A Dynamic Fitness Interval Calculation System and Method Based on Multimodal Signal Fusion Analysis

Abstract

This invention discloses a dynamic fitness interval calculation system and method based on multimodal signal fusion analysis. The system includes a signal acquisition module, an embedded low-power processing module, and a reinforcement learning algorithm module. The signal acquisition module acquires the user's surface electromyography (EMG) signals, blood oxygen saturation signals, and heart rate variability (HRV) signals. The embedded low-power processing module performs noise reduction and feature extraction on the user's EMG, HRV, and HRV signals, and obtains the pre-fatigue level of the EMG signals based on the EMG features. The reinforcement learning algorithm module calculates the user's fitness intervals based on the user's historical exercise data, the pre-fatigue level of the EMG signals, HRV characteristics, and HRV recovery rate characteristics. This system and method solve the problems of subjectivity, static nature, and limited data in interval setting in traditional fitness training, providing a more accurate, efficient, and comfortable interval experience.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fitness equipment and biosignal processing technology, specifically relating to a dynamic fitness interval calculation system and method based on multimodal signal fusion analysis. Background Technology

[0002] In the field of fitness training, the proper setting of intervals is crucial for improving training effectiveness, preventing overtraining, and promoting muscle recovery. However, the two currently used mainstream methods—user subjective judgment and fixed timer mode—both have significant shortcomings, limiting the optimization of training results.

[0003] While users' subjective assessment of rest duration may seem direct and convenient, it actually carries the risk of subjective bias. Each person's fatigue perception threshold is different, and individual differences such as training experience, physical condition, and psychological state can influence this, making it difficult for subjective assessments of fatigue to accurately reflect the true metabolic state of muscles. This mismatch can lead to two extremes: either muscle damage or chronic fatigue accumulation due to overtraining, or insufficient training intensity due to excessive rest, thus affecting the efficiency of muscle growth and performance improvement.

[0004] On the other hand, while fixed-timer models provide a standardized framework for rest periods, seemingly fair and reasonable, they actually ignore the widespread individual differences among trainees. An individual's fitness level, training goals, and even their physical condition on the day can significantly impact the required recovery time. For example, a beginner trainee may need a longer recovery time to adapt to training intensity, while an advanced trainee may require shorter rest periods to pursue higher-intensity training. Fixed-timers cannot flexibly adapt to these dynamic changes, thus significantly reducing the flexibility and effectiveness of training plans.

[0005] Furthermore, the limitations of current smart fitness devices in monitoring training status cannot be ignored. While these devices can track basic data such as heart rate and number of movements, this information falls short in reflecting the neuromuscular drive state and metabolic fatigue. Muscle fatigue is a complex physiological process involving multiple levels, including a decline in neuromuscular coordination, depletion of energy reserves, and accumulation of metabolic products. These deep-seated changes are difficult to fully capture using only surface indicators such as heart rate or movement count. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a dynamic fitness interval calculation system and method based on multimodal signal fusion analysis, which addresses the shortcomings of the existing technology.

[0007] To address the aforementioned technical issues, in a first aspect, a dynamic fitness interval calculation system based on multimodal signal fusion analysis is disclosed, comprising a signal acquisition module, an embedded low-power processing module, and a reinforcement learning algorithm module.

[0008] The signal acquisition module is used to acquire the user's surface electromyography signal, blood oxygen saturation signal, and heart rate variability signal;

[0009] The embedded low-power processing module is electrically connected to the signal acquisition module and is used to perform noise reduction and feature extraction on the user's surface electromyography (EMG) signal, heart rate variability (HRV) signal, and blood oxygen saturation signal to obtain surface EMG statistical features, HRV frequency domain features, and blood oxygen recovery rate features; and to obtain the surface EMG signal pre-fatigue level based on the surface EMG statistical features.

[0010] The reinforcement learning algorithm module is used to store the user's historical exercise data and calculate the user's fitness intervals based on the user's historical exercise data, surface electromyography signal pre-fatigue level, heart rate variability characteristics, and blood oxygen saturation recovery rate characteristics.

[0011] Furthermore, the embedded low-power processing module includes a signal noise reduction unit, a feature extraction unit, and a lightweight model processing unit.

[0012] The signal noise reduction unit is used to remove noise interference from the user's surface electromyography signal, heart rate variability signal and blood oxygen saturation signal to obtain the noise-reduced user surface electromyography signal, heart rate variability signal and blood oxygen saturation signal.

[0013] The feature extraction unit is used to extract amplitude and time-related statistical features from the noise-reduced user surface electromyography signal, extract frequency domain features from the noise-reduced heart rate variability signal, and extract blood oxygen recovery rate features from the noise-reduced blood oxygen saturation signal.

[0014] The lightweight model processing unit is used to obtain the pre-fatigue level of the surface electromyography signal based on the statistical characteristics of surface electromyography.

[0015] Furthermore, the signal denoising unit uses a fourth-order Butterworth bandpass filter to denoise the surface electromyography signal. The passband frequency range is 10-500Hz. Zero phase distortion is achieved through forward and reverse filtering to obtain the denoised user surface electromyography signal.

[0016] The heart rate variability signal was subjected to 0.1-0.4Hz bandpass filtering, moving average filtering and linear fitting to remove baseline drift and obtain the noise-reduced heart rate variability signal;

[0017] An adaptive filter was used to eliminate ambient light interference in the blood oxygen saturation signal, and the arterial blood flow signal was separated by wavelet transform to obtain the denoised blood oxygen saturation signal.

[0018] Furthermore, the feature extraction unit extracts amplitude and time-related statistical features from the denoised user surface electromyography (EMG) signal, including: performing time-domain and frequency-domain analysis on the denoised user EMG signal; calculating the root mean square (RMS), variance, mean absolute value (MAV), and zero-crossing rate (ZC) in the time-domain analysis; and converting the EMG signal from the time domain to the frequency domain through Fast Fourier Transform (FFT) in the frequency domain analysis, thereby revealing the frequency components of the EMG signal.

[0019] The feature extraction unit extracts frequency domain features from the denoised heart rate variability signal, including: performing frequency domain analysis on the denoised heart rate variability signal to obtain low-frequency and high-frequency components, wherein the low-frequency and high-frequency components reflect the activity of the sympathetic and parasympathetic nervous systems, respectively, and the ratio of low-frequency to high-frequency components can be used to assess the balance state of the autonomic nervous system.

[0020] The feature extraction unit extracts the blood oxygen recovery rate feature from the denoised blood oxygen saturation signal, including: calculating blood oxygen saturation based on the ratio of red light to infrared light absorbance, calculating the slope of the blood oxygen saturation recovery curve using a linear regression algorithm, and generating the blood oxygen recovery rate.

[0021] Furthermore, the reinforcement learning algorithm module calculates the user's exercise intervals, including calculating the user's rest time. Based on the pre-fatigue level of surface electromyography (EMG) signals, the ratio of low-frequency to high-frequency components of heart rate variability signals, blood oxygen recovery rate, user subjective ratings, and whether excessive rest has occurred, the reward function R of the reinforcement learning algorithm is designed as follows:

[0022]

[0023] Where α, β, and γ are weighting coefficients used to adjust the relative importance of each component in the reward function R; iPEMGF t+1 and iPEMGF t R represents the pre-fatigue level of surface electromyography signals at time t+1 and time t, respectively. 过度休息惩罚 This represents the impact of rest time exceeding a preset value on the reward function R; R HRV恢复 This represents the influence of the ratio of low-frequency to high-frequency components of the heart rate variability signal on the reward function R. This represents the effect of the blood oxygenation recovery rate on the reward function R.

[0024] The reward function R comprehensively considers user recovery effectiveness, training efficiency, and user satisfaction. By continuously adjusting the suggested rest time, it aims to maximize long-term training results and user experience. Regularly collecting user feedback allows for adjustments to the reward function and rest time of the reinforcement learning algorithm module, more accurately reflecting user preferences and needs.

[0025] Furthermore, the influence value R of the ratio of low-frequency to high-frequency components of the heart rate variability signal on the reward function R. HRV恢复 The formula is as follows:

[0026]

[0027] Where δ is the weighting coefficient for the heart rate variability recovery reward, used to adjust the contribution ratio of the heart rate variability index in the overall reward function. This represents the ratio of low-frequency to high-frequency components of the heart rate variability signal after the user has rested according to the recommended rest period. It represents the ratio of low-frequency to high-frequency components of the heart rate variability signal when the user is in a completely relaxed resting state, and represents the user's ideal recovery target, used to evaluate the recovery effect after rest;

[0028] The influence of the blood oxygenation recovery rate on the reward function R The formula is as follows:

[0029]

[0030] Where ε is the blood oxygen saturation penalty weighting coefficient, used to control the degree of impact of insufficient blood oxygen saturation recovery on the total reward, and Th represents the preset threshold for blood oxygen recovery rate. This indicates the rate at which a user's blood oxygen levels recover after resting for the recommended time.

[0031] Furthermore, the signal acquisition module includes a silicon microneedle array sensor and a dual-wavelength PPG (photoplethysmographic) sensor. The silicon microneedle array sensor is used to acquire surface electromyography signals of the user; the dual-wavelength PPG sensor is used to acquire blood oxygen saturation signals and heart rate variability signals.

[0032] Furthermore, the system also includes a user interaction module, which includes a display screen and a sound output device, for feeding back the rest time generated by the reinforcement learning algorithm module to the user.

[0033] The system architecture consists of the following key layers: signal acquisition layer, edge processing layer, decision layer, and feedback layer.

[0034] Signal Acquisition Layer: The core of this layer lies in integrating a silicon microneedle array sensor into the system. This design not only ensures that users do not experience pain from the inserted electrodes but also effectively prevents slippage due to sweating. The flexible dry electrode array can closely adhere to the surface of human muscles, effectively capturing the weak electrical signals generated by muscle activity, providing a reliable foundation for subsequent data processing and analysis. Furthermore, to further enhance the system's comprehensiveness and accuracy, this layer also integrates a dual-wavelength PPG sensor for acquiring heart rate variability (HRV) and blood oxygen saturation (SpO2) data. HRV reflects the dynamic balance of the autonomic nervous system by monitoring changes in heart rate, while SpO2 assesses the oxygen content in the blood by measuring the reflection of red and infrared light. The PPG sensor is physically isolated from the silicon microneedle array at a certain distance to avoid signal interference, while a flexible fit design ensures close contact between the sensor and the skin, reducing motion artifacts. By acquiring multimodal signals, the system can more comprehensively assess the user's physiological state, providing richer data support for subsequent fatigue analysis and rest recommendations.

[0035] Edge Processing Layer: After signal acquisition, an embedded low-power processing module is used for signal denoising, feature extraction, and lightweight model inference. The innovation at this layer lies in integrating high-performance computing units into a compact device, enabling real-time processing and analysis of sEMG, HRV, and SpO2 signals, significantly improving the timeliness and accuracy of data processing. The lightweight model quantifies the characteristic signals of sEMG, calculating the pre-fatigue level, providing a basis for the decision-making layer to calculate the stability of the pre-fatigue level at adjacent time points. For HRV signals, low-frequency components are first extracted using a 0.1-0.4Hz bandpass filter, and baseline drift is eliminated by combining it with a moving average filter. Then, time-domain features, such as SDNN, RMSSD, and the frequency-domain feature LF / HF power ratio, are calculated to quantify autonomic nerve stress, providing a basis for the decision-making layer to calculate HRV recovery. For SpO2 signals, adaptive filtering is used to eliminate ambient light interference, and wavelet transform is used to separate arterial blood flow signals. Finally, blood oxygen saturation is calculated based on the ratio of red to infrared absorbance, generating blood oxygen recovery rate features, providing a basis for the decision-making layer to calculate SpO2 penalties.

[0036] Decision Layer: This layer introduces a reinforcement learning-based dynamic interval optimization algorithm. This algorithm combines the user's historical exercise data with real-time pre-fatigue levels and intelligently generates personalized rest suggestions through an activation function. This innovation not only enhances the personalization of fitness management but also effectively prevents muscle damage and fatigue accumulation caused by over-exercise by dynamically adjusting rest times. The application of reinforcement learning allows the algorithm to continuously learn and adapt to different users' exercise habits and physical conditions, thereby providing more accurate and effective health management strategies. To fully utilize HRV and SpO2 data, the decision layer includes multi-dimensional parameters such as sEMG pre-fatigue levels, HRV-LF / HF ratio, and SpO2 recovery rate. Through collaborative analysis of multimodal data, the system can generate more accurate rest suggestions, thereby further improving the user's training effect.

[0037] Feedback Layer: To translate the personalized suggestions from the decision-making layer into user-perceptible reminders, such as vibrations or sounds, this design ensures that users receive timely and accurate reminders in different scenarios, thereby effectively adjusting their exercise status and rest time. The introduction of reminders not only improves the user experience but also enhances the effectiveness and practicality of health management.

[0038] Compared with existing technologies, the beneficial effects of this invention are:

[0039] (1) Improve the accuracy of fatigue assessment: By fusing multimodal signals (sEMG, HRV, SpO2) and combining them with reinforcement learning algorithms, the limitations of traditional single data source assessment of fatigue are solved, and the accuracy of fatigue level classification is significantly improved.

[0040] (2) Real-time processing and feedback to enhance user experience: The embedded low-power processing module enables local real-time signal processing and fatigue assessment, ensuring that users receive personalized rest suggestions in a timely manner.

[0041] (3) Personalized enhancement: Based on users' historical data and real-time status, personalized rest time is dynamically generated to avoid the shortcomings of traditional fixed rest time and improve training effect and user experience.

[0042] (4) Cost optimization: Lightweight model design and modular system architecture reduce hardware requirements, making it suitable for portable devices such as smartwatches and reducing user costs.

[0043] (5) Multi-scenario applicability: The system is applicable to a variety of fitness scenarios, including strength training, endurance training, high-intensity interval training, etc. It can provide personalized rest suggestions according to different training types and user status, and has wide applicability. Attached Figure Description

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0045] Figure 1 This is a schematic diagram of the structure of a dynamic fitness interval calculation system based on multimodal signal fusion analysis provided in an embodiment of this application.

[0046] Figure 2 A perspective view of a smartwatch provided in an embodiment of this application;

[0047] Figure 3 A top view of the back of a smartwatch provided in an embodiment of this application;

[0048] Figure 4 An internal perspective view of a smartwatch casing provided in an embodiment of this application.

[0049] Figure 5 The flowchart shows the structure of the lightweight model processing unit in the embedded low-power processing module of the dynamic fitness interval calculation system based on multimodal signal fusion analysis provided in the embodiments of this application.

[0050] Figure 6 This is a schematic diagram of a smartwatch worn by a user, as provided in an embodiment of this application. Detailed Implementation

[0051] The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0052] This invention is applicable to various fitness scenarios, including strength training, endurance training, and high-intensity interval training, and can provide personalized rest suggestions based on different training types and user conditions.

[0053] The first embodiment of this application discloses a dynamic fitness interval calculation system based on multimodal signal fusion analysis, such as... Figure 1 As shown, it includes a signal acquisition module, an embedded low-power processing module, and a reinforcement learning algorithm module.

[0054] The signal acquisition module is used to acquire the user's surface electromyography signal, blood oxygen saturation signal, and heart rate variability signal;

[0055] The embedded low-power processing module is electrically connected to the signal acquisition module and is used to perform noise reduction and feature extraction on the user's surface electromyography (EMG) signal, heart rate variability (HRV) signal, and blood oxygen saturation signal to obtain surface EMG statistical features, HRV frequency domain features, and blood oxygen recovery rate features; and to obtain the surface EMG signal pre-fatigue level based on the surface EMG statistical features.

[0056] The reinforcement learning algorithm module is used to store the user's historical exercise data and calculate the user's fitness intervals based on the user's historical exercise data, surface electromyography signal pre-fatigue level, heart rate variability characteristics, and blood oxygen saturation recovery rate characteristics.

[0057] In this embodiment, the embedded low-power processing module includes a signal noise reduction unit, a feature extraction unit, and a lightweight model processing unit.

[0058] The signal noise reduction unit is used to remove noise interference from the user's surface electromyography signal, heart rate variability signal and blood oxygen saturation signal to obtain the noise-reduced user surface electromyography signal, heart rate variability signal and blood oxygen saturation signal.

[0059] The feature extraction unit is used to extract amplitude and time-related statistical features from the noise-reduced user surface electromyography signal, extract frequency domain features from the noise-reduced heart rate variability signal, and extract blood oxygen recovery rate features from the noise-reduced blood oxygen saturation signal.

[0060] The lightweight model processing unit is used to obtain the pre-fatigue level of the surface electromyography signal based on the statistical characteristics of surface electromyography.

[0061] The signal denoising unit is responsible for removing noise interference from the signal. For electromyography (EMG) signals, a bandpass filter is used to efficiently remove background noise, power line interference, and other non-physiological signal components. Typically, the frequency components of surface EMG signals are concentrated between 10Hz and 500Hz, a range that covers the vast majority of signal components directly related to muscle activity. Therefore, the bandpass filter is designed to allow only signals within this frequency range to pass through, while significantly attenuating or completely blocking signals outside this range. A fourth-order Butterworth bandpass filter is chosen because of its flat amplitude-frequency response and controllable phase distortion within the passband. Zero phase distortion is achieved through forward-backward filtering; a fourth-order Butterworth bandpass filter is constructed using a cascaded biquad filter, suitable for fixed-point arithmetic in embedded systems. For HRV signals, low-frequency components are first extracted using a 0.1-0.4Hz bandpass filter, then the HRV signal is smoothed using a moving average filter; finally, linear fitting is used to remove the linear trend of the signal, further reducing the impact of baseline drift and preparing for subsequent feature extraction. For the SpO2 signal, adaptive filtering is used to eliminate ambient light interference, and wavelet transform is used to separate the arterial blood flow signal.

[0062] The feature extraction unit extracts key features from the denoised signal. The denoised electromyography (EMG), HRV, and SpO2 signals are fed into the feature extraction unit, a core step in the data processing flow. In time-domain analysis, the feature extraction unit primarily focuses on the amplitude and time-dependent statistical properties of the EMG signal. Commonly used feature parameters include: root mean square (RMS), variance, mean absolute value, and zero-crossing rate.

[0063] RMS reflects the average energy level of electromyographic signals and is an important indicator for assessing muscle activity intensity. Its calculation formula is:

[0064]

[0065] Among them, X i RMS is the value of the electromyographic signal at discrete time point i, and N is the total number of sampling points of the signal. An increase in the RMS value usually indicates enhanced muscle activity.

[0066] Variance measures the degree of dispersion of electromyographic signal values ​​relative to their mean, reflecting the fluctuation of the signal. Large variance indicates drastic signal changes, which is associated with instability in muscle activity or fatigue.

[0067] Frequency domain analysis transforms electromyographic (EMG) signals from the time domain to the frequency domain using Fast Fourier Transform (FFT), revealing the frequency components of the signal. Power Spectral Density (PSD) is a key characteristic parameter describing the energy distribution of EMG signals at different frequencies. Peaks on the PSD curve typically correspond to the dominant frequency bands of the EMG signal, reflecting the main frequency components of muscle activity. By comparing PSD curves under different states (e.g., rest, mild activity, heavy activity), changes in the dominant frequency bands can be observed. During muscle fatigue, the frequency components of the EMG signal may change, manifested as energy variations in the PSD curve within a specific frequency range. For example, as muscle fatigue intensifies, the energy in the mid- and high-frequency bands may increase, while the energy in the low-frequency band may relatively decrease. By monitoring the long-term trend of PSD curve changes, muscle fatigue states can be predicted and identified. The frequency components of EMG signals may differ for different muscles or different movements. By comparing PSD curves under different muscles or movements, their differences and commonalities can be analyzed, leading to a deeper understanding of the characteristics of muscle activity.

[0068] For HRV signals, frequency domain analysis primarily focuses on the LF / HF power ratio. Low-frequency (LF, 0.04-0.15Hz) and high-frequency (HF, 0.15-0.4Hz) components reflect the activity of the sympathetic and parasympathetic nervous systems, respectively. The LF / HF ratio can be used to assess the balance of the autonomic nervous system. For the denoised SpO2 signal, oxygen saturation is calculated mainly based on the ratio of red to infrared absorbance. Then, a sliding window linear regression algorithm is used to dynamically calculate the slope of the SpO2 recovery curve, generating a blood oxygen recovery rate characteristic. This recovery rate characterizes the user's metabolic system's reoxygenation capacity and is used to quantitatively assess exercise-induced hypoxia and recovery efficacy.

[0069] The structural flowchart of the lightweight model processing unit is as follows: Figure 5 As shown, this unit integrates a lightweight fatigue classification model for sEMG pre-fatigue level assessment. This model is trained on a large amount of user electromyography (sEMG) signal and fatigue state data, possessing efficient and accurate assessment capabilities. To achieve this goal while reducing computational complexity and extending the smartwatch's battery life, the lightweight fatigue classification model employs lightweight deep learning techniques to classify the pre-fatigue level of electromyography (sEMG) signals. Input data is provided by the feature extraction unit, including key features such as root mean square value, variance, mean absolute value (MAV), zero-crossing rate (ZC), and power spectral density (PSD). The window length is set to 400 milliseconds to ensure sufficient muscle activity information is captured for fatigue assessment. The convolutional layer contains 16 channels, each using a 5-channel convolutional kernel for local feature extraction to learn local patterns and trends in electromyographic signals. The pooling layer employs MaxPooling with a 2-kernel length to reduce the dimensionality of feature data, decrease computation, and retain key information to improve the model's generalization ability. The fully connected layer, consisting of 32 neurons, integrates the features extracted by the convolutional layers and outputs the final pre-fatigue level through a Softmax classifier. Fatigue levels are categorized into three levels: 0 represents normal, 1 represents mild fatigue, and 2 represents severe fatigue. To deploy the model on resource-constrained smartwatches, this embodiment uses the TensorFlow Lite Micro framework for quantization. Specifically, the model undergoes 8-bit integer quantization, reducing its size to less than 50KB while maintaining high accuracy. The quantized model exhibits an inference latency of less than 30 milliseconds on the smartwatch, ensuring the feasibility of real-time fatigue level assessment.

[0070] The reinforcement learning algorithm module and the embedded low-power processing module together constitute the core decision-making part of the dynamic fitness interval calculation system based on multimodal signal fusion analysis provided in this embodiment. The reinforcement learning algorithm module receives processed sEMG pre-fatigue level, HRV-LF / HF ratio, and SpO2 recovery rate from the embedded low-power processing module, and also integrates the user's personalized information. These feature definitions collectively form the basis of the reinforcement learning algorithm input, enabling the algorithm to comprehensively and deeply understand the user's current state and historical trends. The reinforcement learning algorithm module provides a series of discrete suggested rest time options, including 30 seconds, 60 seconds, 90 seconds, and 120 seconds, designed to meet the user's rest needs in different situations. These suggestions are based on the algorithm's comprehensive analysis of the user's current state and possible future states, aiming to maximize the user's recovery effect and training efficiency. The design of the reward function reflects the algorithm's profound insight into the user's actual experience and needs. It comprehensively considers the stability of the electromyographic pre-fatigue level (iPEMGF) at adjacent time points (reflecting the smoothness of muscle activity), the user's subjective rating (reflecting the user's satisfaction with the suggestions), and the penalty for excessive rest (avoiding unnecessary rest time leading to a decrease in training efficiency). The reward function also includes an HRV recovery reward (if the LF / HF ratio approaches the resting state after rest) and a SpO2 penalty (if the blood oxygen recovery rate is below a threshold), to more comprehensively balance muscle recovery, neural regulation, and metabolic efficiency. Specifically, the reward function R consists of five parts:

[0071]

[0072] Where α, β, and γ are weighting coefficients used to adjust the relative importance of each component in the reward function R. iPEMGF t+1 and iPEMGF t R represents the pre-fatigue level of surface electromyography signals at time t+1 and time t, respectively. 过度休息惩罚 This represents the impact of rest time exceeding a preset value on the reward function R. It can be set to a fixed value; the effect only occurs when the rest time exceeds the preset value. R HRV恢复 This represents the influence of the ratio of low-frequency to high-frequency components of the heart rate variability signal on the reward function R. This represents the effect of the blood oxygenation recovery rate on the reward function R.

[0073] HRV reflects changes in heart rate interval and is an important indicator for assessing the state of the autonomic nervous system. A higher HRV generally indicates stronger parasympathetic activity (responsible for relaxation and recovery), suggesting a better recovery state; a lower HRV may indicate dominant sympathetic activity (responsible for stress and activation), suggesting stress or fatigue. In the frequency domain analysis of HRV, low-frequency (LF, 0.04-0.15Hz) and high-frequency (HF, 0.15-0.4Hz) components reflect sympathetic and parasympathetic activity, respectively. A reward mechanism encourages the system to recommend rest periods that help users recover to a resting state, i.e., the LF / HF ratio approaches the resting value. If the LF / HF ratio significantly decreases after rest (approaching the resting state), a positive reward is given. If the LF / HF ratio does not significantly improve or increase, no reward or a smaller reward is given. The influence of the ratio of low-frequency to high-frequency components of the heart rate variability signal on the reward function R is represented by the value R. HRV恢复 The formula is as follows:

[0074]

[0075] δ is the weighting coefficient for the heart rate variability recovery reward, used to adjust the contribution ratio of the heart rate variability index to the overall reward function. This represents the ratio of low-frequency to high-frequency components of the heart rate variability signal after the user has rested according to the recommended rest period. This represents the ratio of low-frequency to high-frequency components of the heart rate variability signal when a user is in a completely relaxed, resting state. It typically needs to be measured and determined beforehand. It represents the user's ideal recovery target and is used to assess the recovery effect after rest.

[0076] SpO2 reflects the oxygen content in the blood, which is normally 95%-100% at rest. After exercise, SpO2 may temporarily decrease and then gradually recover. The blood oxygen recovery rate refers to the speed at which SpO2 returns to resting levels after exercise. A fast recovery rate indicates high metabolic efficiency and strong recovery ability; a slow recovery rate may indicate fatigue or poor metabolism. A penalty mechanism is used to prevent the system's recommended rest time from being too short, leading to insufficient blood oxygen recovery. If the user's SpO2 recovery rate is lower than a preset threshold (e.g., 1% increase per minute) after rest, a penalty is triggered. If the SpO2 recovery rate is higher than the threshold, no penalty is triggered. The impact of the blood oxygen recovery rate on the reward function R is also discussed. The formula is as follows:

[0077]

[0078] ε is the weighting coefficient for blood oxygen saturation penalty, used to control the impact of insufficient blood oxygen saturation recovery on the total reward; Th represents the preset threshold for blood oxygen recovery rate. This indicates the rate at which a user's blood oxygen levels recover after resting for the recommended time.

[0079] For example, the preset SpO2 recovery rate threshold is 1.5% per minute. If the SpO2 recovery rate is 1.0% per minute after a user rests, a penalty is triggered; if the recovery rate is 2.0% per minute, no penalty is triggered.

[0080] In this embodiment, the signal acquisition module includes a silicon microneedle array sensor and a dual-wavelength PPG sensor. The silicon microneedle array sensor is used to acquire the user's surface electromyography signal; the dual-wavelength PPG sensor is used to acquire blood oxygen saturation signal and heart rate variability signal.

[0081] In this embodiment, the system further includes a user interaction module, which includes a display screen and a sound output device, for feeding back the rest time generated by the reinforcement learning algorithm module to the user.

[0082] The system also features a data privacy protection mechanism to ensure the secure storage and transmission of user electromyography signals, heart rate variability data, blood oxygen saturation data, and personal information, thereby enhancing user trust and user experience.

[0083] The second embodiment of this application provides a smartwatch, such as... Figure 2 , Figure 3 and Figure 4 As shown, the system includes a housing 5, a strap 5.1, and the aforementioned dynamic fitness interval calculation system based on multimodal signal fusion analysis. The signal acquisition module is integrated into the bottom surface of the smartwatch housing 5, while the embedded low-power processing module 2 and the reinforcement learning algorithm module 3 are integrated inside the smartwatch housing 5. The embedded low-power processing module 2 can use the embedded low-power chip ESP32. To deploy the lightweight fatigue classification model on a resource-constrained smartwatch, this embodiment uses the TensorFlow Lite Micro framework to quantize the lightweight model processing unit in the embedded low-power processing module 2. Specifically, the lightweight fatigue model undergoes 8-bit integer quantization, which reduces the size of the lightweight fatigue model to less than 50KB while maintaining high accuracy. The inference latency of the quantized lightweight fatigue model on the smartwatch is less than 30 milliseconds, ensuring the feasibility of real-time fatigue level assessment.

[0084] The signal acquisition module includes a silicon microneedle array sensor 1 and a dual-wavelength PPG sensor 4. The silicon microneedle array sensor 1, used to acquire the user's surface electromyography (EMG) signals, is located on the bottom surface of the watch case 5 and is in direct contact with the user's skin. The silicon microneedles are safe and painless, and can efficiently penetrate the skin surface to directly acquire high-quality surface EMG signals. The silicon microneedle array ensures that it does not affect the user's daily activity comfort during long-term wear, while continuously and stably collecting EMG data.

[0085] The dual-wavelength PPG sensor 4 is used to collect heart rate variability (HRV) and blood oxygen saturation (SpO2) data. The dual-wavelength PPG sensor 4 monitors HRV and SpO2 in real time by emitting red light (660nm) and infrared light (940nm) and detecting reflected signals. The dual-wavelength PPG sensor 4 is located on the bottom surface of the watch case 5, physically isolated from the silicon microneedle array sensor 1 by a distance to avoid signal interference. Simultaneously, the watch strap 5.1 ensures close contact between the sensor and the skin, reducing motion artifacts. Through the simultaneous acquisition of multimodal signals, the system can more comprehensively assess the user's physiological state, providing richer data support for subsequent fatigue analysis and rest recommendations.

[0086] The smartwatch also includes a user interaction module 5.2. A strap 5.1 secures the watch to the user's arm, ensuring close contact between the silicon microneedle array sensor 1 and the skin, while also guaranteeing a stable fit. The strap uses a flexible material design to ensure close contact between the dual-wavelength PPG sensor 4 and the skin, reducing motion artifacts and improving signal acquisition stability. The user interaction module 5.2 includes a display screen and a sound output device, used to provide the user with intuitive and easy-to-understand feedback on rest suggestions generated by the reinforcement learning algorithm module 3. The display screen can show rest time, including real-time HRV and SpO2 values ​​and trend graphs, while the sound output device can provide voice reminders to the user when necessary, such as: "Autonomic nervous system tension and slow blood oxygen recovery have been detected; it is recommended to extend the rest to 120 seconds." The smartwatch can be worn on the user's arm, such as... Figure 6 As shown.

[0087] The third embodiment of this application provides a method for calculating dynamic fitness intervals based on multimodal signal fusion analysis, including:

[0088] Collect user surface electromyography signals, blood oxygen saturation signals, and heart rate variability signals;

[0089] Noise reduction and feature extraction were performed on the user's surface electromyography (EMG) signal, heart rate variability (HRV) signal, and blood oxygen saturation signal to obtain surface EMG statistical features, HRV frequency domain features, and blood oxygen recovery rate features; the pre-fatigue level of the surface EMG signal was obtained based on the surface EMG features.

[0090] The user's workout intervals are calculated based on the user's historical exercise data, surface electromyography signal pre-fatigue level, heart rate variability characteristics, and blood oxygen saturation recovery rate characteristics.

[0091] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a dynamic fitness interval calculation method based on multimodal signal fusion analysis, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0092] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0093] This invention provides a dynamic fitness interval calculation system and method based on multimodal signal fusion analysis. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A dynamic fitness interval calculation system based on multimodal signal fusion analysis, characterized in that, It includes a signal acquisition module, an embedded low-power processing module, and a reinforcement learning algorithm module. The signal acquisition module is used to acquire the user's surface electromyography signal, blood oxygen saturation signal, and heart rate variability signal; The embedded low-power processing module is electrically connected to the signal acquisition module and is used to perform noise reduction and feature extraction on the user's surface electromyography signal, heart rate variability signal and blood oxygen saturation signal to obtain surface electromyography statistical features, heart rate variability frequency domain features and blood oxygen recovery rate features. The pre-fatigue level of surface electromyography (EMG) signals is obtained based on the statistical characteristics of surface EMG signals. The reinforcement learning algorithm module is used to store the user's historical exercise data and calculate the user's fitness intervals based on the user's historical exercise data, surface electromyography signal pre-fatigue level, heart rate variability frequency domain characteristics, and blood oxygen recovery rate characteristics. The reinforcement learning algorithm module calculates the user's exercise intervals, including the user's rest time. The reward function R for the reinforcement learning algorithm is designed based on the pre-fatigue level of surface electromyography signals, the ratio of low-frequency to high-frequency components of heart rate variability signals, blood oxygen recovery rate, user subjective ratings, and whether excessive rest has occurred. ， Wherein, α, β, and γ are weighting coefficients used to adjust the relative importance of each part in the reward function R; and These represent the pre-fatigue levels of surface electromyography signals at time t+1 and time t, respectively. This represents the impact of rest time exceeding a preset value on the reward function R; This represents the influence of the ratio of low-frequency to high-frequency components of the heart rate variability signal on the reward function R. This represents the effect of the blood oxygen recovery rate on the reward function R; The influence of the ratio of low-frequency to high-frequency components of the heart rate variability signal on the reward function R. The formula is as follows: ， Where δ is the weighting coefficient for the heart rate variability recovery reward, used to adjust the contribution ratio of the heart rate variability index in the overall reward function. This represents the ratio of low-frequency to high-frequency components of the heart rate variability signal after the user has rested according to the recommended rest period. The ratio of low-frequency to high-frequency components of the heart rate variability signal when a user is in a completely relaxed resting state represents the user's ideal recovery target and is used to assess the recovery effect after rest. The influence of the blood oxygen recovery rate on the reward function R The formula is as follows: ， in, Th represents the weighting coefficient for blood oxygen saturation penalty, used to control the impact of insufficient blood oxygen saturation recovery on the total reward. This indicates the rate at which a user's blood oxygen levels recover after resting for the recommended time.

2. The dynamic fitness interval calculation system based on multimodal signal fusion analysis according to claim 1, characterized in that, The embedded low-power processing module includes a signal noise reduction unit, a feature extraction unit, and a lightweight model processing unit. The signal noise reduction unit is used to remove noise interference from the user's surface electromyography signal, heart rate variability signal and blood oxygen saturation signal to obtain the noise-reduced user surface electromyography signal, heart rate variability signal and blood oxygen saturation signal. The feature extraction unit is used to extract amplitude and time-related statistical features from the noise-reduced user surface electromyography signal, extract frequency domain features from the noise-reduced heart rate variability signal, and extract blood oxygen recovery rate features from the noise-reduced blood oxygen saturation signal. The lightweight model processing unit includes a lightweight fatigue classification model, which is used to obtain the pre-fatigue level of surface electromyography signals based on the statistical characteristics of surface electromyography.

3. The dynamic fitness interval calculation system based on multimodal signal fusion analysis according to claim 2, characterized in that, The signal noise reduction unit uses a fourth-order Butterworth bandpass filter to reduce the noise of the surface electromyography signal. The passband frequency range is 10-500Hz. Zero phase distortion is achieved through forward and reverse filtering to obtain the user's surface electromyography signal after noise reduction. The heart rate variability signal was subjected to 0.1-0.4Hz bandpass filtering, moving average filtering and linear fitting to remove baseline drift and obtain the noise-reduced heart rate variability signal; An adaptive filter was used to eliminate ambient light interference in the blood oxygen saturation signal, and the arterial blood flow signal was separated by wavelet transform to obtain the denoised blood oxygen saturation signal.

4. The dynamic fitness interval calculation system based on multimodal signal fusion analysis according to claim 3, characterized in that, The feature extraction unit extracts amplitude and time-related statistical features from the denoised user surface electromyography (EMG) signal, including: performing time-domain analysis and frequency-domain analysis on the denoised user surface EMG signal; calculating the root mean square value, variance, mean absolute value, and zero-crossing rate in the time-domain analysis; and converting the surface EMG signal from the time domain to the frequency domain through fast Fourier transform in the frequency domain analysis, thereby revealing the frequency components of the surface EMG signal. The feature extraction unit extracts frequency domain features from the denoised heart rate variability signal, including: performing frequency domain analysis on the denoised heart rate variability signal to obtain low-frequency and high-frequency components, wherein the low-frequency and high-frequency components reflect the activity of the sympathetic and parasympathetic nervous systems, respectively, and the ratio of low-frequency to high-frequency components can be used to assess the balance state of the autonomic nervous system. The feature extraction unit extracts the blood oxygen recovery rate feature from the denoised blood oxygen saturation signal, including: calculating blood oxygen saturation based on the ratio of red light to infrared light absorbance, calculating the slope of the blood oxygen saturation recovery curve using a linear regression algorithm, and generating the blood oxygen recovery rate.

5. The dynamic fitness interval calculation system based on multimodal signal fusion analysis according to claim 4, characterized in that, The signal acquisition module includes a silicon microneedle array sensor and a dual-wavelength PPG sensor. The silicon microneedle array sensor is used to acquire surface electromyography signals of the user; the dual-wavelength PPG sensor is used to acquire blood oxygen saturation signals and heart rate variability signals.

6. The dynamic fitness interval calculation system based on multimodal signal fusion analysis according to claim 5, characterized in that, It also includes a user interaction module, which includes a display screen and a sound output device, for feeding back the rest time generated by the reinforcement learning algorithm module to the user.

7. A smartwatch, comprising a casing, a strap, and a dynamic fitness interval calculation system based on multimodal signal fusion analysis as described in any one of claims 1-6, characterized in that, The signal acquisition module is integrated into the bottom of the smartwatch casing, while the embedded low-power processing module and reinforcement learning algorithm module are integrated inside the smartwatch casing.

8. A method for calculating dynamic fitness intervals based on multimodal signal fusion analysis, applied to the dynamic fitness interval calculation system based on multimodal signal fusion analysis as described in any one of claims 1-6, characterized in that, include: Collect user surface electromyography signals, blood oxygen saturation signals, and heart rate variability signals; Noise reduction and feature extraction were performed on the user's surface electromyography (EMG) signal, heart rate variability (HRV) signal, and blood oxygen saturation signal to obtain surface EMG statistical features, HRV frequency domain features, and blood oxygen recovery rate features; the pre-fatigue level of the surface EMG signal was obtained based on the surface EMG features. The user's workout intervals are calculated based on the user's historical exercise data, surface electromyography signal pre-fatigue level, heart rate variability frequency domain characteristics, and blood oxygen recovery rate characteristics.

Images