Metabolic equivalent driven endurance training intensity gradient generation system and method

Abstract

The application discloses a metabolic equivalent driven endurance training intensity gradient generation system and method thereof, and belongs to the technical field of intelligent sports training. The system comprises a non-contact physiological signal acquisition module, a multi-dimensional metabolic state modeling module, a floating threshold generation module, a multi-channel feedback regulation module and an edge abnormality filtering module. Respiratory and heart rate and other multi-dimensional physiological parameters are collected through the fusion of a millimeter wave radar and a PPG sensor. A three-dimensional metabolic state dynamic evaluation model is constructed based on a GRU network, and real-time MET values are output. Intensity adjustment threshold values are adaptively generated according to MET deviation. The exercise intensity is cooperatively regulated through the tactile feedback of intelligent running shoes and the slope adjustment of a treadmill. Abnormal data are filtered at the edge by using a LOF algorithm. The application solves the problems of single physiological signal dimension, static metabolic state evaluation and abnormal data interference in the prior art, and realizes the precise real-time regulation of endurance training intensity.

Description

Metabolic equivalent-driven endurance training intensity gradient generation system and method Technical Field

[0001] This invention relates to the field of intelligent sports training technology, specifically to a metabolic equivalent-driven endurance training intensity gradient generation system and method, which, in particular, achieves refined and personalized dynamic control of exercise intensity during endurance training by integrating non-contact physiological signal monitoring with a deep learning model. Background Technology

[0002] In the prior art, Chinese Patent CN 112023342 B (authorization announcement number, authorization announcement date March 15, 2022, titled "Method for Adjusting Treadmill Speed ​​and Incline, Exercise Training Method and Treadmill") discloses a method for adjusting treadmill speed and incline in real time based on heart rate changes. The technical solution of this patent includes: real-time monitoring of the user's current heart rate during exercise; when the user's current heart rate is outside the target heart rate range, calculating the difference between the current heart rate and the average of the target heart rate range; establishing a polynomial formula based on the user's test heart rate, test intensity, resting heart rate, and resting intensity; calculating the difference between the current intensity and the target intensity of the treadmill based on the difference between the current heart rate and the average of the target heart rate range and the polynomial formula, and obtaining the target intensity; converting the target intensity into target speed and target incline using an intensity-speed / incline conversion formula. This technical solution achieves automatic adjustment of treadmill parameters by establishing a polynomial mapping relationship between heart rate and exercise intensity.

[0003] However, this existing technical solution has the following shortcomings:

[0004] First, regarding the acquisition of physiological signals, this solution relies solely on contact-based heart rate monitoring as a single physiological indicator. In actual endurance training scenarios, when athletes are in the high-intensity interval training phase, this single heart rate indicator exhibits significant lag. Specifically, the response delay of heart rate to changes in exercise intensity is typically 30 to 60 seconds, which means that under rapidly changing training intensities, the system cannot accurately and promptly reflect the athlete's true metabolic state. Furthermore, contact-based heart rate monitoring devices (such as chest straps or wristbands) are prone to signal quality degradation during prolonged exercise due to factors such as sweat and displacement, further reducing the accuracy and reliability of the monitoring.

[0005] Second, regarding the metabolic state assessment mechanism, this scheme employs a polynomial fitting method based on static test data, failing to fully consider the dynamic evolution of metabolic state during exercise. In actual endurance training, an athlete's metabolic state is influenced by multiple factors, including prior fatigue accumulation, current exercise load, and changes in respiratory efficiency. While the polynomial relationship established by this scheme can reflect the heart rate-intensity correspondence under static test conditions, it lacks the ability to model the nonlinear dynamic changes in metabolic state during exercise. Specifically, during endurance training lasting more than 30 minutes, due to glycogen depletion, lactic acid accumulation, and changes in breathing patterns, the actual metabolic equivalent (MET) corresponding to the same heart rate may deviate by more than 15%, leading to a significant decrease in the accuracy of intensity regulation.

[0006] Third, regarding the abnormal data processing mechanism, this solution lacks the ability to identify and filter unconventional physiological events during exercise. In actual training scenarios, athletes may experience brief coughing, sneezing, or deep breathing adjustments, which can cause momentary abnormal fluctuations in heart rate data. Since this solution directly calculates intensity based on real-time heart rate data, when such abnormal events occur, the system may misinterpret them as requiring adjustments to exercise intensity, leading to unnecessary and frequent changes in treadmill parameters. This severely impacts the continuity of training and the athlete's comfort. Quantitative analysis shows that in 60 minutes of endurance training, such abnormal events occur approximately once to twice every 10 minutes, accumulating to approximately 10% to 15% of intensity adjustments being misjudged.

[0007] Therefore, there is an urgent need to develop an endurance training intensity regulation system that can integrate multidimensional physiological signals, accurately assess dynamic metabolic state, and effectively filter abnormal data, in order to meet the higher requirements of modern intelligent sports training for refinement, personalization, and real-time performance. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a metabolic equivalent-driven endurance training intensity gradient generation system and method. Through non-contact multidimensional physiological signal fusion, dynamic modeling of metabolic state based on temporal deep learning, and intelligent anomaly filtering at the edge, the system achieves precise real-time control of exercise intensity during endurance training, solving technical problems such as the single dimension of physiological signals, static metabolic state assessment, and interference from abnormal data in existing technologies.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] The metabolic equivalent-driven endurance training intensity gradient generation system includes: a non-contact physiological signal acquisition module, a multi-dimensional metabolic state modeling module, a floating threshold generation module, a multi-channel feedback regulation module, and an edge anomaly filtering module.

[0011] The non-contact physiological signal acquisition module is used to acquire the respiratory rate and respiratory amplitude of athletes through millimeter-wave radar, and at the same time acquire heart rate and blood oxygen saturation through a wrist photoplethysmography (PPG) sensor, so as to realize the synchronous acquisition of multi-dimensional physiological parameters.

[0012] The multidimensional metabolic state modeling module is used to construct a three-dimensional real-time assessment model of the exerciser's metabolic state based on a gated cyclic unit (GRU) network, taking respiratory amplitude characteristics, heart rate variability (HRV) time-domain characteristics, and exercise power as inputs, and outputting the current metabolic equivalent (MET) value.

[0013] The floating threshold generation module is used to dynamically generate an adaptive intensity adjustment threshold based on the continuous deviation between the preset target MET range and the current actual MET value.

[0014] The multi-channel feedback control module is used to apply gradient tactile cues to the soles of the runner's feet through the built-in tactile motor array of the smart running shoe after detecting that the MET value has continuously deviated from the target range for a preset time. At the same time, it sends an incline adjustment command to the treadmill for intensity compensation.

[0015] The edge-end anomaly filtering module is used to detect and remove abnormal data points caused by unconventional physiological events such as coughing and sneezing in real time on edge computing devices based on the Local Outlier Factor (LOF) algorithm.

[0016] Furthermore, the present invention also provides a metabolic equivalent-driven endurance training intensity gradient generation method. This method is based on the above-mentioned system and includes key steps such as multidimensional physiological signal acquisition, dynamic modeling of metabolic state, adaptive generation of floating threshold, multi-channel collaborative feedback regulation, and intelligent filtering of abnormal data, to ensure precise dynamic control of exercise intensity during endurance training.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] First, by coordinating non-contact millimeter-wave radar with a wrist-worn PPG sensor, the synchronous acquisition of multi-dimensional physiological parameters such as respiration, heart rate, and blood oxygenation is achieved. Compared to existing technologies that rely solely on a single heart rate indicator, the multi-dimensional physiological signal fusion scheme of this invention can more comprehensively reflect the metabolic state of the exerciser. In particular, the respiratory frequency amplitude characteristic responds to changes in exercise intensity approximately 15 to 20 seconds faster than the heart rate, significantly improving the system's real-time performance. Simultaneously, non-contact radar monitoring avoids the signal quality degradation problems caused by sweat and displacement during prolonged exercise with contact-based devices, improving monitoring accuracy by approximately 12%.

[0019] Second, by using three-dimensional dynamic modeling of metabolic state based on GRU network, the temporal evolution of metabolic state during exercise can be captured. Compared with the static polynomial fitting method used in the prior art, the deep learning model of this invention fully considers the influence of dynamic factors such as early fatigue accumulation and changes in respiratory efficiency on metabolic state, reducing the MET value estimation error from 15% in the prior art to less than 5%, significantly improving the accuracy of intensity regulation.

[0020] Third, through a floating threshold adaptive generation mechanism, the system can dynamically adjust the timing of intervention based on the real-time metabolic response characteristics of the exerciser, avoiding the problems of over-intervention or under-intervention that may occur with fixed thresholds. Combined with the dual-channel coordinated control of haptic feedback from smart running shoes and treadmill incline adjustment, it achieves precise, timely, and non-invasive intensity guidance for the exerciser.

[0021] Fourth, the LOF anomaly filtering algorithm deployed at the edge can identify and remove abnormal data points caused by unconventional physiological events such as coughing and sneezing in real time at the data acquisition end, reducing the proportion of misjudgment intensity adjustment from 10% to 15% in the existing technology to less than 2%, which greatly improves the continuity and comfort of the training process.

[0022] In summary, this invention comprehensively solves the key technical problems existing in the intensity regulation of endurance training through an innovative combination of technologies including multidimensional physiological signal fusion, temporal deep learning modeling, adaptive threshold generation, multi-channel collaborative regulation, and edge-end intelligent filtering. It has significant technological progress and practical value. Attached Figure Description

[0023] Figure 1 is a schematic diagram of the overall architecture of the endurance training intensity gradient generation system driven by metabolic equivalent of the present invention.

[0024] Figure 2 is a detailed schematic diagram of the non-contact physiological signal acquisition module of the present invention.

[0025] Figure 3 is a detailed schematic diagram of the multidimensional metabolic state modeling module of the present invention.

[0026] Figure 4 is a flowchart of the floating threshold generation module of the present invention.

[0027] Figure 5 is a detailed schematic diagram of the multi-channel feedback control module of the present invention.

[0028] Figure 6 is a flowchart of the edge-end anomaly filtering module of the present invention.

[0029] Figure 7 is an overall flowchart of the endurance training intensity gradient generation method driven by metabolic equivalent of the present invention. Detailed Implementation

[0030] Please refer to Figures 1-7. The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0031] As shown in Figure 1, the metabolic equivalent-driven endurance training intensity gradient generation system provided by this invention includes: a non-contact physiological signal acquisition module 1, a multi-dimensional metabolic state modeling module 2, a floating threshold generation module 3, a multi-channel feedback regulation module 4, and an edge anomaly filtering module 5. These five modules are interconnected through a data bus and control signal lines, working together to achieve precise dynamic regulation of exercise intensity during endurance training.

[0032] As shown in Figure 2, the non-contact physiological signal acquisition module 1 includes a millimeter-wave radar unit 11 and a wrist PPG sensing unit 12.

[0033] The millimeter-wave radar unit 11 employs a 60GHz frequency-modulated continuous wave (FMCW) radar with an operating wavelength of approximately 5 millimeters, enabling non-contact monitoring of chest cavity movement through clothing. In this embodiment, the radar unit 11 is mounted on the front column of the treadmill, maintaining a distance of 0.5 to 1.5 meters from the exerciser's chest. The radar unit 11 extracts temporal data of chest cavity displacement by emitting millimeter-wave signals and receiving signals reflected back from the human chest cavity. Specifically, the radar unit 11 continuously records the anterior-posterior displacement of the chest cavity at a sampling frequency of 100 Hz, which directly reflects the chest cavity expansion and contraction caused by respiratory movements.

[0034] The acquired time-series data of thoracic displacement were processed. First, a bandpass filter was used to remove DC components and high-frequency noise, retaining the respiratory signal in the 0.1 to 1.0 Hz frequency band. Then, two key features, respiratory rate and respiratory amplitude, were calculated. The respiratory rate was obtained by identifying the main peak frequency in the power spectrum using Fast Fourier Transform (FFT), measured in breaths per minute. The respiratory amplitude was defined as the peak-to-peak value of thoracic displacement within a single respiratory cycle, measured in millimeters. These two features together constitute the respiratory frequency-amplitude feature vector.

[0035] The wrist-worn PPG sensing unit 12 is integrated into the smart bracelet worn by the athlete. It employs dual-wavelength (green and red) LED light sources and a photodetector to monitor heart rate and blood oxygen saturation using photoplethysmography (PPG) technology. The green LED operates at a wavelength of 525 nanometers and is primarily used for heart rate detection; the red LED operates at a wavelength of 660 nanometers and, in conjunction with infrared light, is used for blood oxygen saturation detection. The PPG sensing unit 12 acquires photoelectric signals at a sampling frequency of 50 Hz and extracts the pulse waveform by detecting the periodic changes in light absorption caused by changes in arterial blood vessel volume.

[0036] The PPG signal was preprocessed, including baseline drift correction and motion artifact filtering. Baseline drift was corrected using cubic spline interpolation fitting and subtraction, while motion artifacts were suppressed using an adaptive filtering algorithm. Based on the preprocessed PPG signal, a peak detection algorithm was used to identify each cardiac cycle, and the time interval between two consecutive heartbeats (RR interval) was calculated. Heart rate was defined as 60 seconds divided by the average RR interval, in beats per minute.

[0037] Furthermore, the wrist PPG sensing unit 12 also calculates the temporal characteristics of heart rate variability (HRV). HRV reflects cardiac autonomic regulation and metabolic stress. In this embodiment, the standard deviation (SDNN) is used as the temporal characteristic index of HRV. SDNN is defined as the standard deviation of N consecutive RR intervals, and the calculation formula is:

[0038] ,

[0039] in: This represents the standard deviation of heart rate variability, in milliseconds. The total number of RR intervals used for calculation is, in this embodiment, taken as the number of all heartbeat cycles within the most recent 60 seconds, typically between 50 and 100; For the first The RR interval of one heartbeat, in milliseconds; This is the average of N RR intervals, in milliseconds.

[0040] Blood oxygen saturation ( The oxygen saturation percentage is calculated by analyzing the absorption ratio of red and infrared light. According to Beer-Lambert's law, oxyhemoglobin and deoxyhemoglobin have different absorption coefficients for different wavelengths of light. By measuring the ratio of the light intensity of red and infrared light after passing through tissue and using a calibration curve, the percentage of blood oxygen saturation is obtained. In this embodiment, the normal range for blood oxygen saturation is 95% to 100%. When the exercise intensity is too high, leading to insufficient aerobic metabolism, blood oxygen saturation will drop below 92%.

[0041] The physiological signals collected by the millimeter-wave radar unit 11 and the wrist PPG sensing unit 12 are transmitted to the edge computing device via Bluetooth Low Energy (BLE) protocol. The edge computing device is an embedded computing platform equipped with an ARM processor, deployed in the treadmill's control system. The edge computing device receives data streams from the two sensing units, performs timestamp synchronization and data format conversion, and integrates multi-dimensional physiological parameters into a unified data packet for subsequent module processing.

[0042] Through the above technical solution, the non-contact physiological signal acquisition module 1 achieves simultaneous multidimensional monitoring of the athlete's respiration, heart rate, HRV, and blood oxygen saturation. Compared with existing technologies that rely solely on a single heart rate indicator, the multidimensional physiological parameters provided by this module can more comprehensively reflect the athlete's metabolic state and exercise load level, laying a data foundation for subsequent metabolic state modeling.

[0043] As shown in Figure 3, the multidimensional metabolic state modeling module 2 is constructed based on a gated recurrent unit (GRU) network and is used to convert multidimensional physiological signals into real-time estimations of metabolic equivalent (MET) values. This module includes a feature preprocessing unit 21, a GRU network unit 22, and a MET value output unit 23.

[0044] The feature preprocessing unit 21 receives raw physiological data from the non-contact physiological signal acquisition module 1, including respiratory rate, respiratory amplitude, heart rate, HRV time-domain features (SDNN), and blood oxygen saturation. In addition, this unit also acquires current exercise power data from the treadmill control system. Exercise power is defined as the instantaneous power consumption determined by the treadmill's speed, incline, and the exerciser's weight, measured in watts, and calculated using the following formula:

[0045] ,

[0046] in: Power is measured in watts. The athlete's weight is expressed in kilograms. The acceleration due to gravity is taken as 9.8 meters per second squared. The linear velocity of the treadmill is measured in meters per second. The incline angle of the treadmill is expressed in radians. The value is the rolling friction coefficient, which is approximately 0.01 for treadmills.

[0047] Feature preprocessing unit 21 normalizes the raw data of the above six dimensions to eliminate dimensional differences between different physical quantities. Normalization uses the min-max standardization method, with the following formula:

[0048] ,

[0049] in: These are the normalized eigenvalues, ranging from 0 to 1; These are the original eigenvalues; and These represent the minimum and maximum values ​​of the feature in the training dataset, respectively. Specifically, the normalized intervals for respiratory rate are 10 to 40 breaths per minute, respiratory amplitude is 5 to 30 millimeters, heart rate is 60 to 200 breaths per minute, SDNN is 10 to 100 milliseconds, blood oxygen saturation is 85% to 100%, and exercise power is 50 to 500 watts.

[0050] The normalized six-dimensional feature vectors are arranged in chronological order to form a temporal feature matrix. In this embodiment, a sliding window mechanism is used, with each time window containing data from the most recent 30 seconds, and the window sliding step size is 1 second. Since the sampling frequency is once per second (for downsampling or averaging of physiological signals), each time window contains 30 time steps, and each time step corresponds to a six-dimensional feature vector, forming an input tensor with dimension [missing information]. .

[0051] GRU network unit 22 is the core innovative module of this invention, used for deep learning modeling of temporal feature matrices to capture the temporal dependencies of metabolic states. GRU is an improved recurrent neural network (RNN) structure that effectively solves the gradient vanishing and gradient explosion problems in traditional RNNs in long sequence modeling by introducing update gate and reset gate mechanisms, making it particularly suitable for processing physiological signal time-series data with long-term dependencies.

[0052] The mathematical model of GRU network unit 22 is as follows:

[0053] First, calculate the update gate. :

[0054] ,

[0055] in: For time step The update gate activation value ranges from 0 to 1, and the dimension is... ; It is the Sigmoid activation function. ; To update the gate weight matrix, the dimension is... ; For time step The hidden state vector, with dimension ; For time step The input feature vector has a dimension of In this embodiment ; This represents a vector concatenation operation, where the concatenated vector has dimensions of . ; To update the gate bias vector, the dimension is .

[0056] Then calculate the reset door. :

[0057] ,

[0058] in: For time step The reset gate activation value ranges from 0 to 1, and the dimension is... ; To reset the gate weight matrix, the dimension is... ; To reset the gate bias vector, the dimension is Other symbols have the same meaning as above.

[0059] Next, calculate the candidate hidden state. :

[0060] ,

[0061] in: For time step The candidate hidden states, with dimension . ; The hyperbolic tangent activation function is used. ; Let be the candidate hidden state weight matrix, with dimension . ; Represents element-wise multiplication (Hadamard product); Let be the candidate hidden state bias vector, with dimension . Other symbols have the same meaning as above.

[0062] Finally, calculate the hidden state at the current time step. :

[0063] ,

[0064] in: For time step The final hidden state, with dimension . Update Gate The weighting of new and old information is controlled, when When the value is close to 0, the current hidden state mainly retains historical information; when... When the value approaches 1, the current hidden state primarily adopts the new candidate state.

[0065] In this embodiment, the GRU network unit 22 adopts a two-layer stacked structure, with the first layer having a hidden state dimension. The hidden state dimension of the second layer The input sequence is processed by the first GRU layer, and the output hidden state sequence is used as the input for the second GRU layer. After the second GRU layer processes the entire time window (30 time steps), it takes the hidden state of the last time step. As a feature representation of the entire time window, this vector has a dimension of 32 and encodes the temporal features of the athlete's metabolic state in the most recent 30 seconds.

[0066] The GRU network was trained using supervised learning. Training data consisted of actual MET values ​​measured using laboratory gas metabolism analysis equipment (such as the Cosmed K5 portable metabolic analyzer) as labels. The training samples comprised combinations of different exercise intensities (resting, slow walking, brisk walking, jogging, and sprinting) and individual athlete characteristics (age, gender, weight, and exercise level), totaling 5000 samples. The mean squared error (MSE) was used as the loss function during training.

[0067] ,

[0068] in: The value of the loss function; In this embodiment, the batch size is set to 32. For the first The true MET value of each sample; For the first Predicted MET values ​​for each sample.

[0069] The optimization algorithm uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs. An early stopping mechanism is introduced during training: training stops when the loss function on the validation set fails to decrease for 10 consecutive epochs to prevent overfitting. Furthermore, Dropout regularization is employed, with Dropout layers with a dropout rate of 0.3 placed between GRU layers to further enhance the model's generalization ability.

[0070] MET value output unit 23 receives the feature vector output by GRU network unit 22. This is mapped to MET values ​​through a fully connected layer. The mathematical expression for a fully connected layer is:

[0071] ,

[0072] in: The output is a dimensionless metabolic equivalent value, representing a multiple of the resting metabolic rate. The output layer weight matrix has dimensions of . ; This represents the hidden state at the last time step of the second layer of the GRU network, with a dimension of 32. This is the output layer bias term, a scalar.

[0073] In this embodiment, the typical range of MET values ​​is 1 to 15. The MET value at rest is approximately 1, the MET value for slow walking (4 km / h) is approximately 3 to 4, the MET value for brisk walking (6 km / h) is approximately 5 to 6, the MET value for jogging (8 km / h) is approximately 8 to 9, and the MET value for jogging (above 12 km / h) can reach 12 to 15.

[0074] Through the above technical solution, the multidimensional metabolic state modeling module 2 achieves accurate real-time conversion from multidimensional physiological signals to metabolic equivalents. Compared with the static polynomial fitting method used in existing technologies, the time-series modeling based on deep learning in this module fully considers the dynamic evolution characteristics of metabolic state, and can capture the influence of factors such as previous fatigue accumulation and changes in respiratory efficiency on the current metabolic state. This keeps the average absolute error of MET value estimation within 0.3, with an error rate of less than 5%, which is significantly better than the 15% error rate of existing technologies.

[0075] As shown in Figure 4, the floating threshold generation module 3 is used to dynamically generate intensity adjustment trigger thresholds based on the athlete's real-time metabolic response characteristics and a preset target MET interval. This module includes a target interval setting unit 31, a deviation calculation unit 32, and an adaptive threshold generation unit 33.

[0076] The target interval setting unit 31 sets a reasonable target MET interval based on the athlete's training goals and individual characteristics. For aerobic endurance training, the target MET interval is typically set within a moderate intensity range, i.e., 50% to 70% of the athlete's maximum metabolic capacity. Maximum metabolic capacity is determined through a VO2 max test or an age-based estimation formula. In this embodiment, a modified version of the Karvonen formula is used to calculate the target MET interval:

[0077] ,

[0078] ,

[0079] in: The lower limit of the target MET interval; The upper limit of the target MET interval; This is the resting metabolic equivalent, typically taken as 1. The maximum metabolic equivalent is estimated using the age-based formula. calculate, The athlete's age is expressed in years.

[0080] For example, for a 30-year-old athlete, the estimated maximum metabolic equivalent is... Then the lower limit of the target MET interval is The upper limit of the target MET interval is .

[0081] The deviation calculation unit 32 continuously monitors the deviation between the current actual MET value and the target MET range. The deviation is defined. for:

[0082] ,

[0083] in: The deviation is dimensionless and ranges from 0 to positive infinity. When the MET value is within the target range, This indicates that the MET value deviates from the target range; The current actual MET value is output by the multidimensional metabolic state modeling module 2.

[0084] The deviation calculation unit 32 not only calculates the instantaneous deviation, but also counts the duration of continuous deviation. The MET value is defined as the time during which the MET value remains outside the target range. A sliding counter mechanism is used, checking the MET value once per second. If the current MET value deviates from the target range (…), the MET value is counted. If the MET value returns to the target range, the counter increments by 1; if the MET value returns to the target range ( If the value of the counter is zero, the counter is reset to zero. The value of the counter is the duration of the continuous deviation, in seconds.

[0085] Adaptive threshold generation unit 33 based on deviation and duration of continuous deviation The invention dynamically generates a trigger threshold for intensity adjustment. The innovation lies in that it does not use a fixed deviation duration threshold, but rather adaptively adjusts the trigger duration based on the magnitude of the deviation. Specifically, a trigger duration threshold is defined. for:

[0086] ,

[0087] in: The deviation duration threshold for triggering intensity adjustment, in seconds; The baseline deviation duration is set at 180 seconds (3 minutes), representing the tolerance duration for smaller deviations. The adjustment coefficient is set to 2.0 to control the rate at which the threshold decays with the degree of deviation. This represents the current deviation. The base of the natural logarithm is approximately 2.718.

[0088] This formula reflects the adaptive nature of the floating threshold: when the deviation is small, the trigger threshold is close to the baseline duration of 180 seconds, and the system has a high tolerance for slight deviations, avoiding frequent interventions; when the deviation is large, the trigger threshold drops rapidly, and the system can respond more quickly to significant deviations and adjust the intensity in a timely manner. For example, when When (deviating by 10%), Seconds; when When (deviating by 30%), Seconds; when When (deviating from 50%), Second.

[0089] The adaptive threshold generation unit 33 checks every second whether the adjustment trigger condition is met. When the deviation from the duration continues... Exceeding the current trigger threshold At that time, an intensity adjustment signal is generated and sent to the multi-channel feedback control module 4 to initiate the intensity adjustment process. The intensity adjustment signal includes the current deviation. Information such as deviation from the direction (too high or too low) and suggested adjustment amount.

[0090] Through the above technical solution, the floating threshold generation module 3 achieves adaptive optimization of the intensity adjustment triggering mechanism. Compared with the existing technology that uses a fixed duration threshold (such as a fixed 3 minutes), this module can dynamically adjust the intervention timing according to the exerciser's real-time metabolic response. This avoids frequent interventions caused by oversensitivity and ensures timely response when there is a significant deviation, making the training process smoother and more personalized.

[0091] As shown in Figure 5, after receiving the intensity adjustment signal from the floating threshold generation module 3, the multi-channel feedback control module 4 performs intensity control through a dual-channel collaborative method, including a tactile feedback submodule 41 and a slope adjustment submodule 42.

[0092] The haptic feedback submodule 41 includes a built-in haptic motor array in the smart running shoe, used to apply gradient tactile cues to the sole of the athlete's foot. In this embodiment, the haptic motor array consists of eight miniature linear resonant actuators (LRAs) distributed in the forefoot and heel areas of the insole. An LRA is a high-frequency vibration motor operating at approximately 175 Hz, capable of generating clear and perceptible haptic pulses while consuming less power than a traditional eccentric rotary motor.

[0093] The intensity of tactile feedback is divided into three levels, corresponding to different degrees of deviation:

[0094] When deviation At a distance between 0.1 and 0.2, Level 1 haptic feedback is triggered. Level 1 feedback uses a gentle, single, short pulse lasting 100 milliseconds with an amplitude of 30% of the motor's maximum amplitude. This feedback is relatively weak, providing only a slight reminder to the user and is suitable for minor deviations.

[0095] When deviation At a value between 0.2 and 0.4, secondary haptic feedback is triggered. This secondary feedback uses medium-intensity dual pulses, each lasting 150 milliseconds with a 100-millisecond interval, and an amplitude of 60% of the motor's maximum amplitude. This moderate feedback intensity is sufficient to elicit clear attention from the user and is suitable for moderate deviations.

[0096] When deviation When the value is greater than 0.4, level three haptic feedback is triggered. Level three feedback uses a strong three-pulse sequence, each pulse lasting 200 milliseconds, with an 80-millisecond interval, and an amplitude of 90% of the motor's maximum amplitude. This strong feedback ensures that the user immediately perceives the adjustment signal and is suitable for situations with significant deviations.

[0097] The directionality of tactile feedback is achieved through the activation patterns of the motor array. When the MET value is too low and the intensity needs to be increased, the motors in the forefoot area are activated, prompting the athlete to increase their cadence or stride length; when the MET value is too high and the intensity needs to be decreased, the motors in the heel area are activated, prompting the athlete to slow down their cadence or stride length. Through tactile stimulation of different parts of the sole, the athlete can intuitively understand the direction of intensity adjustment without the need for additional visual or auditory cues, achieving truly non-intrusive feedback.

[0098] The incline adjustment submodule 42 sends an incline adjustment command to the treadmill to perform active intensity compensation. Incline adjustment amount. Calculated based on the current deviation and direction of deviation:

[0099] ,

[0100] in: This represents the slope adjustment amount, expressed as a percentage. The slope adjustment gain coefficient is set to 5, meaning that for every 0.1 increase in deviation, the slope is adjusted by 0.5 percentage points. This represents the current deviation. This is a symbolic function that returns 1 if the expression within the parentheses is greater than 0, -1 if it is less than 0, and 0 if it is equal to 0. The midpoint value of the target MET interval is calculated using the following formula: .

[0101] For example, if the current MET value is too low ( ), deviation The slope adjustment amount is This means reducing the incline by 1.5 percentage points to lessen the exercise load and help the exerciser recover to the target MET range. Conversely, if the current MET value is too high, the incline is increased to increase the exercise load.

[0102] The incline adjustment command is sent to the treadmill's motor control system via the RS-485 serial communication protocol. Upon receiving the command, the treadmill control system drives the incline adjustment motor to smoothly adjust the running belt incline. The incline adjustment rate is limited to 0.5 percentage points per second to avoid discomfort or the risk of falls caused by excessively rapid incline changes.

[0103] The tactile feedback submodule 41 and the slope adjustment submodule 42 operate in a collaborative mode: tactile feedback, as the first response, is triggered immediately upon detecting deviation, drawing the exerciser's attention and prompting adjustment direction; slope adjustment, as the second response, is executed within 5 to 10 seconds after the tactile feedback is activated, objectively altering the exercise load to assist the exerciser in achieving the target intensity. This dual-channel collaborative control mechanism balances the exerciser's subjective autonomy with the system's objective assistance, making intensity control more efficient and user-friendly.

[0104] Furthermore, the multi-channel feedback control module 4 also features an adjustment frequency limit mechanism to prevent excessively frequent interventions from impacting the training experience. After an intensity adjustment is triggered, the system enters a cooldown period of 60 seconds. During this cooldown period, even if a deviation is detected again, a new adjustment will not be triggered immediately; instead, the system will wait until the cooldown period ends before determining whether an adjustment is necessary. This mechanism ensures that the adjustment interval is no less than one minute, avoiding excessively frequent interventions.

[0105] Through the above technical solutions, the multi-channel feedback control module 4 achieves precise, timely, and non-invasive intensity guidance for the exerciser. The gradient design and directional cues of the tactile feedback enhance the information content and comprehensibility of the feedback, while the active compensation of slope adjustment further improves the effectiveness of the control. The dual-channel coordination and frequency limiting mechanism ensure the continuity and comfort of the training process.

[0106] As shown in Figure 6, the edge-end anomaly filtering module 5 is used to detect and remove abnormal data points in real time at the data acquisition end, avoiding interference from unconventional physiological events on intensity regulation decisions. This module includes a sliding window buffer unit 51, a local outlier factor calculation unit 52, and an anomaly discrimination unit 53.

[0107] The sliding window buffer unit 51 maintains a fixed-length data buffer to store the time-series data of MET values ​​collected in the most recent period. In this embodiment, the buffer length is set to 20 data points, corresponding to a 20-second time window (assuming the MET value update frequency is once per second). The sliding window adopts a first-in-first-out (FIFO) queue structure. Whenever a new MET value arrives, it is added to the tail of the queue, while the oldest data at the head of the queue is removed, keeping the buffer length constant.

[0108] The Local Outlier Factor (LOF) calculation unit 52 calculates the outlier degree of each data point in the buffer based on the LOF algorithm. The LOF algorithm is a density-based unsupervised anomaly detection method that identifies outliers with significantly lower densities than their neighbors by comparing the local density of a data point with that of its neighborhood. Compared to global anomaly detection methods, the LOF algorithm can detect local anomalies, making it particularly suitable for processing unevenly distributed time-series data.

[0109] The calculation steps of the LOF algorithm are as follows:

[0110] First step, calculate the... Data points To its first Distance to nearest neighbors. Definition distance for To its first The Euclidean distance between the nearest neighbors. In this embodiment, Setting it to 5 means considering the 5 nearest neighbors for each data point. The Euclidean distance calculation formula is:

[0111] ,

[0112] in: For data points and The Euclidean distance between them; and These are two data points in the MET value sequence, both of which are scalars.

[0113] The second step is to calculate the... The reachable distance of each data point Defined as:

[0114] ,

[0115] in: For data points Compared to of Reachable distance; for of distance; for and The Euclidean distance between them. The reachability distance is introduced to reduce the impact of statistical fluctuations, when... distance When closer, use of Distance serves as the lower bound of reachable distance.

[0116] The third step is to calculate the... Local reachability density of data points :

[0117] ,

[0118] in: For data points Locally achievable density; for of Nearest neighbor set, containing neighbors with The closest One data point; for The number of elements in the nearest neighbor set is equal to In this embodiment, it is 5; the denominator is To all of them The average reachability distance of nearest neighbors, and the local reachability density is defined as the reciprocal of this average reachability distance, reflecting... The data density of the region.

[0119] Step 4, calculate the... Local outlier factors for each data point :

[0120] ,

[0121] in: For data points The local outlier factor is dimensionless; the molecule is All The average locally accessible density of the nearest neighbors reflects the average density level of the neighborhood; the denominator is Locally reachable density; LOF value represents The ratio of the density of a point to the average density of its neighborhood; when the LOF value is close to 1, it indicates... The density is comparable to that of the neighborhood, which is considered normal; when the LOF value is significantly greater than 1, it indicates... If the density is lower than that of the neighborhood, it may be an outlier.

[0122] The anomaly detection unit 53 determines whether the current data point is an outlier based on the calculated LOF value. A LOF threshold is set. When the LOF value of a data point exceeds this threshold, it is identified as an outlier and marked. Data points marked as outliers will not participate in subsequent MET value statistics and intensity adjustment decisions.

[0123] Specifically, when the floating threshold generation module 3 calculates the duration of continuous deviation, the time step corresponding to the anomaly point is skipped and not included in the cumulative deviation duration. After the multidimensional metabolic state modeling module 2 outputs the MET value, the edge anomaly filtering module 5 performs real-time anomaly detection on the MET value. If it is determined to be an anomaly, the MET value is replaced with the most recent non-anomaly MET value in the sliding window to ensure the continuity and rationality of the MET value sequence passed to subsequent modules.

[0124] To further enhance the robustness of anomaly detection, this embodiment also introduces a confidence fusion mechanism. In addition to the LOF algorithm, a statistically based 3σ criterion (three-standard-deviation criterion) is used for auxiliary judgment. The 3σ criterion states that, under normal distribution conditions, a data point deviating from the mean by more than three standard deviations is considered an anomaly if the probability is less than 0.3%. The mean of the MET values ​​within the sliding window is calculated. and standard deviation :

[0125] ,

[0126] ,

[0127] in: The mean of the MET values ​​within the sliding window; The standard deviation of the MET values ​​within the sliding window; The length of the sliding window is 20 in this embodiment; For the first in the window MET value.

[0128] If the current MET value satisfy If the LOF algorithm and the 3σ criterion both determine it as an anomaly, then it is marked as an anomaly. The final anomaly determination result adopts the AND logic of two methods: only when both the LOF algorithm and the 3σ criterion determine it as an anomaly is it finally marked as an anomaly. This conservative strategy effectively reduces the false positive rate.

[0129] Through the above technical solution, the edge-end anomaly filtering module 5 effectively identifies and removes abnormal data caused by unconventional physiological events such as coughing and sneezing. The combination of the LOF algorithm's local anomaly detection capability and the 3σ criterion's statistical anomaly detection capability achieves an anomaly detection accuracy of over 92%. More importantly, this module is deployed on edge computing devices and processes data in real time at the data acquisition end, avoiding the backpropagation of abnormal data and fundamentally eliminating the source of false positives. This reduces the false positive rate from 10% to 15% in existing technologies to less than 2%, significantly improving system reliability and user experience.

[0130] As shown in Figure 1, the overall workflow of the metabolic equivalent-driven endurance training intensity gradient generation system of the present invention is as follows:

[0131] First, the non-contact physiological signal acquisition module 1 continuously collects multi-dimensional physiological parameters such as respiratory rate, heart rate, HRV and blood oxygen saturation of the exerciser, and at the same time obtains exercise power data from the treadmill, and transmits the collected data to the edge computing device via Bluetooth.

[0132] Second, the edge-end anomaly filtering module 5 performs real-time anomaly detection on the collected raw data, removes outlier data points caused by abnormal events such as coughing and sneezing, and outputs a filtered and purified data stream.

[0133] Third, the multidimensional metabolic state modeling module 2 receives the purified multidimensional physiological data and exercise power data, performs time-series modeling through the GRU deep learning network, and outputs the current metabolic equivalent (MET) value to achieve accurate real-time assessment of the exerciser's metabolic state.

[0134] Fourth, the floating threshold generation module 3 compares the current MET value with the preset target MET range, calculates the deviation degree and the duration of continuous deviation, and dynamically generates an adaptive intensity adjustment trigger threshold based on the deviation degree. When the duration of continuous deviation exceeds the trigger threshold, an intensity adjustment signal is issued.

[0135] Fifth, after receiving the intensity adjustment signal, the multi-channel feedback control module 4 immediately activates the tactile feedback submodule, which applies gradient tactile cues to the soles of the athlete's feet through the tactile motor built into the smart running shoe; at the same time, it activates the incline adjustment submodule, which sends an incline adjustment command to the treadmill, and actively changes the exercise load to help the athlete return to the target intensity range.

[0136] Sixth, the system enters a 60-second cooldown period, during which no new intensity adjustments are triggered, waiting for the exerciser's metabolic state to respond to the adjustment. After the cooldown period, the system continues to monitor the MET value; if it still deviates from the target range, the above adjustment process is repeated.

[0137] Through the coordinated operation of the above five modules, the system of the present invention realizes a complete closed loop from multidimensional physiological signal acquisition, abnormal data filtering, accurate assessment of metabolic state, adaptive threshold generation to multi-channel coordinated regulation, ensuring accurate real-time control of exercise intensity during endurance training.

[0138] This invention also provides a method for generating endurance training intensity gradients driven by metabolic equivalents, as shown in Figure 7. This method is based on the above system and includes the following steps:

[0139] Step S1: The exerciser's respiratory rate and respiratory amplitude are collected by non-contact millimeter-wave radar, and heart rate, heart rate variability time-domain characteristics and blood oxygen saturation are collected by wrist PPG sensor. At the same time, the exercise power data of the treadmill is obtained to achieve synchronous collection of multi-dimensional physiological parameters.

[0140] Specifically, millimeter-wave radar monitors chest cavity displacement at a sampling frequency of 100 Hz, and extracts respiratory rate and amplitude after bandpass filtering and spectral analysis; a PPG sensor collects pulse wave signals at a sampling frequency of 50 Hz, and calculates heart rate, RR interval sequence, and HRV time-domain features (SDNN) after preprocessing and peak detection, while calculating blood oxygen saturation through dual-wavelength light absorption ratio; exercise power is calculated based on treadmill speed, incline, and exerciser weight. All data is transmitted to an edge computing device via Bluetooth for timestamp synchronization and format standardization.

[0141] Step S2: Perform edge anomaly filtering on the collected multidimensional physiological data. Based on the Local Outlier (LOF) algorithm and the 3σ statistical criterion, detect and remove outlier data points in real time.

[0142] Specifically, a sliding window buffer of length 20 is maintained to store the historical MET values ​​of the most recent 20 seconds; for each data point within the window, the local outlier factor of its 5 nearest neighbors is calculated, and the window mean and standard deviation are calculated, applying the 3σ criterion; when the LOF value of a data point is greater than 1.5 and deviates from the mean by more than three times the standard deviation, it is identified as an outlier; outliers do not participate in subsequent statistics and decisions, and are replaced by the most recent non-outlier value before the outlier.

[0143] Step S3: Input the anomaly-filtered multidimensional physiological data into the metabolic state modeling model based on the GRU network, and output the current metabolic equivalent (MET) value.

[0144] Specifically, the six-dimensional input features (respiratory rate, respiratory amplitude, heart rate, SDNN, blood oxygen saturation, and exercise power) are subjected to min-max normalization to construct a temporal feature matrix with 30 time steps. The feature matrix is ​​then input into a two-layer GRU network, with the first layer having a hidden state dimension of 64 and the second layer having a hidden state dimension of 32. Dropout regularization is used to prevent overfitting. The hidden state of the second GRU at the last time step is extracted and mapped to the MET value output through a fully connected layer.

[0145] Step S4: Calculate the deviation of the current MET value from the target MET interval and the duration of the deviation, and dynamically generate an adaptive intensity adjustment trigger threshold based on the deviation.

[0146] Specifically, the maximum metabolic equivalent (MET) is estimated based on the athlete's age, and the target MET range (50% to 70% intensity) is determined based on the Karvonen formula; the deviation of the current MET value from the target range is then calculated. A piecewise function is used, with the deviation being the normalized relative deviation; the duration of continuous deviation of the MET value from the target interval is statistically analyzed. According to the exponential decay formula Calculate the adaptive trigger threshold; the greater the deviation, the smaller the trigger threshold, and the faster the response.

[0147] Step S5: When the duration of continuous deviation exceeds the adaptive trigger threshold, multi-channel feedback control is activated. Gradient tactile cues are applied through the smart running shoe haptic motor, and an incline adjustment command is sent to the treadmill at the same time.

[0148] Specifically, the tactile feedback level is determined based on the degree of deviation: Level 1 feedback is a single short pulse (100 ms, 30% amplitude), Level 2 feedback is a double pulse (150 ms, 60% amplitude), and Level 3 feedback is a triple pulse (200 ms, 90% amplitude). The motor activation area is selected based on the direction of deviation: a low MET value activates the forefoot area to indicate acceleration, while a high MET value activates the heel area to indicate deceleration. Simultaneously, the slope adjustment amount is calculated. Send incline adjustment commands to the treadmill, with the adjustment rate limited to 0.5 percentage points per second.

[0149] Step S6: Enter the adjustment cool-down period (60 seconds). During the cool-down period, no new adjustments are triggered, and the exerciser's metabolic state is awaited. After the cool-down period ends, return to step S1 and continue monitoring and regulation.

[0150] Through the above-described methods and steps, this invention achieves precise dynamic control of exercise intensity during endurance training. This method fully integrates innovative technologies such as multidimensional physiological signals, deep learning temporal modeling, adaptive threshold generation, intelligent filtering of abnormal data, and multi-channel collaborative feedback, demonstrating significant technical advantages and practical value compared to existing technologies.

[0151] To verify the effectiveness of the system and method of this invention, a comparative experiment was conducted. Thirty volunteers (aged 25 to 45, half male and half female) were recruited. Each volunteer underwent 60 minutes of endurance training using either an existing technical solution (a polynomial regulation method based on a single heart rate, corresponding to patent CN 112023342 B) or the solution of this invention. Various indicators during the training process were recorded.

[0152] Experimental results show that:

[0153] First, regarding the accuracy of MET value estimation, the average absolute error of the proposed solution is 0.28, and the relative error is 4.2%, which are significantly better than the 0.89 and 14.3% of the prior art solutions. This verifies that multidimensional metabolic state modeling based on GRU networks has higher accuracy than static polynomial fitting.

[0154] Second, regarding the timeliness of intensity regulation, the solution of this invention has an average response delay of 18 seconds to changes in exercise intensity, while the existing technology has a delay of 45 seconds. This is due to the rapid response of respiratory amplitude features to intensity changes and the accurate capture of temporal dependencies by the deep learning model.

[0155] Third, regarding the misjudgment adjustment ratio, the misjudgment rate of the present invention is 1.8%, while that of the prior art is 12.5%. This proves that the edge-end LOF anomaly filtering algorithm is effective in identifying and eliminating irregular physiological events.

[0156] Fourth, in terms of user experience rating (out of 10), the average score of this invention is 8.6, while the existing technology solution scores 6.3. Volunteer feedback indicates that the haptic feedback of this invention is intuitive and effective, making the training process smoother and more comfortable.

[0157] Fifth, regarding the target MET interval retention rate (defined as the percentage of time the MET value remains within the target interval), the present invention achieves 82.4%, while the prior art solution achieves 61.7%. This fully demonstrates that the adaptive threshold generation and multi-channel collaborative regulation mechanism of the present invention can more effectively maintain the athlete's training within the ideal intensity range.

[0158] In summary, the experimental results fully verify the significant advantages of the metabolic equivalent-driven endurance training intensity gradient generation system and method of this invention in terms of accuracy, real-time performance, reliability, and user experience compared with existing technologies, demonstrating good practical value and promising prospects for promotion.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A metabolic equivalent-driven endurance training intensity gradient generation system, characterized in that, include: The non-contact physiological signal acquisition module is used to acquire the respiratory rate and respiratory amplitude of the athlete through millimeter-wave radar, and at the same time to acquire heart rate and blood oxygen saturation through a wrist photoplethysmography sensor. A multidimensional metabolic state modeling module is used to take the respiratory rate, respiratory amplitude, heart rate, time-domain features of heart rate variability, and exercise power as inputs based on a gated recurrent unit network, and output the current metabolic equivalent value; a floating threshold generation module includes a target interval setting unit, a deviation calculation unit, and an adaptive threshold generation unit; The target range setting unit estimates the maximum metabolic equivalent based on the athlete's age and determines a preset target metabolic equivalent range based on the heart rate reserve method. The lower limit is the resting metabolic equivalent plus 50% of the difference between the maximum metabolic equivalent and the resting metabolic equivalent, and the upper limit is the resting metabolic equivalent plus 70% of the difference. The deviation calculation unit calculates the deviation between the current actual metabolic equivalent value and the preset target metabolic equivalent range. When the current actual metabolic equivalent value is lower than the lower limit, the deviation is the difference between the lower limit and the current value divided by the lower limit. When the current actual metabolic equivalent value is higher than the upper limit, the deviation is the difference between the current value and the upper limit divided by the upper limit. The adaptive threshold generation unit dynamically calculates the trigger duration threshold according to the exponential decay formula. The trigger duration threshold is the baseline deviation duration of 180 seconds multiplied by the base of the natural logarithm raised to the power of -2.0 times the deviation. When the continuous deviation duration exceeds the trigger duration threshold, an intensity adjustment signal is generated. The multi-channel feedback control module includes a tactile feedback submodule and a slope adjustment submodule; After receiving the intensity adjustment signal, the tactile feedback submodule applies gradient tactile feedback to the sole of the athlete's foot through the tactile motor array built into the smart running shoe; The slope adjustment submodule calculates the slope adjustment amount based on the deviation degree and deviation direction, and sends the slope adjustment command to the treadmill via the RS-485 serial communication protocol; The edge-end anomaly filtering module includes a sliding window caching unit, a local outlier factor calculation unit, and an anomaly detection unit; The sliding window cache unit maintains a first-in-first-out queue with a length of 20 data points, storing the time-series data of metabolic equivalent values ​​for the most recent 20 seconds. The local outlier calculation unit calculates the local outlier factor of each data point in the sliding window cache unit for its 5 nearest neighbors, and determines the degree of outlier by calculating the ratio of the local reachability density of the data point to the average local reachability density of its neighborhood. The anomaly detection unit simultaneously adopts the local outlier criterion and the three-standard-deviation statistical criterion. When the local outlier factor value of a data point is greater than 1.5 and deviates from the window mean by more than three standard deviations, it is determined to be an outlier and replaced with the nearest non-outlier value.

2. The metabolic equivalent-driven endurance training intensity gradient generation system according to claim 1, characterized in that: The non-contact physiological signal acquisition module includes a millimeter-wave radar unit and a wrist photoplethysmography (PPG) sensor unit. The millimeter-wave radar unit uses a 60GHz frequency-modulated continuous wave radar to monitor the chest displacement of the exerciser at a sampling frequency of 100 Hz. It extracts the respiratory rate and respiratory amplitude through bandpass filtering and fast Fourier transform. The wrist PPG sensor unit uses a dual-wavelength LED light source and a photodetector to acquire pulse wave signals at a sampling frequency of 50 Hz. It calculates the heart rate through a peak detection algorithm, calculates the time-domain characteristics of heart rate variability through the standard deviation of continuous RR intervals, and calculates the blood oxygen saturation through the ratio of red light to infrared light absorption.

3. The metabolic equivalent-driven endurance training intensity gradient generation system according to claim 1, characterized in that: The multidimensional metabolic state modeling module includes a feature preprocessing unit, a gated recurrent unit network unit, and a metabolic equivalent value output unit. The feature preprocessing unit performs minimum-maximum normalization on the respiratory rate, respiratory amplitude, heart rate, temporal features of heart rate variability, blood oxygen saturation, and exercise power to construct a temporal feature matrix; The gated recurrent unit network unit adopts a two-layer stacked structure. The first layer has a hidden state dimension of 64, and the second layer has a hidden state dimension of 32. The temporal feature matrix is ​​modeled temporally through update gate, reset gate, and candidate hidden state mechanism. The metabolic equivalent value output unit maps the hidden state of the last time step of the second layer of the gated recurrent unit network unit to the current metabolic equivalent value through a fully connected layer.

4. The metabolic equivalent-driven endurance training intensity gradient generation system according to claim 1, characterized in that: In the multi-channel feedback control module, the tactile feedback submodule includes a tactile motor array composed of eight miniature linear resonant actuators, distributed in the forefoot and heel areas of the insole. The tactile feedback level is determined according to the deviation. Level 1 feedback is a single short pulse lasting 100 milliseconds with an amplitude of 30% of the maximum amplitude; Level 2 feedback is a double pulse lasting 150 milliseconds with an amplitude of 60% of the maximum amplitude; and Level 3 feedback is a triple pulse lasting 200 milliseconds with an amplitude of 90% of the maximum amplitude. In the slope adjustment submodule, the slope adjustment amount is equal to 5 times the deviation multiplied by the sign function of the current metabolic equivalent value and the midpoint value of the target metabolic equivalent interval.

5. The metabolic equivalent-driven endurance training intensity gradient generation system according to claim 4, characterized in that: The haptic feedback submodule provides directional cues through the activation mode of the motor array. When the current actual metabolic equivalent value is too low and the intensity needs to be increased, the motor in the forefoot area is activated; when the current actual metabolic equivalent value is too high and the intensity needs to be decreased, the motor in the heel area is activated. The slope adjustment submodule limits the slope adjustment rate to 0.5 percentage points per second. The multi-channel feedback control module has an adjustment frequency limiting mechanism. After an intensity adjustment is triggered, a cooldown period of 60 seconds is initiated, during which no new adjustments are triggered.

6. The metabolic equivalent-driven endurance training intensity gradient generation system according to claim 3, characterized in that: The training of the gated recurrent unit network unit adopts a supervised learning method, and the training data is measured by a gas metabolism analysis device to obtain the actual metabolic equivalent value as a label; The training process uses mean squared error as the loss function, and the Adam optimizer is used as the optimization algorithm. The learning rate is set to 0.001, the batch size is 32, and the training epochs are 100. An early stopping mechanism is introduced during training. Training stops when the loss function on the validation set does not decrease for 10 consecutive epochs. Dropout regularization with a dropout rate of 0.3 is used to prevent overfitting.

7. The metabolic equivalent-driven endurance training intensity gradient generation system according to claim 1, characterized in that: The exercise power is calculated based on the treadmill's speed, incline, and the exerciser's weight. The calculation formula is the sum of the exerciser's weight multiplied by the gravitational acceleration, the treadmill's linear velocity, the sine of the incline angle, and the rolling friction coefficient multiplied by the cosine of the incline angle. The heart rate variability time-domain characteristic is the standard deviation of the continuous RR interval, reflecting the heart's autonomic nervous system regulation function and metabolic stress state. The edge computing device is an embedded computing platform equipped with an ARM processor, deployed in the treadmill's control system, receiving data streams from the millimeter-wave radar unit and the wrist photoplethysmography (PPG) sensor unit, and performing timestamp synchronization and data format conversion.

8. A method for generating intensity gradients in endurance training driven by metabolic equivalents, characterized in that, The endurance training intensity gradient generation system based on the metabolic equivalent driven by any one of claims 1 to 7 Includes the following steps: S1: Collects the exerciser's respiratory rate and respiratory amplitude through non-contact millimeter-wave radar, and collects heart rate, heart rate variability time-domain characteristics and blood oxygen saturation through wrist photoplethysmography (PPG) sensor, while also acquiring the treadmill's exercise power data. S2: Perform edge-end anomaly filtering on the collected multidimensional physiological data, maintain a sliding window buffer with a length of 20 data points, calculate the local outlier factor of each data point's 5 nearest neighbors within the window, and simultaneously calculate the window mean and standard deviation. When the local outlier factor value of a data point is greater than 1.5 and deviates from the window mean by more than three times the standard deviation, it is identified as an outlier and replaced with the nearest non-outlier value; S3: Perform min-max normalization on the anomaly-filtered multidimensional physiological data, construct a time-series feature matrix containing 30 time steps, input the metabolic state modeling model based on a gated recurrent unit network, and output the current metabolic equivalent value; S4: Estimate the maximum metabolic equivalent based on the athlete's age, determine the preset target metabolic equivalent range based on the heart rate reserve method, calculate the deviation and duration of the current metabolic equivalent value from the preset target metabolic equivalent range, and dynamically generate an adaptive trigger duration threshold based on the exponential decay formula. The trigger duration threshold is the base deviation duration of 180 seconds multiplied by the base of the natural logarithm raised to the power of -2.0 times the deviation. S5: When the duration of the continuous deviation exceeds the trigger duration threshold, activate multi-channel feedback control, determine the tactile feedback level based on the deviation, apply gradient tactile cues through the tactile motor of the smart running shoe, calculate the incline adjustment amount based on the deviation and the direction of deviation, and send an incline adjustment command to the treadmill. S6: Enter a 60-second adjustment cooldown period. No new adjustments are triggered during the cooldown period. After the cooldown period ends, return to step S1 to continue monitoring and control.

Images