A portable fitness tracker identification method and system

Abstract

This invention discloses a portable fitness tracker identification method and system, comprising: synchronously acquiring raw pulse wave signals, motion posture parameters, and preliminary heart rate data through the tracker; separating the effective and interfering components of the signal using a motion artifact elimination loop network; extracting multidimensional physiological feature parameters using a photoplethysmography pulse wave analysis model; calculating fatigue correlation probability values ​​using a muscle fatigue probability prediction model in conjunction with motion parameters; and performing cross-dimensional data correlation fusion using a full-dimensional physiological-motor fusion analysis platform to construct a physiological-motor state mapping relationship and output the identification results. Simultaneously, through detailed processes such as time domain segmentation and feature point extraction, parameter adaptation and iterative correction, and data alignment and correlation fusion, the accuracy of feature analysis and the reliability of fusion are improved, achieving accurate identification of exercise intensity, fatigue level, and physiological adaptability, adapting to portable fitness tracker application scenarios, and ensuring comprehensive and accurate monitoring.

Description

Technical Field

[0001] This invention relates to the field of fitness tracker identification technology, and in particular to a portable fitness tracker identification method and system. Background Technology

[0002] With the popularization of the national fitness concept and the development of wearable device technology, portable fitness trackers have become core devices for monitoring exercise status and physiological parameters due to their portability and real-time monitoring advantages. The demand for accurate identification of physiological signals during exercise is becoming increasingly urgent. In exercise scenarios, human physiological signals are easily interfered with by factors such as changes in posture and muscle activity, requiring accurate analysis through multi-dimensional parameter fusion. Photoplethysmography (PPG), as a key carrier reflecting cardiovascular status and exercise load, combined with the comprehensive identification of posture parameters and muscle fatigue status, has become a core direction for improving the reliability of fitness tracker monitoring. Related technologies have been widely applied in scenarios such as sports health management and exercise intensity regulation, necessitating the construction of an integrated identification solution that balances signal interference removal, in-depth feature analysis, and multi-dimensional data fusion.

[0003] Existing technologies have two significant shortcomings: First, motion artifact removal methods lack adaptability to dynamic changes in motion posture. They rely solely on single filtering or simple algorithms to separate interference components, making it difficult to effectively address complex interference under different motion intensities. This limits the accuracy of pulse wave feature extraction, thereby affecting the accuracy of subsequent physiological parameter analysis. Second, the fusion analysis of physiological parameters and motion state lacks a comprehensive correlation mechanism. It often relies on a single type of parameter or a simple weighting method for judgment, failing to establish a deep mapping relationship between physiological characteristics, motion posture, and fatigue state. Consequently, the recognition results cannot fully reflect the dynamic changes in the human body's state during exercise, making it difficult to meet the actual needs of precise motion monitoring and health management. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a portable sports wristband identification method and system.

[0005] The technical solution adopted in this invention is a portable fitness tracker identification method, characterized by the following steps: S1, acquiring raw pulse wave signals during motion using a photoplethysmography (PPG) sensor mounted on the portable fitness tracker, simultaneously acquiring motion posture parameters and preliminary heart rate detection data output by the tracker's built-in three-axis accelerometer and gyroscope, and detecting ambient light levels, monitoring light exposure levels under different spectra, including blue light exposure; S2, inputting the raw pulse wave signals and motion posture parameters into a motion artifact elimination recurrent network, separating the effective components of the pulse wave from motion interference components through a multi-layer convolution and recurrent feedback mechanism; S3, calling the PPG pulse wave analysis... The model extracts features from the artifact-free pulse wave signal and analyzes it to obtain multi-dimensional physiological feature parameters such as pulse wave propagation time, peak amplitude, and waveform slope. In step S4, the multi-dimensional physiological feature parameters and motion posture parameters are input into the muscle fatigue probability prediction model, and the probability value associated with muscle fatigue state is calculated by combining the motion duration parameter. In step S5, a cross-dimensional data association and fusion analysis is performed on the muscle fatigue probability value, pulse wave feature parameters, and motion posture parameters through a full-dimensional physiological-motor fusion analysis platform to construct a physiological-motor state mapping relationship. In step S6, based on the fusion analysis results, the motion state recognition results are output, including motion intensity level, muscle fatigue degree, and physiological state suitability judgment.

[0006] Furthermore, the expression for the photoplethysmography (PPG) analytical model is as follows: ,in, This is the analyzed three-dimensional pulse wave feature matrix. This is the pulse wave signal attenuation coefficient. These are the photosensitive area parameters of the photoplethysmography sensor. The fundamental angular frequency of the pulse wave. This is the phase offset. This is a correction factor for conduction velocity. For parameters related to the elastic modulus of blood vessels, The two-dimensional Laplace operator for the original pulse wave signal. It is the amplitude gain factor. For parameters related to the motion amplitude of the accelerometer, For signal acquisition period, It is the second derivative signal of the pulse wave.

[0007] Furthermore, the expression for the muscle fatigue probability prediction model is as follows: ,in, This represents the probability value of muscle fatigue. This represents the fusion value of pulse wave characteristics. It is the motion attitude parameter matrix (including triaxial acceleration and angular velocity components). For parameters related to exercise intensity, For the duration of the movement, These are probability weighting coefficients. This is a fatigue accumulation correction factor. The rate of change of motion posture. This is the attitude influence coefficient. This is the intensity cumulative factor. The gradient values ​​are the motion attitude parameters.

[0008] Furthermore, the expression for the motion artifact removal recurrent network is: ,in: The output signal after artifact removal at time t, The convolution kernel weight matrix is... These are the recurrent layer weight parameters. for Real-time network output, The input layer weight matrix, The input for fusing the original pulse wave and motion parameters at time t is... For the loop layer bias term, The integral coefficients for the time window. The length of the cyclic feedback time window. This is the integral weight matrix. for Input signals at all times, This represents the artifact suppression coefficient. To adapt weights to motion parameters, Let be the feature matrix of motion artifacts at time t. This is the activation function.

[0009] Furthermore, the fusion model expression of the full-dimensional physiological motion fusion analysis platform is as follows: ,in, To integrate the analysis results, This is the pulse wave characteristic parameter vector. This represents the probability value of muscle fatigue. Let be the motion attitude parameter vector, and () be the matrix tensor product operation. The diagonal matrix of motion parameters To integrate the weighting coefficients, These are the vector magnitudes of the pulse wave characteristics and motion parameters, respectively. It is the identity matrix. It is the vector inner product.

[0010] Furthermore, the expression for the portable fitness tracker's identification parameter adaptation model is as follows: ,in, To identify the fit value, For exercise intensity level parameters, For heart rate detection parameters, For motion posture fusion parameters, For the adaptation coefficient, For the gradient of heart rate changes, For the identification period, The rate of change of motion posture.

[0011] Further, step S3 includes the following sub-steps: S31, based on the feature extraction logic of the photoplethysmography pulse wave analytical model, the artifact-free pulse wave signal is divided into time domain segments, and multiple analysis windows are divided according to the signal period, with each window corresponding to a single pulse wave period; S32, within each analysis window, the pulse wave rising edge, peak point, and falling edge are located to calibrate feature points, and the time interval and amplitude difference between each feature point are calculated to obtain the basic parameters of pulse wave conduction time and peak amplitude; S33, the pulse wave waveform is smoothed using a curve fitting algorithm, and the slope values ​​of the rising and falling edges of the waveform are solved to extract waveform curvature change feature parameters; S34, the time domain feature parameters and waveform feature parameters are combined according to a preset dimension to construct a multidimensional physiological feature parameter matrix, providing input data for subsequent fusion analysis.

[0012] Further, step S4 includes the following sub-steps: S41, standardizing and mapping the multidimensional physiological feature parameters obtained in S3 to convert them into an adaptive input format for the muscle fatigue probability prediction model, and simultaneously performing vector normalization on the movement posture parameters; S42, inputting the adapted physiological feature parameters, normalized movement posture parameters, and movement duration parameters into the model, and performing feature weighting and nonlinear transformation through the model's built-in multi-layer computing unit; S43, calculating the muscle fatigue correlation factor based on the transformed feature data, adjusting the factor weights in conjunction with the exercise intensity parameter, and obtaining a preliminary fatigue probability value; S44, iteratively correcting the preliminary fatigue probability value through a model feedback mechanism to eliminate deviations caused by abnormal parameters, and outputting the final muscle fatigue state correlation probability value.

[0013] Further, S5 includes the following sub-steps: S51, using the parameter adaptation module of the full-dimensional physiological-motor fusion analysis platform, the data format of muscle fatigue probability values, pulse wave characteristic parameters, and movement posture parameters are unified and aligned in dimensions to construct a cross-type data set; S52, the platform's correlation analysis unit is invoked to calculate the correlation coefficient matrix between different types of parameters based on preset physiological-motor correlation rules, and to identify strongly correlated parameter combinations; S53, the platform's fusion calculation unit performs weighted fusion on the strongly correlated parameter combinations, and a multi-dimensional data fusion algorithm is used to generate a comprehensive feature vector; S54, a physiological-motor state mapping relationship model is constructed based on the comprehensive feature vector to determine the physiological parameter adaptation range under different movement states.

[0014] A portable fitness tracker identification system, applied to a portable fitness tracker identification method, includes: a photoplethysmography-motion parameter synchronous acquisition unit, electrically connected to the photoplethysmography sensor, three-axis accelerometer, and gyroscope of the portable fitness tracker, for synchronously acquiring raw pulse wave signals, motion posture parameters, and preliminary heart rate detection data; a motion artifact intelligent separation unit, signal-connected to the photoplethysmography-motion parameter synchronous acquisition unit, with a built-in motion artifact elimination loop network, performing separation processing of effective pulse wave components and motion interference components; and a multi-dimensional pulse wave feature analysis unit, signal-connected to the motion artifact intelligent separation unit, equipped with a photoplethysmography pulse wave analysis model, for analyzing the artifact-free pulse wave signal. Feature extraction and analysis; a muscle fatigue probability calculation unit, which is connected to the multi-dimensional pulse wave feature analysis unit and the photoplethysmography-motion parameter synchronous acquisition unit, calculates the probability value associated with muscle fatigue state through a muscle fatigue probability prediction model; a full-dimensional data fusion analysis unit, which is connected to the muscle fatigue probability calculation unit, the multi-dimensional pulse wave feature analysis unit, and the motion artifact intelligent separation unit, constructs a physiological-motor state mapping relationship; and a recognition result output unit, which is connected to the full-dimensional data fusion analysis unit, outputs the exercise intensity level, muscle fatigue degree, and physiological state suitability judgment result based on the fusion analysis result. Each unit interacts and works collaboratively through the data transmission bus built into the wristband.

[0015] Beneficial Effects: This invention proposes a portable fitness tracker identification method and system. By simultaneously acquiring pulse wave signals and motion posture parameters, and utilizing a specialized recurrent network to separate motion interference components, it solves the problems of incomplete artifact elimination and weak adaptability to dynamic motion in traditional methods, laying a high-quality data foundation for subsequent feature extraction. It leverages a targeted pulse wave analysis model to deeply mine multi-dimensional physiological features, combined with a muscle fatigue probability prediction model to achieve accurate quantification of fatigue state. Furthermore, a comprehensive fusion analysis platform establishes a deep mapping relationship between physiological parameters and motion state, breaking through the limitations of existing technologies with single-parameter analysis and simple fusion mechanisms. Six functional units work collaboratively via a bus to achieve a closed-loop processing from signal acquisition, artifact elimination, feature analysis to fusion judgment. This ensures both the continuity and efficiency of data processing, and comprehensively improves the accuracy and comprehensiveness of the identification results through multi-model collaboration and cross-dimensional fusion analysis. It can accurately output exercise intensity, fatigue level, and physiological adaptability judgment, fully meeting the actual needs for precise monitoring and dynamic adaptation in sports and health management scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a portable fitness tracker identification method includes the following steps: S1 collects raw pulse wave signals during exercise through the photoplethysmography sensor mounted on the portable sports bracelet, simultaneously acquires motion posture parameters and preliminary heart rate detection data output by the built-in three-axis accelerometer and gyroscope, and detects ambient light levels to monitor light exposure levels under different spectra, including blue light exposure. Specifically, step S1 achieves synchronous acquisition of multi-source data, providing comprehensive and accurate raw data support for subsequent analysis. In practice, the photoplethysmography sensor on the portable fitness tracker uses a combination of visible and near-infrared light sources with a wavelength range of 500-600 nanometers. A constant current drives a light-emitting diode to emit light. After the light penetrates the skin surface, it is reflected by the blood in the blood vessels and received by a photodiode, which converts it into an electrical signal. This electrical signal is the raw pulse wave signal during exercise. The signal sampling frequency is set to 200Hz to ensure the capture of subtle waveform changes in the pulse wave. Meanwhile, the built-in triaxial accelerometer collects acceleration data along the X, Y, and Z axes at a sampling frequency of 100Hz, with a measurement range of ±8g and a resolution of 1mg, accurately capturing displacement and vibration information during human movement. The gyroscope also outputs triaxial angular velocity data at a sampling frequency of 100Hz, with a measurement range of ±2000° / s and a resolution of 0.1° / s, reflecting rotational changes in posture. Preliminary heart rate detection data is initially calculated through periodic recognition of the photoplethysmography sensor signal, with the sampling interval consistent with the pulse wave signal. The three are time-stamped through the built-in synchronization trigger module of the wristband, ensuring the time consistency of the pulse wave signal, posture parameters, and preliminary heart rate detection data collected at the same time. The collected data is temporarily stored in the wristband's local cache module with a cache capacity of no less than 16MB to avoid data loss.

[0019] S2, input the original pulse wave signal and motion posture parameters into the motion artifact elimination recurrent network, and separate the effective components of the pulse wave and motion interference components through the network's multi-layer convolution and recurrent feedback mechanism; Specifically, step S2 efficiently separates motion artifacts, ensuring the purity of the pulse wave signal and laying the foundation for subsequent feature extraction. During implementation, the raw pulse wave signal and motion posture parameters collected in step S1 are first converted to 32-bit floating-point data, then combined into an input data matrix according to timestamp order, and input into the motion artifact elimination recurrent network. This network consists of 3 convolutional layers, 2 recurrent layers, and 1 fusion output layer. The convolutional layers use 3×3 kernels with 16, 32, and 64 kernels respectively, extracting features from the input data through a sliding window operation to capture local signal features. The recurrent layers employ a long short-term memory network structure, with 128 hidden units per layer, using a gating mechanism to remember long-term dependencies and a recurrent feedback mechanism to continuously correct the signal separation results. During network training, an adaptive moment estimation optimization algorithm is used, with a learning rate of 0.001 and 500 iterations, optimizing the model by minimizing the mean square error between the effective signal and the separated signal. During signal separation, the network analyzes the correlation between motion posture parameters and pulse wave signals to identify interference components that are synchronized with changes in acceleration and angular velocity, i.e., motion artifacts. Then, it uses an inverse suppression mechanism to remove these artifacts from the original pulse wave signal, and finally outputs the effective pulse wave components after removing motion interference components. The signal-to-noise ratio of the separated effective signal is improved by more than 30% compared with the original signal, ensuring the accuracy of subsequent feature analysis.

[0020] S3, call the photoplethysmography pulse wave analytical model to extract features from the artifact-free pulse wave signal, and analyze to obtain multi-dimensional physiological feature parameters such as pulse wave propagation time, peak amplitude, and waveform slope; Specifically, step S3 uses a photoplethysmography (PPG) pulse wave analytical model to extract deep features of the effective components of the pulse wave, and analyzes the multidimensional parameters reflecting the human physiological state. In practice, the PPG pulse wave analytical model pre-installed in the wristband processor is first invoked. This model is built based on the physiological conduction mechanism of the pulse wave. For the pulse wave signal after artifact removal, the signal is first segmented and divided into individual analysis units according to the pulse wave period. Each unit contains a complete pulse wave waveform. Subsequently, the model uses a peak detection algorithm to locate the systolic peak, diastolic trough, and dicrotic peak of each pulse wave waveform. The time interval between the systolic peak and diastolic trough is calculated, which is the pulse wave conduction time, with a measurement accuracy of 1 ms. The difference between the signal amplitude corresponding to the systolic peak and the baseline amplitude is calculated to obtain the peak amplitude, with a resolution of 0.1 mV. A linear fitting algorithm is used to fit the rising edge (from the diastolic trough to the systolic peak) and falling edge (from the systolic peak to the diastolic trough) of the pulse wave, and the slope of the fitted line is solved to obtain the waveform slope parameter, with the unit being mV / ms. In addition, the model also extracts feature parameters such as rise time, fall time, and dicrotic wave amplitude ratio of the pulse wave. The rise time is the time span from the diastolic trough to the systolic peak, the fall time is the time span from the systolic peak to the diastolic trough, and the dicrotic wave amplitude ratio is the ratio of the peak amplitude of the dicrotic wave to the peak amplitude of the systolic wave. All extracted feature parameters are organized into a multidimensional physiological feature parameter matrix according to a preset dimension. The matrix dimension is N×8 (N is the number of pulse wave cycles analyzed), which provides comprehensive feature input for subsequent calculation of muscle fatigue probability.

[0021] S4. Input the multidimensional physiological feature parameters and movement posture parameters into the muscle fatigue probability prediction model, and calculate the probability value associated with muscle fatigue state by combining the movement duration parameter. Specifically, step S4, based on multidimensional feature parameters and motion parameters, uses a muscle fatigue probability prediction model to quantitatively assess fatigue status. During implementation, the multidimensional physiological feature parameters obtained in step S3 are first normalized to the [0,1] interval. Simultaneously, the acceleration and angular velocity data in the motion posture parameters are statistically analyzed to calculate statistics such as average acceleration and maximum angular velocity during the exercise. These are then combined with user-preset exercise type parameters (e.g., running, cycling, strength training) to determine the exercise intensity-related parameters. Subsequently, the normalized multidimensional physiological feature parameters, motion posture statistics, exercise intensity-related parameters, and exercise duration parameters are input into the muscle fatigue probability prediction model. This model has built-in fatigue assessment coefficients corresponding to different exercise types, and probability calculations are achieved through weighted summation and nonlinear transformation. The exercise duration parameter is recorded by the wristband's built-in timing module with an accuracy of 1 second, starting from when the user begins exercise and triggers data collection. During the model calculation process, the input parameters are first assigned different weights through a feature weighting layer, with pulse wave propagation time weighted at 0.3, peak amplitude weighted at 0.2, average acceleration weighted at 0.2, and motion duration weighted at 0.3. Then, the weighted features are mapped to the probability space through an activation function to output the probability value associated with muscle fatigue state. The probability value ranges from [0,1], with a larger value indicating a higher degree of muscle fatigue. Fixed-point arithmetic optimization is used during the calculation process to ensure real-time performance on the wristband embedded processor, with a single calculation taking no more than 10ms.

[0022] S5 uses a full-dimensional physiological-motor fusion analysis platform to perform cross-dimensional data association and fusion analysis on muscle fatigue probability values, pulse wave characteristic parameters, and movement posture parameters, and constructs a physiological-motor state mapping relationship. Specifically, step S5 utilizes a full-dimensional physiological-motor fusion analysis platform to achieve deep fusion of multiple types of parameters, constructing a mapping relationship between physiological and motor states. In practice, the platform first acquires the muscle fatigue probability value output from step S4, the multi-dimensional pulse wave feature parameter matrix from step S3, and the raw motion posture data from step S1 through the data receiving module. The data preprocessing module unifies the format and aligns the dimensions, converting all parameters to the same data dimension (e.g., a 128-dimensional feature vector). Interpolation algorithms are used to supplement missing data, ensuring data integrity. Subsequently, the platform's correlation analysis module, based on the Pearson correlation coefficient algorithm, calculates the correlation coefficient between the muscle fatigue probability value and each pulse wave feature parameter and motion posture parameter. A correlation coefficient threshold of 0.6 is set, and strongly correlated parameter combinations are selected, such as the combination of the muscle fatigue probability value with pulse wave conduction time and average acceleration. Next, the fusion calculation module employs a weighted average fusion algorithm to fuse strongly correlated parameter combinations, assigning different weights to parameters based on their importance: muscle fatigue probability value has a weight of 0.4, pulse wave feature parameters have a weight of 0.3, and movement posture parameters have a weight of 0.3. A dynamic weight adjustment mechanism is introduced during the fusion process, adjusting the weights of each parameter in real time according to the exercise intensity; the higher the exercise intensity, the higher the weight of the movement posture parameter. Finally, through the mapping relationship construction module, based on the fused comprehensive feature vector and a mapping model trained using massive amounts of exercise and physiological state sample data, a physiological-exercise state mapping relationship is constructed. This clarifies the exercise intensity level and muscle fatigue degree range corresponding to different comprehensive feature vectors. The mapping model is constructed using a support vector machine algorithm, achieving a classification accuracy of over 95%, providing a basis for the final recognition result output.

[0023] S6 outputs motion state recognition results based on the fusion analysis results, including motion intensity level, muscle fatigue level, and physiological state adaptability judgment.

[0024] Specifically, step S6, based on the fusion analysis results of step S5, accurately outputs comprehensive motion state recognition results, providing users with intuitive motion monitoring feedback. In practice, it first reads the physiological-motion state mapping results output by the full-dimensional physiological-motion fusion analysis platform. This result includes the exercise intensity level, muscle fatigue level range, and physiological state adaptability index corresponding to the comprehensive feature vector. Exercise intensity levels are divided into three levels: low, medium, and high, determined by parameters such as average acceleration, exercise duration, and pulse wave frequency in the comprehensive feature vector. Low intensity corresponds to an average acceleration less than 0.5g, exercise duration less than 30min, and a pulse wave frequency of 60-90 beats / minute; medium intensity corresponds to an average acceleration of 0.5-1.5g, exercise duration 30-60min, and a pulse wave frequency of 90-120 beats / minute; and high intensity corresponds to an average acceleration greater than 1.5g, exercise duration greater than 60min, and a pulse wave frequency greater than 120 beats / minute. Muscle fatigue levels are categorized based on muscle fatigue probability values: 0-0.3 indicates mild fatigue, 0.3-0.7 indicates moderate fatigue, and 0.7-1.0 indicates severe fatigue. Physiological adaptability is assessed based on the matching degree between pulse wave characteristic parameters and exercise intensity. When pulse wave conduction time is stable and peak amplitude fluctuation is less than 10%, adaptability is considered good; when pulse wave conduction time fluctuation is greater than 15% and peak amplitude fluctuation is greater than 20%, adaptability is considered poor. The recognition results are output via the wristband's display screen in a combination of text and icons, and can also be synchronized to a linked mobile app via Bluetooth. The output frequency is once per minute, ensuring users can monitor their exercise and physiological status in real time. The accuracy of the recognition results has been verified to be over 92% in actual testing, meeting the practical needs of exercise and health monitoring.

[0025] Preferably, the expression for the photoplethysmography (PPG) analytical model is: ,in, This is the analyzed three-dimensional pulse wave feature matrix. This is the pulse wave signal attenuation coefficient. These are the photosensitive area parameters of the photoplethysmography sensor. The fundamental angular frequency of the pulse wave. This is the phase offset. This is a correction factor for conduction velocity. For parameters related to the elastic modulus of blood vessels, The two-dimensional Laplace operator for the original pulse wave signal. It is the amplitude gain factor. For parameters related to the motion amplitude of the accelerometer, For signal acquisition period, It is the second derivative signal of the pulse wave.

[0026] Specifically, the photoplethysmography (PPG) pulse wave analytical model is used to deeply mine multidimensional physiological features from the artifact-free pulse wave signal, providing accurate feature support for subsequent analysis. In implementation, the model first acquires the pulse wave signal after motion artifact removal, combining the photosensitive area parameter of the PPG sensor (fixed to 2 square millimeters according to the wristband hardware design), the signal attenuation coefficient set to 0.85, and the vascular elastic modulus correlation parameter preset to an adjustable range of 1.2-1.8 based on the physiological data of users of different age groups. The model incorporates the basic angular frequency and phase shift of the pulse wave into dynamic adjustment through joint calculations in the time and spatial domains. The initial value of the angular frequency is set according to a resting heart rate of 60 beats / minute, and the phase shift is corrected in real time according to the movement posture, with a correction step of 0.01 radians. The amplitude gain factor is set to 0.7, and the signal acquisition period is fixed at 0.005 seconds. By integrating the second derivative of the pulse wave and processing the original pulse wave signal using a two-dimensional Laplace operator, a three-dimensional pulse wave feature matrix is ​​constructed. During implementation, the model runs on the 32-bit microprocessor built into the wristband, using fixed-point arithmetic optimization. The time taken for a single analysis is no more than 8ms. The analyzed feature parameters include conduction time and peak amplitude, with measurement accuracies of 1ms and 0.1mV, respectively, ensuring high-quality physiological feature input for subsequent muscle fatigue prediction.

[0027] Preferably, the expression for the muscle fatigue probability prediction model is: ,in, This represents the probability value of muscle fatigue. This represents the fusion value of pulse wave characteristics. It is the motion attitude parameter matrix (including triaxial acceleration and angular velocity components). For parameters related to exercise intensity, For the duration of the movement, These are probability weighting coefficients. This is a fatigue accumulation correction factor. The rate of change of motion posture. This is the attitude influence coefficient. This is the intensity cumulative factor. The gradient values ​​are the motion attitude parameters.

[0028] Specifically, when implementing the muscle fatigue probability prediction model, it first receives multi-dimensional physiological feature fusion values. These fusion values ​​are obtained by weighting parameters such as pulse wave propagation time and peak amplitude in a ratio of 0.4:0.6. The motion posture parameter matrix includes real-time data and statistics of triaxial acceleration and angular velocity. The motion intensity correlation parameter is calculated based on the mean acceleration, with a value range of 0-1. The motion duration is accurately recorded by the wristband timing module with a time resolution of 1 second, and the probability weight coefficient is set to 0.6. The fatigue accumulation correction factor is adjusted according to the type of exercise: 0.7 for running, 0.6 for cycling, and 0.8 for strength training. The model calculates the product of the rate of change of motion posture and the square of the motion duration, combines it with the ratio of the pulse wave feature fusion value to the square root of the motion parameters, and obtains the preliminary fatigue probability through nonlinear transformation. Then, it is corrected by multiplying the gradient value of the motion posture parameters by the intensity accumulation coefficient (valued at 0.5), which is dynamically adjusted according to the motion intensity. During implementation, the model adopts a segmented calculation strategy, updating the fatigue probability value every 5 seconds. The calculation process consumes no more than 20% of the wristband's processor resources, and the output fatigue probability value is strictly controlled within the range of 0-1 to ensure the accuracy and reliability of the quantification results.

[0029] Preferably, the expression for the motion artifact removal recurrent network is: ,in: The output signal after artifact removal at time t, The convolution kernel weight matrix is... These are the recurrent layer weight parameters. for Real-time network output, The input layer weight matrix, The input for fusing the original pulse wave and motion parameters at time t is... For the loop layer bias term, The integral coefficients for the time window. The length of the cyclic feedback time window. This is the integral weight matrix. for Input signals at all times, This represents the artifact suppression coefficient. To adapt weights to motion parameters, Let be the feature matrix of motion artifacts at time t. This is the activation function.

[0030] Specifically, the motion artifact elimination recurrent network is used to separate the effective components from motion interference in the original pulse wave signal, ensuring signal purity. During implementation, the original pulse wave signal and motion parameters are first fused into the input. The input data is converted to 32-bit floating-point format, and the convolutional kernel weight matrix is ​​initialized to a 3×3 size. The recurrent layer weight parameters are initialized using the Xavier initialization method, with the initial value of the input layer weight matrix ranging from -0.01 to 0.01. The recurrent layer bias term is fixed at 0.05, the ReLU activation function is used, the time window integral coefficient is set to 0.3, and the recurrent feedback time window length is set to 20 data points based on the sampling frequency. The artifact suppression coefficient is set to 0.8, and the motion parameter adaptation weight matrix is ​​dynamically adjusted according to the motion intensity: 0.2 for low-intensity motion, 0.4 for medium-intensity motion, and 0.6 for high-intensity motion. The motion artifact feature matrix is ​​obtained from acceleration and angular velocity data through feature extraction, including information such as vibration frequency and amplitude changes. After the network training is completed, it is embedded in the wristband firmware. During operation, it captures local signal features through multi-layer convolution and combines recurrent layers to remember long-term dependencies. It outputs the artifact-free signal every 10ms. The effective signal-to-noise ratio of the separated signal is improved by more than 30% compared with the original signal, ensuring the accuracy of subsequent feature extraction.

[0031] Preferably, the fusion model expression of the full-dimensional physiological motion fusion analysis platform is: ,in, To integrate the analysis results, This is the pulse wave characteristic parameter vector. This represents the probability value of muscle fatigue. Let be the motion attitude parameter vector, and () be the matrix tensor product operation. The diagonal matrix of motion parameters To integrate the weighting coefficients, These are the vector magnitudes of the pulse wave characteristics and motion parameters, respectively. It is the identity matrix. It is the vector inner product.

[0032] Specifically, the fusion model of the full-dimensional physiological-motor fusion analysis platform is used to achieve deep correlation and fusion of multiple types of parameters, constructing a physiological-motor state mapping relationship. During model implementation, the pulse wave feature parameter vector (1×8 dimension), muscle fatigue probability value, and motion posture parameter vector (1×6 dimension) are first obtained. In the fusion weight coefficients, α is set to 0.4, β to 0.3, γ to 0.1, δ to 0.1, ε to 0.05, and ω to 0.05, with the identity matrix dimension consistent with the feature vector dimension. The model achieves spatial fusion of pulse wave features and motion parameters through matrix tensor product operations. It utilizes a diagonal matrix to transform motion posture parameters, enhancing the weight adaptability of the parameters. Vector magnitude normalization is then used to avoid the influence of parameter magnitude differences. Vector inner product calculation is used to quantify the correlation between pulse wave features and motion parameters. The correlation result is used to dynamically adjust the fusion weights; when the correlation is higher than 0.7, the corresponding parameter weight is increased by 20%. During implementation, the model runs on the platform's fusion computing unit, adopts a parallel computing architecture, and the time for a single fusion is no more than 10ms. The output fusion parsing result is a 128-dimensional comprehensive feature vector, which comprehensively covers physiological and motion state information, providing comprehensive data support for the subsequent construction of mapping relationships.

[0033] Preferably, the expression for the portable fitness tracker's identification parameter adaptation model is: ,in, To identify the fit value, For exercise intensity level parameters, For heart rate detection parameters, For motion posture fusion parameters, For the adaptation coefficient, For the gradient of heart rate changes, For the identification period, The rate of change of motion posture.

[0034] Specifically, the portable fitness tracker's parameter adaptation model optimizes the adaptability of recognition parameters and improves overall recognition accuracy. During model implementation, the exercise intensity level parameter is divided based on average acceleration: low intensity (0.2), medium intensity (0.6), and high intensity (1.0). The heart rate detection parameter uses real-time heart rate data with an accuracy of 1 beat / minute. The motion posture fusion parameter consists of the mean and variance of three-axis acceleration and angular velocity, totaling six dimensions. In the adaptation coefficients, λ, μ, ν, and ξ are set to 0.3, 0.2, 0.3, and 0.2, respectively. The heart rate change gradient is calculated from five consecutive heart rate data points, with a fixed recognition period of one minute. The model calculates the product of the exercise intensity level and the square of the heart rate, combined with the product of the motion posture fusion parameter and the heart rate change gradient, to obtain the basic adaptation value. Then, it accumulates the product of the motion posture change rate and the exercise intensity through integral calculation, and corrects the result by combining the square root ratio of heart rate, motion posture, and exercise intensity. During implementation, the model receives the output parameters of each module in real time and updates the recognition fit value every 30 seconds. The fit value ranges from 0 to 1. When the fit value is higher than 0.8, it is determined that the current parameter set has good fit and the recognition result is highly reliable. When it is lower than 0.5, the parameter weights are automatically adjusted and recalculated to ensure the accuracy and stability of the recognition result.

[0035] Preferably, step S3 includes the following sub-steps: S31, based on the feature extraction logic of the photoplethysmography pulse wave analytical model, the artifact-free pulse wave signal is divided into time domain segments, and multiple analysis windows are divided according to the signal period, with each window corresponding to a single pulse wave period; S32, within each analysis window, the pulse wave rising edge, peak point, and falling edge are located to calibrate feature points, and the time interval and amplitude difference between each feature point are calculated to obtain the basic parameters of pulse wave conduction time and peak amplitude; S33, the pulse wave waveform is smoothed using a curve fitting algorithm, and the slope values ​​of the rising and falling edges of the waveform are solved to extract waveform curvature change feature parameters; S34, the time domain feature parameters and waveform feature parameters are combined according to a preset dimension to construct a multidimensional physiological feature parameter matrix, providing input data for subsequent fusion analysis.

[0036] Specifically, the multidimensional pulse wave feature extraction process in step S3 achieves accurate feature analysis through four sub-steps. In step S31, based on the feature extraction logic of the photoplethysmography pulse wave analytical model, the artifact-free pulse wave signal is uniformly segmented in the time domain. The segmentation window length is set to 2 seconds, adapted to the pulse wave cycle (0.5-1.5 seconds), ensuring that each window contains 1-4 complete pulse wave cycles. The window overlap rate is set to 50% to avoid missing feature information. In step S32, within each analysis window, an adaptive threshold method is used to locate the pulse wave rising edge start point, systolic peak point, diastolic trough point, and dicrotic peak point. The pulse wave propagation time is obtained by calculating the time difference between adjacent feature points, with an accuracy of 1ms. The peak amplitude is obtained by calculating the amplitude difference between the peak and trough points, with a resolution of 0.1m. V; S33 uses a cubic spline curve fitting algorithm to smooth the pulse wave waveform, with the fitting error controlled within 5%. The waveform slope value is obtained by solving the first derivative of the fitted curve at the rising and falling edges, with the unit being mV / ms. At the same time, the rate of change of curvature at the inflection point of the curve is calculated to extract the waveform curvature feature parameters. S34 combines the time domain feature parameters (conduction time, rise time, fall time) and waveform feature parameters (peak amplitude, slope, rate of change of curvature) according to the preset 8-dimensional feature dimensions to construct an N×8-dimensional physiological feature parameter matrix (N is the number of analysis windows). The matrix data is stored in 32-bit floating-point format to provide standardized input data for the subsequent muscle fatigue probability prediction model.

[0037] Preferably, step S4 includes the following sub-steps: S41, standardizing and mapping the multidimensional physiological feature parameters obtained in S3 to convert them into an adaptive input format for the muscle fatigue probability prediction model, and simultaneously performing vector normalization on the movement posture parameters; S42, inputting the adapted physiological feature parameters, normalized movement posture parameters, and movement duration parameters into the model, and performing feature weighting and nonlinear transformation through the model's built-in multi-layer computing unit; S43, calculating the muscle fatigue correlation factor based on the transformed feature data, adjusting the factor weights in conjunction with the exercise intensity parameter, and obtaining a preliminary fatigue probability value; S44, iteratively correcting the preliminary fatigue probability value through a model feedback mechanism to eliminate deviations caused by abnormal parameters, and outputting the final muscle fatigue state correlation probability value.

[0038] Specifically, the muscle fatigue probability calculation process in step S4 achieves precise quantification of fatigue state through four sub-steps. In step S41, the multidimensional physiological feature parameters obtained in step S3 are first mapped to the [0,1] interval through linear transformation. The transformation coefficients are calibrated based on massive sample data. Simultaneously, the motion posture parameters (three-axis acceleration and angular velocity) are vector normalized using the maximum-minimum method to convert parameter values ​​to the same order of magnitude. In step S42, the adapted 8-dimensional physiological feature parameters, 6-dimensional normalized motion posture parameters, and 1-dimensional motion duration parameters (accuracy 1 second) are input into the muscle fatigue probability prediction model. The model has three fully connected computational units, with 64, 32, and 16 neurons in each layer, respectively. The input parameters are weighted and summed using a weight matrix (initial value range -0.02 to 0.02), and then converted through a nonlinear activation function. High-dimensional feature vector; S43 calculates the muscle fatigue correlation factor based on the transformed high-dimensional feature vector. This factor is obtained by weighting the contribution values ​​of physiological features and exercise parameters in a ratio of 0.6:0.4. The exercise intensity parameter (value 0-1) is adjusted by adjusting the weight coefficient to correct the factor size. The higher the exercise intensity, the higher the weight of the exercise parameter contribution value is to 0.5; S44 uses the built-in iterative correction mechanism of the model, with a maximum of 5 iterations. Each iteration adjusts the parameter weights based on the deviation between the previous calculation result and the historical sample data. The deviation threshold is set to 0.01. When the deviation is less than the threshold, the iteration stops. The final output is a muscle fatigue state correlation probability value in the range of [0,1]. The time taken for a single iteration is no more than 2ms, ensuring overall computational efficiency.

[0039] Preferably, step S5 includes the following sub-steps: S51, using the parameter adaptation module of the full-dimensional physiological-motor fusion analysis platform, the data format of muscle fatigue probability values, pulse wave characteristic parameters, and movement posture parameters are unified and aligned in dimensions to construct a cross-type data set; S52, the platform's correlation analysis unit is invoked to calculate the correlation coefficient matrix between different types of parameters based on preset physiological-motor correlation rules, and to identify strongly correlated parameter combinations; S53, the platform's fusion calculation unit performs weighted fusion on the strongly correlated parameter combinations, and a multi-dimensional data fusion algorithm is used to generate a comprehensive feature vector; S54, a physiological-motor state mapping relationship model is constructed based on the comprehensive feature vector to determine the physiological parameter adaptation range under different movement states.

[0040] Specifically, the full-dimensional data fusion and analysis process in step S5 constructs a physiological-motor state mapping relationship through four sub-steps. In S51, the parameter adaptation module of the full-dimensional physiological-motor fusion and analysis platform receives muscle fatigue probability values ​​(1D), an 8D pulse wave feature parameter matrix, and 6D motion posture parameters. A data format conversion algorithm is used to unify all parameters into a 128-dimensional feature vector format. Missing data is supplemented using linear interpolation at a 100ms interval to ensure data integrity. In S52, the platform's correlation analysis unit is invoked to calculate the correlation coefficients between different types of parameters based on the Pearson correlation coefficient algorithm. The calculation window length is set to 10 seconds, resulting in a 15×15-dimensional correlation coefficient matrix (15 being the total parameter dimension). A correlation coefficient threshold of 0.6 is set, and combinations of parameters strongly correlated with muscle fatigue probability values, pulse wave conduction time, and average acceleration are selected. In S53, the platform's fusion calculation unit... The model performs weighted fusion of strongly correlated parameter combinations. The initial weights are allocated as follows: physiological feature parameter 0.3, movement posture parameter 0.3, and fatigue probability value 0.4. A dynamic weight adjustment mechanism is introduced. For each level increase in exercise intensity (low → medium → high), the weight of the movement posture parameter increases by 0.1. A weighted average fusion algorithm is used to generate a 128-dimensional comprehensive feature vector. Based on the comprehensive feature vector, combined with a training dataset containing more than 100,000 exercise and physiological state samples, a physiological-exercise state mapping model is constructed using the support vector machine algorithm. The model kernel function is a radial basis function, the penalty coefficient is set to 1.0, and the classification error is controlled within 5%. The model clarifies the exercise intensity level and muscle fatigue degree range corresponding to different comprehensive feature vectors, providing the core basis for the recognition result output in step S6.

[0041] The photoplethysmography (PPG) pulse wave analytical model is an algorithmic model used in this invention for in-depth mining of physiological information from pulse waves. It is a feature extraction tool built based on the PPG sensing principle and signal analysis logic. Its implementation process requires first receiving the effective pulse wave signal after motion artifact elimination, dividing it into 2-second analysis windows with a 50% overlap, ensuring each window contains 1-4 complete pulse wave cycles; using an adaptive threshold method to locate key feature points such as the rising edge start point and the systolic peak point, calculating time-domain parameters such as pulse wave propagation time (accuracy 1ms) and peak amplitude (resolution 0.1mV); using a cubic spline curve to fit the waveform (fitting error ≤5%), solving the first derivative to obtain the waveform slope, and calculating the rate of change of curvature at inflection points to obtain waveform features; finally, combining the time-domain and waveform feature parameters into an 8-dimensional physiological feature parameter matrix, stored in 32-bit floating-point format. This model accurately extracts multidimensional features reflecting the human cardiovascular state from the pulse wave signal after artifact removal, providing standardized input for subsequent muscle fatigue prediction. It breaks through the limitations of traditional pulse wave analysis that only focuses on a single parameter. By analyzing multidimensional features, it improves the comprehensiveness and accuracy of physiological information extraction, laying a high-quality data foundation for judging exercise state and physiological adaptability, and ensuring the scientificity and reliability of the recognition results.

[0042] The muscle fatigue probability prediction model is an algorithmic model that quantifies muscle fatigue during human movement. It is a nonlinear probability calculation tool that integrates physiological characteristics and motion parameters. Its implementation process requires first linearly transforming the 8-dimensional physiological characteristic parameters (mapping them to the [0,1] interval), and then normalizing the 6-dimensional motion posture parameters (three-axis acceleration and angular velocity) using the minimax method. The adapted parameters and the motion duration parameters (accuracy 1 second) are input into a model with three built-in fully connected computational units (64, 32, and 16 neurons respectively). The model is then weighted and summed using a weight matrix with initial values ​​ranging from -0.02 to 0.02, and converted into a high-dimensional feature vector through a nonlinear activation function. The contribution values ​​of physiological characteristics and motion parameters are weighted at a ratio of 0.6:0.4 to obtain a fatigue correlation factor. When the exercise intensity increases, the weight of the motion parameters is adjusted to 0.5. The calculation results are optimized through an iterative correction mechanism (up to 5 iterations, deviation threshold of 0.01, and single iteration time ≤2ms), finally outputting the fatigue probability value in the [0,1] interval. This model is based on multi-source parameters to accurately quantify muscle fatigue, realizing a quantitative mapping from characteristic parameters to fatigue state. It solves the shortcomings of traditional fatigue assessment that rely on subjective judgment or single parameters. Through multi-dimensional parameter fusion and iterative optimization, it improves the objectivity and accuracy of fatigue assessment, provides core basis for exercise intensity regulation and health risk warning, and helps to realize personalized exercise monitoring.

[0043] The motion artifact elimination recurrent network is a network model that separates the effective components of the pulse wave signal from motion interference. It is a signal processing architecture that combines convolution and recurrent feedback mechanisms. The implementation process requires first converting the original pulse wave signal and motion posture parameters into 32-bit floating-point data and combining them into an input data matrix. The network consists of 3 convolutional layers (3×3 convolutional kernels, with 16, 32, and 64 kernels), 2 recurrent layers (128 hidden units), and 1 fusion output layer. The convolutional layers capture local signal features through a sliding window, and the recurrent layers use a long short-term memory structure to remember long-term dependencies. The recurrent layer weights are initialized using Xavier, with the initial values ​​of the input layer weights ranging from -0.01 to 0.01, and the recurrent layer bias term set to 0.05. The ReLU activation function is selected. The integral coefficient of the time window is set to 0.3, the length of the recurrent feedback time window is 20 data points, and the artifact suppression coefficient is 0.8. The motion parameters are dynamically adjusted to adapt the weights according to the motion intensity (0.2, 0.4, and 0.6 for low, medium, and high intensities, respectively). The model is optimized by minimizing the mean square error (learning rate 0.001, 500 iterations), and an artifact-free signal is output every 10ms. This network removes interference components such as vibration and displacement from pulse wave signals in motion scenarios, retaining effective physiological signals. It overcomes the limitations of traditional filtering methods in dealing with dynamic motion interference. Through the synergistic effect of convolution and recurrent feedback, it improves the accuracy of artifact elimination, increasing the signal-to-noise ratio by more than 30%. This provides a high-purity signal foundation for subsequent feature analysis and fatigue prediction, ensuring the stability of the overall recognition scheme.

[0044] The all-dimensional physiological-motor fusion analysis platform is a processing platform that realizes deep correlation and mapping of multiple types of parameters. It integrates data adaptation, correlation analysis, fusion calculation, and mapping construction into a unified system module. Its implementation process requires the parameter adaptation module to first receive muscle fatigue probability values, 8-dimensional pulse wave feature parameters, and 6-dimensional motion posture parameters, which are then uniformly converted into a 128-dimensional feature vector. Linear interpolation (100ms interval) is used to supplement missing data. The correlation analysis unit then uses the Pearson correlation coefficient algorithm (10-second calculation window) to generate a 15×15-dimensional correlation coefficient matrix, filtering strongly correlated parameter combinations with a threshold of 0.6. The fusion calculation unit assigns initial weights based on physiological characteristics (0.3), motion posture (0.3), and fatigue probability (0.4). For each increase in exercise intensity, the weight of the motion posture parameter is increased by 0.1. A weighted average algorithm is used to generate a 128-dimensional comprehensive feature vector. Based on a training dataset of over 100,000 samples, a mapping model (classification error ≤5%) is constructed using a radial basis function kernel and a support vector machine algorithm with a penalty coefficient of 1.0, clarifying the correspondence between the comprehensive feature vector and exercise intensity and fatigue level. This platform breaks through the limitations of isolated analysis of physiological and exercise parameters, constructs cross-dimensional data mapping relationships, and improves data utilization through full-dimensional fusion and deep correlation. It enables the recognition results to fully reflect the dynamic relationship between human movement and physiological state, solving the problems of simple fusion mechanism and single recognition dimension in traditional methods. It provides core support for accurately outputting exercise intensity level, fatigue level and physiological adaptability judgment, and meets the comprehensive needs of sports health monitoring.

[0045] like Figure 2As shown, a portable fitness tracker identification system is applied to a portable fitness tracker identification method. The system includes: a photoplethysmography-motion parameter synchronous acquisition unit, electrically connected to the photoplethysmography sensor, three-axis accelerometer, and gyroscope of the portable fitness tracker, used to synchronously acquire raw pulse wave signals, motion posture parameters, and preliminary heart rate detection data; a motion artifact intelligent separation unit, signal-connected to the photoplethysmography-motion parameter synchronous acquisition unit, with a built-in motion artifact elimination loop network to perform separation processing of the effective components of the pulse wave and motion interference components; and a multi-dimensional pulse wave feature analysis unit, signal-connected to the motion artifact intelligent separation unit, equipped with a photoplethysmography pulse wave analysis model to analyze the artifact-free pulse wave signal. Feature extraction and analysis are performed; the muscle fatigue probability calculation unit is connected to the multi-dimensional pulse wave feature analysis unit and the photoplethysmography-motion parameter synchronous acquisition unit, respectively, and calculates the probability value associated with muscle fatigue state through the muscle fatigue probability prediction model; the full-dimensional data fusion analysis unit is connected to the muscle fatigue probability calculation unit, the multi-dimensional pulse wave feature analysis unit and the motion artifact intelligent separation unit, respectively, and constructs the physiological-motor state mapping relationship; the recognition result output unit is connected to the full-dimensional data fusion analysis unit, and outputs the exercise intensity level, muscle fatigue degree and physiological state adaptability judgment result based on the fusion analysis result. Each unit interacts and works collaboratively through the data transmission bus built into the wristband.

[0046] A portable fitness tracker identification method and system employs a synchronous acquisition mechanism, integrating multi-source data from a photoplethysmometer, accelerometer, and gyroscope to achieve comprehensive capture of physiological signals and motion parameters, providing rich data support for subsequent analysis. Motion artifact elimination is achieved through a dedicated recurrent network, relying on multi-layer convolution and recurrent feedback mechanisms to significantly improve the accuracy of signal separation in complex motion scenarios, effectively preserving the effective features of the pulse wave. Furthermore, targeted analytical and predictive models are used to deeply mine the multi-dimensional features of the pulse wave and quantify muscle fatigue state, achieving a progressive analysis from signal to feature to state, thus enhancing the scientific rigor of the identification results.

[0047] At the system level, a modular design constructs a full-process architecture, with six functional units each performing their respective functions and coordinating efficiently through a bus to ensure closed-loop operation of data from acquisition, processing, analysis to output, balancing processing efficiency and reliability. The multi-dimensional fusion and analysis platform breaks through the limitations of single-parameter analysis, establishing a deep correlation between physiological characteristics, movement posture, and fatigue state, achieving efficient fusion and mapping of cross-dimensional data, and allowing the recognition results to more comprehensively reflect human movement and physiological state. Each model and unit is optimized for the application scenarios of portable wristbands, adapting to the hardware resources and usage environment of the wristbands, ensuring recognition accuracy while taking into account portability and practicality, and is suitable for health monitoring needs in various sports scenarios.

[0048] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A portable fitness tracker identification method, characterized in that, Includes the following steps: S1. The raw pulse wave signal during movement is collected using the photoplethysmography (PPG) sensor mounted on the portable fitness tracker. Simultaneously, motion posture parameters and preliminary heart rate data are acquired from the tracker's built-in three-axis accelerometer and gyroscope. Ambient light levels are also detected, monitoring light exposure levels under different spectra, including blue light exposure. S2. The raw pulse wave signal and motion posture parameters are input into a motion artifact removal recurrent network. The network uses multi-layer convolution and recurrent feedback mechanisms to separate the effective pulse wave components from motion interference components. S3. The PPG pulse wave analytical model is used to extract features from the artifact-free pulse wave signal. The process involves: S4, obtaining multidimensional physiological characteristic parameters such as pulse wave conduction time, peak amplitude, and waveform slope; S5, inputting these parameters and movement posture parameters into a muscle fatigue probability prediction model, and calculating the probability value associated with muscle fatigue state by combining the movement duration parameter; S6, using a full-dimensional physiological-motor fusion analysis platform to perform cross-dimensional data association and fusion analysis on the muscle fatigue probability value, pulse wave characteristic parameters, and movement posture parameters, and constructing a physiological-motor state mapping relationship; and S7, outputting movement state recognition results based on the fusion analysis results, including movement intensity level, muscle fatigue degree, and physiological state suitability judgment.

2. The portable fitness tracker identification method according to claim 1, characterized in that, The expression for the analytical model of the photoplethysmography pulse wave is: in, The analyzed three-dimensional pulse wave feature matrix. This is the pulse wave signal attenuation coefficient. These are the photosensitive area parameters of the photoplethysmography sensor. The fundamental angular frequency of the pulse wave. This is the phase offset. This is a correction factor for conduction velocity. For parameters related to the elastic modulus of blood vessels, The two-dimensional Laplace operator for the original pulse wave signal. It is the amplitude gain factor. For parameters related to the motion amplitude of the accelerometer, For signal acquisition period, This is the second derivative signal of the pulse wave.

3. The portable fitness tracker identification method according to claim 1, characterized in that, The expression for the muscle fatigue probability prediction model is as follows: ,in, This represents the probability value of muscle fatigue. This represents the fusion value of pulse wave characteristics. This is a motion attitude parameter matrix, including triaxial acceleration and angular velocity components. For parameters related to exercise intensity, For the duration of the movement, These are probability weighting coefficients. This is a fatigue accumulation correction factor. The rate of change of motion posture. This is the attitude influence coefficient. This is the intensity cumulative factor. The gradient values ​​are the motion attitude parameters.

4. The portable fitness tracker identification method according to claim 1, characterized in that, The expression for the motion artifact removal recurrent network is: ,in: The output signal after artifact removal at time t, The convolution kernel weight matrix is... These are the recurrent layer weight parameters. for Real-time network output, The input layer weight matrix, The input for fusing the original pulse wave and motion parameters at time t is... For the loop layer bias term, The integral coefficients for the time window are... The length of the cyclic feedback time window. This is the integral weight matrix. for Input signals at all times, This represents the artifact suppression coefficient. To adapt weights to motion parameters, Let be the feature matrix of motion artifacts at time t. This is the activation function.

5. The portable fitness tracker identification method according to claim 1, characterized in that, The fusion model expression of the full-dimensional physiological motion fusion analysis platform is as follows: ,in, To integrate the analysis results, This is the pulse wave characteristic parameter vector. This represents the probability value of muscle fatigue. Let be the motion attitude parameter vector, and () be the matrix tensor product operation. The diagonal matrix of motion parameters To integrate the weighting coefficients, These are the vector magnitudes of the pulse wave characteristics and motion parameters, respectively. It is the identity matrix. It is the dot product of vectors.

6. The portable fitness tracker identification method according to claim 1, characterized in that, The expression for the portable fitness tracker's identification parameter adaptation model is as follows: ,in, To identify the fit value, For exercise intensity level parameters, For heart rate detection parameters, For motion posture fusion parameters, For the adaptation coefficient, For the gradient of heart rate changes, For the identification period, The rate of change of motion posture.

7. The portable fitness tracker identification method according to claim 1, characterized in that, S3 includes the following sub-steps: S31, based on the feature extraction logic of the photoplethysmography pulse wave analytical model, the artifact-free pulse wave signal is divided into time domain segments, and multiple analysis windows are divided according to the signal period, with each window corresponding to a single pulse wave period; S32, within each analysis window, the pulse wave rising edge, peak point, and falling edge are located to calibrate feature points, and the time interval and amplitude difference between each feature point are calculated to obtain the basic parameters of pulse wave conduction time and peak amplitude; S33, the pulse wave waveform is smoothed using a curve fitting algorithm, and the slope values ​​of the rising and falling edges of the waveform are solved to extract the waveform curvature change feature parameters; S34, the time domain feature parameters and waveform feature parameters are combined according to a preset dimension to construct a multi-dimensional physiological feature parameter matrix, providing input data for subsequent fusion analysis.

8. The portable fitness tracker identification method according to claim 1, characterized in that, S4 includes the following sub-steps: S41, standardizing and mapping the multidimensional physiological feature parameters obtained in S3 to convert them into an adaptive input format for the muscle fatigue probability prediction model, and simultaneously performing vector normalization on the movement posture parameters; S42, inputting the adapted physiological feature parameters, normalized movement posture parameters, and movement duration parameters into the model, and performing feature weighting and nonlinear transformation through the model's built-in multi-layer computing unit; S43, calculating the muscle fatigue correlation factor based on the transformed feature data, adjusting the factor weights in conjunction with the exercise intensity parameter, and obtaining a preliminary fatigue probability value; S44, iteratively correcting the preliminary fatigue probability value through the model feedback mechanism, eliminating the bias caused by abnormal parameters, and outputting the final muscle fatigue state correlation probability value.

9. The portable fitness tracker identification method according to claim 1, characterized in that, S5 includes the following sub-steps: S51, using the parameter adaptation module of the full-dimensional physiological-motor fusion analysis platform, the data format of muscle fatigue probability values, pulse wave characteristic parameters, and movement posture parameters are unified and aligned in dimensions to construct a cross-type data set; S52, the platform's correlation analysis unit is invoked to calculate the correlation coefficient matrix between different types of parameters based on preset physiological-motor correlation rules, and to identify strongly correlated parameter combinations; S53, the platform's fusion calculation unit performs weighted fusion on the strongly correlated parameter combinations and uses a multi-dimensional data fusion algorithm to generate a comprehensive feature vector; S54, based on the comprehensive feature vector, a physiological-motor state mapping relationship model is constructed to determine the physiological parameter adaptation range under different movement states.

10. A portable fitness tracker identification system, characterized in that, This system is applied to a portable fitness tracker identification method as described in claim 1, comprising: a photoplethysmography-motion parameter synchronous acquisition unit, electrically connected to the photoplethysmography sensor, three-axis accelerometer, and gyroscope of the portable fitness tracker, for synchronously acquiring raw pulse wave signals, motion posture parameters, and preliminary heart rate detection data; a motion artifact intelligent separation unit, signal-connected to the photoplethysmography-motion parameter synchronous acquisition unit, with a built-in motion artifact elimination loop network, performing separation processing of effective pulse wave components and motion interference components; and a multi-dimensional pulse wave feature analysis unit, signal-connected to the motion artifact intelligent separation unit, equipped with a photoplethysmography pulse wave analysis model, for feature extraction of the artifact-free pulse wave signal. The system comprises three parts: a muscle fatigue probability calculation unit, which is connected to the multi-dimensional pulse wave feature analysis unit and the photoplethysmography-motion parameter synchronous acquisition unit, respectively, and calculates the probability value associated with muscle fatigue state through a muscle fatigue probability prediction model; a full-dimensional data fusion analysis unit, which is connected to the muscle fatigue probability calculation unit, the multi-dimensional pulse wave feature analysis unit, and the motion artifact intelligent separation unit, respectively, and constructs a physiological-motor state mapping relationship; and a recognition result output unit, which is connected to the full-dimensional data fusion analysis unit, and outputs the exercise intensity level, muscle fatigue degree, and physiological state suitability judgment result based on the fusion analysis result. Each unit interacts and works collaboratively through the data transmission bus built into the wristband.

Images