A sleep quality detection and evaluation method for electronic watches

Abstract

The present application provides an electronic watch sleep quality detection and evaluation method, comprising: the electronic watch uses an optimized heart rate monitoring unit to periodically collect the user's heart rate data through a photoelectric volumetric blood flow sensor, and after detecting a turning over action, the photoelectric volumetric blood flow sensor increases the frequency of collecting the heart rate data; the electronic watch performs abnormal recognition processing on the heart rate data during turning over, analyzes the variability of the heart rate data, identifies non-physiological heart rate fluctuations, cleans and screens the heart rate data during turning over, and removes abnormal values; the electronic watch combines the screened heart rate data with the user's historical sleep data, establishes a user sleep quality evaluation model, evaluates the user's sleep quality, generates a sleep report, and provides personalized sleep suggestions.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for detecting and evaluating sleep quality of an electronic watch. Background Art

[0002] In the process of electronic watches evaluating sleep quality, the watch needs to evaluate the user's sleep status through heart rate monitoring. However, such an evaluation system has a chain of several technical problems. The user's turning over during sleep will cause the tightness of the contact between the watch and the skin to change, which in turn affects the accuracy of the heart rate sensor's detection of the pulse signal. The different angles and strengths of the arms when turning over will change the degree of pressure on the skin by the watch strap and the stable position of the watch, which will cause deviations in the acquisition of the heart rate signal. If the user's turning over action and its interference with the heart rate data collection cannot be effectively distinguished, it is difficult to ensure the accuracy of the heart rate data, thereby affecting the evaluation of sleep quality. One problem is how to monitor and distinguish the user's turning over action in real time without interfering with the user's natural sleep pattern, and judge the impact of these actions on heart rate monitoring, and how to compensate for the change in skin looseness caused by turning over in real time to ensure that the acquisition of heart rate data is not interfered with by the turning over action. The algorithm for dynamically monitoring the turning over action needs to accurately identify the angle, frequency and intensity of turning over, distinguish its impact on the arm skin looseness and the watch wearing position, and correct the heart rate signal error caused by it in real time, which is of great significance for achieving accurate sleep quality evaluation. Summary of the invention

[0003] The present invention provides a method for detecting and evaluating sleep quality of an electronic watch, which mainly includes:

[0004] The three-axis acceleration sensor built into the electronic watch monitors the acceleration data of the wrist movement caused by turning over during sleep in real time. It determines whether there is a turning over movement through the acceleration data, calculates and records the turning over information, including the turning over speed, angle and direction, and stores the collected turning over information in the electronic watch memory and synchronizes it to the cloud at regular intervals.

[0005] After detecting the turning over action, the photoelectric volumetric pulse blood flow sensor emits a test beam to the skin surface, measures the amount of reflected light from the test beam, and uses the support vector machine model to evaluate skin relaxation and blood flow changes, and associates the turning over action to record the effect of the turning over action on skin compression and the effect on local skin blood circulation;

[0006] The cloud dynamically evaluates the acceleration data and the measurement results of the photoelectric volumetric pulse blood flow sensor and processes them through a fusion algorithm module, including comparing the data changes before and after each turning over action, calculating the tightness of the electronic watch's contact with the skin, to determine whether the wearing position of the electronic watch has changed, and recording the timestamp of the wearing position change;

[0007] The cloud uses the turning over information, skin laxity data and watch wearing position information to establish a turning over impact model, analyze the association between different turning over actions and skin laxity and watch wearing position, and analyze the impact on heart rate collection, so that the electronic watch can adjust the weight coefficient in the heart rate monitoring unit, so that the heart rate monitoring unit can dynamically adjust the heart rate signal processing parameters based on the turning over action, and optimize the heart rate monitoring unit's collection of heart rate signals;

[0008] The electronic watch uses an optimized heart rate monitoring unit to periodically collect the user's heart rate data through a photoelectric volumetric blood flow sensor. After detecting a turning over action, the photoelectric volumetric blood flow sensor increases the frequency of collecting heart rate data.

[0009] The electronic watch performs abnormal recognition processing on the heart rate data during turning over, analyzes the variability of the heart rate data, identifies non-physiological heart rate fluctuations, cleans and screens the heart rate data during turning over, and removes abnormal values;

[0010] The electronic watch combines the filtered heart rate data with the user's historical sleep data to build a user sleep quality assessment model, evaluate the user's sleep quality, generate a sleep report, and provide personalized sleep recommendations.

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

[0012] The present invention provides an electronic watch sleep quality detection and evaluation method, which realizes high-precision monitoring of the user's sleep behavior, deeply understands the user's sleep dynamics based on the monitoring data, and provides a detailed data basis for subsequent sleep quality evaluation; the application of the photoelectric volumetric pulse blood flow sensor, combined with the machine learning model, can not only evaluate the impact of turning over on skin pressure, but also monitor the real-time changes in blood circulation, which provides a new dimension for analyzing changes in the body's microenvironment during sleep, and helps to more comprehensively understand the impact of sleep on the human physiological state; by processing the turning over information, skin condition data and watch wearing position information in the cloud, the heart rate monitoring unit can be dynamically adjusted. The working parameters of the unit ensure that the accuracy of heart rate monitoring is not affected by external factors such as turning over; the optimization process improves the accuracy and adaptability of heart rate monitoring; the frequency of heart rate data collection is increased during turning over, combined with abnormal recognition processing, to effectively eliminate non-physiological heart rate fluctuations caused by turning over, ensure the purity of heart rate data, and improve the effectiveness and reliability of data analysis; integrate heart rate data during turning over and historical sleep data to establish a user-specific sleep model, which can more accurately evaluate sleep quality and provide customized sleep improvement suggestions accordingly; this personalized analysis helps users understand their own sleep conditions and take targeted measures to improve sleep quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flowchart of a method for detecting and evaluating sleep quality of an electronic watch.

[0014] Figure 2 The present invention is a schematic diagram of a sleep quality detection and evaluation method for an electronic watch.

[0015] Figure 3 It is another schematic diagram of the electronic watch sleep quality detection and evaluation method of the present invention. DETAILED DESCRIPTION

[0016] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0017] like Figure 1-3 In this embodiment, a method for detecting and evaluating sleep quality of an electronic watch may specifically include:

[0018] S101. The three-axis acceleration sensor built into the electronic watch monitors the acceleration data of the wrist movement of the user due to turning over during sleep in real time, and determines whether there is a turning over movement through the acceleration data, so as to calculate and record the turning over information, including the turning over rate, angle and direction, and stores the collected turning over information in the electronic watch memory and synchronizes it to the cloud at regular intervals.

[0019] According to the acceleration data monitored by the real-time three-axis acceleration sensor, filtering is used to remove noise interference and obtain wrist movement information; according to the wrist movement information, it is determined whether there is a turning movement through a pattern recognition algorithm to obtain the time point of the turning movement; for the identified turning movement, the rate, angle and direction of the turning movement are determined by analyzing the change trend and peak value of the acceleration data to obtain the turning information; the obtained turning information is stored in the electronic watch memory in chronological order to form a turning movement sequence of the user during sleep; the turning information stored in the electronic watch is regularly uploaded to the cloud server through the wireless communication module for data synchronization and backup; after receiving the turning information uploaded by the electronic watch, the cloud server analyzes it in combination with other sleep data of the user, evaluates the sleep quality of the user, generates a sleep quality report, and obtains relevant sleep improvement prompt information based on the sleep quality report, and promotes the sleep quality report and the relevant sleep improvement prompt information to the user.

[0020] Exemplarily, the built-in three-axis acceleration sensor of the electronic watch monitors the acceleration data of the user's wrist during sleep in real time, and removes high-frequency noise interference by low-pass filtering the acceleration data to extract effective wrist movement information. According to the wrist movement information, the mean, variance, peak and other features of the acceleration are extracted, and these features are used to train the support vector machine model to realize the classification of wrist movements, determine whether there is a turning action, and obtain the time point of the turning action. For the identified turning action, the turning rate is calculated by integrating the acceleration data, and the turning angle and direction are calculated by the direction change of the acceleration data to obtain detailed turning information. The obtained turning information is stored in the electronic watch memory in chronological order to form a turning action sequence during the user's sleep. The electronic watch regularly uploads the stored turning information to the cloud server through the Bluetooth low-power module for data synchronization and backup to ensure the security and reliability of the data. After the cloud server receives the turning information uploaded by the electronic watch, it combines the user's other sleep data and uses the association rule mining algorithm to find the correlation and regularity between the user's sleep data, evaluate the user's sleep quality, and generate a sleep report. Users can access the cloud server through mobile applications or web pages to view their sleep turning data and sleep quality reports. The cloud server generates personalized sleep improvement suggestions and guidance based on the user's sleep data and turning situation, helping users understand their sleep conditions and improve sleep quality. The three-axis accelerometer built into the electronic watch collects the acceleration data of the user's wrist in real time at a sampling frequency of 50Hz, and each sampling point contains three axial acceleration values. The collected acceleration data is preprocessed, and a 20Hz low-pass filter is used to remove high-frequency noise interference and extract effective wrist movement information. Then, feature extraction is performed on the filtered acceleration data, and the acceleration mean, variance and peak value of each axis are calculated to form a 9-dimensional feature vector. The support vector machine model is used to classify the feature vector. The SVM model learns the optimal decision boundary through the training set, and divides the wrist movement into two categories: turning movement and non-turning movement, with an accuracy rate of more than 95%. For the identified turning action, the acceleration data of the three axes are integrated in a trapezoidal manner to calculate the change in wrist speed during the turning process and obtain the turning rate; the rotation angle and direction of the wrist during the turning process are calculated by the change in the direction of the acceleration data in three-dimensional space to obtain the turning angle and direction. The time point, speed, angle and direction of the identified turning action are stored in the flash memory of the electronic watch in chronological order, and each turning action record occupies 20 bytes of space. When the flash memory space of the electronic watch reaches 80%, the data is uploaded to the cloud server for backup through the low-power Bluetooth module, with a transmission rate of 1Mbps to ensure data security and reliability.The cloud server uses April association rule mining algorithm to analyze the correlation between the user's turning over data and sleep quality, find the frequently occurring turning over patterns and the corresponding sleep quality scores, and generate a personalized sleep report. Users can view the sleep report through the mobile app to understand their sleep status, and provide relevant sleep improvement tips, such as adjusting sleeping posture, optimizing sleeping environment, etc., to help users improve their sleep quality.

[0021] S102. After detecting the turning over action, the photoelectric volumetric pulse blood flow sensor emits a test beam to the skin surface, measures the amount of reflected light of the test beam, and uses a support vector machine model to evaluate skin relaxation and blood flow changes, and associates the turning over action to record the impact of the turning over action on skin compression and the impact on local skin blood circulation.

[0022] After monitoring the user's turning over action, the built-in photoelectric volumetric pulse blood flow sensor of the electronic watch emits a test light beam to the user's skin surface, and obtains optical characteristic data of the skin surface by measuring the change in the amount of light reflected back from the skin surface; based on the obtained optical characteristic data, a pre-trained support vector machine model is used to evaluate the skin's relaxation and local blood flow change to obtain a skin relaxation index and a blood flow change index; the evaluated skin relaxation index and blood flow change index are correlated with the monitored turning over action to determine the influence of the turning over action on the skin compression degree and the local skin blood circulation, and obtain the correlation rule between the turning over action and the skin state; the correlation analysis result is used to calculate the skin relaxation index and blood flow change index. The association rules between the turning over action and the skin condition are stored in the memory of the electronic watch to form a mapping relationship between the user's turning over habit and the skin health information; the electronic watch regularly uploads the stored association rules between the turning over action and the skin condition to the cloud server through the wireless communication module, and summarizes and analyzes them with the data of other users to obtain the influence of the turning over action on skin health; the cloud server generates sleeping posture optimization suggestions and skin care plans based on the influence of the turning over action and skin health obtained by the summary analysis, combined with personalized user data, and pushes the sleeping posture optimization suggestions and the skin care plans to the user, reminding the user to improve his sleeping habits and prevent skin compression and blood circulation problems.

[0023] Exemplarily, after the built-in photoelectric volumetric pulse blood flow sensor of the electronic watch detects the user's turning over action, it immediately emits a test beam to the user's skin surface, and obtains the optical property data of the skin surface, including reflectivity, absorptivity and scattering rate, by measuring the change in the amount of light reflected back from the skin surface. According to the acquired skin optical property data, a pre-trained support vector machine model is used, a radial basis kernel function is selected, and the skin optical property data is used as the model input. The skin laxity index and blood flow change index are obtained through model evaluation. The model is optimized through a large amount of training sample data to improve the accuracy of the evaluation. The skin laxity index and blood flow change index obtained by the evaluation are associated with the monitored turning over action. The Apriori association rule mining algorithm is used to mine the association rules between the turning over action and the skin state index according to the co-occurrence frequency between the two, such as "many turning over times and large angles → high skin laxity", etc., to judge the impact of the turning over action on the skin compression degree and local blood circulation. The association rules between turning over and skin condition obtained from the association analysis are stored in the memory of the electronic watch to form a mapping relationship between the user's turning over habits and skin health conditions, which is used for subsequent skin health assessments and sleep optimization suggestions. The electronic watch regularly uploads the stored association rules between turning over and skin conditions to the cloud server through the wireless communication module, and summarizes and analyzes them with the data of other users. By increasing the sample size, the accuracy and universality of association rule mining are improved. Based on the influence of turning over and skin health obtained from the summary analysis, the cloud server combines preset rules and personalized user data, uses rule reasoning and personalized recommendation algorithms to generate sleeping posture optimization suggestions for users, such as side sleeping, supine lying, etc., and skin care plans, such as using specific skin care products, massage, etc., and pushes them to users through mobile phone applications to guide users to improve their sleeping habits and prevent skin compression and blood circulation problems. The built-in photoelectric capacitance pulse blood flow sensor of the electronic watch uses a dual-wavelength LED light source of 660nm and 940nm with an emission power of 1mW. Within 50ms after detecting the user's turning over, the reflected light signal on the surface of the user's wrist skin is collected at a sampling frequency of 100Hz to obtain a reflected light intensity time series containing 20 data points. By performing a fast Fourier transform on the reflected light intensity time series, the frequency domain characteristics of the reflected light signal, including the fundamental frequency amplitude, second harmonic amplitude and phase, are extracted as skin optical property data. The pre-trained support vector machine model is used to evaluate the skin optical property data. The model uses a radial basis kernel function and determines the optimal parameters such as C=10,γ=0.1 through 5-fold cross validation. The average accuracy on the test set reaches more than 95%. The model outputs skin laxity indicators such as 0-100 points and blood flow change indicators such as -50% to +50%.The Apriori association rule mining algorithm was used to set the minimum support to 0.2 and the minimum confidence to 0.8 to mine the association rules between turning over actions and skin condition indicators, and the top-5 rules with the highest confidence were obtained, such as "Number of turning over>=10 times / hour&&Turning over angle>45 degrees→skin laxity decreases by 5 points". The association rules were converted into production rules in the form of IF-THEN and stored in the rule base of the electronic watch. The electronic watch uploads the rule base to the cloud server through the low-power Bluetooth module every day. The cloud server aggregates the rule bases of 1,000 users and uses the FP-Growth algorithm to mine frequent item sets. The support threshold is set to 0.05 and the confidence threshold is set to 0.7 to obtain association rules with strong universality. The cloud server matches the association rules with the preset rules to generate personalized sleep posture adjustment suggestions, such as "left side lying, turning over angle controlled within 30 degrees", etc., and reminds users through push notifications of mobile phone applications.

[0024] S103. The cloud dynamically evaluates the acceleration data and the measurement results of the photoelectric volumetric pulse blood flow sensor, and processes them through a fusion algorithm module, including comparing the data changes before and after each turning over action, calculating the tightness of the contact between the electronic watch and the skin, to determine whether the wearing position of the electronic watch has changed, and recording the timestamp of the wearing position change.

[0025] The acceleration data and the photoelectric volumetric blood flow sensor measurement results sent by the electronic watch are obtained, and the acceleration data and the photoelectric volumetric blood flow sensor measurement results are aligned according to the timestamp to form time series data; the time series data is preprocessed by using the Kalman filter algorithm to remove noise interference and obtain data features; the data features are cascaded through the fusion algorithm module to construct a multimodal feature vector; the multimodal feature vector is dynamically evaluated by using a sliding time window to obtain the statistical features in each time window; by comparing the changes in the statistical features of adjacent time windows, it is determined whether a turning action occurs. If the change of the statistical feature exceeds a preset threshold, it is determined that a turning action is detected, and the timestamp of the turning action is recorded; the rate of change of the contact area between the electronic watch and the skin is calculated based on the acceleration data and photoelectric volumetric pulse blood flow data before and after each turning action; if the rate of change exceeds a preset ratio, it is determined that the wearing position of the electronic watch has changed, and the timestamp of the wearing position change is recorded; the timestamp of the turning action is associated with the timestamp of the wearing position change to obtain the wearing position change caused by the turning action; the associated turning action and wearing position change data are statistically analyzed to obtain the turning habits and watch wearing habits, so as to form a user behavior portrait.

[0026] Exemplarily, the cloud server receives the acceleration data from the electronic watch and the measurement results of the photoelectric volumetric pulse blood flow sensor, aligns the two types of data according to the timestamp, and forms a unified time series data. The time series data is preprocessed, and the Kalman filter algorithm is used to remove noise interference and extract stable data features. Through the fusion algorithm module, the acceleration data and the photoelectric volumetric pulse blood flow data are respectively extracted with feature vectors, and then the two feature vectors are spliced ​​into a high-dimensional feature vector to construct a multimodal feature vector as the input of the dynamic evaluation. The dynamic evaluation module uses the sliding time window technology to perform segmented analysis on the multimodal feature vector. The length of each time window is 10 seconds, and the sliding step is 5 seconds. The statistical features in each time window, such as mean, variance, kurtosis, skewness, etc., are extracted. These features can reflect the overall distribution and change trend of the data. By comparing the statistical feature changes of adjacent time windows and combining thresholds such as mean change greater than 0.5g and variance change greater than 0.2g2, these thresholds are obtained through statistical analysis of a large amount of real data to determine whether a turning action has occurred. If the feature change exceeds the threshold, it is considered that a turning action has occurred and the timestamp of the turning action is recorded. According to the acceleration data and photoelectric volumetric blood flow data before and after the turning action, by analyzing the amplitude change of the photoelectric volumetric blood flow signal, the change rate of the contact area between the electronic watch and the skin is estimated. If the change rate exceeds 20%, it is determined that the wearing position of the electronic watch has changed, and the timestamp of the wearing position change is recorded. The identified turning action timestamp and the wearing position change timestamp are compared. If the difference between the two timestamps is less than 2 seconds, the two events are associated to obtain the wearing position change caused by the turning action. The associated turning action and wearing position change data are statistically analyzed to extract the user's behavioral characteristics, such as turning frequency, turning amplitude, wearing position change frequency, etc., to form the user's turning habits and watch wearing habits data. The electronic watch collects acceleration data and photoelectric volume pulse blood flow data at a frequency of 50Hz. Each data contains three-axis acceleration values ​​and pulse blood flow values, and the data volume is 3000 per minute. After receiving the data, the cloud server first aligns the timestamps, sorts the two types of data according to the timestamps, and linearly interpolates the missing data to form a complete time series data. Then, the second-order Kalman filter algorithm is used to filter the data at a frequency of 10Hz to remove high-frequency noise and motion artifacts and extract stable data features. 10-dimensional feature vectors are extracted from the filtered acceleration data and pulse blood flow data, including mean, variance, kurtosis, skewness, etc., and then the two feature vectors are spliced ​​into a 20-dimensional multimodal feature vector. A sliding time window with a length of 10 seconds and a step size of 5 seconds is used to perform segmented analysis on the multimodal feature vector, and each time window contains 500 data points.Calculate the statistical features such as mean, variance, kurtosis, skewness, etc. for the data in each time window, and compare them with the features of the previous and next windows. If the mean change is greater than 0.5g and the variance change is greater than 0.2g2, it is judged that a turning action has occurred, and the timestamp of the turning action is recorded. At the same time, by analyzing the amplitude change of the pulse blood flow signal before and after turning over, if the amplitude change exceeds 30%, it is judged that the wearing position of the electronic watch has changed, and the timestamp of the wearing position change is recorded. Perform time series analysis on the turning action timestamp and the wearing position change timestamp. If the time difference between the two timestamps is less than 2 seconds, they are associated as the same event, indicating that the turning action causes the change of the wearing position. Perform statistical analysis on the associated event sequence, and obtain the user's turning frequency of 10 times per hour, the turning amplitude of 30°-60°, and the wearing position change frequency of 1 time per hour, forming the user's turning habit and watch wearing habit data, providing a basis for subsequent sleep quality analysis and health monitoring.

[0027] S104. The cloud uses the turning over information, skin laxity data and watch wearing position information to establish a turning over impact model, analyzes the correlation between different turning over actions and skin laxity and watch wearing position, and analyzes the impact on heart rate collection, so that the electronic watch can adjust the weight coefficient in the heart rate monitoring unit, so that the heart rate monitoring unit dynamically adjusts the heart rate signal processing parameters based on the turning over action, and optimizes the heart rate monitoring unit's collection of heart rate signals.

[0028] The turning-over information, skin looseness data and watch wearing position information sent by the electronic watch are obtained, and classified according to user information and timestamp to form a user behavior database; feature extraction is performed on the data in the user behavior database to obtain turning-over action features, skin looseness features and watch wearing position features; a turning-over influence model is established using a support vector machine algorithm, with the turning-over action features, skin looseness features and watch wearing position features as inputs and the heart rate collection quality as output, and the weight coefficients of the skin looseness features and the watch wearing position features on the heart rate collection quality are obtained through training; different turning-over actions and skin looseness are analyzed according to the turning-over influence model. The correlation between the turning over influence model and the watch wearing position is obtained to obtain the law affecting the heart rate collection; the analysis result of the turning over influence model is converted into the parameter adjustment strategy of the heart rate monitoring unit to generate a parameter adjustment matrix; the parameter adjustment matrix is ​​sent to the heart rate monitoring unit of the electronic watch, and the heart rate monitoring unit searches for the parameter adjustment coefficient according to the detected turning over action, skin laxity and watch wearing position, and dynamically adjusts the heart rate signal processing parameters; the collected heart rate signal is processed using the optimized heart rate signal processing parameters; the optimized heart rate data is sent to the cloud server, and the cloud server performs quality assessment on the heart rate data, and uses the assessment result to iteratively optimize the turning over influence model.

[0029] Exemplarily, the cloud server receives the turning information, skin looseness data and watch wearing position information from the electronic watch, and classifies these data according to user information and timestamp to form a user behavior database. Feature extraction is performed on the data in the behavior database to obtain the type, amplitude, frequency and other characteristics of the turning action, the mean and variance of the skin looseness, and the change of the watch wearing position. The support vector machine algorithm is adopted, the radial basis kernel function is selected, and the optimal penalty coefficient C and kernel function parameter γ are determined by grid search and cross-validation. The turning action features, skin looseness features and watch wearing position features are used as inputs to establish an initial support vector machine model, and the heart rate collection quality is used as output to obtain the weights of the influence of different feature combinations on the heart rate collection quality, and obtain the final turning influence model. According to the turning influence model, the correlation between different types of turning actions and skin looseness and watch wearing position is analyzed to obtain a quantitative influence rule, such as when the turning amplitude is greater than 60° and the duration exceeds 10 seconds, the watch wearing position deviates from the standard position by more than 20°, and the heart rate collection quality decreases by more than 30%. The analysis results of the turning-over influence model are converted into the parameter adjustment strategy of the heart rate monitoring unit to generate a parameter adjustment matrix. The rows of the matrix represent the turning-over action type, the columns represent the combination of skin looseness and watch wearing position, and the matrix elements represent the adjustment coefficients of the heart rate signal processing parameters in the corresponding situation. The parameters include filter order, cutoff frequency, gain coefficient, etc. These parameters are closely related to the quality of heart rate acquisition. For example, the higher the filter order, the better the smoothness of the heart rate signal, but the real-time performance will decrease. The cloud server sends the parameter adjustment matrix to the heart rate monitoring unit of the electronic watch. The heart rate monitoring unit searches for the corresponding adjustment coefficients in the parameter adjustment matrix according to the turning-over action, skin looseness and watch wearing position detected in real time, and dynamically adjusts the heart rate signal processing parameters. The heart rate monitoring unit applies the optimized heart rate signal processing parameters to the heart rate acquisition process, and performs filtering, denoising and other processing on the collected original heart rate signal to obtain high-quality heart rate data with high signal-to-noise ratio and small baseline drift, thereby improving the accuracy and reliability of heart rate monitoring. The electronic watch regularly uploads high-quality heart rate data to the cloud server. The cloud server uses the high-quality heart rate data, combined with the user's personal information and health records, to evaluate the user's health status, predict disease risks and send abnormal condition alerts, and provide users with personalized health management services. At the same time, the cloud server evaluates the quality of the heart rate data, feeds the evaluation results back to the turning-over influence model, iteratively optimizes the turning-over influence model, and continuously improves the prediction performance of the turning-over influence model and the effectiveness of the parameter adjustment strategy. The cloud server receives electronic watch data from 1,000 users every day. On average, each user generates 100 turning information, 500 skin laxity data, and 50 watch wearing position information every day.Feature extraction was performed on these data to obtain 12 features of turning over, such as the number of turning over, average turning over amplitude, maximum turning over angle, etc., 5 features of skin laxity, such as the mean, maximum, and minimum values ​​of laxity, and 3 features of the watch wearing position, such as the wearing position deviation and deviation duration. The support vector machine algorithm was used to search for the optimal parameter combination on a grid with a penalty coefficient C of 10 to the power of -2 to 102 and a kernel function parameter γ of 2-3 to 23 through 5-fold cross validation, and the average accuracy of the optimal model was 85%. The model was applied to the feature data of each user to obtain the weight matrix of the influence of turning over, skin laxity, and watch wearing position on the quality of heart rate collection. For example, the weight of turning over more than 20 times / day and the average turning amplitude greater than 45° is -0.3, indicating a significant negative impact on the quality of heart rate collection. According to the weight matrix, the parameter adjustment strategy of the heart rate monitoring unit is generated, and a 3D matrix is ​​designed. The dimensions are 5 types of turning action types, 3 levels of skin looseness, and 4 levels of watch wearing position deviation, with a total of 60 combinations. Each combination corresponds to a set of adjustment coefficients for heart rate signal processing parameters, such as increasing the filter order by 20%, reducing the cutoff frequency by 10%, etc. The strategy matrix is ​​sent to the electronic watch. When it is detected that the user turns over 18 times / day, the average turning amplitude is 50°, the skin looseness is 2 levels, and the watch wearing position deviation is 3 levels, the corresponding parameter adjustment coefficient is obtained by looking up the table, and the filter order for heart rate signal processing is increased from 10th order to 12th order, and the cutoff frequency is reduced from 0.5Hz to 0.45Hz. The heart rate signal is processed, and the signal-to-noise ratio of the output heart rate data is increased from 20dB to 25dB, the baseline drift is reduced from ±5% to ±3%, and the accuracy of heart rate monitoring is increased from 95% to 98%. The cloud server receives 50GB of high-quality heart rate data every day, assesses the health status of the data, and uses anomaly detection algorithms to identify that 5% of users are at risk of arrhythmia, tachycardia, etc., generates health reports and pushes them to users. At the same time, it fine-tunes the parameters of the turning over impact model to increase the average accuracy of the model to.

[0030] S105. The electronic watch uses the optimized heart rate monitoring unit to periodically collect the user's heart rate data through the photoelectric volumetric blood flow sensor; after detecting the turning over action, the photoelectric volumetric blood flow sensor increases the frequency of collecting the heart rate data.

[0031] After the heart rate monitoring unit of the electronic watch is optimized, the sampling rate, filter order and cutoff frequency of the photoelectric volumetric blood flow sensor are dynamically adjusted according to the heart rate signal processing parameters sent from the cloud; the heart rate data of the user is continuously collected; by analyzing the data of the three-axis acceleration sensor, the occurrence time, duration and amplitude of the turning action are identified, and if it is detected that the duration of the turning action exceeds the preset time threshold and the amplitude of the turning action is greater than the preset amplitude threshold, the collection frequency of the heart rate data is increased; after the working time of the photoelectric volumetric blood flow sensor in the high sampling rate mode reaches the preset threshold, the high-resolution heart rate data collected is extracted, the heart rate variability index is calculated, and the extracted feature data is uploaded to the cloud server, and according to the influencing rules and quantitative indicators analyzed by the cloud server, the sampling rate strategy of the photoelectric volumetric blood flow sensor in different scenarios is dynamically adjusted, and the optimized sampling rate strategy and heart rate monitoring results are regularly uploaded to the cloud for the user to view.

[0032] For example, after optimization, the heart rate monitoring unit of the electronic watch dynamically adjusts the sampling rate, filter order, cutoff frequency, etc. of the photoelectric volumetric pulse blood flow sensor according to the heart rate signal processing parameters sent from the cloud, and continuously collects the user's heart rate data at a sampling rate of 60 times per minute under normal conditions. The heart rate monitoring unit analyzes the collected heart rate data in real time, and determines whether the data quality meets the preset threshold requirements by calculating the statistical characteristics of the heart rate data such as the mean, variance, and peak value. If the data quality decreases, such as the signal-to-noise ratio is lower than 20dB, where the signal-to-noise ratio calculation formula is SNR=10log10(Ps / Pn), where Ps is the signal power and Pn is the noise power, the filter order and cutoff frequency are automatically adjusted, and the data is processed twice until the data quality meets the standard. The heart rate monitoring unit also monitors the user's turning over action. By analyzing the data of the three-axis acceleration sensor, the characteristics of the turning over action, such as the occurrence time, duration, and amplitude of the turning over action, are identified. According to the statistical analysis of a large amount of user data, it is found that when the turning over action lasts for more than 10 seconds and the amplitude is greater than 45°, the heart rate changes significantly, and it is necessary to increase the sampling rate for key monitoring. Therefore, when a turning over action that meets the above conditions is detected, the photoelectric volumetric pulse blood flow sensor is triggered to increase the sampling rate from 60 times per minute to 120 times per minute to obtain heart rate data with higher time resolution. After the photoelectric volumetric pulse blood flow sensor works in the high sampling rate mode for 30 seconds, the collected high-resolution heart rate data is feature extracted, and the heart rate variability indicators such as SDNN and RMSSD are calculated to reflect the dynamic changes of the heart rate during the turning over process. The sampling time is set to 30 seconds because according to the calculation requirements of the heart rate variability indicator, the sampling time needs to be at least 30 seconds to obtain enough data points for feature extraction, and the extracted feature data is uploaded to the cloud server. After receiving the high-resolution heart rate feature data, the cloud server compares and analyzes it with the heart rate feature data at the normal sampling rate. Through the support vector machine algorithm, the radial basis kernel function is used to determine the penalty coefficient C and the kernel function parameter γ through grid search and cross-validation, and the association model between heart rate changes and turning over actions is trained to obtain the influence of turning over actions on heart rate and quantitative indicators. The heart rate monitoring unit dynamically adjusts the sampling rate strategy of the photoelectric volumetric pulse blood flow sensor according to the influence rules and quantitative indicators obtained by cloud analysis. Different sampling rate combinations are used in different scenarios such as sleeping at night, sitting for a long time, and strenuous exercise. For example, a basic sampling rate of 30 times per minute and a turning over trigger sampling rate of 90 times per minute are used during sleep at night to take into account both data quality and energy consumption. These combinations are obtained through cluster analysis of a large amount of user data. After receiving the optimized parameters sent from the cloud, the heart rate monitoring unit of the electronic watch adjusted the sampling rate of the photoelectric volumetric pulse blood flow sensor from 50 times per minute to 60 times per minute, increased the filter order from 10th to 12th, and reduced the cutoff frequency from 0.5Hz to 0.4Hz.The heart rate monitoring unit conducts a quality assessment on the collected heart rate data every 30 seconds. By calculating the root mean square error, signal-to-noise ratio and other indicators of the data, when the RMSE exceeds 10% or the SNR is less than 15dB, the data quality is judged to be degraded, and the filter order is automatically increased to 15, the cutoff frequency is reduced to 0.3Hz, and the quality assessment cycle is shortened to 15 seconds. The heart rate monitoring unit uses a weighted sliding average algorithm to filter the acceleration data with a 10-second window. When the acceleration amplitude of three consecutive windows exceeds 0.5g and the duration exceeds 12 seconds, it is determined that a turning action has occurred, triggering the photoelectric volumetric pulse blood flow sensor to increase the sampling rate to 120 times per minute, and then restore the normal sampling rate after 35 seconds. The high-sampling heart rate data is extracted through wavelet transform to obtain 20 characteristic parameters such as energy, kurtosis, and frequency domain entropy, and uploaded to the cloud. The cloud uses the support vector machine algorithm to classify high and low sampling rate data, and determines the optimal parameters C = 10 and γ = 0.1 through 5-fold cross validation, and obtains that the heart rate change amplitude when turning over is more than 30% higher than the normal state. The cloud further uses the K-means clustering algorithm to divide the heart rate data of 1,000 users into five scenes, including sleep, sedentary, and exercise. The optimal sampling rate combination for each scene is: sleep, such as 30 times / minute, 100 times / minute, sedentary, such as 45 times / minute, 80 times / minute, exercise, such as 90 times / minute, 150 times / minute. The heart rate monitoring unit dynamically switches the sampling rate combination according to the current scene and the occurrence of the turning over action, maintains a high sampling rate within 1 minute after the turning over action occurs, and then restores the basic sampling rate of the corresponding scene. The optimized heart rate monitoring solution can increase the collection frequency of heart rate data by 20%, improve data quality by 15%, and reduce energy consumption.

[0033] S106. The electronic watch performs abnormal recognition processing on the heart rate data during the turning over period, analyzes the variability of the heart rate data, identifies non-physiological heart rate fluctuations, cleans and screens the heart rate data during the turning over period, and removes abnormal values.

[0034] When the turning over action is detected, the statistical features of the heart rate data in the preset time before and after the turning over are extracted to form a heart rate variability feature vector; a machine learning algorithm is used to perform abnormality detection on the heart rate variability feature vector, identify non-physiological heart rate fluctuation patterns, and mark the heart rate fluctuation patterns as abnormal values; the abnormal values ​​are processed by data cleaning and interpolation algorithms to remove the abnormal values; the cleaned heart rate data are reorganized in chronological order to form a heart rate change curve during the turning over period; a classification algorithm is used to match the heart rate change pattern feature space to identify the normal heart rate change pattern during the turning over period, and a time-frequency analysis is performed on the heart rate change segments that deviate from the normal pattern to construct a multi-dimensional feature vector of the abnormal segments; a clustering algorithm is used to cluster the abnormal segments to identify different types of abnormal heart rate change patterns; the heart rate abnormality detection and classification results are fed back to the user, and health advice is provided according to the abnormality type; at the same time, the abnormal heart rate data during the turning over period is uploaded to the cloud server, and the heart rate abnormality detection module is optimized by the incremental learning algorithm.

[0035] Exemplarily, the electronic watch continuously monitors the user's heart rate data. When a turning action is detected, the heart rate data of 1 minute before and after the turning is used as an analysis window, and the statistical features of the heart rate data in the window are extracted, including mean, variance, kurtosis, skewness, etc., to form a heart rate variability feature vector. The reason for selecting 1 minute as the analysis window is that according to medical research, the turning process of a normal person is usually completed within 1 minute. Selecting the data of 1 minute before and after the turning process can fully capture the heart rate changes during the turning process. The density-based local outlier detection algorithm is used to perform anomaly detection on the heart rate variability feature vector. The algorithm determines whether the data point is an outlier by calculating the density difference between each data point and its neighboring data points. The LOF value of the outlier is significantly higher than the normal value, and non-physiological heart rate fluctuation patterns, such as heart rate mutations and arrhythmias, are identified and marked as outliers. For the identified abnormal heart rate values, data cleaning and interpolation algorithms are used for processing, such as median filtering, Hampel filtering, etc., to eliminate the influence of outliers, and normal heart rate data at the position of outliers are filled by spline interpolation and other methods to ensure the continuity and integrity of the heart rate data. The cleaned heart rate data are reorganized in chronological order to form a heart rate change curve during turning over, and the morphological features of the curve, such as slope, curvature, inflection point, etc., are extracted to construct the feature space of the heart rate change pattern. The support vector machine algorithm is used to classify the heart rate change pattern feature space, and the radial basis kernel function is used to determine the optimal penalty coefficient C and kernel function parameter γ through grid search and cross-validation. The heart rate change pattern feature space is matched with the pre-marked normal heart rate change pattern, and the normal heart rate change pattern during turning over is identified. The heart rate change segment that deviates from the normal pattern is marked as an abnormal segment. The quantitative index of deviation is that the Euclidean distance from the normal pattern exceeds 1.5 times the standard deviation, or the heart rate change slope exceeds the normal range such as ±20 beats / min. Further time-frequency analysis is performed on the abnormal fragments. Wavelet transform and other methods are used to extract the time domain and frequency domain features of the abnormal fragments, such as energy distribution and spectrum entropy, to construct a multi-dimensional feature vector of the abnormal fragments. Density-based clustering algorithms such as DBSCAN are used to cluster the abnormal fragments to identify different types of abnormal heart rate change patterns. Time series databases such as InfluxDB are used to store abnormal heart rate data. Abnormal data is quickly retrieved and aggregated through tags and indexes to facilitate subsequent analysis and mining. The abnormal heart rate data is uploaded to the cloud server and summarized and analyzed with the abnormal heart rate data of other users. Incremental learning is performed through the online stochastic gradient descent algorithm. The parameters of the anomaly detection module are optimized through small batch gradient updates to adapt to new abnormal heart rate patterns and improve the real-time and adaptability of the anomaly detection module. The results of heart rate abnormality detection and classification are fed back to the user to remind the user that there may be a risk of abnormal heart rate during turning over, and corresponding health advice and medical advice are provided according to the type of abnormality.The electronic watch continuously collects the user's heart rate data at a frequency of 100 times per minute. When the root mean square value of the acceleration data is detected to exceed 0.8g and the duration exceeds 5 seconds, it is determined that a turning action has occurred. A total of 120 heart rate data points are automatically extracted for 1 minute before and after turning. The eight statistical features such as mean, variance, kurtosis and skewness are calculated to form the heart rate variability feature vector. The LOF algorithm is applied to the feature vector for anomaly detection. The number of neighborhood samples k=10 and the anomaly threshold are set to 2.5. The data points with LOF values ​​exceeding 2.5 are marked as outliers, and the corresponding heart rate value range is 40-160 beats / min. The outliers are median filtered with a sliding window length of 5, and then cubic spline interpolation is used to fill the outlier position to obtain the cleaned heart rate time series. The cleaned heart rate series is piecewise linearly fitted, with 20 data points as a segment, and six features such as slope, curvature and inflection point position of each segment are extracted to form the heart rate change pattern feature space. The support vector machine was used to perform binary classification on the feature space, and the radial basis kernel function was selected. The penalty coefficient C = 10 and the kernel function parameter γ = 0.1 were determined through 5-fold cross validation. The segments whose Euclidean distance exceeded 1.5 times the standard deviation of the normal mode mean or whose heart rate change slope exceeded ±20 beats / min were marked as abnormal segments. Wavelet transform was performed on the abnormal segments, and the db4 wavelet basis was used to decompose the 4 layers. Eight features such as energy distribution and spectrum entropy of each layer were extracted. The DBSCAN clustering algorithm was used to cluster the abnormal patterns, and the neighborhood radius ε = 0.5 and the minimum number of samples MinPts = 5 were set to divide the abnormal segments into 3 categories. The InfluxDB time series database was used to store abnormal heart rate data. Each data contained fields such as timestamp, heart rate value, abnormal tag and abnormal category, and indexed by user ID and date. When new abnormal heart rate data is received, the small batch SGD algorithm is used to update the anomaly detection module, with a batch size of 50, a learning rate of 0.01, and a regularization parameter of 0.001. The performance of the anomaly detection module is evaluated every 10 batches, and the iteration is stopped when the F1 value is improved by less than 0.1%. Finally, the anomaly detection results are pushed to the user, providing information such as the occurrence time, duration, and severity of the abnormal heart rate, and matching the corresponding health guidance and medical advice according to the abnormal category.

[0036] S107. The electronic watch combines the screened heart rate data with the user's historical sleep data to establish a user sleep quality assessment model, assess the user's sleep quality, generate a sleep report, and provide personalized sleep advice.

[0037] The historical sleep data of the user is obtained, and a sleep-heart rate data pair is formed with the screened heart rate data, and feature engineering is performed on the sleep-heart rate data pair; an unsupervised learning algorithm is used to perform cluster analysis on the feature vectors to obtain the corresponding relationship between the user's sleep stage and the heart rate, and a user sleep quality assessment model is established according to the corresponding relationship; a supervised learning algorithm is used to train the user sleep quality assessment model, and the hyperparameters of the user sleep quality assessment model are optimized through cross-validation and grid search to obtain an optimized user sleep quality assessment model; the sleep-heart rate data pair is input into the optimized user sleep quality assessment model to obtain a sleep quality score, and a sleep quality trend graph is generated; according to the sleep quality score and personal information, corresponding sleep suggestions are matched from the sleep knowledge module, a personalized sleep report is generated, and the report is pushed to the user; the user's sleep-heart rate data pair and the sleep quality score are regularly uploaded to the cloud server, the sleep quality assessment model is updated in the cloud by a transfer learning algorithm, and the updated sleep quality assessment model parameters are sent to the electronic watch to achieve continuous optimization of the sleep quality assessment model.

[0038] Exemplarily, the electronic watch reads the user's sleep data for the past week from the local storage, including the statistical features of sleep events such as the time of falling asleep, the duration of sleep, the number of tossing and turning, the number of body movements, the number of apnea, etc. These events can be identified by analyzing the fluctuation characteristics of acceleration data and heart rate data, and aligned with the screened heart rate data according to the timestamp to form a sleep-heart rate data pair. Feature engineering is performed on the sleep-heart rate data pair to extract sleep time statistical features such as the time of falling asleep, the time after waking up, the duration of sleep, etc., sleep event statistical features and heart rate statistical features such as the average heart rate, the maximum heart rate, the heart rate variability, etc., and construct a feature vector. The K-means clustering algorithm is used to perform cluster analysis on the feature vectors, and the optimal number of clusters K is determined by the elbow rule or the silhouette coefficient. According to the medical experts' prior knowledge of the heart rate and sleep event features of different sleep stages, the clustering results are mapped to the sleep stages such as light sleep, deep sleep, and rapid eye movement to establish a corresponding relationship. The support vector machine algorithm is used to train the user sleep quality assessment model, with the sleep-heart rate feature vector as input and the user's subjective sleep quality score as output. The user's subjective sleep quality score is collected through the questionnaire survey function of the mobile phone APP. The questionnaire includes dimensions such as sleep satisfaction, sleep depth, and wake-up refreshment. The user completes the questionnaire score after getting up every day. The model hyperparameters are optimized through cross-validation and grid search to obtain the optimal sleep quality assessment model. The user's sleep-heart rate data pairs for the past week are input into the trained sleep quality assessment model to obtain the daily sleep quality score and generate a sleep quality trend chart. The trend chart uses time as the horizontal axis and the sleep quality score as the vertical axis. It uses a line chart to show the user's sleep quality change trend in the past week, and uses different colors to mark different sleep quality levels, such as excellent, good, general, poor, etc. According to the user's sleep quality assessment results, combined with the user's age, gender, physical condition and other personal information, the corresponding sleep recommendations are matched from the sleep knowledge base. The sleep knowledge module contains the sleep health standards of different groups, the diagnosis and treatment of common sleep problems, and lifestyle recommendations for sleep health care. The knowledge is organized and stored in a structured and semantic way to generate a personalized sleep report. The generated sleep report is pushed to the user through the mobile phone APP. The user can view his or her sleep quality trend, sleep stage distribution and personalized sleep recommendations, and can also feedback his or her subjective sleep experience and score through the APP. The electronic watch regularly uploads the user's sleep-heart rate data and score to the cloud server. The cloud server aggregates the data of multiple users and uses transfer learning algorithms such as Fine-tune to fine-tune the sleep quality assessment model on a larger data scale to improve the robustness and generalization ability of the model. The updated model parameters are regularly sent to the electronic watch to achieve continuous optimization and personalized adaptation of the model.At 6 a.m. every morning, the electronic watch reads the user's sleep data for the past week from local storage, including 10 time statistical features such as falling asleep, waking up time, and sleep duration, and 8 event statistical features such as the number of turning over, the number of body movements, and the number of apnea. Sleep events are identified through algorithms such as threshold judgment of acceleration data and peak detection of heart rate data. The sleep data is aligned with the filtered heart rate data according to the timestamp, and 12 features such as average heart rate, maximum heart rate, and heart rate variability are extracted to form a 30-dimensional sleep-heart rate feature vector. After the feature vector is normalized by Z-score, the K-means clustering algorithm is used to calculate the optimal number of clusters K=4 through the elbow rule, corresponding to the four stages of light sleep, deep sleep, rapid eye movement, and wakefulness. According to the prior knowledge provided by medical experts, the clustering results are matched with the sleep stage labels with an accuracy of 85%. The sleep quality assessment model was trained using a support vector machine algorithm, with sleep-heart rate feature vectors as input and subjective sleep quality questionnaire scores of 10 dimensions such as sleep satisfaction, sleep depth, and wake-up refreshment completed by users through a mobile phone APP every day, with 5-level scores for each dimension as output. Through 5-fold cross-validation and 10×10 grid search, the optimal model parameters C=10 and γ=0.01 were obtained, with an average accuracy of 80%. The user's sleep-heart rate data for the past week was input into the trained model to obtain a daily sleep quality score of 0-100 points, and a line graph with time as the horizontal axis and score as the vertical axis was generated. The score was mapped to 4 levels such as excellent ≥90 points, good ≥80 points, average ≥60 points, and poor <60 points, marked in the graph with 4 colors: red, orange, yellow, and green. According to the user's sleep quality level, combined with the user's age, gender, BMI and other personal information, the 10 most relevant suggestions are matched from the knowledge base containing 500 sleep suggestions, such as "maintain a regular sleep schedule" and "avoid using electronic devices before bedtime", etc., to generate a personalized sleep report. Users can view the report in the mobile APP and provide feedback and ratings on the effectiveness of the suggestions. The electronic watch uploads the accumulated sleep-heart rate data (about 500KB) and user feedback data (about 50KB) to the cloud server every week. The cloud server aggregates the data of 1,000 users, uses the transfer learning algorithm to fine-tune the sleep quality assessment model, and adds new sleep suggestions to the knowledge base. The updated sleep quality assessment model parameters (about 2MB) are sent to the electronic watch every month to achieve continuous optimization of the sleep quality assessment model.

[0039] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features with similar functions disclosed in the present application (but not limited to) to form a technical solution.

Claims

1. A method for detecting and evaluating sleep quality of an electronic watch, characterized in that: The method comprises: The three-axis acceleration sensor built into the electronic watch monitors the acceleration data of the wrist movement caused by turning over during sleep in real time. It determines whether there is a turning over movement through the acceleration data, calculates and records the turning over information, including the turning over speed, angle and direction, and stores the collected turning over information in the electronic watch memory and synchronizes it to the cloud at regular intervals. After detecting the turning over action, the photoelectric volumetric pulse blood flow sensor emits a test beam to the skin surface, measures the amount of reflected light from the test beam, and uses the support vector machine model to evaluate skin relaxation and blood flow changes, and associates the turning over action to record the effect of the turning over action on skin compression and the effect on local skin blood circulation; The cloud dynamically evaluates the acceleration data and the measurement results of the photoelectric volumetric blood flow sensor, and processes them through a fusion algorithm module, including comparing the data changes before and after each turning over action, calculating the tightness of the electronic watch's contact with the skin to determine whether the wearing position of the electronic watch has changed, and recording the timestamp of the wearing position change, including: aligning the acceleration data sent by the electronic watch and the measurement results of the photoelectric volumetric blood flow sensor according to the timestamp to form time series data; Using Kalman filtering algorithm to remove noise interference from the time series data to obtain data features; The data features are cascaded through a fusion algorithm module to construct a multimodal feature vector; Dynamically evaluating the multimodal feature vector using a sliding time window to obtain statistical features within each time window; By comparing the statistical feature changes of adjacent time windows, it is determined whether a turning action occurs. If the statistical feature changes exceed a preset threshold, it is determined that a turning action is detected and a timestamp of the turning action is recorded; Calculate the change rate of the contact area between the electronic watch and the skin based on the acceleration data and photoelectric volume flow data before and after each turning action. If the change rate exceeds a preset ratio, it is determined that the wearing position of the electronic watch has changed, and the timestamp of the wearing position change is recorded. Associating the timestamp of the turning over action with the timestamp of the wearing position change to obtain the wearing position change caused by the turning over action; Perform statistical analysis on the associated turning-over action and wearing position change data to obtain turning-over habits and watch wearing habits, forming a user behavior profile; The cloud uses the turning over information, skin laxity data and watch wearing position information to establish a turning over impact model, analyze the association between different turning over actions and skin laxity and watch wearing position, and analyze the impact on heart rate collection, so that the electronic watch can adjust the weight coefficient in the heart rate monitoring unit, so that the heart rate monitoring unit can dynamically adjust the heart rate signal processing parameters based on the turning over action, and optimize the heart rate monitoring unit's collection of heart rate signals; The electronic watch uses an optimized heart rate monitoring unit to periodically collect the user's heart rate data through a photoelectric volumetric blood flow sensor. After detecting a turning over action, the photoelectric volumetric blood flow sensor increases the frequency of collecting heart rate data. The electronic watch performs abnormal recognition processing on the heart rate data during turning over, analyzes the variability of the heart rate data, identifies non-physiological heart rate fluctuations, cleans and screens the heart rate data during turning over, and removes abnormal values; The electronic watch combines the filtered heart rate data with the user's historical sleep data to build a user sleep quality assessment model, evaluate the user's sleep quality, generate a sleep report, and provide personalized sleep recommendations.

2. The method according to claim 1, wherein: The three-axis acceleration sensor built into the electronic watch monitors the acceleration data of the wrist movement of the user's wrist caused by turning over during sleep in real time, and determines whether there is a turning over movement through the acceleration data, so as to calculate and record the turning over information, including the turning over speed, angle and direction, and store the collected turning over information in the electronic watch memory, and synchronize it to the cloud at regular intervals, including: According to the acceleration data monitored in real time by the three-axis acceleration sensor, filtering is used to remove noise interference to obtain wrist movement information; According to the wrist motion information, determining whether there is a turning over motion by means of a pattern recognition algorithm, and obtaining a time point when the turning over motion occurs; For the identified turning action, determining the speed, angle and direction of the turning action by analyzing the change trend and peak value of the acceleration data, and obtaining the turning information; The acquired turning over information is stored in the memory of the electronic watch in chronological order to form a turning over action sequence of the user during sleep; The turning information stored in the electronic watch is uploaded to the cloud periodically through the wireless communication module for data synchronization and backup; After the cloud receives the turning over information uploaded by the electronic watch, it analyzes it in combination with other sleep data of the user, evaluates the user's sleep quality, generates a sleep quality report, and obtains relevant sleep improvement prompt information based on the sleep quality report, and promotes the sleep quality report and the relevant sleep improvement prompt information to the user.

3. The method according to claim 1, wherein: After detecting the turning over action, the photoelectric volumetric pulse blood flow sensor emits a test beam to the skin surface, measures the amount of reflected light of the test beam, uses a support vector machine model to evaluate skin relaxation and blood flow changes, and associates the turning over action to record the effect of the turning over action on skin compression and the effect on local skin blood circulation, including: The built-in photoelectric volumetric pulse blood flow sensor of the electronic watch emits a test light beam to the user's skin surface after detecting the user's turning over action, and obtains the optical characteristic data of the skin surface by measuring the change in the amount of light reflected back from the skin surface; Based on the acquired optical characteristic data, a pre-trained support vector machine model is used to evaluate the skin laxity and local blood flow change to obtain a skin laxity index and a blood flow change index; The skin laxity index and blood flow change index obtained by evaluation are correlated with the monitored turning action to determine the influence of the turning action on the skin compression degree and local skin blood circulation, and obtain the correlation rule between the turning action and the skin condition; The association rules between the turning over action and the skin condition obtained by the association analysis are stored in the memory of the electronic watch to form a mapping relationship between the user's turning over habit and the skin health information; The electronic watch regularly uploads the stored association rules between turning over action and skin condition to the cloud through the wireless communication module, and summarizes and analyzes the data of other users to obtain the influence of the turning over action on skin health; The cloud generates sleeping posture optimization suggestions and skin care plans based on the summarized and analyzed influence of turning over movements on skin health, combined with personalized user data, and pushes the sleeping posture optimization suggestions and skin care plans to users, reminding them to improve their sleeping habits and prevent skin compression and blood circulation problems.

4. The method according to claim 1, wherein: The cloud uses the turning over information, skin laxity data and watch wearing position information to establish a turning over influence model, analyze the association between different turning over actions and skin laxity and watch wearing position, and analyze the influence on heart rate collection, so that the electronic watch can adjust the weight coefficient in the heart rate monitoring unit, so that the heart rate monitoring unit can dynamically adjust the heart rate signal processing parameters based on the turning over action, and optimize the heart rate monitoring unit to collect the heart rate signal, including: Obtaining the turning over information, skin laxity data and watch wearing position information sent by the electronic watch, classifying them according to user information and timestamps, and forming a user behavior database; Extracting features from the data in the user behavior database to obtain turning-over action features, skin laxity features, and watch wearing position features; A support vector machine algorithm is used to establish a turning over influence model, with the turning over action feature, the skin looseness feature and the watch wearing position feature as input, and the heart rate collection quality as output, and the influence weight coefficient of the skin looseness feature and the watch wearing position feature on the heart rate collection quality is obtained through training; Analyze the correlation between different turning movements and skin laxity and watch wearing position according to the turning-over influence model to obtain the rules affecting heart rate collection; Converting the analysis result of the turning over influence model into a parameter adjustment strategy of the heart rate monitoring unit to generate a parameter adjustment matrix; The parameter adjustment matrix is ​​sent to the heart rate monitoring unit of the electronic watch, and the heart rate monitoring unit searches for parameter adjustment coefficients according to the detected turning over action, skin looseness and watch wearing position, and dynamically adjusts the heart rate signal processing parameters; Using optimized heart rate signal processing parameters to process the collected heart rate signal; The optimized heart rate data is sent to the cloud, and the cloud performs quality assessment on the heart rate data, and uses the assessment result to iteratively optimize the turning over influence model.

5. The method according to claim 1, wherein: The electronic watch uses the optimized heart rate monitoring unit to periodically collect the user's heart rate data through the photoelectric volumetric blood flow sensor. After detecting the turning over action, the photoelectric volumetric blood flow sensor increases the frequency of collecting the heart rate data, including: After the heart rate monitoring unit of the electronic watch is optimized, the sampling rate, filter order and cutoff frequency of the photoelectric volumetric pulse blood flow sensor are dynamically adjusted according to the heart rate signal processing parameters sent from the cloud; Continuously collecting heart rate data of the user; By analyzing the data of the three-axis acceleration sensor, the occurrence time, duration and amplitude of the turning over action are identified, and if it is detected that the duration of the turning over action exceeds a preset time threshold and the amplitude of the turning over action is greater than a preset amplitude threshold, the frequency of collecting heart rate data is increased; After the working time of the photoelectric volumetric blood flow sensor in the high sampling rate mode reaches a preset threshold, feature extraction is performed on the collected high-resolution heart rate data, and the heart rate variability index is calculated. The extracted feature data is uploaded to the cloud, and according to the influencing rules and quantitative indicators obtained by the cloud analysis, the sampling rate strategy of the photoelectric volumetric blood flow sensor in different scenarios is dynamically adjusted, and the optimized sampling rate strategy and heart rate monitoring results are regularly uploaded to the cloud for the user to view.

6. The method according to claim 1, wherein: The electronic watch performs abnormal recognition processing on the heart rate data during turning over, analyzes the variability of the heart rate data, identifies non-physiological heart rate fluctuations, cleans and screens the heart rate data during turning over, and removes abnormal values, including: When the turning over action is detected, the statistical features of the heart rate data in a preset time before and after the turning over are extracted to form a heart rate variability feature vector; Using a machine learning algorithm to perform anomaly detection on the heart rate variability feature vector, identify non-physiological heart rate fluctuation patterns, and mark the heart rate fluctuation patterns as abnormal values; Processing the outliers by data cleaning and interpolation algorithms to remove the outliers; The cleaned heart rate data are reorganized in chronological order to form a heart rate change curve during turning over; Using a classification algorithm to match the heart rate change pattern feature space, identifying the normal heart rate change pattern during the turning over period, performing time-frequency analysis on the heart rate change segments that deviate from the normal pattern, and constructing a multi-dimensional feature vector of the abnormal segments; Clustering the abnormal segments using a clustering algorithm, identifying different types of abnormal heart rate change patterns, feeding back abnormal heart rate detection and classification results to the user, and providing health advice based on the abnormality type; At the same time, the abnormal heart rate data during turning over is uploaded to the cloud, and the abnormal heart rate detection module is optimized through an incremental learning algorithm.

7. The method according to claim 1, wherein: The electronic watch combines the screened heart rate data with the user's historical sleep data to establish a user sleep quality assessment model, assess the user's sleep quality, generate a sleep report, and provide personalized sleep suggestions, including: Acquire the historical sleep data of the user, form a sleep-heart rate data pair with the filtered heart rate data, and perform feature engineering on the sleep-heart rate data pair; An unsupervised learning algorithm is used to perform cluster analysis on feature vectors to obtain a corresponding relationship between the user's sleep stage and heart rate, and a user sleep quality assessment model is established based on the corresponding relationship; Using a supervised learning algorithm to train the user sleep quality assessment model, optimizing the hyperparameters of the user sleep quality assessment model through cross-validation and grid search, and obtaining an optimized user sleep quality assessment model; Inputting the sleep-heart rate data pair into the optimized user sleep quality assessment model to obtain a sleep quality score and generate a sleep quality trend graph; According to the sleep quality score and personal information, corresponding sleep suggestions are matched from the sleep knowledge module, a personalized sleep report is generated, and pushed to the user; The user's sleep-heart rate data pair and the sleep quality score are regularly uploaded to the cloud, a transfer learning algorithm is used to update the user's sleep quality assessment model in the cloud, and the updated user's sleep quality assessment model parameters are sent to the electronic watch to achieve continuous optimization of the user's sleep quality assessment model.

Images