A multi-array multi-modal fusion-based precise heart rate calculation method

Abstract

This invention discloses a method for accurate heart rate calculation based on multi-source multimodal fusion, relating to the fields of biosignal processing and health monitoring technology. The method includes: simultaneously acquiring multi-source array data such as biosignals, environmental information, and motion status, and constructing a time-aligned dataset; preprocessing and extracting features from the data to obtain multi-dimensional features such as pulse peak interval, R-wave interval, environmental interference level, and exercise intensity level; identifying the current dynamic scene based on features and adaptively assigning heart rate calculation weights to photoelectric and electrocardiogram signals; calculating two preliminary heart rate values ​​in parallel and performing weighted fusion; combining individual feature calibration and a time-series trend model to perform anomaly judgment and correction on the preliminary heart rate values, and outputting the final heart rate value and confidence score. This method, through multi-source fusion and scene-adaptive weighting, achieves multi-scene adaptive, anti-interference, and personalized accurate heart rate monitoring.

Description

Technical Field

[0001] This invention relates to the field of biosignal processing and health monitoring technology, specifically to a method for accurate heart rate calculation based on multi-array multimodal fusion. Background Technology

[0002] Heart rate monitoring is a key means of health management, disease early warning and exercise assessment. Its accuracy directly affects the reliability of physiological state interpretation. With the development of wearable devices and smart medical and elderly care technologies, heart rate calculation methods are constantly evolving, from relying on single signals in the early stage to multi-source fusion. However, it still faces challenges in terms of accuracy and stability in complex dynamic scenarios.

[0003] Traditional heart rate calculation mainly relies on two types of physiological signals: photoplethysmography (PPG) signals and electrocardiogram (ECG) signals. PPG-based methods collect light intensity reflection arrays caused by changes in subcutaneous blood volume using photoelectric sensors, and extract heart rate information through filtering and peak detection. This approach offers advantages such as non-invasiveness and ease of integration, making it a mainstream solution for consumer-grade smart wearable devices. However, PPG signals are highly susceptible to interference from changes in ambient light, sensor fit, skin characteristics (such as skin color and stratum corneum thickness), and motion artifacts. The signal-to-noise ratio decreases significantly during user activity, leading to jumps or deviations in heart rate calculations. Especially in non-steady-state scenarios such as daily activities and rehabilitation training for the elderly, the coupling effect of exercise intensity and environmental interference often results in heart rate errors exceeding ±10 beats per minute from a single PPG signal, severely limiting its application in precise health monitoring.

[0004] Heart rate calculation based on electrocardiogram (ECG) signals acquires ECG potential arrays through electrodes, identifies R waves, and calculates RR intervals. It is considered a high-precision standard in clinical and sports medicine. ECG signals have good resistance to motion interference and individual consistency. However, its acquisition requires maintaining good contact between the electrodes and the skin. Long-term wear can easily cause discomfort. In addition, the device has large power consumption and size, making it difficult to widely apply in lightweight, long-lasting wearable devices. ECG signals may still be affected by electromyographic noise, power frequency interference, and baseline drift, especially during dynamic activities, where signal quality may fluctuate.

[0005] To alleviate the limitations of single signals, existing technologies attempt to introduce an array of accelerometers for motion interference compensation. This involves identifying motion intensity using triaxial acceleration data and filtering or empirically correcting photoplethysmography (PPG) or electrocardiogram (ECG) signals. While this method improves heart rate stability during exercise to some extent, it still has significant shortcomings: First, the compensation models are mostly based on fixed thresholds or static mappings, lacking the ability to adaptively adjust to complex scenarios (such as sudden changes in illumination, loosening of the device, and individual differences). Second, deep fusion and cross-validation of multi-source signals at the feature level are not achieved, making it impossible to fundamentally distinguish between real physiological changes and external interference. Third, the judgment and correction of abnormal heart rate values ​​rely heavily on simple threshold rules, without considering historical trends and individual physiological baselines, resulting in large fluctuations and low reliability of the output results.

[0006] Existing heart rate calculation methods still have significant shortcomings in terms of resistance to complex interference, dynamic scene adaptation, and individualized calibration, making it difficult to meet the core requirements of accurate, stable, and real-time heart rate data in smart healthcare scenarios. Especially among the elderly, whose activity patterns are diverse, physiological states are easily changed, and wearing environments are complex, there is an urgent need for a high-precision heart rate calculation method that can integrate multi-dimensional information, adapt to environmental and movement changes, and fit individual physiological characteristics to improve the reliability and practicality of health monitoring. Summary of the Invention

[0007] To address the issues mentioned in the background art regarding heart rate calculation, such as weak anti-interference capability, poor adaptability to dynamic scenarios, and inaccurate individualization, this application provides a method for accurate heart rate calculation based on multi-array multimodal fusion.

[0008] The above-mentioned objective of this application is achieved through the following technical solution:

[0009] A method for accurate heart rate calculation based on multi-array multimodal fusion, comprising the following steps:

[0010] Collect multi-source array data from users to construct a multi-array dataset;

[0011] The multi-array dataset is preprocessed, and multi-dimensional features are extracted from the preprocessed arrays;

[0012] Based on the extracted multi-dimensional features, the current dynamic scene is identified, and computational weights are adaptively assigned to different biological signal arrays according to the identified dynamic scene.

[0013] Based on the biological signals in the multi-array dataset, at least two preliminary heart rate values ​​are calculated in parallel, and then weighted and fused according to the assigned calculation weights to obtain a preliminary fused heart rate value.

[0014] Based on a preset individual characteristic calibration model and historical heart rate change trends, the preliminary fused heart rate value is anomaly detected and corrected, and the final heart rate value and corresponding confidence score are output.

[0015] By adopting the above technical solution, user data is collected synchronously and multiple datasets are constructed. Targeted preprocessing and feature extraction are performed on different types of data, preserving key information of the original signal while reducing noise interference. By constructing a scene classification model and matching the current scene based on exercise intensity and environmental interference level, adaptive allocation of computational weights is achieved, allowing the heart rate calculation path to be dynamically adjusted according to the actual scene. Then, preliminary heart rate values ​​are calculated in parallel and weighted fusion is performed, fully utilizing the complementarity of photoplethysmography (PPG) signals and electrocardiogram (ECG) signals to further improve the reliability of heart rate values. Finally, through individual feature calibration and historical heart rate trend modeling, anomaly detection and correction are performed on the preliminary fused heart rate values, and heart rate results including confidence scores are output. This not only ensures the accuracy of the output data but also provides users with quantitative indicators of data credibility.

[0016] In a preferred embodiment, this application can be further configured as follows: the collection of multi-source array data from users to construct a multi-array dataset includes:

[0017] Simultaneously acquire biological signal arrays, environmental arrays, and motion state arrays;

[0018] The biosignal array includes a photoelectric sensor light intensity array and an electrocardiogram analog signal array; the environment array includes an ambient light intensity array and a sensor fit array; the motion state array is a three-dimensional array of an accelerometer.

[0019] A timestamp alignment algorithm is used to unify the biosignal array, environment array, and motion state array onto the same time axis to construct the multi-array dataset.

[0020] By adopting the above technical solution, a multi-source synchronous data system integrating "biological signals, environmental state, and motion information" is constructed, and a high-precision timestamp alignment algorithm is used to establish a spatiotemporally consistent multidimensional information field. This provides a reliable data foundation for subsequent multimodal feature fusion and cross-validation, fundamentally eliminating information misalignment and fusion deviation caused by sensor sampling delays or asynchronous data streams. By simultaneously capturing physiological signals and their surrounding environment and motion context, the system gains the ability to directly observe interference sources, creating the prerequisite for accurate interference identification and filtering at the algorithm level.

[0021] In a preferred embodiment, this application can be further configured as follows: preprocessing the multi-array dataset and extracting multi-dimensional features from the preprocessed array includes:

[0022] The light intensity array of the photoelectric sensor is subjected to bandpass filtering and normalization, and the pulse peak interval feature is extracted from it.

[0023] Wavelet transform is performed on the ECG analog signal array to remove electromyographic and power frequency interference, and R-wave interval features are extracted from it.

[0024] The ambient light intensity array is smoothed using a sliding window mean, and the environmental interference level characteristics are determined based on the smoothed data.

[0025] Based on the three-dimensional array of the accelerometer, the synthetic acceleration value is calculated, and the motion intensity level features are divided and extracted according to the synthetic acceleration value.

[0026] By employing the above technical solutions, targeted preprocessing and feature extraction strategies are used for different source arrays. Bandpass filtering and normalization are applied to photoelectric signals, eliminating baseline drift and ambient light interference while preserving effective pulse wave components. Wavelet transform is used on electrocardiogram signals to effectively remove electromyographic interference and power frequency interference, highlighting clear R-wave characteristics. Ambient light intensity is smoothed and its interference level is assessed. Acceleration data is synthesized and calculated to classify motion intensity, enabling a quantitative assessment of the impact of the external environment and one's own movement on heart rate signals. This provides accurate and efficient information input for subsequent scene recognition, weight allocation, and anomaly detection.

[0027] In a preferred embodiment, this application can be further configured as follows: based on the extracted multi-dimensional features, the current dynamic scene is identified, and computational weights are adaptively assigned to different biosignal arrays according to the identified dynamic scene, including:

[0028] Construct a predefined scene classification model, which classifies dynamic scenes into at least five categories: resting without interference, resting with interference, slight activity, moderate activity, and vigorous activity.

[0029] Based on the extracted motion intensity level features and environmental interference level features, the scene classification model is matched to determine the current monitoring scene;

[0030] Based on the current monitoring scenario, a first weight W1 is assigned to the heart rate calculation path based on the light intensity array of the photoelectric sensor, and a second weight W2 is assigned to the heart rate calculation path based on the ECG analog signal array, and the sum of the first weight W1 and the second weight W2 is 1.

[0031] By adopting the above technical solution, and by constructing a refined scene classification model and performing dynamic matching based on real-time features, it can "determine" the user's current state—whether they are resting quietly, experiencing light interference, or engaging in activities of varying intensities. Based on this accurate judgment, it adaptively adjusts the level of trust in different physiological signal sources. For example, it relies more on signals with stronger noise resistance in interfering environments, or adopts signal paths optimized through motion compensation during strenuous activities. This dynamic weight allocation mechanism achieves intelligent matching between the calculation strategy and the physical scene, ensuring that heart rate fusion calculation can always be performed using the current optimal signal combination method under any complex conditions. This overcomes the inherent contradiction that a single algorithm cannot take into account the performance of multiple scenarios at the source, and achieves the robustness of high-precision heart rate monitoring across all scenarios.

[0032] In a preferred embodiment, this application can be further configured such that: the adaptive allocation of computational weights based on different biosignal arrays according to the identified dynamic scene includes: a first weight W1 allocated to the heart rate calculation path based on the photoelectric sensor light intensity array, and a second weight W2 allocated to the heart rate calculation path based on the ECG analog signal array, satisfying the following rules:

[0033] In a quiet, undisturbed environment, W2 is greater than W1;

[0034] In a resting, disturbed scenario, the weight of W2 is greater than its value in a resting, undisturbed scenario.

[0035] In scenarios with minimal activity, the weights of W1 and W2 are equal or close.

[0036] In moderate activity scenarios, W1 is greater than W2;

[0037] In scenarios involving intense activity, the weight of W1 is greater than its value in scenarios involving moderate activity.

[0038] By adopting the above technical solutions, this series of rules is not simply fixed, but constitutes a coherent and progressive adaptive adjustment strategy. It can continuously tilt computing resources toward more reliable and stable signal sources according to the upgrading of scenario complexity, thereby building a dynamic defense system at the algorithm level, effectively suppressing the errors that may be introduced by different interference sources, and ensuring the consistency and accuracy of heart rate calculation across the entire spectrum of scenarios from resting to strenuous activity.

[0039] In a preferred embodiment, this application can be further configured as follows: based on the biosignals in the multi-array dataset, at least two preliminary heart rate values ​​are calculated in parallel, and weighted fusion is performed according to the assigned calculation weights to obtain a preliminary fused heart rate value, specifically including:

[0040] An adaptive threshold peak detection algorithm is used to identify adjacent peak points and calculate the peak time interval for the pulse peak interval feature sequence. The time interval is then converted into a heart rate value and used as the first heart rate value HR1.

[0041] The R-wave peak points are identified using the differential threshold method on the R-wave interval feature sequence. The time interval between adjacent R-wave peak points is calculated, and the time interval is converted into a heart rate value as the second heart rate value HR2.

[0042] The preliminary fused heart rate value HR_temp is calculated using the following fusion formula: HR_temp = W1 × HR1 + W2 × HR2.

[0043] By adopting the above technical solution, and processing photoplethysmography (PPG) signals and electrocardiogram (ECG) signals in parallel, and extracting heart rate features using adaptive threshold peak detection and differential thresholding methods respectively, the physiological complementarity of the two signals is preserved, and the systematic error caused by the failure of a single algorithm is avoided by using independent calculation paths. The introduction of dynamic weights in the fusion formula enables the final output to adapt to scene changes. In the resting state, it relies more on the anti-interference characteristics of the ECG signal, while in the exercise state, it enhances the real-time response capability of the photoelectric signal, thereby achieving a balance between accuracy and robustness in heart rate calculation across all scenarios.

[0044] In a preferred embodiment, this application may be further configured such that, before performing anomaly detection and correction on the preliminary fused heart rate value, an individual characteristic calibration step is included:

[0045] Based on standard heart rate data collected from users in a resting state and synchronized biological signal arrays, an initial individual calibration parameter library containing skin light reflectance coefficient and electrocardiogram signal amplitude coefficient is calculated and established.

[0046] During continuous monitoring, a dynamic calibration mechanism is triggered based on a preset cycle or changes in the user's physiological state. The parameters in the individual calibration parameter library are updated by combining recent resting heart rate data and its changing trends.

[0047] By adopting the above technical solution, and by constructing an initial individual calibration parameter library and establishing a dynamic update mechanism, it is possible to accurately capture the unique physiological differences among different users, such as the degree of influence of skin light reflection characteristics on photoelectric signals, and the correlation between ECG signal amplitude and cardiac activity intensity. This calibration method based on individual physiological baselines effectively eliminates signal response deviations caused by differences in skin color, subcutaneous fat thickness, and cardiac electrophysiological characteristics among users, so that subsequent heart rate calculations no longer rely on a general model, but truly achieve personalized adaptation. At the same time, the dynamic calibration mechanism, by continuously tracking the changing trend of the user's resting heart rate, can promptly identify physiological parameter drifts caused by aging, changes in health status, or environmental adaptation, and automatically adjust calibration parameters to ensure the long-term effectiveness of the individualized model, thereby providing a highly reliable benchmark reference for anomaly judgment and correction.

[0048] In a preferred embodiment, this application may be further configured such that, before performing anomaly detection and correction on the fused preliminary heart rate value, a time-series trend modeling step is included:

[0049] Based on the sliding window, a time series sequence is constructed from the preliminary fused heart rate value HR_temp, and the slope, mean, and standard deviation of the heart rate change within the window are calculated.

[0050] Construct a heart rate trend model to identify heart rate change patterns and determine whether there are sudden changes in physiological state or continuous abnormal fluctuations.

[0051] By adopting the above technical solution and constructing a time-series trend model, the short-term dynamic characteristics of heart rate data can be captured. For example, the slope can reflect the acceleration or deceleration trend of heart rate, while the mean and standard deviation quantify the fluctuation amplitude and dispersion. Combined with the local analysis capability of the sliding window, it can effectively distinguish between physiological heart rate fluctuations (such as recovery after exercise) and pathological abnormal changes (such as arrhythmia), providing multi-dimensional judgment basis for subsequent abnormal correction. By continuously updating the window data, the model can dynamically adapt to the heart rate change patterns under different monitoring scenarios, avoiding misjudgments caused by user activity mode switching or environmental interference, thereby improving the stability of results while ensuring detection sensitivity.

[0052] In a preferred embodiment, this application can be further configured such that: the anomaly detection and correction of the fused preliminary heart rate value includes:

[0053] By combining the heart rate change pattern output from time-series trend modeling with a preset anomaly detection threshold, abnormal points in the preliminary fused heart rate value HR_temp are identified.

[0054] When an anomaly is identified, cross-validation is performed based on the pulse peak interval characteristics and R wave interval characteristics in the biosignal array to determine whether the cause of the anomaly is signal interference or a real physiological change.

[0055] If the problem is confirmed to be caused by signal interference, a prediction algorithm based on time-series trends is used to correct the anomalies and output the corrected heart rate value HR_cal.

[0056] If the change is determined to be a genuine physiological change, the heart rate value is retained, and the relevant parameters in the individual characteristic calibration model are updated.

[0057] By adopting the above technical solution and combining time-series trend and multimodal signal cross-validation mechanisms, the source of heart rate abnormalities can be accurately distinguished as either spurious fluctuations caused by external interference or genuine changes in the user's physiological state. When an abnormality is detected, bidirectional verification is first performed using the independent characteristics of photoelectric and electrocardiogram signals. If both show abnormalities and the time-series trend model indicates non-physiological changes, it is determined to be interference and prediction correction is initiated. Conversely, if both signal characteristics and the trend model support physiological changes (such as a sustained increase in heart rate after exercise), the original value is retained and individual calibration parameters are updated synchronously. This hierarchical judgment strategy avoids erroneous corrections caused by the failure of a single signal and achieves adaptive tracking of long-term physiological changes in users through dynamic parameter updates, ultimately outputting heart rate results that are both accurate and physiologically reasonable.

[0058] In a preferred embodiment, this application can be further configured such that the output of the final heart rate value and the corresponding confidence score includes:

[0059] The algorithms for feature extraction, weight allocation, heart rate calculation, and anomaly correction are lightweighted, and fixed-point arithmetic is used to reduce computational complexity.

[0060] Control the delay of the overall calculation process to achieve real-time calculation and output of heart rate;

[0061] The final heart rate value and corresponding confidence score are output at a preset frequency. The confidence score is obtained by comprehensive calculation based on signal quality, scene matching degree and timing consistency.

[0062] When the confidence score is lower than the set threshold, a sensor wearing status reminder or an environment adjustment prompt will be triggered.

[0063] By adopting the above technical solutions, and through algorithm lightweighting and fixed-point arithmetic optimization, the system's dependence on hardware resources is reduced, enabling heart rate calculation to run efficiently on low-power devices. At the same time, the calculation latency is controlled within the millisecond range, meeting the needs of real-time monitoring. The preset frequency output ensures the continuity and regularity of heart rate data, making it easy for users to track heart rate change trends. The confidence score provides users with intuitive feedback on data reliability by quantifying signal quality, scene matching degree, and temporal consistency. When the score is lower than the threshold, the system automatically triggers wearing status reminders or environmental adjustment suggestions to guide users to optimize monitoring conditions, thereby further improving data quality.

[0064] In summary, this application includes at least one of the following beneficial technical effects:

[0065] 1. By synchronously acquiring data from multiple source arrays and aligning it with timestamps, combined with targeted preprocessing and multi-dimensional feature extraction, a data foundation with strong anti-interference capabilities and comprehensive information was constructed. Furthermore, based on dynamic scene recognition and adaptive allocation of calculation weights, and by utilizing parallel computation and weighted fusion of multiple signals, the shortcomings of traditional single-signal methods in terms of accuracy drop under environmental interference, motion state, and individual differences were effectively overcome, ensuring accurate and stable heart rate calculation across all scenarios from resting to strenuous activity.

[0066] 2. By introducing individual feature calibration and time-series trend modeling, it can not only personalize the model based on the user's initial physiological parameters, but also dynamically update the model parameters based on changes in physiological state during long-term use; combined with an intelligent anomaly handling mechanism based on trend analysis and signal cross-validation, it can accurately distinguish between signal interference and real physiological changes, making the monitoring results more consistent with the user's long-term physiological characteristics.

[0067] 3. By lightweighting and optimizing the entire algorithm and fixed-point operation, the system complexity and power consumption are reduced while maintaining high computational accuracy. This enables end-to-end low-latency real-time heart rate output. By outputting a confidence score that integrates multiple dimensions of evaluation and providing proactive interactive prompts, the continuity and reliability of data output are ensured, while also enhancing user trust and experience. Attached Figure Description

[0068] Figure 1 This is a flowchart of an embodiment of a heart rate accurate calculation method based on multi-array multimodal fusion according to this application;

[0069] Figure 2 This is a flowchart illustrating step S20 in an embodiment of a heart rate accuracy calculation method based on multi-array multimodal fusion according to this application.

[0070] Figure 3This application presents a flowchart of a parallel calculation and fusion process for a heart rate accuracy calculation method based on multi-array multimodal fusion.

[0071] Figure 4 This is a flowchart illustrating the heart rate data calibration and anomaly handling process of a heart rate accuracy calculation method based on multi-array multimodal fusion proposed in this application.

[0072] Figure 5 This is a flowchart illustrating step S50 in an embodiment of a heart rate accuracy calculation method based on multi-array multimodal fusion according to this application. Detailed Implementation

[0073] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] In one embodiment, such as Figure 1 As shown, this application discloses a method for accurate heart rate calculation based on multi-array multimodal fusion, which specifically includes the following steps:

[0075] S10: Collect multi-source array data from users and construct a multi-array dataset.

[0076] In this embodiment, multi-source array data refers to the sequence of raw sensor data collected synchronously from the user, which contains at least three types of information: biosignals, environmental information, and motion status; multi-array dataset refers to a spatiotemporally unified structured data set formed by aligning these heterogeneous data with timestamps.

[0077] Specifically, biosignal arrays, environmental arrays, and motion state arrays are collected simultaneously; the specific composition of each array is defined, and a timestamp alignment algorithm is used to unify multi-source data onto the same timeline to complete the construction of the dataset.

[0078] S20: Preprocess the multi-array dataset and extract multi-dimensional features from the preprocessed array.

[0079] In this embodiment, preprocessing refers to removing ambient light interference, baseline drift, electromyography, and power frequency interference from the original data using specific signal processing methods; extracting multi-dimensional features refers to quantifying and calculating key indicators that can characterize heart rate cycle, exercise intensity, and environmental interference level from the preprocessed data.

[0080] Specifically, bandpass filtering and normalization are performed on the light intensity array of the photoelectric sensor to extract the pulse peak interval feature; wavelet transform is performed on the ECG analog signal array to remove interference and extract the R-wave interval feature; sliding window smoothing is performed on the ambient light intensity array to determine the environmental interference level feature; and synthetic acceleration values ​​are calculated based on the three-dimensional array of the accelerometer sensor to classify and extract the motion intensity level feature.

[0081] S30: Based on the extracted multi-dimensional features, identify the current dynamic scene, and adaptively assign computational weights to different biological signal arrays according to the identified dynamic scene.

[0082] In this embodiment, identifying the current dynamic scene refers to classifying the user's state into a predefined scene category based on characteristics such as exercise intensity and environmental interference level; adaptively allocating calculation weights refers to dynamically adjusting the contribution ratio of each biological signal in the final heart rate calculation according to the differences in signal-to-noise ratio and reliability of each biological signal under different scenes.

[0083] Specifically, firstly, a predefined scene classification model is constructed, which divides dynamic scenes into at least five categories: resting without interference, resting with interference, slight activity, moderate activity, and strenuous activity. Based on the extracted features of exercise intensity level and environmental interference level, the model is matched to determine the current monitoring scene. Finally, according to the current scene, a first weight W1 is assigned to the heart rate calculation path based on the light intensity array of the photoelectric sensor, and a second weight W2 is assigned to the heart rate calculation path based on the ECG analog signal array, and W1+W2=1 is satisfied.

[0084] S40: Based on the biological signals in the multi-array dataset, at least two preliminary heart rate values ​​are calculated in parallel, and weighted fusion is performed according to the assigned calculation weights to obtain a preliminary fused heart rate value.

[0085] In this embodiment, parallel computing refers to simultaneously and independently extracting heart rate values ​​from both photoelectric and electrocardiogram (ECG) biological signals; weighted fusion refers to combining two preliminary heart rate values ​​into a more reliable intermediate result based on computational weights adaptively allocated according to the scenario.

[0086] Specifically, based on the threshold peak detection algorithm, peak identification is performed on the pulse wave signal in the light intensity array of the photoelectric sensor, the timestamps of adjacent peak points are recorded and the interval is calculated, and the first heart rate value HR1 is obtained by using the heart rate conversion formula. Simultaneously, the differential threshold method is used to process the ECG analog signal array, and the second heart rate value HR2 is obtained by detecting the R wave peak point and calculating the time interval between adjacent peaks. Based on the scene weights W1 and W2 determined in step S30, the preliminary fused heart rate value is calculated by using the weighted formula HR_temp = W1×HR1 + W2×HR2.

[0087] S50: Based on the preset individual characteristic calibration model and historical heart rate change trend, perform anomaly judgment and correction on the preliminary fused heart rate value, and output the final heart rate value and the corresponding confidence score.

[0088] In this embodiment, the individual characteristic calibration model is a database that stores personalized parameters such as the user's skin light reflectance coefficient and electrocardiogram signal amplitude coefficient; the historical heart rate change trend refers to the recent heart rate change pattern established through time series analysis; anomaly judgment and correction refers to identifying and processing non-physiological heart rate fluctuations caused by motion artifacts or environmental interference; and confidence score refers to a comprehensive index that quantifies the reliability of heart rate results.

[0089] Specifically, baseline calibration is performed on the initial fused heart rate values ​​based on individual characteristic calibration parameters. Anomalies are identified by combining a time-series trend model and cross-validation is performed based on the original photoelectric and electrocardiogram characteristics to distinguish between interference and real physiological changes. According to the validation results, anomalies caused by interference are corrected using a trend prediction algorithm, while data for real changes are retained and the calibration model is updated. Finally, the processed heart rate sequence is optimized in a lightweight manner, and results with a comprehensive confidence score are output in real time. When the score is too low, a device or environment adjustment prompt is triggered.

[0090] In one embodiment, step S10, namely, collecting multi-source array data from users and constructing a multi-array dataset, includes:

[0091] S11: Synchronously acquire biological signal arrays, environmental arrays, and motion state arrays.

[0092] In this embodiment, synchronous acquisition refers to configuring a unified hardware clock source for all sensors, so that data points from different sensors are stamped with timestamps originating from the same reference, ensuring spatiotemporal consistency of data; the biosignal array contains raw photoelectric and electrocardiogram data reflecting cardiovascular activity; the environmental array contains external illumination and device wearing status data that affect signal quality; and the motion state array contains triaxial acceleration data reflecting the intensity of body movement.

[0093] Specifically, all sensor modules are activated and simultaneously collect data at a predefined sampling frequency: photoelectric sensors acquire skin reflection light intensity sequences to form photoelectric arrays, ECG electrodes acquire body surface potential difference sequences to form ECG arrays, ambient light sensors acquire ambient light intensity sequences, wear detection circuits acquire fit signals, and accelerometers acquire X, Y, and Z axis acceleration sequences; all data are appended in real time with microsecond-level timestamps based on the system clock.

[0094] S12: The biosignal array includes a photoelectric sensor light intensity array and an electrocardiogram analog signal array; the environment array includes an ambient light intensity array and a sensor fit array; the motion state array is a three-dimensional array of an accelerometer.

[0095] In this embodiment, the photoelectric sensor light intensity array refers to the time series of skin reflected light intensity, including red and infrared light channels, collected by a photodiode; the electrocardiogram analog signal array refers to the simulated waveform sequence of body surface potential difference collected by a contact electrode; the ambient light intensity array refers to the digital sequence of visible light intensity output by the ambient light sensor; the sensor fit array refers to the fit status score sequence generated by detecting the optical coupling degree between the sensor and the skin; and the three-dimensional array of the accelerometer refers to the time series set containing acceleration values ​​along three orthogonal axes: X, Y, and Z.

[0096] Specifically, the photoelectric array records the red and infrared light intensity values ​​corresponding to each sampling moment; the electrocardiogram array records the simulated electrocardiogram voltage value corresponding to each sampling moment; the ambient light array records the ambient light brightness value at each sampling moment; the fit array records the score value characterizing the tightness of the fit based on the reflected light reference or impedance value; and the acceleration array records the acceleration component values ​​in three spatial directions at each sampling moment.

[0097] S13: Using a timestamp alignment algorithm, the biosignal array, environment array, and motion state array are unified to the same time axis to construct the multi-array dataset.

[0098] In this embodiment, the timestamp alignment algorithm refers to a data processing method that maps multiple data sequences with different acquisition start times and sampling frequencies to the same high-resolution time reference line through time interpolation or resampling techniques; the multi-array dataset refers to a structured data table that has been time-aligned and contains values ​​of all sensor dimensions at each time point.

[0099] Specifically, using the system master clock as a reference, the original timestamps of all arrays are calibrated for error; the data of each array are resampled to a unified millisecond-level time grid using linear interpolation; the photoelectric intensity value, ECG voltage value, ambient light intensity value, fit score value, and triaxial acceleration value at the same time after resampling are integrated into a complete record and arranged in chronological order to form the final multi-array data table.

[0100] In one embodiment, such as Figure 2 As shown, in step S20, namely, preprocessing the multi-array dataset and extracting multi-dimensional features from the preprocessed array, the following steps are included:

[0101] S21: Perform bandpass filtering and normalization on the light intensity array of the photoelectric sensor, and extract the pulse peak interval feature from it.

[0102] In this embodiment, bandpass filtering refers to using a Butterworth filter to retain light intensity variation components that match the human pulse frequency, suppressing baseline drift and ambient light interference; normalization processing refers to scaling the filtered light intensity data to a uniform data value to eliminate signal amplitude differences caused by individual skin color or tightness of the clothing; pulse peak interval characteristics refer to the time difference between the main peaks of adjacent pulse waves, which is the core periodic parameter for calculating heart rate.

[0103] Specifically, a Butterworth bandpass filter (passband 0.5-8 Hz) is applied to the original light intensity array to filter out low-frequency motion artifacts and high-frequency electronic noise; the filtered signal is normalized to its maximum and minimum values; an adaptive threshold peak detection algorithm is used to identify the position of the pulse wave main peak on the normalized curve; and the time interval between consecutive main peaks is calculated to form a pulse peak interval feature sequence.

[0104] S22: Perform wavelet transform on the ECG analog signal array to remove electromyographic and power frequency interference, and extract R-wave interval features from it.

[0105] In this embodiment, wavelet transform refers to a time-frequency analysis method that decomposes the electrocardiogram (ECG) signal into components of different scales (frequency) and effectively separates the target signal from the interference components through selective reconstruction; electromyography (EMG) and power frequency interference refer to random high-frequency noise (usually >5Hz) generated by muscle contraction and fixed-frequency noise (50Hz / 60Hz) introduced by the power grid, respectively; R-wave interval characteristics refer to the time difference between adjacent R-wave peaks in the ECG signal, which is the gold standard periodic parameter for heart rate calculation.

[0106] Specifically, the original ECG voltage array is subjected to discrete wavelet transform to decompose the signal into multiple sub-bands; the coefficients of the high-frequency sub-band representing electromyographic noise and the specific sub-band representing power frequency interference are set to zero; the remaining sub-bands are reconstructed to obtain the denoised ECG signal; the R-wave peak point is accurately located in the reconstructed signal using the differential thresholding method; the time interval between consecutive R waves is calculated to form an R-wave interval characteristic sequence.

[0107] S23: Perform sliding window mean smoothing on the ambient light intensity array, and determine the environmental interference level characteristics based on the smoothed data.

[0108] In this embodiment, sliding window mean smoothing refers to moving a fixed-length window in chronological order on the data sequence and replacing the original value of the window center point with the arithmetic mean of the data within the window to suppress abrupt noise; environmental interference level characteristics refer to numerical indicators that quantify the impact of ambient light on the stability of photoelectric heart rate measurement.

[0109] Specifically, the original ambient light intensity array is smoothed by a sliding window with a window length of 10 sampling points to obtain a stable light intensity sequence; the fluctuation standard deviation of the stable sequence is calculated, and it is mapped to a discrete environmental interference level according to a preset threshold.

[0110] S24: Calculate the synthetic acceleration value based on the three-dimensional array of the acceleration sensor, and classify and extract the motion intensity level features according to the synthetic acceleration value.

[0111] In this embodiment, the synthesized acceleration value refers to the total acceleration amplitude obtained by vector synthesis of each axial component of the three-dimensional acceleration signal, which is used to characterize the overall motion intensity of the body; the motion intensity level feature refers to the discrete level used to describe the user's activity state, which is divided according to the amplitude range of the synthesized acceleration value.

[0112] Specifically, for the X, Y, and Z axis acceleration values ​​at each sampling time, the composite acceleration value is calculated according to the formula a=√(ax²+ay²+az²), and the gravitational acceleration component is subtracted. The average value of the composite acceleration value within the continuous time window is compared with a preset threshold to classify the exercise intensity level: for example, an average value less than 0.1 is resting, 0.1-0.3 is light activity, 0.3-0.6 is moderate activity, and greater than 0.6 is strenuous activity. The level corresponding to the current window is output as the exercise intensity level feature.

[0113] In one embodiment, such as Figure 3 As shown, in step S30, namely, identifying the current dynamic scene based on the extracted multi-dimensional features, and adaptively assigning computational weights to different biosignal arrays according to the identified dynamic scene, the process includes:

[0114] S31: Construct a predefined scene classification model, wherein the scene classification model divides dynamic scenes into at least five categories: resting without interference, resting with interference, slight activity, moderate activity, and intense activity.

[0115] In this embodiment, the scene classification model refers to a system based on feature threshold rules, which is used to automatically determine the monitoring scene category in which the user is located based on the input motion and environmental characteristics; dynamic scene refers to the real-time state that affects the quality of heart rate signal and is jointly determined by the intensity of user activity and external environmental conditions.

[0116] Specifically, a two-dimensional combination of exercise intensity level and environmental disturbance level is defined, and a mapping rule is established for five scenarios: when the exercise intensity is at rest and the environmental disturbance level is low, it corresponds to the "resting without disturbance" scenario; when the exercise intensity is at rest and the environmental disturbance level is medium or high, it corresponds to the "resting with disturbance" scenario; when the exercise intensity is light activity, it corresponds to the "light activity" scenario; when the exercise intensity is moderate activity, it corresponds to the "moderate activity" scenario; and when the exercise intensity is vigorous activity, it corresponds to the "vigorous activity" scenario. This mapping rule is then solidified as a scenario classification model.

[0117] S32: Based on the extracted motion intensity level features and environmental interference level features, match the scene classification model to determine the current monitoring scene.

[0118] In this embodiment, matching refers to comparing the combination of real-time motion intensity and environmental interference level with the mapping rules in the predefined scene classification model to find the corresponding scene category label; the current monitored scene refers to the classification result of the user's state determined in the most recent decision window.

[0119] Specifically, the system obtains the motion intensity level feature value (e.g., "light activity") and the environmental disturbance level feature value (e.g., "low") within the current time window; uses these two feature values ​​as inputs to look up the mapping table defined in the scene classification model; and outputs the corresponding scene label based on the matching result. For example, if the motion intensity is "light activity" and the environmental disturbance is "low", then the current monitoring scene is determined to be "light activity".

[0120] S33: Based on the current monitoring scenario, assign a first weight W1 to the heart rate calculation path based on the photoelectric sensor light intensity array, and assign a second weight W2 to the heart rate calculation path based on the electrocardiogram analog signal array, and the sum of the first weight W1 and the second weight W2 is 1.

[0121] In this embodiment, weight allocation refers to assigning different fusion coefficients to the heart rate calculation paths of the two types of physiological signals (photoplethysmography signal and electrocardiogram signal) based on the differences in signal-to-noise ratio and reliability in the current scenario. W1 and W2 represent the reliability weights of the photoplethysmography signal path and the electrocardiogram signal path, respectively, and their sum is fixed at 1 to ensure the normalization of the fusion result.

[0122] Specifically, based on the current monitoring scenario label determined in step S32, a heart rate calculation path is assigned. For example, in the "resting without interference" scenario, W1=0.4, W2=0.6; in the "resting with interference" scenario, W1=0.3, W2=0.7; in the "mild activity" scenario, W1=0.5, W2=0.5; in the "moderate activity" scenario, W1=0.6, W2=0.4; and in the "vigorous activity" scenario, W1=0.7, W2=0.3. The corresponding W1 and W2 values ​​for this scenario are then output to the fusion calculation module.

[0123] In one embodiment, such as Figure 3 As shown, in step S33, the first weight W1 assigned to the heart rate calculation path based on the photoelectric sensor light intensity array, and the second weight W2 assigned to the heart rate calculation path based on the ECG analog signal array, satisfy the following rules:

[0124] In a quiet, undisturbed environment, W2 > W1;

[0125] In a resting, disturbed scenario, the weight of W2 is greater than its value in a resting, undisturbed scenario.

[0126] In scenarios with minimal activity, the weights of W1 and W2 are equal or close.

[0127] In moderate activity scenarios, W1 > W2;

[0128] In scenarios involving intense activity, the weight of W1 is greater than its value in scenarios involving moderate activity.

[0129] In this embodiment, the weight allocation rule refers to a set of pre-set logical constraints based on the typical performance of photoelectric and electrocardiogram signals under different physiological states and environmental interferences, in order to ensure that the fusion weight can dynamically and reasonably reflect the real-time reliability of each signal source.

[0130] Specifically, in a resting, undisturbed environment, the ECG signal is stable and free of motion artifacts, so it is given a higher weight (e.g., W2=0.6, W1=0.4), satisfying W2 > W1. When environmental interference exists, the photoelectric signal is more susceptible to influence, so its weight needs to be further reduced (e.g., W2=0.7, W1=0.3), making the W2 value higher than its 0.6 value in the undisturbed environment. During mild activity, the reliability of the two types of signals tends to be balanced, so W1≈W2 (e.g., both are 0.5). After entering moderate activity, the electromyographic interference of motion on the ECG signal increases, while the motion-compensated photoelectric signal is relatively more reliable, so W1 > W2 (e.g., W1=0.6, W2=0.4). During strenuous activity, the quality of the ECG signal further decreases, and the photoelectric path weight needs to be further increased (e.g., W1=0.7, W2=0.3), satisfying the W1 value being greater than its 0.6 value during moderate activity.

[0131] In one embodiment, such as Figure 3 As shown, in step S40, that is, based on the biosignals in the multi-array dataset, at least two preliminary heart rate values ​​are calculated in parallel, and weighted fusion is performed according to the assigned calculation weights to obtain a preliminary fused heart rate value, specifically including:

[0132] S41: An adaptive threshold peak detection algorithm is used on the pulse peak interval feature sequence to identify adjacent peak points and calculate the peak time interval, and the time interval is converted into a heart rate value as the first heart rate value HR1;

[0133] S42: The R-wave peak points are identified using the differential threshold method on the R-wave interval feature sequence, the time interval between adjacent R-wave peak points is calculated, and the time interval is converted into a heart rate value as the second heart rate value HR2.

[0134] S43: The preliminary fusion heart rate value HR_temp is calculated using the following fusion formula: HR_temp = W1 × HR1 + W2 × HR2.

[0135] In this embodiment, parallel computing refers to simultaneously performing independent heart rate calculations on both photoelectric and electrocardiogram feature sequences; weighted fusion refers to combining the two heart rate estimates into a single, more reliable intermediate heart rate result based on dynamically allocated weight coefficients.

[0136] Specifically, a sliding analysis is performed on the pulse peak interval feature sequence. Whenever two adjacent peak points are detected, their time difference (in seconds) is calculated and converted into a heart rate value per minute using the formula HR1 = 60 / time difference, and updated in real time. A similar operation is performed on the R wave interval feature sequence, detecting the time interval between adjacent R wave peaks and converting it into HR2. The real-time weights W1 and W2 obtained in step S33 and the real-time calculated HR1 and HR2 are substituted into the formula HR_temp = W1 × HR1 + W2 × HR2 to output the preliminary fused heart rate value at the current moment.

[0137] In one embodiment, such as Figure 4 As shown, before step S50, that is, before the abnormality judgment and correction of the preliminary fused heart rate value, an individual characteristic calibration step is also included:

[0138] S441: Based on standard heart rate data collected from the user in a resting state and a synchronized array of biological signals, calculate and establish an initial individual calibration parameter library containing skin light reflectance coefficient and electrocardiogram signal amplitude coefficient.

[0139] In this embodiment, standard heart rate data refers to the heart rate value measured in the user's resting state and used as a benchmark reference; the initial individual calibration parameter library refers to the data structure that stores the user's personalized physiological characteristic parameters.

[0140] Specifically, the system guides users to wear the device while at rest, simultaneously collecting their standard heart rate data along with the light intensity array from the photoelectric sensor and the ECG analog signal array collected by the device itself. Based on the statistical relationship between the standard heart rate value and the peak amplitude of the photoelectric signal, the system calculates the user's skin light reflectance coefficient. Based on the statistical relationship between the standard heart rate value and the R-wave amplitude of the ECG signal, the system calculates the ECG signal amplitude coefficient. These two coefficients and other relevant statistics are stored as a user-specific initial individual calibration parameter library.

[0141] S442: During continuous monitoring, a dynamic calibration mechanism is triggered based on a preset cycle or changes in the user's physiological state. The parameters in the individual calibration parameter library are updated by combining recent resting heart rate data and its changing trends.

[0142] In this embodiment, the dynamic calibration mechanism refers to the program logic by which the system actively or passively recalculates and adjusts personalized parameters during long-term use; the preset cycle refers to the calibration strategy executed at fixed time intervals (such as every 7 days); changes in user physiological state refer to physiological events that may affect signal characteristics, identified through features such as heart rate variability and activity patterns; and recent resting heart rate data and its changing trend refer to the heart rate sequence and its statistical regularity measured by the user in a resting state over a recent period of time.

[0143] Specifically, when a preset period (e.g., 7 days) is reached or the system detects a significant deviation in the user's baseline resting heart rate, the calibration process is automatically triggered. The device-calculated heart rate data and its corresponding raw biological signal array are extracted from the user's resting state within the most recent period (e.g., 24 hours) before the trigger. The trend of heart rate changes and the stability of signal characteristics during this period are analyzed. The skin light reflectance coefficient and ECG signal amplitude coefficient are recalculated based on the new data. The individual calibration parameter library is updated with the newly calculated parameters, replacing or smoothing the original parameter values ​​to adapt the model to the user's current physiological state.

[0144] In one embodiment, such as Figure 4 As shown, before step S50, before the abnormality judgment and correction of the fused preliminary heart rate value, a time-series trend modeling step is also included:

[0145] S451: Construct a time series of the preliminary fused heart rate value HR_temp based on a sliding window, and calculate the slope, mean, and standard deviation of the heart rate change within the window.

[0146] In this embodiment, time series construction refers to arranging continuous heart rate values ​​in chronological order to form a time series data that can be analyzed; a sliding window refers to a fixed-length data segment that moves forward on the time series to extract local time series features; the slope, mean, and standard deviation are used to characterize the rate of change, average level, and degree of fluctuation of heart rate within the window period, respectively.

[0147] Specifically, using the HR_temp sequence as input, a sliding window covering the most recent 10 heart rate values ​​(approximately 10 seconds of data) is employed. After each window slide, all heart rate values ​​within the window are extracted to form a subsequence. A first-order linear fit is performed on this subsequence, and the slope of the fitted line is the slope of the heart rate change. The arithmetic mean of the subsequence is calculated as the mean heart rate within the window. The standard deviation of the subsequence is calculated as the standard deviation of heart rate fluctuation. These three feature values ​​are updated in real time and together constitute the output of a trend model describing the short-term dynamics of the current heart rate.

[0148] S452: Construct a heart rate trend model to identify heart rate change patterns and determine whether there are sudden changes in physiological state or continuous abnormal fluctuations.

[0149] In this embodiment, the heart rate trend model refers to a judgment logic based on recent heart rate time-series statistical characteristics, used to assess whether the current heart rate change conforms to the expected physiological pattern; the heart rate change pattern refers to a specific dynamic form jointly characterized by features such as slope, mean, and standard deviation; physiological state mutation refers to a rapid and physiologically consistent heart rate jump caused by exercise, emotion, or changes in physical state; and continuous abnormal fluctuations refer to random, violent, and physiologically inconsistent heart rate fluctuations caused by signal interference.

[0150] Specifically, the heart rate trend model will be constructed by inputting the slope, mean, and standard deviation calculated in real time in step S451: if the absolute value of the slope continuously exceeds the set threshold and the mean changes directionally, it will be identified as a sudden change in physiological state (such as from rest to exercise); if the standard deviation continuously exceeds the set threshold and the slope has no consistent direction, and the heart rate value shows irregular and drastic jumps in a short period of time, it will be identified as continuous abnormal fluctuations.

[0151] In one embodiment, such as Figure 5 As shown, in step S50, the abnormality judgment and correction of the preliminary fused heart rate value includes:

[0152] S51: Combine the heart rate change pattern output by time-series trend modeling with the preset abnormal judgment threshold to identify abnormal points in the preliminary fused heart rate value HR_temp.

[0153] In this embodiment, an anomaly refers to a heart rate data point in the HR_temp sequence that deviates significantly from the predicted range of the heart rate trend model or does not conform to physiological laws due to signal interference (such as motion artifacts or sudden environmental changes); the anomaly judgment threshold refers to a quantitative standard set based on physiological knowledge or experimental data.

[0154] Specifically, the current HR_temp value and the current heart rate change pattern label (such as "normal", "physiological mutation", or "abnormal fluctuation") output by the time-series trend model are obtained. If it is "normal", a relatively narrow reasonable range around the mean heart rate (such as mean ± 5 beats / minute) is set as the threshold. If it is "physiological mutation", the range is appropriately widened (such as mean ± 10 beats / minute) to accommodate physiological changes. If it is "abnormal fluctuation", a stricter threshold based on the rate of change of adjacent points is used. The current HR_temp value is compared with the corresponding reasonable range. If it exceeds the range, the HR_temp value at that moment is marked as an outlier.

[0155] S52: When an anomaly is detected, cross-validation is performed based on the pulse peak interval characteristics and R wave interval characteristics in the biosignal array to determine whether the cause of the anomaly is signal interference or a real physiological change.

[0156] In this embodiment, cross-validation refers to simultaneously comparing and analyzing two independent heart rate cycle estimates derived from photoelectric signals (pulse peak interval) and electrocardiogram signals (R wave interval) to observe their consistency in order to infer the root cause of the abnormality; signal interference refers to signal distortion caused by external factors such as motion artifacts, environmental noise, and poor device contact; physiological real change refers to the real heart rate fluctuations caused by changes in user movement, emotions, or physical condition that conform to physiological mechanisms.

[0157] Specifically, the corresponding time window for the occurrence of the anomaly is located; the synchronous pulse peak interval feature sequence and R wave interval feature sequence are extracted from this time window; these two feature sequences are converted into independent photoelectric heart rate value sequences and electrocardiogram heart rate value sequences, respectively; the consistency of the values ​​of the preliminary fused heart rate value, photoelectric heart rate value and electrocardiogram heart rate value at the anomaly point is compared; if the photoelectric heart rate value and the electrocardiogram heart rate value at this moment are both significantly deviated from the preliminary fused heart rate value HR_temp and are close to each other, it is likely to be judged as a real physiological change; if only one path (usually the optical circuit) has a significant jump while the other path (usually the cardiac circuit) remains stable, or if the two feature sequences themselves are severely missing or disordered within this window, it is judged to be caused by signal interference.

[0158] S53: If it is confirmed that the problem is caused by signal interference, the abnormal point is corrected by a prediction algorithm based on time-series trends, and the corrected heart rate value HR_cal is output.

[0159] S54: If the change is determined to be a genuine physiological change, retain the heart rate value and update the relevant parameters in the individual characteristic calibration model.

[0160] In this embodiment, "correction" refers to the algorithmic repair of distorted data caused by signal interference; "retention" refers to the adoption of real physiological fluctuation data; and "parameter update" refers to the optimization of individual models using real physiological change data to achieve adaptive evolution of the system.

[0161] Specifically, if cross-validation determines that the cause of the anomaly is signal interference, a prediction algorithm based on time-series trends is used to extrapolate a correction value HR_cal based on the trend of normal heart rate data before and after the anomaly point, and this value is used to replace the output of the anomaly point HR_temp. If it is determined to be a real physiological change, the anomaly point HR_temp is directly retained as a valid output. At the same time, the heart rate and biosignal feature data within a window before and after this change event are used to recalculate or weight and update the relevant physiological response coefficients in the individual calibration parameter library, so that the model continuously fits the user's physiological characteristics.

[0162] In one embodiment, such as Figure 5 As shown, in step S50, the output of the final heart rate value and the corresponding confidence score includes:

[0163] S55: The algorithm for each module, including feature extraction, weight allocation, heart rate calculation, and anomaly correction, is lightweighted by using fixed-point arithmetic to reduce computational complexity.

[0164] In this embodiment, algorithm lightweighting refers to reducing computational resource consumption and memory usage by optimizing algorithm structure, simplifying calculation steps, and reducing numerical precision requirements; fixed-point arithmetic refers to performing arithmetic operations on data types with integers or a fixed number of decimal places, which can significantly reduce the processing latency and power consumption of embedded or mobile devices compared to floating-point arithmetic.

[0165] Specifically, in the feature extraction module, floating-point coefficients in the filtering and transformation algorithms are converted into fixed-point coefficients, and intermediate variables are stored as fixed-point numbers; in the weight allocation module, the logical judgment of scene matching and weight mapping is simplified to a table lookup operation; in the heart rate calculation and anomaly correction module, the multiplication and division operations in peak detection, time interval calculation, and trend prediction are optimized to a combination of fixed-point shifting and addition. Through these optimizations, the overall computational complexity and power consumption are reduced while ensuring that the core algorithm logic remains unchanged.

[0166] S56: Controls the delay of the overall calculation process to achieve real-time calculation and output of heart rate.

[0167] In this embodiment, control delay refers to constraining and optimizing the entire link time from data acquisition, processing, calculation to final output to ensure that the system response time meets the real-time requirements; real-time calculation and output means that within each processing cycle, the system can complete all calculations and give the current heart rate result within an acceptable time window (usually ≤1 second).

[0168] Specifically, a pipelined processing architecture is implemented to enable modules such as data acquisition, preprocessing, and feature extraction to work in parallel to shorten waiting time; a maximum allowable processing time is set for each module, and lightweight optimization is used to ensure that it is completed within this time limit; an end-to-end time monitoring mechanism is established from sensor data input to heart rate value output to ensure that the total processing time of a single session is stably controlled within the set target (such as 1 second), thereby achieving low latency and continuous real-time output of heart rate.

[0169] S57: Output the final heart rate value and corresponding confidence score at a preset frequency. The confidence score is obtained by comprehensive calculation based on signal quality, scene matching degree and timing consistency.

[0170] S58: When the confidence score is lower than the set threshold, trigger a sensor wearing status reminder or an environment adjustment prompt.

[0171] In this embodiment, the preset frequency refers to the fixed time interval (e.g., once per second) at which the system provides data updates; the confidence score is a comprehensive indicator that quantifies the credibility of the current heart rate result, with a score range typically between 0 and 100; signal quality refers to the clarity and signal-to-noise ratio of the original photoelectric and electrocardiogram signals; scene matching degree refers to the clarity of the current dynamic scene classification and the rationality of the weight allocation; and temporal consistency refers to the degree of agreement between the current heart rate value and its historical trend.

[0172] Specifically, the system outputs the processed final heart rate value along with the calculated confidence score at a fixed frequency of 1Hz. When the output confidence score is consistently lower than a preset threshold (such as 60 points), the system determines that the current monitoring conditions are unfavorable and automatically issues targeted prompts to the user, such as "Please adjust the device" or "Please move to a stable lighting environment," through device screen display, vibration, or voice broadcast.

[0173] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0174] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. The actual structure is not limited to this. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A method for accurate heart rate calculation based on multi-array multimodal fusion, characterized in that, The accurate heart rate calculation method based on multi-array multimodal fusion includes the following steps: Collect multi-source array data from users to construct a multi-array dataset; The multi-array dataset is preprocessed, and multi-dimensional features are extracted from the preprocessed arrays; Based on the extracted multi-dimensional features, the current dynamic scene is identified, and computational weights are adaptively assigned to different biological signal arrays according to the identified dynamic scene. Based on the biological signals in the multi-array dataset, at least two preliminary heart rate values ​​are calculated in parallel, and then weighted and fused according to the assigned calculation weights to obtain a preliminary fused heart rate value. Based on a preset individual characteristic calibration model and historical heart rate change trends, the preliminary fused heart rate value is anomaly detected and corrected, and the final heart rate value and corresponding confidence score are output.

2. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 1, characterized in that, The process of collecting multi-source array data from users to construct a multi-array dataset includes: Simultaneously acquire biological signal arrays, environmental arrays, and motion state arrays; The biosignal array includes a photoelectric sensor light intensity array and an electrocardiogram analog signal array; the environment array includes an ambient light intensity array and a sensor fit array; the motion state array is a three-dimensional array of an accelerometer. A timestamp alignment algorithm is used to unify the biosignal array, environment array, and motion state array onto the same time axis to construct the multi-array dataset.

3. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 2, characterized in that, The preprocessing of the multi-array dataset and the extraction of multi-dimensional features from the preprocessed arrays include: The light intensity array of the photoelectric sensor is subjected to bandpass filtering and normalization, and the pulse peak interval feature is extracted from it. Wavelet transform is performed on the ECG analog signal array to remove electromyographic and power frequency interference, and R-wave interval features are extracted from it. The ambient light intensity array is smoothed using a sliding window mean, and the environmental interference level characteristics are determined based on the smoothed data. Based on the three-dimensional array of the accelerometer, the synthetic acceleration value is calculated, and the motion intensity level features are divided and extracted according to the synthetic acceleration value.

4. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 3, characterized in that, The process of identifying the current dynamic scene based on the extracted multi-dimensional features, and adaptively assigning computational weights to different biosignal arrays according to the identified dynamic scene, includes: Construct a predefined scene classification model, which classifies dynamic scenes into at least five categories: resting without interference, resting with interference, slight activity, moderate activity, and vigorous activity. Based on the extracted motion intensity level features and environmental interference level features, the scene classification model is matched to determine the current monitoring scene; Based on the current monitoring scenario, a first weight W1 is assigned to the heart rate calculation path based on the light intensity array of the photoelectric sensor, and a second weight W2 is assigned to the heart rate calculation path based on the ECG analog signal array, and the sum of the first weight W1 and the second weight W2 is 1.

5. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 4, characterized in that, The adaptive allocation of calculation weights based on different biosignal arrays according to the identified dynamic scene includes: a first weight W1 allocated to the heart rate calculation path based on the photoelectric sensor light intensity array, and a second weight W2 allocated to the heart rate calculation path based on the ECG analog signal array, satisfying the following rules: In a quiet, undisturbed environment, W2 is greater than W1; In a resting, disturbed scenario, the weight of W2 is greater than its value in a resting, undisturbed scenario. In scenarios with minimal activity, the weights of W1 and W2 are equal or close. In moderate activity scenarios, W1 is greater than W2; In scenarios involving intense activity, the weight of W1 is greater than its value in scenarios involving moderate activity.

6. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 4, characterized in that, The process of calculating at least two preliminary heart rate values ​​in parallel based on the biological signals in the multi-array dataset, and then weighting and fusing them according to the assigned calculation weights to obtain a preliminary fused heart rate value, specifically includes: An adaptive threshold peak detection algorithm is used to identify adjacent peak points and calculate the peak time interval for the pulse peak interval feature sequence. The time interval is then converted into a heart rate value and used as the first heart rate value HR1. The R-wave peak points are identified using the differential threshold method on the R-wave interval feature sequence. The time interval between adjacent R-wave peak points is calculated, and the time interval is converted into a heart rate value as the second heart rate value HR2. The preliminary fused heart rate value HR_temp is calculated using the following fusion formula: HR_temp = W1 × HR1 + W2 × HR2.

7. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 1, characterized in that, Before performing anomaly detection and correction on the preliminary fused heart rate value, an individual characteristic calibration step is also included: Based on standard heart rate data collected from users in a resting state and synchronized biological signal arrays, an initial individual calibration parameter library containing skin light reflectance coefficient and electrocardiogram signal amplitude coefficient is calculated and established. During continuous monitoring, a dynamic calibration mechanism is triggered based on a preset cycle or changes in the user's physiological state. The parameters in the individual calibration parameter library are updated by combining recent resting heart rate data and its changing trends.

8. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 1, characterized in that, Before performing anomaly detection and correction on the initial fused heart rate value, a time-series trend modeling step is also included: Based on the sliding window, a time series sequence is constructed from the preliminary fused heart rate value HR_temp, and the slope, mean, and standard deviation of the heart rate change within the window are calculated. Construct a heart rate trend model to identify heart rate change patterns and determine whether there are sudden changes in physiological state or continuous abnormal fluctuations.

9. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 1, characterized in that, The process of anomaly detection and correction of the preliminary fused heart rate value includes: By combining the heart rate change pattern output from time-series trend modeling with a preset anomaly detection threshold, abnormal points in the preliminary fused heart rate value HR_temp are identified. When an anomaly is identified, cross-validation is performed based on the pulse peak interval characteristics and R wave interval characteristics in the biosignal array to determine whether the cause of the anomaly is signal interference or a real physiological change. If the problem is confirmed to be caused by signal interference, a prediction algorithm based on time-series trends is used to correct the anomalies and output the corrected heart rate value HR_cal. If the change is determined to be a genuine physiological change, the heart rate value is retained, and the relevant parameters in the individual characteristic calibration model are updated.

10. The method for accurate heart rate calculation based on multi-array multimodal fusion according to claim 1, characterized in that, The output of the final heart rate value and the corresponding confidence score includes: The algorithms for feature extraction, weight allocation, heart rate calculation, and anomaly correction are lightweighted, and fixed-point arithmetic is used to reduce computational complexity. Control the delay of the overall calculation process to achieve real-time calculation and output of heart rate; The final heart rate value and corresponding confidence score are output at a preset frequency. The confidence score is obtained by comprehensive calculation based on signal quality, scene matching degree and timing consistency. When the confidence score is lower than the set threshold, a sensor wearing status reminder or an environment adjustment prompt will be triggered.

Images