A smart watch-based exercise heart rate monitoring method and device

Abstract

The present application relates to the field of heart rate monitoring, and particularly relates to a method and device for monitoring heart rate during exercise based on a smart watch. The method comprises the following steps: obtaining geographical position information of a user based on a smart watch, identifying a user exercise scene, and generating an exercise scene type; collecting original heart rate detection parameters, performing time series heart rate feature analysis, and constructing a heart rate feature set; performing deep exercise mode recognition on the exercise scene type and the heart rate feature set, and obtaining a user exercise mode; performing global heart rate situation awareness on the heart rate feature set, and constructing a global heart rate situation map; and performing comprehensive risk assessment and self-adaptive early warning decision-making on the global heart rate situation map according to the user exercise mode. The present application realizes self-adaptive analysis of a user heart rate state, and improves user exercise safety assessment efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of heart rate monitoring, and more particularly to a method and device for monitoring exercise heart rate based on a smartwatch. Background Technology

[0002] Heart rate monitoring is a core function of smartwatches, playing a crucial role, especially in the field of sports. Heart rate not only reflects a user's immediate health status but also accurately reflects key information such as exercise intensity, physical condition, fatigue level, and recovery progress. By monitoring and analyzing users' heart rate data in real time, smartwatches can provide feedback on users' status during exercise, helping them avoid overexertion, adjust their exercise plans in a timely manner, and provide important data for evaluating and optimizing exercise results.

[0003] However, with the widespread application of smartwatches in the sports and health field, traditional heart rate monitoring methods can no longer meet the multiple demands for personalization, real-time performance, and accuracy. Existing optical sensor-based heart rate monitoring technologies, while providing some real-time data, are often limited by sensor accuracy, external interference, and wearing position, making it difficult to provide stable and accurate heart rate data in complex exercise environments. Furthermore, most existing smartwatch heart rate monitoring systems only provide single heart rate data, failing to analyze multi-dimensional information such as exercise intensity and fatigue levels in real time. They also lack dynamic adjustments and personalized suggestions for different exercise scenarios, resulting in inaccurate monitoring results and making it difficult to provide users with comprehensive and scientific sports and health guidance. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method and device for monitoring exercise heart rate based on a smartwatch, thereby resolving at least one of the aforementioned technical issues.

[0005] To achieve the above objectives, the present invention provides a method for monitoring exercise heart rate based on a smartwatch, comprising the following steps:

[0006] Step S1: Based on the user's geographical location information obtained from the smartwatch, the user's movement scene is identified, thereby generating the movement scene type;

[0007] Step S2: Collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set;

[0008] Step S3: Perform deep recognition of exercise patterns based on exercise scene type and heart rate feature set to obtain the user's exercise pattern;

[0009] Step S4: Perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map;

[0010] Step S5: Perform a comprehensive risk assessment and adaptive early warning decision based on the user's exercise pattern of the global heart rate status map.

[0011] This specification provides a smartwatch-based heart rate monitoring device for performing the smartwatch-based heart rate monitoring method described above, comprising:

[0012] The scene recognition module is used to obtain the user's geographical location information based on the smartwatch, identify the user's movement scene, and generate the movement scene type.

[0013] The chip analysis module is used to collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set.

[0014] The motion recognition module is used to perform deep recognition of motion patterns based on motion scene type and heart rate feature set to obtain the user's motion pattern.

[0015] The heart rate situation awareness module is used to perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map.

[0016] The heart rate risk assessment module is used to perform comprehensive risk assessment and adaptive early warning decisions on the global heart rate status map based on the user's exercise pattern.

[0017] The beneficial effects of this invention are specifically as follows: By acquiring the user's geographical location information (GPS data), the smartwatch can identify exercise scenarios (such as running, road cycling, mountain biking, walking, etc.) in real time, thereby providing more accurate background information for subsequent heart rate analysis. Each exercise scenario has a different impact on heart rate, therefore accurate scenario identification helps in the customization of subsequent analysis. Different exercise scenarios (such as running and cycling) may elicit different heart rate responses. During running, heart rate fluctuations may be related to running speed and incline; while during cycling, heart rate may be closely related to leg strength and cycling posture. By accurately identifying the scenario, the system can adjust the heart rate monitoring algorithm to cope with different exercise needs. Furthermore, by collecting raw heart rate detection parameters (such as heart rate, heart rate variability, instantaneous heart rate, etc.) and performing time-series analysis, details of heart rate fluctuations can be captured. For example, during high-intensity exercise, the amplitude and frequency of heart rate fluctuations may reflect different physiological states. This information helps to more accurately analyze the body's response during exercise. The time-series characteristics of heart rate can reveal the user's exercise adaptability, fatigue state, and potential health risks. By constructing personalized heart rate feature sets, the system can tailor heart rate monitoring plans for each user to better reflect their exercise load and recovery status. Combining geolocation information (exercise scenario) and heart rate features, the smartwatch can perform deep recognition of the user's exercise patterns. For example, it can not only identify whether the user is running or cycling, but also further identify aspects such as exercise intensity, recovery status, and exercise efficiency. This in-depth analysis helps users comprehensively understand their exercise performance. Based on deeply recognized exercise patterns, the smartwatch can provide users with more targeted exercise feedback. For example, if the user is conducting high-intensity running training, the watch may suggest reducing speed or increasing recovery time; while during low-intensity walking, it may suggest maintaining the current intensity. The global heart rate status map, by combining the user's heart rate feature set, exercise scenario, and exercise pattern, can comprehensively perceive the user's heart rate status. Whether during or after exercise, the system can clearly display the overall trend of heart rate changes and mark key nodes of abnormal heart rate fluctuations. The global heart rate status map not only helps users track their current heart rate status, but can also issue warnings at critical moments. For example, if the heart rate suddenly increases or decreases sharply, the system will mark the abnormal fluctuation and remind the user to rest or adjust the intensity of exercise. Through real-time graphical display, users can clearly see their heart rate status at a glance, avoiding exercise injuries caused by ignoring heart rate changes. Based on the user's exercise pattern and global heart rate status graph, the system can perform personalized exercise risk assessments. Since everyone's exercise tolerance is different, the system will comprehensively consider factors such as the user's exercise pattern and heart rate fluctuations to assess whether the exercise is excessive and whether there are risks such as abnormal heart rates. Through intelligent algorithms, the system can automatically adjust the warning threshold based on real-time heart rate changes and exercise patterns.For example, the system can set a higher upper limit for heart rate for an experienced athlete, while a lower threshold may be set for an average user. Furthermore, the system can provide dynamic warnings and decisions, such as slowing down the exercise, pausing the workout, or offering recovery suggestions, to prevent the body from being burdened by excessively high or low heart rates. Comprehensive risk assessment not only helps users adjust in real time during exercise but also provides comprehensive health management advice before and after exercise. This personalized health management not only improves exercise performance but also effectively prevents sports injuries and enhances exercise safety. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of a smartwatch-based method for monitoring exercise heart rate according to the present invention.

[0019] Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1.

[0020] Figure 3 This is a detailed flowchart illustrating the implementation steps of step S2;

[0021] Figure 4 This is a flowchart illustrating the detailed implementation steps of step S3. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] This application provides a method and device for monitoring exercise heart rate based on a smartwatch. The execution entity of the method and device includes, but is not limited to, mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio / image management system, an information management system, and a cloud data management system.

[0024] Please see Figures 1 to 4 This invention provides a method for monitoring exercise heart rate based on a smartwatch, comprising the following steps:

[0025] Step S1: Based on the user's geographical location information obtained from the smartwatch, the user's movement scene is identified, thereby generating the movement scene type;

[0026] Step S2: Collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set;

[0027] Step S3: Perform deep recognition of exercise patterns based on exercise scene type and heart rate feature set to obtain the user's exercise pattern;

[0028] Step S4: Perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map;

[0029] Step S5: Perform a comprehensive risk assessment and adaptive early warning decision based on the user's exercise pattern of the global heart rate status map.

[0030] In the embodiments of the present invention, see Figure 1 The diagram below illustrates the steps of a smartwatch-based exercise heart rate monitoring method according to the present invention. In this example, the steps of the smartwatch-based exercise heart rate monitoring method include:

[0031] Step S1: Based on the user's geographical location information obtained from the smartwatch, the user's movement scene is identified, thereby generating the movement scene type;

[0032] In this embodiment, the high-precision GNSS positioning module built into the smartwatch (supporting multiple systems such as GPS / GLONASS / Galileo) is used to acquire real-time latitude and longitude coordinates and altitude information at a sampling frequency of 1 Hz-5 Hz during the user's movement. First, the system calculates the user's moving speed (m / s) and speed change (m / s²) using the sliding window method, and combines this with trajectory curvature analysis to determine whether the user is moving in a straight line, circling a curve, or moving uphill or downhill. For example, in a 10-minute outdoor running experiment, if the speed is stable at 2.5-3.0 m / s and the altitude change is less than 5 m, it is initially determined to be jogging on flat ground; if the trajectory curvature is high and the turning is frequent, it may be jogging on an indoor basketball court or park path. At the same time, the system also obtains the temperature, humidity, wind speed, and weather conditions of the current location by calling a third-party meteorological API, and performs fusion analysis with the illuminance values ​​(lux) collected by the light sensor to construct an environmental feature vector. By using machine learning classification models (such as random forest or XGBoost), the system integrates trajectory features, speed features, and environmental features in a multi-dimensional manner to achieve scene recognition, such as "running in an indoor gym," "cycling in an outdoor park," and "hiking in the mountains." Experimental data shows that this method achieves an accuracy of 94.2% in a sample set containing 12 different sports scenes.

[0033] Step S2: Collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set;

[0034] In this embodiment, the raw heart rate signal (in bpm) during user movement is continuously acquired using the photoplethysmography (PPG) sensor of the smartwatch at a sampling rate of 25 Hz to 50 Hz. The raw signal may be affected by motion artifacts, wrist sway, unstable skin contact, etc., therefore preprocessing is required first. This includes 0.5-4 Hz bandpass filtering to remove noise below the respiratory rate and high frequencies, and peak detection based on an adaptive threshold method to correct missed or false heartbeats. Then, the attitude change rate is calculated using triaxial acceleration data from the inertial measurement unit (IMU), and dynamic gain correction is applied to the heart rate signal to reduce attitude-induced heart rate deviation. Next, the corrected heart rate data is divided into multi-scale time windows (e.g., 15 s, 60 s, 300 s), and statistical features (mean, standard deviation, coefficient of variation), dynamic features (peak value, trough value, heart rate range), and time-frequency features (power spectral density based on FFT) are extracted to form a heart rate feature vector.

[0035] Step S3: Perform deep recognition of exercise patterns based on exercise scene type and heart rate feature set to obtain the user's exercise pattern;

[0036] In this embodiment, exercise scene information is fused with heart rate features to identify the user's specific exercise mode (such as "interval running," "steady-paced long-distance running," "uphill sprinting," "endurance cycling," etc.). The system first inputs the scene type obtained from S1 as a context label into a deep neural network model (such as a dual-input LSTM network), and inputs the heart rate feature set from S2 as temporal features into another branch, finally merging the outputs in the fusion layer. The LSTM model can capture the temporal dependence of heart rate changes; for example, the heart rate during interval running shows a "sawtooth" rise and fall, while steady-paced long-distance running shows a flat, high-plateau state. In experiments, for 10 different exercise modes (including multiple exercise methods in the same scene), the method achieved an average recognition accuracy of 92.7% under 5-fold cross-validation. In addition, the model can also detect exercise mode switching, such as switching from jogging on flat ground to uphill sprinting, with a switching delay of approximately 3-5 seconds. This deep recognition capability is crucial for subsequent heart rate risk assessment and adaptive warning.

[0037] Step S4: Perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map;

[0038] In this embodiment, a secondary stability analysis is performed on the heart rate feature set, including heart rate rise time (the time required to reach the target heart rate range from a resting state), stable duration (the duration the heart rate is maintained within the target range), and recovery time (the time it takes for the heart rate to return to a resting state after exercise). For example, in a 10km run by a 30-year-old subject, the rise time was 210 seconds, the stable duration was 36 minutes, and the recovery time was 420 seconds. Subsequently, heart rate fluctuation pattern recognition is performed, including indicators such as peak-trough cycle, fluctuation frequency, and amplitude changes. Wavelet transform is used to analyze the multi-scale dynamic features of heart rate, extracting pattern labels such as "stable," "oscillating," and "step-like rise." Finally, a global heart rate trend map is constructed by combining heart rate change trend mining (identifying rising, falling, stable, and oscillating states) to visualize the dynamic heart rate throughout the exercise.

[0039] Step S5: Perform a comprehensive risk assessment and adaptive early warning decision based on the user's exercise pattern of the global heart rate status map.

[0040] In this embodiment, the exercise pattern obtained in S3 is matched and analyzed with the global heart rate status map in S4. For example, in the "high-intensity interval running" mode, the upper limit of the allowable heart rate is higher, but the threshold requirements for the rate of rise and recovery time are stricter; while in the "endurance cycling" mode, more attention is paid to the risk of load accumulation in the medium-to-high heart rate range for a long time. Risk assessment adopts a multi-parameter weighted method, and the weights are automatically adjusted according to the exercise mode. Taking a set of experiments as an example, the system assessed the risk index of 15 subjects for 30 minutes of training in different modes. The results showed that when the risk level was high (index > 75), users had a high risk of cardiovascular load if they continued to exercise. The warning strategies include vibration prompts, voice reminders, screen alerts, etc., and the trigger conditions can be automatically adjusted according to individual heart rate responses. For example, when it is detected that a user is sprinting uphill in hot weather and the heart rate has exceeded the personalized safety limit of 10 bpm, the system will issue a warning in advance. This adaptive warning mechanism can reduce the duration of users in high-risk states by more than 30%, significantly improving exercise safety.

[0041] In this embodiment, see Figure 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include:

[0042] Based on smartwatches, user location information and ambient light signals are obtained;

[0043] Calculate the user's real-time altitude and latitude / longitude coordinates based on the user's geographic location information;

[0044] Based on the user's geographic location information, meteorological parameters for the current location are collected to obtain the meteorological parameters for the current location;

[0045] Perform environmental meteorological analysis on meteorological parameters at the current location to generate environmental meteorological characteristics;

[0046] The ambient light intensity is calculated from the ambient light signal, and the intensity variation is analyzed to obtain an ambient light intensity distribution map.

[0047] The system identifies user movement scenarios by analyzing real-time altitude, latitude and longitude coordinates, environmental meteorological characteristics, and ambient light intensity distribution maps, thereby generating movement scenario types.

[0048] In this embodiment, the smartwatch's built-in GNSS (Global Navigation Satellite System) module, such as GPS, BeiDou, or GLONASS, receives satellite signals in real time to obtain the user's geographical location information. To improve positioning accuracy, multi-constellation joint positioning is typically used, and A-GPS (Assisted GPS) technology is employed when the signal is weak, downloading ephemeris data via cellular networks to accelerate positioning. In the experiment, the positioning sampling frequency can be set to 1Hz (once per second) to ensure the continuity and timeliness of location data in motion scenarios. Simultaneously, the smartwatch's built-in light sensor (usually a visible light sensor based on photodiodes) synchronously collects the current lighting environment signal. This signal is generally recorded in lux, a physical quantity of light intensity, and the sampling frequency can be consistent with the location sampling. To reduce light measurement fluctuations caused by the wearer's wrist movements, a simple low-pass filter can be added to the watch firmware, such as using an IIR filter with a cutoff frequency of 0.5Hz to smooth the light signal. Finally, the raw data output in this step includes a timestamp, latitude and longitude, raw light intensity values, and positioning quality indicators (such as HDOP value). These data will serve as the foundational input for subsequent altitude calculations, meteorological data collection, and illumination intensity analysis. The user's real-time altitude and precise latitude and longitude are calculated using the raw three-dimensional coordinates (longitude, latitude, and ellipsoidal height) from the GNSS positioning data. The altitude value provided by the GNSS system is ellipsoidal height, which needs to be converted to orthometric height (what we commonly refer to as altitude) using geoid models such as EGM96 or EGM2008. In the experiment, an embedded geographic coordinate transformation algorithm can be used. Inputting latitude, longitude, and ellipsoidal height, the algorithm uses bilinear interpolation to find the corresponding geoid difference in the Earth's gravity potential model, achieving an accuracy of ±0.5 meters. In dynamic scenarios (such as running or cycling), altitude data may fluctuate due to GNSS multipath effects. Therefore, Kalman filtering can be used for smoothing. The filtering parameters can be set to 0.1 m² for process noise and 2 m² for measurement noise to balance response speed and stability. Latitude and longitude coordinates are obtained by directly parsing GNSS NMEA protocol data and then converting it from WGS84 to the required projected coordinate system (such as GCJ-02 or UTM). The final output of this step is a time-synchronized, accurate latitude, longitude, and altitude sequence, providing spatial reference for subsequent location-based meteorological data matching and scene recognition.

[0049] Meteorological parameters for the current location are obtained through two methods: the first is in-situ measurement using the smartwatch's built-in environmental sensors (such as temperature, humidity, and barometric pressure sensors); the second is querying the latest meteorological data based on latitude, longitude, and altitude via a mobile network by calling an online meteorological service API (such as the National Meteorological Center or OpenWeatherMap). In the experiment, it is recommended to request meteorological data every 60 seconds to balance real-time performance and power consumption. The collected meteorological parameters typically include: temperature (°C), relative humidity (%RH), atmospheric pressure (hPa), wind speed (m / s), wind direction (°), and precipitation (mm / h). If local sensors are used, the smartwatch's MEMS barometer (resolution 0.01hPa, corresponding to an altitude change of approximately 0.08m) combined with a temperature sensor (accuracy ±0.3°C) and a humidity sensor (accuracy ±2%RH) can be used for real-time data acquisition. Under experimental conditions, a moving buffer can also be set to perform a moving average of the meteorological data from the past 5 minutes to reduce the interference of instantaneous fluctuations on subsequent analysis. The output of this step is a structured meteorological parameter dataset, stored in association with geographic coordinates and altitude. Feature extraction is performed on the collected meteorological parameters to form environmental meteorological feature vectors that can be used for subsequent pattern recognition. First, continuous variables such as temperature, humidity, and atmospheric pressure are statistically analyzed, including mean, standard deviation, and rate of change. For example, in the experiment, the rate of temperature change can be defined as ΔT / Δt, and a significant change in ambient temperature is considered to occur when ΔT / Δt > 0.2℃ / min. Second, wind speed and direction are vectorized, converting wind direction to radians and calculating the average values ​​of the north-south and east-west components to obtain the wind field characteristics. For precipitation, short-term accumulation (e.g., the past 15 minutes) can be used to determine whether precipitation is occurring. To address the influence of different dimensions of meteorological parameters, all features need to be normalized, for example, using Z-score standardization to ensure a mean of 0 and a standard deviation of 1.

[0050] Calibration is performed using the raw signal (in Lux) from the light sensor to correct errors caused by sensor angle variations, environmental occlusion, and temperature drift. In the experiment, multi-point calibration can be performed under known light intensity environments (e.g., standard laboratory lighting at 500 Lux) to establish a linear or polynomial mapping relationship between the light signal and the actual light intensity. The calibrated light intensity data is arranged in a time series and spatially mapped using user location data. To analyze the trend of light intensity changes, the rate of change of light intensity within a sliding window can be calculated. For example, with a 10-second window, if the rate of change exceeds 50 Lux / s, it indicates that the user has entered or left a significantly dark environment (such as a tunnel or tree-lined path). In experimental visualization, light intensity is bound to latitude and longitude coordinates, and spatial interpolation algorithms (such as IDW inverse distance weighted interpolation or Kriging interpolation) are used to generate two-dimensional or three-dimensional light intensity distribution maps. These maps visually reflect the differences in light intensity at different locations. To improve image smoothness, the interpolation radius can be set to 50 meters, and Gaussian blurring is used to reduce the influence of outliers. The multi-source features obtained in the previous steps are fused and input into the motion scene recognition model to determine the type of motion environment the user is in. Common scene types include: outdoor running in sunny weather, outdoor cycling in cloudy or rainy weather, mountain hiking, and indoor gym training. Supervised machine learning algorithms (such as random forest, support vector machine (SVM), or lightweight neural networks) can be used for classification in the experiment. First, the altitude change rate, latitude and longitude movement speed, environmental meteorological feature vectors, and light intensity distribution features are combined into a unified feature vector; then, the classification model is trained using a pre-collected labeled dataset (e.g., 1000 hours of motion data from different scenes). To improve recognition accuracy, the features can be temporally modeled, for example, using a Long Short-Term Memory (LSTM) network to capture time dependencies. In the experimental verification, the dataset is divided into a training set (70%) and a test set (30%), and five-fold cross-validation is used to evaluate the model performance. The results show that after fusing altitude change, meteorological features, and light distribution information, the accuracy of distinguishing between indoor and outdoor scenes can exceed 95%, and the recognition accuracy for specific motion types (such as mountain cycling and plain running) is improved by about 8%.

[0051] In this embodiment, see Figure 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include:

[0052] The raw heart rate detection parameters are obtained by continuously collecting heart rate data using the pulse sensor built into the smartwatch.

[0053] The smartwatch collects the user's body posture parameters using its inertial sensing unit.

[0054] The user's body posture parameters are used to calculate the rate of change of posture, in order to generate the rate of change of posture.

[0055] Dynamic gain heart rate parameters are generated by applying attitude disturbance to the original heart rate detection parameters based on the attitude change rate.

[0056] Multi-scale time window division and feature statistics were performed on the dynamic gain heart rate parameters to construct a heart rate feature set.

[0057] In this embodiment, the smartwatch primarily relies on a built-in photoplethysmography (PPG) sensor for continuous heart rate acquisition. The PPG sensor typically consists of a green LED (wavelength approximately 525nm) and a photodiode receiver, detecting changes in blood volume with each heartbeat by transmitting light through the skin. Under experimental conditions, to balance power consumption and accuracy, the PPG sampling frequency can be set to 50Hz. This captures the complete waveform details of each heartbeat and allows subsequent noise reduction and peak detection algorithms to accurately identify the pulse signal. The raw light intensity signal is affected by motion artifacts, skin color, wearing tightness, and ambient light interference. Therefore, a front-end analog filter (such as a 0.5Hz~5Hz bandpass filter) needs to be added to the sensor firmware to remove DC components and high-frequency noise. After digitization by the ADC, the signal is stored as a sequence of raw pulse waveforms with a timestamp for alignment with subsequent attitude data. In the experiment, the accuracy of PPG heart rate acquisition can be verified by synchronous acquisition and comparison with medical-grade electrocardiogram (ECG) equipment. Typically, the error can be controlled within ±2 bpm in the resting state, while further algorithmic compensation is needed to maintain accuracy during movement. An inertial sensing unit (IMU) typically consists of a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, which can continuously acquire the user's wrist movement information in space. In the experiment, to ensure data synchronization with the PPG signal, the IMU sampling frequency can be set to 50Hz, consistent with the heart rate sampling frequency, facilitating fusion analysis. The accelerometer is used to capture the user's linear acceleration changes, the gyroscope records angular velocity information, and the magnetometer can be used for heading reference. During acquisition, the data from these three types of sensors are fused using quaternion or direction cosine matrix (DCM) algorithms to calculate the user's wrist attitude angles (such as pitch, roll, and yaw) in real time. To reduce drift error, complementary filtering (cutoff frequency 0.98) or extended Kalman filtering (process noise covariance Q = 0.001, measurement noise covariance R = 0.03) can be used to stabilize the attitude calculation in the experiment. The obtained attitude parameters not only reflect the user's exercise type (e.g., running, weightlifting, yoga), but can also be used in subsequent steps to determine the degree of interference to the PPG signal, making it one of the core inputs for dynamic heart rate correction.

[0058] The rate of change of posture is an important indicator that measures the magnitude of change in a user's posture per unit time. It can be used to assess motion intensity and the degree of interference with PPG signals. The specific calculation method involves performing first-order differences on continuously sampled posture angles (Pitch, Roll, Yaw) and then calculating their Euclidean norm, for example, Δθ = √[ + + The attitude change amplitude at each sampling point is obtained. In the experiment, to reduce the impact of high-frequency jitter, a moving average filter with a window length of 0.5 seconds can be applied before calculating the rate of change. The unit of attitude change rate is usually ° / s (angle change per second). During running, this value may remain between 50° / s and 100° / s, while at rest it will be below 5° / s. To accommodate users with different exercise intensities, a rate of change threshold can be set in the experimental design; for example, a rate of change exceeding 30° / s is considered significant, at which point the PPG signal may be significantly affected by motion artifacts. The final output attitude change rate sequence will be aligned with the original heart rate data for subsequent dynamic gain adjustment.

[0059] PPG signals are prone to baseline drift and waveform distortion at high posture change rates, leading to deviations in heart rate calculations. Therefore, dynamic gain adjustment of the original heart rate parameters based on the posture change rate is necessary to reduce the impact of motion artifacts. Specifically, a mapping relationship between posture change rate and heart rate error is established. For example, in experiments, the error gain coefficient k(Δθ) can be obtained through regression analysis. When the posture change rate increases, stronger smoothing and anti-interference processing is applied to the heart rate signal. The implementation of dynamic gain can be divided into two steps: First, the posture change rate is mapped to a gain factor (range 0.5~1.0), where a higher posture change rate corresponds to a lower gain to suppress transient abnormal peaks. Second, this gain factor is introduced into the heart rate calculation to perform a weighted average of the current and historical PPG peak intervals (RR intervals), thereby smoothing the output heart rate value. Experiments show that after adopting the dynamic gain method, the heart rate error during high-speed running (posture change rate 80° / s) can be reduced from ±15 bpm to ±6 bpm, significantly improving measurement stability during exercise.

[0060] Dynamic gain heart rate data is transformed into a feature set suitable for motion monitoring and pattern recognition. Multi-scale time window segmentation refers to analyzing heart rate data in segments at different time lengths (e.g., 5 seconds, 30 seconds, 1 minute, 5 minutes), capturing both instantaneous changes (short time window) and reflecting long-term trends (long time window). Within each time window, a series of statistical features are extracted, such as mean heart rate, standard deviation of heart rate (SDNN), coefficient of variation of heart rate (CV), maximum / minimum heart rate, and rate of heart rate rise. In experiments, frequency domain features, such as the low-frequency / high-frequency power ratio (LF / HF), can also be calculated to assess autonomic nervous system activity. To adapt to real-time monitoring needs, a sliding window calculation method can be used, updating features every 1 second. The resulting heart rate feature set can be used not only for exercise load assessment but also as input for motion scene recognition or health risk warning models. Experimental results show that multi-scale feature fusion improves accuracy in exercise type classification by approximately 7% compared to single-time-scale analysis.

[0061] In this embodiment, the specific steps for performing multi-scale time window division and feature statistics on the dynamic gain heart rate parameters to construct a heart rate feature set are as follows:

[0062] The dynamic gain heart rate parameters are divided into multiple time windows to generate heart rate monitoring parameters for multiple time windows; the heart rate monitoring parameters for multiple time windows include a 15-second short cycle window, a 60-second medium cycle window, and a 300-second long cycle window;

[0063] Calculate the mean heart rate, standard deviation, and coefficient of variation of the heart rate monitoring parameters to obtain the baseline heart rate statistics;

[0064] The maximum and minimum heart rate values ​​of heart rate monitoring parameters in multiple time windows are detected window by window, and the peak and trough values ​​of heart rate in different cycles are extracted.

[0065] The dynamic range and fluctuation amplitude of heart rate are calculated based on the peak and trough values ​​of heart rate.

[0066] The heart rate monitoring parameters are decomposed into time and frequency components, and the power spectral density is calculated to obtain power spectral densities at multiple scales.

[0067] Power fluctuation difference analysis was performed based on power spectral density to obtain heart rate variability characteristics;

[0068] A heart rate feature set is constructed by fitting the time-series distribution of the baseline heart rate statistics, heart rate variability characteristics, heart rate dynamic range, and fluctuation amplitude.

[0069] In this embodiment, the dynamic gain heart rate parameters, corrected for attitude disturbances, are segmented according to different time scales to simultaneously capture short-term, real-time, and long-term heart rate variation characteristics. In the experiment, three typical time windows were selected: 15 seconds (short cycle), 60 seconds (medium cycle), and 300 seconds (long cycle). The 15-second short-cycle window reflects instantaneous heart rate fluctuations and is suitable for monitoring acute responses during exercise; the 60-second medium-cycle window can be used to observe stable activity states; and the 300-second long-cycle window reflects the overall trend and endurance changes during exercise. To ensure real-time performance and smoothness, a sliding window method is used. For example, the 15-second window updates every 1 second, the 60-second window every 5 seconds, and the 300-second window every 30 seconds. Each window retains a complete data sequence of the dynamic gain heart rate and is linked to a timestamp to ensure accurate alignment of data across different scales. In experimental testing, this multi-scale segmentation method significantly enhances the adaptability of subsequent feature extraction to different exercise patterns. For example, in comparing sprinting and long-distance running, short-cycle windows can promptly reflect heart rate spikes, while long-cycle windows smooth out short-term fluctuations, better revealing changes in endurance levels. Basic heart rate statistics are calculated within each window, including the mean HR, standard deviation (SD), and coefficient of variation (CV). The mean reflects the overall heart rate level during that time period, the standard deviation measures the dispersion of heart rate within that window, and the coefficient of variation is the ratio of the standard deviation to the mean, used to normalize the degree of heart rate fluctuation. In the experiment, taking a 15-second window as an example, if the heart rate range for this data segment is 140–150 bpm, the mean is approximately 145 bpm, the standard deviation is approximately 3 bpm, and the coefficient of variation is approximately 0.021, indicating that the heart rate is relatively stable during this time period. However, in high-intensity interval training, the standard deviation of the short-cycle window can rise to over 8 bpm, and the coefficient of variation approaches 0.05, demonstrating significant volatility. To ensure the accuracy of statistical calculations, outliers (such as heart rates outside the physiological range: <40 bpm or >220 bpm) can be removed before data calculation. Baseline statistics are not only the raw input for subsequent feature fusion but can also be directly used for simple exercise intensity assessments. For example, in training monitoring, a sustained increase in the CV of a long-period window may indicate fatigue or increased cardiovascular load in the exerciser.

[0070] Heart rate extremes were detected within each time window to obtain the maximum (peak) and minimum (trough) heart rates during that period, reflecting extreme states of heart rate fluctuations. In the experiment, this detection was achieved by traversing the heart rate sequence within the window and combining it with simple extreme value judgment rules (such as selecting the maximum and minimum values ​​from smoothed data). To prevent false extremes caused by transient noise, a 3-point moving average filter was first applied to the heart rate sequence. Taking a 60-second window as an example, if an athlete transitions from jogging to sprinting within this time period, the peak might reach 175 bpm, and the trough 145 bpm, a difference of 30 bpm, reflecting a significant change in exercise intensity. In a 300-second window, the difference between the peak and trough might be smaller, indicating higher long-term heart rate stability. By extracting peaks and troughs for short, medium, and long periods respectively, a basis can be provided for subsequent calculations of dynamic range and fluctuation amplitude. Furthermore, in exercise recovery monitoring, a faster recovery rate from the trough may indicate better cardiopulmonary recovery ability in the user.

[0071] Heart rate dynamic range is defined as the difference between the peak and trough values ​​within the same time window, while fluctuation amplitude can be further defined as the dynamic range divided by the average heart rate within that window as a percentage. In experiments, dynamic range can intuitively reflect the intensity of heart rate changes during a given period. For example, a dynamic range of up to 20 bpm within a short 15-second window indicates significant heart rate fluctuations; while within a long 300-second window, the dynamic range might only be 8 bpm, indicating a more stable exercise state. Calculating fluctuation amplitude facilitates comparisons between individuals because it eliminates differences in absolute heart rate levels. For instance, two users with the same dynamic range of 10 bpm might have average heart rates of 100 bpm and 180 bpm respectively, with fluctuation amplitudes of 10% and 5.5%, reflecting different relative volatility. In training monitoring experiments, high fluctuation amplitude was found to frequently occur during high-intensity interval training or rapid interval running, while low fluctuation amplitude was more common in steady-state endurance training. These two indicators are not only used for subsequent feature set construction but also serve as reference conditions for abnormal heart rate alarms in real-time monitoring.

[0072] The purpose of time-frequency decomposition is to simultaneously map changes in the heart rate sequence to both the time and frequency domains, revealing the energy distribution of different frequency components across various time periods. In experiments, wavelet transforms (such as the Daubechies 4th-order wavelet) or short-time Fourier transforms (STFTs) can be used to decompose the heart rate signal into low-frequency (LF, 0.04–0.15 Hz) and high-frequency (HF, 0.15–0.4 Hz) components. The power spectral density (PSD) can be calculated based on the Welch method, using a 256-point FFT with 50% overlap and Hanning window weighting to obtain a smooth and stable spectral estimate. The PSD of the heart rate signal in different time windows is calculated separately, thus obtaining a multi-scale spectral energy distribution. Experimental results show that the high-frequency components in short-period windows are sensitive to changes in respiratory rhythm, the low-frequency components in medium-period windows are related to sympathetic nerve activity, while long-period windows reflect overall autonomic nervous system balance. Multiscale PSD can not only be used to determine exercise status, but also help distinguish different types of heart rate fluctuation sources. For example, aerobic endurance training and strength interval training have significant differences in spectral characteristics.

[0073] After obtaining multi-scale power spectral density, further power fluctuation difference analysis can be performed to extract heart rate variability (HRV) features. Specific methods include calculating LF power, HF power, and the LF / HF ratio. LF reflects the combined action of the sympathetic and parasympathetic nervous systems, HF primarily reflects parasympathetic activity, and the LF / HF ratio is often used to assess autonomic nervous system balance. In experiments, if the LF / HF ratio significantly increases after high-intensity exercise (e.g., from 1.5 to 4.0), it indicates sympathetic dominance; if the LF / HF ratio gradually decreases to 1-2 during the recovery period, it indicates good recovery. Furthermore, the percentage change in energy of PSD at different frequency bands can be calculated to analyze the stability of heart rate fluctuation patterns. Difference analysis can use analysis of variance (ANOVA) or non-parametric tests to compare power distribution differences at different windows or different exercise stages. The resulting HRV feature set can directly reflect changes in physiological state during exercise monitoring and has practical value in fatigue detection and training load assessment. The previously obtained basic heart rate statistics (mean, standard deviation, CV), heart rate variability features (LF, HF, LF / HF, etc.), dynamic range, and fluctuation amplitude are integrated into a multi-dimensional feature vector, and then fitted with a time-series distribution. Time-series distribution fitting can employ methods such as sliding time window regression, exponential smoothing, or polynomial fitting to capture the trend of feature changes over time. For example, within a 300-second window, a cubic polynomial can be used to fit the curve of the mean heart rate, thereby observing the rising, plateauing, and falling phases of heart rate during long-term training. In the experiment, the magnitude of the fitting residuals can be used as an indicator of feature stability; smaller residuals indicate stable feature changes. The final constructed heart rate feature set will serve as the core output of the exercise heart rate monitoring system, which can be used not only for real-time state classification (such as sprint, endurance, recovery) but also as input for training load prediction models and health risk warning algorithms. Through multiple experiments, this feature set construction method based on multi-source feature fusion and time-series fitting can improve the accuracy of exercise state recognition by approximately 10% or more.

[0074] In this embodiment, reference Figure 4 The diagram below illustrates the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include:

[0075] Calculate the user's movement speed and rate of change based on the user's geographic location information;

[0076] Quantitative assessment of exercise intensity is performed based on heart rate feature set, user movement speed, and change in movement rate to obtain exercise intensity assessment value;

[0077] Based on the rate of change of posture, real-time user actions are inferred to obtain user action inference data.

[0078] Real-time deep recognition of movement patterns is performed on the type of movement scene, inferred user action data, and exercise intensity assessment value to obtain the user's movement pattern.

[0079] In this embodiment, the user's moving speed and rate of change of movement (acceleration) are calculated using real-time geographic location information (latitude and longitude coordinates and timestamp) collected by the smartwatch's GNSS module. Moving speed can be calculated by dividing the geographic distance between two points by the time interval. The Haversine formula or Vincenty formula is recommended for distance calculation to improve accuracy over long distances and at high latitudes. In the experiment, the GNSS positioning frequency is recommended to be set to 1Hz or higher (e.g., 5Hz) to obtain more timely location data. In outdoor running scenarios, 1Hz sampling can control the speed error within ±0.2m / s. The rate of change of movement is calculated by the difference between adjacent speed values ​​and divided by the time interval, in m / s², reflecting the user's acceleration or deceleration trend. To suppress the impact of positioning jitter on speed calculation, the location data can be processed first by Kalman filtering (process noise covariance Q=0.01, measurement noise covariance R=5) or low-pass filtering (cutoff frequency 0.5Hz). The resulting sequences of movement speed and rate of change will be synchronized with data such as heart rate feature sets and posture change rates, providing a dynamic basis for exercise intensity assessment and pattern recognition. Physiological indicators (heart rate feature sets) are combined with external kinematic indicators (velocity, acceleration) to construct a quantitative exercise intensity model. The heart rate feature set may include mean heart rate, coefficient of variation, dynamic range, and HRV, while velocity and acceleration reflect the mechanical load of the exercise. In experiments, a weighted fusion model can be used to synthesize these indicators. For example, weight parameters w1, w2, and w3 can be set to correspond to the contribution ratios of heart rate, velocity, and acceleration (e.g., 0.5:0.3:0.2), and then the comprehensive exercise intensity score can be calculated. To ensure the model's individualized adaptability, the user's maximum heart rate (MaxHR) and resting heart rate (RestHR) can be referenced to convert the heart rate indicators into percentage heart rate reserve (%HRR), thereby eliminating the influence of different ages and genders. In experimental verification, using a heart rate + velocity + acceleration fusion model, the correlation coefficient between the exercise intensity assessment value and the actual measured metabolic equivalents (METs) reached 0.92, which is significantly better than a single heart rate or velocity index. The final output exercise intensity assessment value is a continuous quantity (such as a 0-10 scale or METs value), which can directly reflect the exercise load level.

[0080] The system infers real-time user actions based on their pattern characteristics. The rate of change of posture is calculated from the speed of change of posture angle (° / s) measured by the IMU, and can distinguish different types of limb movements. In the experiment, it is first necessary to establish posture change rate templates for different movements (such as running, walking, cycling, weightlifting, etc.). For example, in running, the posture change rate waveform exhibits periodic peaks (arm swing rhythm), with amplitudes between 60° / s and 100° / s; walking is between 20° / s and 50° / s with a longer period; cycling has a relatively low posture change rate (10° / s-30° / s) and changes slowly. For real-time inference, a sliding time window (12 seconds) can be used to calculate the mean, standard deviation, and frequency features of the posture change rate, and action discrimination can be performed using a classification model (such as K-nearest neighbors KNN, random forest, or lightweight convolutional neural network). In the experiment, using a 3-second window + random forest classifier, an action recognition accuracy of over 92% can be achieved across multiple movement modes. The motion inference data will be output in the form of category labels or probability distributions, serving as an important input for subsequent motion pattern recognition.

[0081] This system fuses multi-source information to achieve high-precision recognition of user movement patterns. Input information includes: movement scene type (e.g., indoor running, outdoor cycling, mountain hiking), user action inference data (action categories and their probabilities), and exercise intensity assessment values ​​(continuous load indicators). To simultaneously utilize the temporal dependence and nonlinear relationships of these features, a deep learning-based temporal classification model, such as a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN), can be used in the experiment. The model input is a multi-source feature sequence from the past few seconds (e.g., 10-second historical data), and the output is the current movement pattern category. Training data can come from labeled multi-scene exercise data collection experiments, such as heart rate, posture, and speed data collected in various environments like treadmills, outdoor tracks, and mountain trails. Experimental results show that after fusing heart rate features, action inference, and scene type, the LSTM model achieves a movement pattern recognition accuracy of over 96%, significantly outperforming the performance of single-modality recognition. Ultimately, the system can output the user's movement pattern in real time, such as "outdoor interval running," "indoor strength training," and "mountain biking," providing accurate data support for personalized exercise monitoring and health guidance.

[0082] In this embodiment, step S4 includes the following steps:

[0083] Perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map;

[0084] The global heart rate situation awareness specifically involves: performing a secondary stability analysis on the heart rate feature set to obtain heart rate stability features; the secondary stability analysis includes heart rate rise time, stability duration, and recovery time;

[0085] Heart rate fluctuation pattern recognition is performed on the heart rate feature set, and the variation patterns of heart rate peak and trough period, fluctuation frequency, and amplitude are analyzed to generate heart rate fluctuation features.

[0086] By mining the heart rate change pattern of the heart rate feature set, a heart rate trend pattern is obtained; the heart rate change pattern mining includes identifying heart rate rise, fall, stabilization and oscillation.

[0087] Based on the adaptive heart rate assessment criteria, adaptive heart rate standard analysis was performed on heart rate stability characteristics, heart rate fluctuation characteristics, and heart rate status patterns, and heart rate change perception was performed to construct a global heart rate status map.

[0088] The system collects the user's wrist skin temperature and predicts the user's multi-step heart rate based on the global heart rate status map, thereby constructing a user heart rate prediction map.

[0089] In this embodiment, the goal of secondary stability analysis is to accurately identify the phased characteristics of heart rate changes within global situational data. First, the heart rate rise time is defined as the time required to transition from a resting or low heart rate zone to the target heart rate zone (e.g., reaching 70% of maximum heart rate), reflecting the cardiovascular system's response speed to exercise load. The stability duration is the length of time the heart rate is maintained within a high-level zone (e.g., within the target heart rate ± 5 bpm range), reflecting exercise endurance or training intensity. Recovery time is the time required to return from the high-level zone to the resting zone, assessing cardiopulmonary recovery capacity. In experiments, this can be achieved by setting thresholds; for example, the target heart rate zone can be calculated based on an individual's maximum heart rate (MaxHR) and resting heart rate (RestHR). Taking a running experiment as an example, a user's rise time during the sprint phase was 45 seconds, the stability duration was 90 seconds, and the recovery time was 120 seconds. The variation of these values ​​between different training days can serve as a basis for assessing training adaptability. Multiple experiments have verified that stability characteristics are significantly correlated with athletes' aerobic endurance levels (correlation coefficient R > 0.85). Heart rate fluctuation pattern recognition aims to analyze the peak-valley distribution, fluctuation frequency, and amplitude variation patterns of heart rate during exercise. Specifically, it involves peak-valley detection of the heart rate sequence, using a method combining first-order derivative zero-crossing points with threshold judgment to extract the inflection points of each significant heart rate rise and fall. The fluctuation frequency is calculated by averaging the periods between peaks, while the amplitude is the difference between adjacent peaks and valences. In experiments, during high-intensity interval training (HIIT), the heart rate peak-valley frequency can reach 0.02-0.05 Hz (20-50 seconds per cycle), with amplitude typically between 15-25 bpm; while in low-intensity steady-state running, the frequency is lower (0.005-0.01 Hz), and the amplitude is less than 8 bpm. To more stably identify patterns, the dominant frequency component of the heart rate signal can be analyzed in the frequency domain and combined with time-domain features for classification. In practical applications, fluctuation characteristics can be used to detect training types; for example, high amplitude and high frequency often indicate interval training, while low amplitude and low frequency often indicate endurance training.

[0090] The purpose of heart rate variability analysis is to abstract continuous heart rate changes into several typical patterns: rising (continuously accelerating), falling (continuously slowing down), stable (minor changes), and oscillation (periodic fluctuations). In experiments, initial classification can be achieved by setting thresholds for the rate of change (e.g., >0.5 bpm / s for rising, <-0.5 bpm / s for falling, and less than 0.1 bpm / s for stable). Oscillation patterns require a combination of fluctuation frequency and amplitude thresholds for determination. The analysis process can use a sliding window (e.g., 30 seconds) to continuously differentiate heart rate variability and generate time period labels. For example, in a cycling experiment, a user's heart rate showed a continuous rising trend during the uphill phase, entered a stable state during flat riding, exhibited short-period oscillations during the sprint phase, and finally decreased and recovered during the gliding phase. Combining these pattern analysis with the exercise scenario can construct a dynamic heart rate trajectory during exercise, providing a basis for personalized training recommendations. The heart rate assessment threshold is dynamically adjusted based on the user's historical data, individual physiological parameters (e.g., age, gender, resting heart rate, maximum heart rate), and current exercise type. For example, for the same exercise intensity, beginners may reach high loads at lower heart rates, while well-trained individuals may remain within a sustainable range at higher heart rates. In the experiment, the target heart rate range and fluctuation characteristic thresholds were updated by calculating the mean and standard deviation of the user's training data over the past week. The analyzed stability characteristics, fluctuation characteristics, and situational patterns were mapped onto a global situational graph, with different states distinguished by color, curve thickness, etc., making the graph both intuitive and quantifiable. The resulting global heart rate situational graph not only displays the current state in real time but also reflects the user's heart rate variation patterns over a period of time.

[0091] Building upon global heart rate situational awareness, wrist skin temperature is introduced as an additional physiological feature. Skin temperature and heart rate are correlated during exercise; for example, increased body temperature is often accompanied by an increased heart rate, especially during prolonged exercise or in high-temperature environments. In the experiment, the temperature sensor sampling frequency was set to 1Hz, with a measurement accuracy within ±0.1℃. Temperature data was aligned with the time axis of the global heart rate situational map and used as one of the input variables to construct a multi-step (e.g., 30 seconds, 60 seconds, 120 seconds) heart rate prediction model. The prediction method can employ a multivariate time series model (e.g., LTM neural network or ARIMAX model), with inputs including past heart rate characteristics, temperature data, and exercise scenario information, and outputs heart rate values ​​for multiple future time points. In experimental validation, incorporating skin temperature reduced the mean absolute error (MAE) of heart rate prediction by approximately 12%. The resulting heart rate prediction map can be used for applications such as training intensity planning and early warning during exercise.

[0092] In this embodiment, the specific steps for collecting the user's wrist skin temperature and predicting the user's multi-step heart rate based on the global heart rate trend map to construct the user's heart rate prediction map are as follows:

[0093] Based on the user's wrist skin temperature collected by a smartwatch;

[0094] Analyze the surface temperature changes of the user's wrist skin to obtain a temperature change characteristic curve;

[0095] Based on the temperature change characteristic curve and the global heart rate status map, a body temperature-heart rate coupling analysis was performed to generate a correlation between body temperature and heart rate changes.

[0096] Based on the correlation between body temperature and heart rate changes, multi-step heart rate prediction is performed on users to construct a user heart rate prediction map.

[0097] In this embodiment, the user's wrist skin temperature is continuously collected using a skin temperature sensor built into the smartwatch. Skin temperature sensors are typically thermistors or infrared thermistors, and their working principle is to measure temperature based on thermal radiation or contact heat conduction from the skin surface. In the experimental setup, the sampling frequency is recommended to be between 1Hz and 5Hz to ensure the capture of rapid temperature changes without excessively consuming the watch's battery resources. The sensor accuracy should be controlled within ±0.1℃, and the measurement range is generally between 20℃ and 45℃, covering the entire range of resting, active, and ambient temperature variations. To reduce environmental interference, the watch's back cover and the skin contact surface should be made of a material with good thermal conductivity, and preliminary filtering (such as a first-order low-pass filter) should be performed in the software to suppress instantaneous noise caused by arm swinging, sweating, or wind cooling. In actual data collection, for example, during a 30-minute outdoor run, the user's wrist temperature rose from 32.4℃ at rest to 34.8℃ during the middle of the run, and finally slowly dropped back to 33.0℃ after the run. This temperature change curve provides basic data support for subsequent analysis. Data preprocessing includes removing outliers (such as data points with significant temperature jumps exceeding ±2℃ within a short period) and smoothing (using a moving average method with a recommended window size of 5-10 seconds) to reduce the impact of measurement errors on the analysis results. Subsequently, time series analysis is used to extract the rising, stable, and falling phases of temperature change. These phases correspond to the thermal stress response, thermal equilibrium, and recovery processes during exercise. For example, temperature change rate thresholds can be set (e.g., >0.02℃ / s for the rising phase, <-0.02℃ / s for the falling phase, and the rest for the stable phase) to divide the entire exercise process into different temperature change intervals. Temperature change characteristic curves obtained through piecewise fitting (such as polynomial fitting or exponential smoothing fitting) can more intuitively reflect the dynamic changes in body surface temperature over time. In a cycling experiment in a high-temperature environment (28℃), the temperature change characteristic curve exhibited a typical pattern of rapid rise—high-level maintenance—slow decline, showing high temporal consistency with the heart rate change curve.

[0098] The temperature change curve was aligned with the global heart rate trend map on the time axis, ensuring that the two sets of data corresponded at the same point in time. Then, the correlation coefficient (e.g., Pearson correlation coefficient) was calculated at different time delays to analyze the immediacy and lag of body temperature changes on heart rate changes. Experiments showed that skin temperature rise typically lags heart rate rise by approximately 30–60 seconds, especially during high-intensity exercise, where this lag effect is more pronounced. Furthermore, cross-correlation function (CCF) and analysis of covariance (ANCOVA) can be used to quantify the coupling strength under different exercise states. For example, in an interval running experiment, the maximum correlation coefficient between body temperature and heart rate reached 0.87, with a delay of 45 seconds; while in low-intensity walking, the correlation coefficient was only 0.45, with a delay of almost 0. This indicates that the degree of coupling between body temperature and heart rate differs significantly under different exercise modes, and this information can be used to enhance the accuracy of heart rate prediction models. After obtaining the correlation between body temperature and heart rate changes, it can be used as one of the core input variables to construct a multi-step heart rate prediction model. The prediction objective is to predict heart rate values ​​at multiple future time points (e.g., 30 seconds, 60 seconds, and 120 seconds later) based on current and past body temperature and heart rate data. Methodologically, multivariate time series prediction models can be employed, such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), or traditional ARIMAX models, using body temperature change rate, heart rate change rate, and exercise scenario characteristics as multidimensional inputs. Experimental validation shows that incorporating body temperature data significantly improves prediction accuracy. For example, in an indoor running experiment, the mean absolute error (MAE) of LSTM prediction without body temperature data was 4.2 bpm, while incorporating body temperature data reduced the MAE to 3.5 bpm, a decrease of approximately 16.7%. The final user heart rate prediction graph is displayed with time as the horizontal axis, overlaying the predicted and actual heart rate curves for easy and intuitive assessment of prediction accuracy. This prediction function not only provides exercisers with training rhythm suggestions but also issues early warnings of high heart rate risks, enhancing exercise safety.

[0099] In this embodiment, step S5 includes the following steps:

[0100] Obtain user personal information; analyze user age, physical fitness level, and health status based on user personal information to obtain personalized user characteristics;

[0101] An adaptive heart rate assessment standard was derived by analyzing user exercise patterns and personalized user characteristics.

[0102] A comprehensive risk assessment of the user's heart rate prediction graph is performed based on the adaptive heart rate assessment standard to obtain the heart rate risk assessment result;

[0103] Adaptive early warning decisions are made based on heart rate risk assessment results, and an adaptive early warning strategy is constructed.

[0104] In this embodiment, basic user information, including name, gender, date of birth, height, and weight, is collected via a smartwatch and a companion app. This information is then combined with the user's historical exercise data (such as average exercise frequency, average exercise time, and maximum heart rate over the past month) and health records (such as past illnesses, hypertension, and diabetes) to create a comprehensive data profile. Age is automatically calculated by subtracting the current date from the birth date to obtain an accurate value (in years). Fitness levels can be assessed using indicators such as Cooper's 12-minute run test results and estimated VO2 max. For example, VO2 max can be estimated using a heart rate-velocity regression model: in the experimental sample, the average VO2 max for a healthy 30-year-old male was 45 mL / kg / min, while for a healthy 50-year-old male it was 38 mL / kg / min. Health status assessment combines a simple questionnaire (such as the PAR-Q health questionnaire) with imported medical data to determine three levels: "healthy," "sub-healthy," and "high-risk." Finally, the system combines age, fitness level, and health status into a personalized user feature vector, providing a data foundation for subsequent adaptive heart rate standard development. After extracting personalized user features, the system performs correlation analysis with real-time detected exercise modes (such as running, cycling, strength training, and hiking) to generate an adaptive heart rate assessment standard suitable for the current exercise state. Traditional heart rate zone divisions are mostly based on a percentage of maximum heart rate (HRmax = 220 - age), but this method ignores individual differences. This step introduces personalized parameters such as the user's VO2 max, resting heart rate (HRrest), and heart rate reserve (HRR = HRmax - HRrest), and uses the Karvonen formula to dynamically calculate the target heart rate zone for exercise. For example, for a 35-year-old user with HRrest = 60 bpm and a high VO2max, the target heart rate zone can be set to 145-170 bpm in a moderate-to-high intensity running mode, while for a user of the same age but with average health, it can be lowered to 135-160 bpm. In experimental verification, users using the adaptive heart rate standard showed a 12% reduction in average perceived fatigue index (RPE) under the same exercise conditions, and their heart rate fluctuations were closer to the target zone, indicating that the adaptive standard can better match individual physiological responses.

[0105] The predicted heart rate curve is compared and analyzed against the adaptive heart rate assessment criteria time-by-time to identify potential heart rate risks. Assessment indicators include: the proportion of time exceeding the safe upper limit, the slope of a rapid heart rate increase (e.g., >3 bpm / s), and the duration of prolonged periods near maximum heart rate (e.g., >85% HRmax for more than 5 minutes). Risk assessment can employ a multi-indicator comprehensive scoring system, such as setting a risk index from 0 to 100, with the proportion of exceeding limits, the slope exceeding the limit, and sustained high heart rate accounting for 40%, 30%, and 30% of the weight, respectively. In a 40-minute interval running experiment, a user's heart rate exceeded the adaptive standard upper limit for 18% of the sprint phase, resulting in a risk index of 78 (high risk), and the system would mark this user as having a high risk of cardiovascular overload. Another user of the same age but with better physical fitness only exceeded the limit for 5% of the time, with a risk index of 42 (low risk). This risk assessment based on the predictive graph can provide early warnings, avoiding post-exercise analysis. Instead of a one-size-fits-all fixed threshold alarm, the system develops dynamic, personalized early warning strategies based on real-time risk levels. First, it categorizes risk indices into three levels: low risk (0.50), medium risk (51-75), and high risk (76-100). Differentiated warning strategies are implemented for each risk level. For example, in a low-risk state, a vibration alert is only triggered if the user's heart rate exceeds the safe range by more than 2 minutes; in a medium-risk state, a vibration and screen notification immediately alert the user to slow down or rest when the heart rate exceeds the limit or the rate of increase is too steep; in a high-risk state, in addition to immediate alerts, emergency contacts or medical assistance functions are invoked. The strategy also incorporates an exercise mode correction factor. For instance, in a running mode during hot weather, the system lowers the warning trigger threshold by 5 bpm to accommodate higher cardiovascular stress. In one experiment, users using the adaptive warning strategy experienced a 37% reduction in the average duration of heart rate exceeding the limit compared to the fixed threshold strategy, significantly improving exercise safety. This adaptive mechanism effectively balances safety and athletic performance, making it particularly suitable for long-term training and cardiac health monitoring scenarios.

[0106] In this embodiment, a smartwatch-based heart rate monitoring device is provided for performing the smartwatch-based heart rate monitoring method described above, including:

[0107] The scene recognition module is used to obtain the user's geographical location information based on the smartwatch, identify the user's movement scene, and generate the movement scene type.

[0108] The chip analysis module is used to collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set.

[0109] The motion recognition module is used to perform deep recognition of motion patterns based on motion scene type and heart rate feature set to obtain the user's motion pattern.

[0110] The heart rate situation awareness module is used to perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map.

[0111] The heart rate risk assessment module is used to perform comprehensive risk assessment and adaptive early warning decisions on the global heart rate status map based on the user's exercise pattern.

[0112] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0113] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for monitoring exercise heart rate based on a smartwatch, characterized in that, Includes the following steps: Step S1: Based on the user's geographical location information obtained from the smartwatch, the user's movement scene is identified, thereby generating the movement scene type; Step S2: Collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set; Step S3: Perform deep recognition of exercise patterns based on exercise scene type and heart rate feature set to obtain the user's exercise pattern; Step S4: Perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map; Step S5: Perform a comprehensive risk assessment and adaptive early warning decision based on the user's exercise pattern of the global heart rate status graph; The specific steps of step S1 are as follows: Based on smartwatches, user location information and ambient light signals are obtained; Calculate the user's real-time altitude and latitude / longitude coordinates based on the user's geographic location information; Based on the user's geographic location information, meteorological parameters for the current location are collected to obtain the meteorological parameters for the current location; Perform environmental meteorological analysis on meteorological parameters at the current location to generate environmental meteorological characteristics; The ambient light intensity is calculated from the ambient light signal, and the intensity variation is analyzed to obtain an ambient light intensity distribution map. The system identifies user movement scenarios by analyzing real-time altitude, latitude and longitude coordinates, environmental meteorological characteristics, and ambient light intensity distribution maps, thereby generating movement scenario types. The specific steps of step S2 are as follows: The raw heart rate detection parameters are obtained by continuously collecting heart rate data using the pulse sensor built into the smartwatch. The smartwatch collects the user's body posture parameters using its inertial sensing unit. The user's body posture parameters are used to calculate the rate of change of posture, in order to generate the rate of change of posture. Dynamic gain heart rate parameters are generated by applying attitude disturbance to the original heart rate detection parameters based on the attitude change rate. Multi-scale time window segmentation and feature statistics are performed on dynamic gain heart rate parameters to construct a heart rate feature set; Specifically, step S3 involves the following steps: Calculate the user's movement speed and rate of change based on the user's geographic location information; Quantitative assessment of exercise intensity is performed based on heart rate feature set, user movement speed, and change in movement rate to obtain exercise intensity assessment value; Based on the rate of change of posture, real-time user actions are inferred to obtain user action inference data. Real-time deep recognition of motion patterns is performed on motion scene type, user action inference data and motion intensity assessment value to obtain user motion patterns; The specific steps of step S4 are as follows: Perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map; The global heart rate situation awareness specifically involves: performing a secondary stability analysis on the heart rate feature set to obtain heart rate stability features; the secondary stability analysis includes heart rate rise time, stability duration, and recovery time; Heart rate fluctuation pattern recognition is performed on the heart rate feature set, and the variation patterns of heart rate peak and trough period, fluctuation frequency, and amplitude are analyzed to generate heart rate fluctuation features. By mining the heart rate change pattern of the heart rate feature set, a heart rate trend pattern is obtained; the heart rate change pattern mining includes identifying heart rate rise, fall, stabilization and oscillation. Based on the adaptive heart rate assessment criteria, adaptive heart rate standard analysis was performed on heart rate stability characteristics, heart rate fluctuation characteristics, and heart rate status patterns, and heart rate change perception was performed to construct a global heart rate status map. The system collects the user's wrist skin temperature and predicts the user's multi-step heart rate based on the global heart rate status map, thereby constructing a user heart rate prediction map.

2. The method for monitoring exercise heart rate based on a smartwatch according to claim 1, characterized in that, The specific steps for performing multi-scale time window division and feature statistics on dynamic gain heart rate parameters to construct a heart rate feature set are as follows: The dynamic gain heart rate parameters are divided into multiple time windows to generate heart rate monitoring parameters for multiple time windows; the heart rate monitoring parameters for multiple time windows include a 15-second short cycle window, a 60-second medium cycle window, and a 300-second long cycle window; Calculate the mean heart rate, standard deviation, and coefficient of variation of the heart rate monitoring parameters to obtain the baseline heart rate statistics; The maximum and minimum heart rate values ​​of heart rate monitoring parameters in multiple time windows are detected window by window, and the peak and trough values ​​of heart rate in different cycles are extracted. The dynamic range and fluctuation amplitude of heart rate are calculated based on the peak and trough values ​​of heart rate. The heart rate monitoring parameters are decomposed into time and frequency components, and the power spectral density is calculated to obtain power spectral densities at multiple scales. Power fluctuation difference analysis was performed based on power spectral density to obtain heart rate variability characteristics; A heart rate feature set is constructed by fitting the time-series distribution of the baseline heart rate statistics, heart rate variability characteristics, heart rate dynamic range, and fluctuation amplitude.

3. The method for monitoring exercise heart rate based on a smartwatch according to claim 1, characterized in that, The specific steps for collecting the user's wrist skin temperature and predicting the user's multi-step heart rate based on the global heart rate status map to construct the user's heart rate prediction map are as follows: Based on the user's wrist skin temperature collected by a smartwatch; Analyze the surface temperature changes of the user's wrist skin to obtain a temperature change characteristic curve; Based on the temperature change characteristic curve and the global heart rate status map, a body temperature-heart rate coupling analysis was performed to generate a correlation between body temperature and heart rate changes. Based on the correlation between body temperature and heart rate changes, multi-step heart rate prediction is performed on users to construct a user heart rate prediction map.

4. The method for monitoring exercise heart rate based on a smartwatch according to claim 1, characterized in that, The specific steps of step S5 are as follows: Obtain user personal information; analyze user age, physical fitness level, and health status based on user personal information to obtain personalized user characteristics; An adaptive heart rate assessment standard was derived by analyzing user exercise patterns and personalized user characteristics. A comprehensive risk assessment of the user's heart rate prediction graph is performed based on the adaptive heart rate assessment standard to obtain the heart rate risk assessment result; Adaptive early warning decisions are made based on heart rate risk assessment results, and an adaptive early warning strategy is constructed.

5. A sports heart rate monitoring device based on a smartwatch, characterized in that, For performing the smartwatch-based heart rate monitoring method as described in claim 1, including: The scene recognition module is used to obtain the user's geographical location information based on the smartwatch, identify the user's movement scene, and generate the movement scene type. The chip analysis module is used to collect raw heart rate detection parameters, perform time-series heart rate feature analysis, and construct a heart rate feature set. The motion recognition module is used to perform deep recognition of motion patterns based on motion scene types and heart rate feature sets to obtain the user's motion pattern. The heart rate situation awareness module is used to perform global heart rate situation awareness on the heart rate feature set and construct a global heart rate situation map. The heart rate risk assessment module is used to perform comprehensive risk assessment and adaptive early warning decisions on the global heart rate status map based on the user's exercise pattern.

Images