while meeting the frequency reduction requirements. Generate a tightness adjustment instruction, and when it is detected that the wearable device starts data collection, adjust the tightness level to the highest. The fuzzy control algorithm is used to convert the highest tightness level into a control variable for the actuator, such as the motor speed, the deformation of the shape memory alloy, etc. By establishing a fuzzy rule base of tightness-control variables, the Mamdani reasoning method is used to map the tightness level, a linguistic variable, to a fuzzy membership, and then through fuzzy synthesis, the fuzzy output is defuzzified using the centroid method to obtain the specific actuator control variable value. The form of the fuzzy rule base is "IF tightness level isHigh THEN motor speed isFast AND alloy deformation isLarge". By defining appropriate membership functions and inference rules, a smooth mapping from tightness level to execution control quantity is achieved. When the data collection task of the wearable device is detected to be completed, a tightness callback instruction is immediately generated to reduce the tightness level from the highest value to the value closest to the optimal tightness level. By building a three-layer feedforward neural network, with tightness level as input and user comfort score as output, a nonlinear regression model of tightness level and user comfort is established. The network structure adopts the "5-10-1" setting, that is, 5 neurons in the input layer, 10 neurons in the hidden layer, and 1 neuron in the output layer. The activation function uses the hyperbolic tangent function tanh. The mean square error is used as the loss function, and the gradient descent method is used to iteratively optimize the network weights. By setting hyperparameters such as learning rate and momentum factor, the convergence speed and stability of the training process are controlled to minimize the user discomfort when the tightness level deviates from the optimal value. The callback tightness level is used as the input of the neural network, and the optimal callback tightness level is predicted through forward propagation calculation, so that it is closest to the user's optimal tightness range. The data collection instructions and tightness adjustment instructions are integrated to generate a unified control instruction sequence in chronological order. The instruction sequence includes attributes such as instruction type, execution time, parameter value, etc. By constructing a directed acyclic graph model of the instruction sequence, a discrete event simulation algorithm, such as the event scheduling method, is used to evaluate the timeliness and coordination of instruction execution. For each node in the graph, a corresponding event object is generated, including attributes such as event type, timestamp, priority, etc., and the event is inserted into the event queue in timestamp order. Through the event loop mechanism, events are gradually triggered and processed, the status and clock are updated, and the found instruction conflicts and errors are detected and corrected in a timely manner, such as adjusting the trigger time and priority of the event, to ensure the feasibility and consistency of the control instruction sequence.The control instruction sequence is sent to the embedded control module of the wearable device. The control module is based on the real-time operating system FreeRTOS, and encapsulates the operations such as instruction parsing, parameter setting, and execution control into independent task threads. The synchronous communication between threads is realized through mechanisms such as semaphores and message queues, and exception handling and fault-tolerant control are realized through mechanisms such as interrupt service routines and watchdog timers. The producer-consumer model is adopted to parse the instruction sequence into parameter setting tasks and execution control tasks, and store them in the task queue respectively. The task scheduler dynamically allocates CPU time and resources according to the priority and time constraints of the tasks, coordinates the concurrent execution of tasks, and ensures real-time and reliability. During the task execution process, the task status is monitored and exception handling is realized through hook functions and callback mechanisms to ensure the smooth completion of instructions. Through the embedded wireless communication module, the MQTT protocol is used to upload the execution status, user feedback, exception log and other information of the wearable device to the cloud server in real time. The collected user data and device working data are preprocessed and feature extracted to construct a multidimensional time series data set. Incremental learning algorithms such as online sequential extreme learning machine (OS-ELM) are used to model and incrementally update the data in real time to dynamically optimize the control strategy. OS-ELM randomly generates hidden layer weights and uses a sequential learning strategy to update the output layer weights online without retraining the entire network. It is suitable for processing streaming data and concept drift problems. By setting the forgetting factor and regularization parameters, the plasticity and stability of the model are balanced to track the changing trends of user behavior patterns and device performance characteristics. Generate control strategy improvement instructions and send them to wearable devices to form a closed-loop feedback control, continuously improving the level of intelligence and user experience. At the same time, cross-validation and A / B testing methods are used to evaluate the optimization effect and support adaptive adjustment and decision optimization of the online learning process.

[0037] Specifically, through statistical analysis of the historical wearing data of sensitive users, it is found that the optimal tightness level is 1.5, the highest acceptable tightness level is 2.5, and the optimal data collection frequency is 10Hz. Based on the ternary quadratic polynomial regression model y=a0+a1x1+a2x2+a3x1 2 +a4x2 2 +a5x1x2 fits the relationship between sensitivity, tightness level and acquisition frequency, where y is the acquisition frequency, x1 is the sensitivity, and x2 is the tightness level. The least squares method is used to estimate the regression coefficients a0 = 15.2, a1 = -3.1, a2 = -4.6, a3 = 0.8, a4 = 1.2, a5 = -0.5, and the model determination coefficient R 2 =0.92. Substituting the user's sensitivity value of 0.8 into the model, it is predicted that the optimal acquisition frequency is 12Hz, and the default acquisition frequency needs to be reduced from 20Hz to 8Hz. Using the fractional calculus model D ax f(t)=t n-a / Γ(n-α)·(d n f(t) / dt n) describes the frequency reduction process, taking α = 0.5, n = 2, generating the acquisition instruction of increasing the sampling period from 0.05s to 0.08s, reducing the buffer size from 1KB to 0.6KB, and extending the data packaging time interval from 1s to 1.5s. When the start of data acquisition is detected, according to the fuzzy rule "IF tightness level isHighTHEN motor speed isFastAND alloy deformation isLarge", the membership of linguistic_variable tightness level is set to μ(High) = 0.8, and the fuzzy output μ(Fast) = 0.7, μ(Large) = 0.6 is obtained through Mamdani reasoning, and then the center of gravity method is used for defuzzification, and the motor speed control value is 800rpm, the alloy deformation control value is 3mm, and the actuator is driven to increase the tightness level to 2.3. After data collection, the trained three-layer neural network is used to predict the best callback tightness level. The network structure is "5-10-1", with 5 neurons in the input layer corresponding to the user's age, gender, BMI, heart rate and blood pressure, 10 neurons in the hidden layer, and 1 neuron in the output layer to give the predicted tightness level. The learning rate is set to 0.01, the momentum factor is 0.9, and after training 1000 epochs, the average prediction error on the test set is 0.15. The user's characteristics are input into the neural network, and the forward propagation calculation obtains that the best callback tightness level is 1.7. The generated acquisition instructions and compactness instructions are integrated in chronological order to construct a directed acyclic graph of instruction sequences. The discrete event simulation is performed using the event scheduling method. Each instruction event contains type, timestamp and priority attributes, such as {type: "acquisition instruction", timestamp: "00:05", priority: 2}, {type: "compactness instruction", timestamp: "00:03", priority: 1}, etc. Events are scheduled according to the principle of minimum timestamp and highest priority. The total simulation time is 15 minutes, the average instruction execution delay is 0.5 seconds, and the priority conflict rate is 2%. The embedded control module is based on the FreeRTOS operating system, and the received instruction sequence is parsed into acquisition tasks and compactness tasks, which enter the task queues with priorities 2 and 1 respectively. The task scheduler adopts a preemptive priority scheduling algorithm to allocate independent cores for compactness tasks. The acquisition tasks are executed concurrently in a time slice rotation mode. The time slice length is set to 50ms, and the context switching time is less than 1ms. The synchronization between tasks is achieved through binary semaphores, asynchronous communication is achieved through ring buffers, and task-level and system-level exception handling is achieved through hardware timers and watchdog mechanisms to ensure real-time and reliability. The wearable device sends the collected user data and device status data to the cloud server in JSON format through the MQTT protocol with a reliability level of QoS1, with a transmission delay of less than 500ms and a packet loss rate of less than 0.1%.The server uses the OS-ELM algorithm to perform online modeling on the received multi-dimensional time series data. The input layer dimension is 10, including features such as heart rate, blood oxygen, and number of steps. 100 hidden layer neurons are initialized, and input weights in the interval [-1,1] are randomly generated. According to the sequential block learning strategy, the output weights are updated every 1 minute. The regularization factor C = 10, the forgetting factor λ = 0.9, and after 6 hours of continuous learning, the root mean square error of the model drops below 0.2. Based on the prediction results of the OS-ELM model, optimization instructions for adjusting parameters such as sampling rate, tightness level, and data fusion strategy are generated and sent to the wearable device for execution to achieve end-to-end closed-loop control. After cross-validation and A / B testing of 10 users, user satisfaction increased by 15%, battery life was extended by 20%, and average monitoring accuracy was improved.

[0038] Step S106: Determine the recommended wearing position and recommended tightness level through a regression model in combination with the minimum monitoring accuracy corresponding to the target indicator. Dynamically adjust the recommended wearing position and recommended tightness by real-time monitoring of user feedback information.

[0039] According to different application scenarios and monitoring targets, such as sports heart rate monitoring, sleep blood oxygen monitoring, etc., determine the minimum monitoring accuracy requirements of the corresponding physiological parameters as a benchmark for evaluating the wearing position and tightness level. By referring to industry specifications such as the IEEE P1708 heart rate monitoring wearable device standard and the ISO 80601-2-61 pulse oximeter standard, the minimum monitoring accuracy threshold is quantitatively set. The sensor array is used to collect physiological parameter data at different wearing positions and tightness levels, such as wearing the device in an adjustable tightness manner on the wrist, upper arm, ankle and other parts, and synchronously recording electrocardiogram, photoelectric volume pulse wave, acceleration and other signals, and the collected data is cleaned, denoised and feature extracted to build a high-quality monitoring accuracy sample set. For different physiological parameters, a multivariate nonlinear regression model is established, such as support vector regression with polynomial kernel function, with wearing position, tightness level, user physiological characteristics, etc. as input features, and the monitoring accuracy of physiological parameters as output target. The minimum angle regression method based on L1 regularization is used for feature selection, and the model hyperparameters are optimized through grid search. The model generalization performance is evaluated using cross-validation to obtain the nonlinear regression equation of monitoring accuracy with respect to wearing position and tightness level. Based on the regression model, a multi-objective optimization problem under the constraint of monitoring accuracy is constructed. With wearing position and tightness level as decision variables, monitoring accuracy, wearing comfort, equipment stability, etc. as optimization targets, a weighted objective function is established, and the TOPSIS method is used to normalize and assign weights to multiple targets. The value range of wearing position and tightness level is set as a constraint condition, and it is converted into a multivariate nonlinear programming problem with inequality constraints. Genetic algorithms are used to solve the problem. The wearing position and tightness level are encoded into binary gene strings. The optimal solution is iteratively searched through genetic operations such as selection, crossover, and mutation. The fitness function comprehensively considers the weighted performance of monitoring accuracy and other objectives. After 1000 generations of iteration, the optimal wearing position and tightness level combination that meets the minimum monitoring accuracy requirement is output. A user feedback collection module is developed to push questionnaires through mobile phone APP to collect users' subjective scores on the current wearing position and tightness level. The scoring dimensions include comfort, pressure, activity restriction, etc. The higher the score, the better the experience. At the same time, the built-in pressure sensors and heart rate sensors of wearable devices are used to detect the user's physiological feedback in real time, extract characteristic indicators such as skin surface pressure distribution and heart rate variability, and use machine learning algorithms to classify and quantify the user's fatigue, stress level and other states, and construct a multi-dimensional user feedback vector. The item-based collaborative filtering algorithm is used to calculate the similarity of different users' preferences for wearing position and tightness level, generate a user-wearing preference matrix, and predict their acceptance of new wearing solutions based on their historical preferences. The comprehensive performance of the current recommendation scheme is evaluated by combining the weighted score of physiological parameter monitoring accuracy and user feedback.A dynamic adjustment strategy based on deep reinforcement learning is constructed, with user feedback vector and monitoring accuracy as state space, adjustment values ​​of wearing position and tightness level as action space, and comprehensive performance improvement value as reward function. The DQN algorithm is used to establish the state-action value function approximation model, and the Q network is constructed with the long short-term memory network (LSTM) as the backbone. The state features are input and the Q value of each action is output. The ε-greedy strategy is used to balance exploration and utilization, generate experience replay data for offline training of the network, and use the target network to alleviate overfitting. Through continuous trial and error and learning, the mapping strategy from state to action is optimized to achieve adaptive dynamic adjustment of wearing position and tightness level in real-time environment. After each adjustment, the optimization performance is evaluated and recorded, user experience feedback on the new solution is collected, key improvement points are extracted, and the state representation, action design and reward function of the reinforcement learning model are iteratively updated. At the same time, the performance difference before and after optimization is objectively evaluated through methods such as A / B testing, and tuning knowledge and data are continuously accumulated to form a self-perfecting closed-loop optimization mechanism. Under the premise of ensuring monitoring accuracy, user comfort is maximized, and adaptive wear optimization of human-machine collaboration is achieved.

[0040] Specifically, according to the IEEE P1708 standard, the accuracy of heart rate monitoring is defined as the root mean square error between the measured value and the reference value, and the measurement error of static heart rate is required to be no more than ±3 times / minute. By wearing pressure-adjustable photoelectric volume pulse wave sensors at six different locations such as the wrist, upper arm, and ankle of 100 subjects, 24-hour dynamic heart rate data were collected with three levels of tightness: loose, medium, and tight. A medical-grade ECG monitor was used to collect baseline heart rate data simultaneously. The 5 million collected heart rate sample points were preprocessed by detrending and bandpass filtering to extract time domain and frequency domain features. A support vector regression model with a polynomial kernel function K(x,x')=(γ*xT*x'+r)d was used, where the kernel function parameters γ=0.1, r=1, and d=3. The minimum angle regression method was used to select the 8 most relevant feature subsets from 50 features. The model parameters C=10, ε=0.1 were optimized by 5-fold cross validation, and the mean absolute error on the test set was 1.5 times / minute. The monitoring accuracy target was set to ±2 times / minute. Combined with the comfort and stability targets, a three-objective weighted function f = 0.5×acc+0.3×comf+0.2×stab was constructed, where acc represents the accuracy score, comf represents the comfort score, and stab represents the stability score. The weight coefficient was determined by the analytic hierarchy process. The genetic algorithm was used to search for the optimal solution. The six wearing positions and three tightness levels were encoded into a 9-bit binary string. The population size was set to 50, the crossover probability was 0.8, and the mutation probability was 0.1. After 1000 iterations, the optimal wearing solution was obtained as medium tightness on the upper arm. At this time, the comprehensive objective function value was 0.93. A questionnaire was used to conduct subjective scoring of 1000 users. The comfort was evaluated in three dimensions: uniformity of pressure distribution, pressure sensation, and activity hindrance. The stability was evaluated by the success rate of heart rate and blood oxygen measurement. 50,000 user feedback vectors were formed. The Pearson correlation coefficient was used to calculate the similarity between users. The collaborative filtering algorithm based on items predicted the user's preference for different wearing solutions. For new users, similar users were matched according to their personal attribute characteristics to generate personalized recommendation solutions. The LSTM-DQN model was built to dynamically adjust the wearing position and tightness. The state quantity included 12 dimensions such as user feedback score, monitoring accuracy, physiological pressure characteristics, etc. The action space included adjusting the position or tightness up / down by 1~2 levels. The reward function was the percentage of comprehensive performance improvement, and a fault-tolerant penalty item was set. After 10,000 training iterations, the average reward value increased from 0.2 to 0.85. The converged policy network was able to recommend the optimal fine-tuning solution in real time according to the user status. The strategy was applied to continuously optimize the wearing solutions of 100 users, and the average monitoring accuracy increased by 2%, the comfort score increased by 15%, and the user satisfaction reached 95%. Through A / B testing, the optimization group and the control group were tracked and compared for 1 month. The device wearing compliance of the optimization group was 30% higher, and the continuous monitoring time was extended by 6 hours / day, which verified the effectiveness of the adaptive optimization mechanism.

[0041] Step S107, apply the long short-term memory neural network to analyze the user's personalized data, build a model to predict the accuracy of biological signals under different usage conditions, compare the model prediction results with the actual collected data, calculate the mean square error, and dynamically adjust the wearing strategy according to the error size until the error converges to an acceptable range.

[0042] Collect multi-dimensional personalized data such as historical wearing records, physiological parameters, and activity types of users, and use data enhancement techniques such as random rotation and mirror flipping to expand sample diversity. Clean, align and standardize the data, balance the category distribution of data in different scenarios through upsampling and downsampling, extract key features that reflect users' wearing habits, such as wearing time, adjustment frequency, comfort feedback, etc., construct a time series feature matrix, and randomly divide it into training set, validation set and test set. Design the architecture of the long short-term memory neural network model, the number of input layer nodes is consistent with the feature dimension, the number of output layer nodes is consistent with the biological signal type, and it contains an LSTM hidden layer. Use the random search algorithm to optimize the hyperparameters, traverse the learning rate, number of hidden units, batch size and other parameter combinations within the specified range, and select the model configuration with the smallest validation set error. Extract the user behavior feature representation by training the LSTM autoencoder as the pre-training initialization parameter of the subsequent model. Use the backpropagation algorithm to train the model, use the mean square error as the loss function, and introduce the L2 regularization term to control the model complexity. Use the Adam optimizer to adaptively adjust the learning rate, and set the gradient clipping threshold to 1.0 to avoid gradient explosion. Dropout regularization is applied to the input layer and LSTM layer with probabilities of 0.2 and 0.5, respectively, to alleviate overfitting. After each epoch, the model performance is evaluated on the validation set. If there is no improvement for 5 consecutive times, the early stopping mechanism is triggered and the model parameters with the best performance are saved. The dataset is divided using 10-fold cross-validation, and the training is repeated 10 times. The prediction results of each model are averaged to improve the robustness of the model. The trained LSTM model is used to predict the accuracy of biological signal acquisition in different usage scenarios based on the user's real-time wear data, environmental conditions, physiological state and other conditions. The sequential attention mechanism is introduced to dynamically adjust the attention weights of different time steps through the mask matrix to highlight the impact of key moments. Causal convolution is applied in the time dimension to effectively capture the long-term dependencies of user behaviors. The confidence of the output results of the model is calibrated, and the uncertainty interval is set according to the historical prediction deviation to improve the interpretability of the prediction. Obtain the biosignal data collected by the device actually worn by the user, perform feature engineering on the data, extract indicators corresponding to the model prediction output, calculate multiple evaluation indicators such as the mean square error and mean absolute error between the predicted value and the true value, evaluate the model prediction performance from multiple angles, and identify the time periods and usage conditions with large prediction errors. Trigger the adaptive model adjustment mechanism when the model prediction error exceeds the threshold for multiple consecutive times. Use the grid search algorithm to search for the optimal configuration in the hyperparameter space such as the number of LSTM units, learning rate, and regularization parameters, and avoid overfitting through early stopping and model integration. Prune and quantize the LSTM model, remove redundant connections and low-importance parameters, and quantize the weights to 8-bit fixed-point numbers to reduce computational overhead while ensuring performance. Fine-tune the parameters of the updated model on the training set to narrow the gap between the predicted value and the true value.According to the feedback information of the prediction error, combined with domain knowledge and user feedback, the rules and reward functions for dynamic optimization of the wearable strategy are formulated, and the optimization problem is modeled as a Markov decision process. The state space includes user attributes, device parameters, environmental factors, etc., the action space includes adjusting the wearing position, tightness, sampling frequency, etc., and the reward function comprehensively considers factors such as signal accuracy and wearing comfort. Applying reinforcement learning algorithms such as DeepQ-Network, by exploring and utilizing balance, learn the optimal wearable parameter adjustment strategy and generate targeted optimization suggestions. Design an online strategy update mechanism based on incremental learning, and regularly fine-tune the strategy network according to the newly collected user data and feedback to improve the adaptability and effectiveness of the strategy. Construct a closed loop of human-computer interaction feedback, push wearable optimization suggestions through smartphone APP, collect users' subjective ratings of the optimization effect, record the behavior logs of users' active adjustment of the wearable mode, use active learning algorithms to screen samples with the most significant model improvement effect, and add them to the training set to participate in model iteration. Based on user portraits and behavior analysis, customize personalized wearable optimization strategies for different user groups to improve the pertinence and acceptance of optimization. Continuously monitor the changes in the prediction error after optimization and draw a learning curve. When the error decreases below a certain threshold after multiple consecutive iterations, it is judged to have converged to an acceptable level, completing the adaptive wearable strategy optimization task.

[0043] Specifically, the wearing records of 1,000 users for three consecutive months were extracted from the wearable device database. Each record contained 10 dimensional features such as user ID, timestamp, acceleration, heart rate, and number of steps. The nearest neighbor interpolation was performed on the missing values, and the outliers were filtered based on the box plot. The SMOTE algorithm was used to upsample the minority class samples to balance the number of samples in each category. 20 statistical features reflecting user habits such as wearing time and heart rate variability were extracted, and combined with the user's attribute features such as age, gender, and BMI to form a 40-dimensional feature vector, which was normalized using Min-Max. The ratio of the training set, validation set, and test set was 8:1:1, and the data of each user was randomly allocated in a ratio of 7:2:1. The LSTM model was built using the Keras framework. The input layer contained 40 nodes, the LSTM layer contained 128 units, and the tanh activation function was used. The output layer contained 5 nodes, which respectively represented the predicted values ​​of 5 biological signals such as blood pressure and heart rate. 50 rounds of random search were performed in the parameter space with the number of hidden units [64, 128, 256], the learning rate [0.001, 0.005, 0.01], and the batch size [32, 64, 128]. The optimal hyperparameter combination was selected based on the mean absolute error of the validation set. The model was trained using the Adam optimizer with an initial learning rate of 0.005. The learning rate was decayed to 0.5 times the original value every 10 epochs, and the L2 regularization coefficient was set to 0.01. The model was trained for 50 epochs. When the validation set error did not decrease for 5 consecutive times, the early stop was triggered. The model parameters with the lowest validation set error were taken as the final model. The five biological signals in the test set were predicted, and the mean absolute errors were 2 times / minute, 5 mg, 1%, 5 steps / minute, and 0.5 degrees Celsius, respectively. The sequence attention mechanism was introduced. The similarity between the query vector and the key vector was calculated by dot product to obtain the attention weight. The hidden state of the LSTM output was weighted summed to generate a feature representation focused on the key moment. Causal convolution is used to extract the long-term dependency of user behavior in the time dimension. The convolution kernel size is 24 hours, the step length is 1 hour, and the dilation rate is 2. The dilated convolution is used to increase the receptive field. Platt calibration is performed on the model prediction value, and the confidence interval is estimated according to the prediction error distribution. When the proportion of actual values ​​falling within the interval exceeds 95%, the confidence is considered high. In the case where the prediction error exceeds 3 times the average error level for 3 consecutive days, the model is triggered to adjust adaptively. The hyperparameters such as the number of LSTM units and learning rate are optimized through grid search, and the weights of the hidden layer and output layer are pruned. Based on the Hessian matrix analysis, the parameters of the top 20% of the sensitivity are retained, and the rest are truncated to 0 to compress the model scale. Fine-tune the updated model for 10 epochs to reduce the prediction error to within 50% of the original.The DQN algorithm is applied to optimize the wearing strategy. The state space includes 20 features such as user type, historical error, and current parameters. The action space includes adjusting the wearing position and tightness by 1 to 5 gears. The reward function is set as the weighted sum of the improvement in prediction accuracy and the comfort of user feedback. After 10,000 trainings, the average reward value converged from 0.2 to 0.85. The optimal strategy is applied to new users, prompting them to wear the bracelet 2 cm above the wrist bone, adjust the tightness to 3 gears, increase the sampling frequency by 20%, reduce the prediction error by 15 percentage points, and improve user satisfaction.

[0044] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A health monitoring method integrating multimodal bioinformation, characterized in that: The method comprises: When a user wears a wearable device for the first time, a multimodal sensor is deployed to collect the user's actual biosignals and record the tightness level and location information of the wearable device; Collect users' tightness perception of wearable devices at different positions and corresponding to different tightness levels, and associate tightness perception with tightness level and position information; including dividing tightness perception into 5 levels, collecting users' subjective evaluation of wearing comfort, and obtaining users' subjective tightness perception data at different tightness levels; collecting users' pressure distribution data at the contact point between the wearable device and the skin at different tightness levels, aligning and calibrating the subjective tightness perception score with the objective pressure distribution measurement results, establishing a quantitative relationship between the two, and establishing a quantitative relationship model between tightness level and pressure distribution; obtaining users' body surface contour data, combining with the three-dimensional structure model of the wearable device, and obtaining theoretical tightness levels at different positions through the contact area and positive pressure distribution at different wearing positions; collecting tightness perception feedback from different users, selecting the user attributes with the greatest correlation with tightness perception, building a decision tree model, selecting the optimal splitting attribute through the information gain index, and generating a tree structure; The monitoring accuracy of the wearable device is calculated based on the actual biological signals and wearable device data. A regression analysis model is established based on the tightness perception, tightness level, and location information of the current user to describe the correlation between different tightness levels and location information and the monitoring accuracy of the wearable device. Receive the target index and personal tightness threshold corresponding to the wearable device input by the user, combine the personal tightness threshold and the best tightness level corresponding to the user's best tightness feeling, determine the personal tightness level range, and evaluate the matching degree of the wearable device to the user based on the preset adjustable range of the wearable device; including obtaining the user's functional requirements and expected physiological monitoring indicators for the wearable device, while allowing the user to customize the comfort range to form a personal tightness threshold; adopt a fuzzy comprehensive evaluation method to construct a multi-level fuzzy evaluation matrix with tightness feeling as the target layer, physiological parameters, physical pressure and wearing parts as the criterion layer, and different tightness levels as the solution layer; determine the weight coefficient and membership function of each evaluation indicator, use the weighted average method to perform multi-level fuzzy synthesis operation, and obtain the quantitative mapping relationship between the user's subjective tightness feeling and the objective tightness level; analyze the user's tightness feedback data, determine the optimal clustering number k through the elbow rule and silhouette coefficient evaluation, use the Euclidean distance to measure the similarity between the tightness feeling feature vectors, iteratively optimize the clustering center, and obtain the best tightness feeling mode for different groups of people; Determine the minimum monitoring accuracy according to the target indicator, and determine the highest tightness level corresponding to the minimum monitoring accuracy through the regression model. If the highest tightness level is within the personal tightness level range, the user is identified as a tolerant user; if the highest tightness level is outside the personal tightness level range, the user is identified as a sensitive user; Combined with the minimum monitoring accuracy corresponding to the target indicator, the recommended wearing position and recommended tightness level are determined through the regression model, and the recommended wearing position and recommended tightness are dynamically adjusted through real-time monitoring of user feedback information; Long short-term memory neural networks are used to analyze user personalized data, and a model is built to predict the accuracy of biological signals under different usage conditions. The model prediction results are compared with the actual collected data, the mean square error is calculated, and the wearing strategy is dynamically adjusted according to the error size until the error converges to an acceptable range.

2. The method according to claim 1, characterized in that When the user wears the wearable device for the first time, the multimodal sensor is deployed to collect the user's actual biological signals and record the tightness level and location information of the wearable device, including: Obtain the user's basic personal information, preliminarily estimate the user's physiological parameter range, and provide a reference benchmark for the actual biological signals collected; Calculate the optimal tightness level and wearing position, obtain the initial wearing parameters, guide users to wear correctly, and improve the accuracy of collected signals; Detect the user's motion status, identify and filter out relatively static time periods, collect biological signals within the time period, extract characteristic parameters of the signals, obtain basic physiological indicators of the user in a static state, and establish a personal health baseline; Continuously record the changes in biological signals when the user wears the device at different tightness levels and positions, extract the statistical characteristics and spectral characteristics of the signal, calculate the mean, variance and correlation coefficient indicators of the characteristic parameters, set thresholds, judge the validity and reliability of the collected data, and adjust the working parameters of the sensor accordingly, and detect the tightness and offset of the wearable device in real time; The biosignal data collected when the user wears the device for the first time is bound to the user’s basic personal information and uploaded to the cloud platform.

3. The method according to claim 1, characterized in that The monitoring accuracy of the wearable device is calculated based on the actual biological signal and the wearable device data, and a regression analysis model is established in combination with the tightness perception, tightness level, and location information of the current user to describe the correlation between different tightness levels and location information and the monitoring accuracy of the wearable device, including: Collect the user's physiological signals and synchronously record the motion data of the user's wearable device; Use convolutional neural network deep learning models to perform quality assessment and anomaly detection on physiological signals; According to the signal quality evaluation results, high-quality physiological signal segments are screened out, and the screened high-quality physiological signal segments are compared with the monitoring results of the reference medical equipment, and the measurement errors and deviations of the wearable device on different physiological indicators are calculated to obtain a quantitative evaluation of the monitoring accuracy; By means of questionnaire survey and human-computer interaction, the subjective comfort scores of users at different tightness levels and wearing positions were collected, and the corresponding relationship between tightness perception and user physiological parameters was established.

4. The method according to claim 1, characterized in that: Determine the minimum monitoring accuracy according to the target indicator, and determine the maximum tightness level corresponding to the minimum monitoring accuracy through the regression model. If the maximum tightness level is within the personal tightness level range, the user is identified as a tolerant user; if the maximum tightness level is outside the personal tightness level range, the user is identified as a sensitive user; including: Determine the target indicators for monitoring various physiological parameters based on the application scenarios and functional positioning of wearable devices; The physiological parameter monitoring data of different subjects under different tightness levels, materials and wearing positions were collected, and the data were cleaned, outliers processed and normalized. A multivariate linear regression model was constructed with the tightness level, material characteristics and wearing position of the wearable device as independent variables and the monitoring error of the physiological parameters as the dependent variable. The stepwise regression method was used for feature selection, and the regression coefficient was analyzed for significance, and the quantitative relationship between tightness level, material characteristics, wearing position and monitoring accuracy was obtained. Substitute the monitoring target indicators into the regression equation, calculate the minimum tightness level required to achieve the target accuracy, determine the reasonable tightness level margin, and finally obtain the maximum tightness level threshold to ensure monitoring accuracy; The highest tightness level threshold is compared with the user's personal tightness level range. If the highest threshold falls within the personal range, the user is marked as tolerant, indicating that the user can adapt to a higher wearing tightness to achieve better monitoring accuracy; If the highest threshold exceeds the upper limit of the personal range, the user will be identified as sensitive, indicating that the user is more sensitive to wearing pressure and needs to prioritize wearing comfort.

5. The method according to claim 4, characterized in that For tolerant users, adjustment instructions for the wearable device are generated to adjust the tightness level of the wearable device to a level higher than the maximum tightness level and closest to the optimal tightness level. If the minimum monitoring accuracy is lower than the threshold, the adjustment instructions include reducing the data collection frequency of the wearable device to extend its service life; if the minimum monitoring accuracy is higher than the threshold, the adjustment instructions include increasing the data collection frequency of the wearable device.

6. The method according to claim 4, characterized in that For the sensitive user, a control instruction of the wearable device is generated to reduce the data collection frequency of the wearable device, and adjust the tightness level of the wearable device to the highest tightness level when the wearable device is collecting data, and adjust the tightness level of the wearable device to be closest to the optimal tightness level when the wearable device is not collecting data.

Images