A non-invasive sugar metabolism monitoring method and system based on multi-source body surface signal fusion

Abstract

The present application relates to the field of blood glucose monitoring, and specifically discloses a non-invasive glucose monitoring method and system based on multi-source body surface signal fusion, which comprises the following steps: pre-acquiring multi-frequency impedance and dual-wavelength reflectance to construct blood perfusion equivalent index, applying low-frequency micro-current heat stimulation to regulate to an ideal window when deviating, collecting multi-wavelength signals to generate characteristic vectors after stabilization, extracting multiple indexes on the cloud to compensate and deduct background, purifying each skin layer characteristics, splicing characteristic vectors for metabolic inversion and lag phase estimation, generating preliminary calibrated blood glucose values, generating monitoring values by secondary calibration based on historical incremental deviation and LED aging compensation vector, and returning to the terminal side to adjust heat stimulation parameters to generate sampling strategies and dormancy thresholds; the system comprises functionally adapted modules. The present application realizes non-invasive precise monitoring of glucose metabolism based on multi-source body surface signal fusion, suppresses long-term drift and adapts to individual differences, and the error is stable within ±0.5mmol / L.

Description

Technical Field

[0001] This invention relates to the field of blood glucose monitoring technology, specifically to a non-invasive glucose monitoring method and system based on multi-source body surface signal fusion. Background Technology

[0002] Non-invasive glucose metabolism monitoring is an important direction for the long-term management of diabetes. Current technologies mainly have the following shortcomings: Existing blood glucose monitoring devices generally rely on invasive sampling data, and their algorithms depend entirely on blood samples. They cannot directly use skin surface signals to infer blood glucose levels, resulting in poor user compliance.

[0003] Existing non-invasive monitoring solutions mostly use a single signal dimension or a simple parallel connection of multiple sensors. They do not consider the dynamic modulation of the optical scattering baseline and impedance baseline by skin temperature and humidity as coupling state variables, nor do they consider the partial volume effect aliasing of multi-wavelength optics between skin layers. They treat multi-wavelength signals as simple splicing of data in the same dimension, resulting in the loss of layer-specific information, weak anti-interference ability, and low measurement accuracy.

[0004] Existing solutions use fixed algorithm models without dynamic correction, cannot adapt to individual differences, and suffer from error accumulation and severe drift over long-term use, resulting in poor individual adaptability.

[0005] The existing scheme does not take into account the tissue fluid as an intermediate layer between blood and body surface signals. Its glucose concentration has a gradient distribution and dynamic diffusion delay from the dermis to the stratum corneum. Direct mapping leads to systematic delays during dynamic tracking.

[0006] Existing solutions employ fixed sampling periods and fixed sleep strategies, resulting in continuous high-frequency sampling, high CPU load, and no sleep frequency reduction strategy, leading to high device power consumption and short battery life.

[0007] Existing solutions mostly rely on external terminals for model inference. Embedded terminals have poor independent computing capabilities, fail when the connection is lost, and have weak offline functionality and insufficient stability.

[0008] To address the aforementioned problems, this invention proposes a non-invasive method and system for monitoring glycemic metabolism based on the fusion of multi-source body surface signals. Through an edge-cloud collaborative architecture, it achieves accurate monitoring that is non-invasive, low-power, highly stable, and supported by high computing power. Summary of the Invention

[0009] The technical problem to be solved by this invention is that existing non-invasive glucose metabolism monitoring suffers from large skin interference, distortion due to multi-source signal aliasing, untrackable individual dynamic delays, and severe long-term drift, resulting in low accuracy and short battery life.

[0010] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A non-invasive method for monitoring glucose metabolism based on multi-source body surface signal fusion, specifically including the following steps: S1. Response monitoring triggers, pre-collects temperature, humidity, multi-frequency impedance and dual-wavelength reflectivity to construct the blood perfusion equivalent index, when it deviates, low-frequency microcurrent thermal stimulation is applied through impedance electrodes and the recovery curve is tracked until it falls into the ideal window, after stabilization, multi-wavelength signals are collected and pre-processed to generate a standardized feature vector and uploaded. S2. The cloud-based system establishes a physical state decoupling model based on standardized feature vectors, extracts multidimensional interference baselines for compensation and elimination and background absorption subtraction, and extracts metabolic thermal hysteresis and ion migration features; it purifies the pure optical features of each cortex based on the interlayer unmixing parameter matrix and splices them to generate a multi-source fusion feature vector; it performs cross-physical field metabolic inversion and hysteresis phase dynamic estimation, calculates the drift factor and state correction coefficient to generate a preliminary calibrated blood glucose value. S3. The cloud performs secondary calibration of the initial blood glucose value based on the user's historical incremental deviation mode and LED aging spectral drift compensation vector to generate glucose metabolism monitoring values ​​and send them back; the terminal adjusts the thermal stimulation parameters based on the recovery time change rate, generates a spatiotemporal sparse sampling strategy and a dormancy threshold based on the trend of glucose metabolism monitoring values, and enters deep dormancy by reducing the frequency and waiting to be woken up.

[0011] Based on the above technical solution, the present invention can be further improved as follows.

[0012] Furthermore, S1 specifically includes: S11. Response monitoring triggers, initiates low-resolution pre-scan, collects skin temperature, skin humidity, multi-frequency skin impedance and optical reflectivity signals of two characteristic wavelengths, and constructs real-time blood perfusion equivalent index based on the weighted fusion of impedance dual-frequency ratio and optical dual-wavelength ratio; S12. Compare the real-time blood perfusion equivalent index with the preset ideal perfusion equivalent window. If there is a deviation, apply low-frequency microcurrent thermal stimulation through the skin impedance electrode. After heating, enter the cooling period and continuously track the dynamic recovery curve of the blood perfusion equivalent index. S13. When multiple consecutive sampling points of the dynamic recovery curve fall into the ideal window of perfusion equivalent, the perfusion equivalent index is determined to be stable, the cooling period ends and the system seamlessly switches to high-resolution formal acquisition. Simultaneously and in parallel, optical reflectivity signals of six characteristic wavelengths, as well as temperature, humidity, and impedance signals, are acquired to generate a multi-source body surface signal dataset. S14. Perform joint preprocessing on the dataset, including moving average filtering, wavelet denoising, outlier removal, signal normalization, and baseline correction, to generate a standardized feature vector. The baseline correction uses a sliding window polynomial fitting to subtract the trend term. The sliding window duration is 20s to retain signal components in the preset physiological frequency band above 0.05Hz. Record the recovery time in the dynamic recovery curve and the cumulative runtime of each wavelength LED light source. Upload the standardized feature vector, recovery time, and cumulative runtime to the cloud server.

[0013] Furthermore, S12 specifically includes: S121. Compare the real-time blood perfusion equivalent index with a preset ideal perfusion equivalent window, wherein the ideal window includes an upper threshold and a lower threshold; S122. When the real-time blood perfusion equivalent index is lower than the lower threshold or higher than the upper threshold, it is determined that the deviation is from the ideal window. The embedded main control unit starts the low-frequency microcurrent thermal stimulation output. The power and duration of the thermal stimulation are dynamically adjusted according to the degree of deviation. The greater the degree of deviation, the higher the power and the longer the duration. S123. After the low-frequency microcurrent thermal stimulation ends, the cooling period begins. Skin impedance signals and dual-wavelength optical reflectivity signals are continuously acquired at a rate no lower than the sampling frequency of the pre-scanning stage. The blood perfusion equivalent index is recalculated, and a dynamic recovery curve is generated.

[0014] Furthermore, in S2, the cloud establishes a physical state decoupling model based on standardized feature vectors, extracts multidimensional interference baselines for compensation and elimination and background absorption subtraction, and extracts metabolic thermal hysteresis and ion migration features, specifically including: S211. A three-dimensional coupled interference field of skin stratum corneum glycosylation-hydration-impedance and a discrete step model of sweat electrolyte pulses are established in the cloud. The discrete step time of sweat electrolyte pulses of skin humidity signal, microvascular automatic fluctuation baseline of the power spectrum peak of skin temperature signal in preset frequency band, glycosylation interference baseline of low frequency drift component of skin impedance signal, and time-varying thickness of electrolyte film jointly estimated by impedance high frequency component and humidity transient response are extracted from the standardized feature vector. S212. The skin temperature signal is first-order differential and combined with the decay time constant after thermal stimulation and the steady-state temperature difference to estimate the skin-tissue fluid heat conduction hysteresis time constant as a metabolic heat hysteresis feature. Ion migration features are extracted by frequency domain decomposition of impedance signal. S213. Based on discrete step time, perform step response compensation on the impedance signal, based on the time-varying thickness of the electrolyte film, perform interface impedance normalization correction on the compensated impedance signal to eliminate sweat abrupt change interference, based on the microvascular automatic fluctuation baseline, perform microcirculation perfusion phase modulation elimination on the optical reflectivity signal, based on the glycosylation interference baseline, perform stratum corneum background absorption compensation on the optical reflectivity signal, and generate physical state decoupling feature vector.

[0015] Furthermore, in S2, the purification of pure optical features of each cortical layer based on the interlayer unmixing parameter matrix and the splicing of these features to generate a multi-source fusion feature vector specifically includes: S221. Based on the correlation between the differential penetration depth and multiple scattering path of six characteristic wavelengths in skin tissue, a model for interlayer aliasing deconvolution and memory effect decorrelation is established. S222. Input the optical reflectivity signal in the physical state decoupled feature vector into the interlayer aliasing deconvolution and memory effect decorrelation model, use the statistical correlation of wavelength memory effect to purify the interlayer signal, and extract the pure optical features of each cortex. S223. The pure optical features of each cortex are spliced ​​with the metabolic thermal hysteresis features, ion migration features, and microvascular automatic fluctuation baseline to form a multi-source fusion feature vector.

[0016] Furthermore, S222 specifically includes: S222a. Construct a multi-input multi-output blind deconvolution model framework, defining each wavelength measurement signal as the observation value after aliasing matrix transformation of different cortical source signals; S222b. By calculating the mutual information function between each wavelength channel, the statistical correlation of the memory effect between wavelengths is estimated, a decorrelation constraint term is established, and a complete objective function of the interlayer aliasing deconvolution and memory effect decorrelation model is formed. S222c. The nonnegative matrix factorization algorithm is used to iteratively solve the objective function of the interlayer aliasing deconvolution and memory effect decorrelation model to obtain the unmixing matrix; S222d. The six-channel measurement signal is linearly transformed using the demixing matrix to separate the waveform estimates of each cortical source signal and output the optical feature components of the epidermis, superficial dermis and deep dermis.

[0017] Furthermore, in S2, cross-physical field metabolic inversion and hysteresis phase dynamic estimation are performed, and the drift factor and state correction coefficient are calculated to generate a preliminary calibrated blood glucose value, specifically including: S231. Input the multi-source fusion feature vector into the pre-trained multivariate linear regression and random forest fusion model to output the initial blood glucose estimate; S232. The metabolic thermal hysteresis feature is used as the thermal diffusion state component in the state vector, the initial blood glucose estimate is used as the observation output, and the peak amplitude of the power spectral density of the microvascular automatic fluctuation baseline is mapped to the microcirculation perfusion rate through a preset perfusion rate conversion coefficient, which is used as the time-varying gain coefficient in the state transition equation. A noisy approximate observation output with tissue fluid glucose concentration as the hidden state and the initial blood glucose estimate as the tissue fluid glucose concentration is constructed. The blood glucose concentration is obtained by delaying the tissue fluid glucose concentration through a hysteresis time constant. S233. An augmented state Kalman filter algorithm is used to augment the lag time constant into the state vector and perform joint recursive estimation with the tissue fluid glucose concentration, dynamically outputting the tissue fluid glucose concentration estimation sequence and the individualized tissue fluid-blood glucose lag phase. S234. Read the tissue fluid glucose concentration estimation sequence, combine it with the individualized tissue fluid-blood glucose lag phase, and convert it into a blood glucose concentration estimation sequence through a preset lag compensation model. Then, combine the user's historical blood glucose data sequence and the current skin temperature and skin humidity, and use the exponential weighted moving average method to calculate the long-term physiological drift factor and the short-term state correction coefficient. Multiply the blood glucose concentration estimation sequence with the long-term physiological drift factor and the short-term state correction coefficient to generate a preliminary calibrated blood glucose value.

[0018] Furthermore, S3 specifically includes: S31. The cloud server calculates the LED aging spectral drift compensation vector based on the cumulative running time of each wavelength LED light source and the preset aging decay curve; and calculates the individualized incremental calibration factor based on the incremental deviation pattern of the user's historical data sequence. S32. Correct the deviation of the initial calibrated blood glucose value using an individualized incremental calibration factor; based on the regression coefficient matrix of the multivariate linear regression sub-model, decompose the corrected blood glucose value into an optical feature-dependent blood glucose estimation component and a non-optical feature-dependent estimation component; use the LED aging spectral drift compensation vector to restore the amplitude of the optical feature-dependent blood glucose estimation component, and then weight and fuse the restored optical feature-dependent estimation component and the non-optical feature-dependent estimation component to generate a glucose metabolism monitoring value; and then send the glucose metabolism monitoring value, hysteresis phase, and individualized incremental calibration factor back to the embedded main control unit. S33. The embedded main control unit updates the historical sliding window based on the recovery time, calculates the rate of change of the recovery time, and dynamically adjusts the low-frequency microcurrent thermal stimulation power and duration of the next cycle; based on the trend of changes in sugar metabolism monitoring values, hysteresis phase and communication quality, it generates a spatiotemporal sparse sampling strategy and a communication interruption fault-tolerant sleep threshold. S34. Set a wake-up timer according to the spatiotemporal sparse sampling strategy, turn off the power of the sensor and communication module, reduce the frequency to enter deep sleep, and wait for the next cycle to wake up.

[0019] Furthermore, S33 specifically includes: S331. The embedded main control unit stores the recovery time recorded in each acquisition cycle into a fixed-length historical sliding window. When the window is full, the latest value is used to overwrite the oldest value. The mean and standard deviation of the recovery time within the window are calculated, and the recovery time change rate is defined as the ratio of the deviation of the current recovery time from the mean to the standard deviation. S332. When the rate of change of recovery time is lower than the first threshold, it is determined that the blood flow response tends to be stable, and the power of the low-frequency microcurrent thermal stimulation in the next cycle is reduced and the duration is shortened by a preset step size; when the rate of change of recovery time is higher than the second threshold, it is determined that the blood flow response degenerates, and the thermal stimulation power is increased and the duration is extended by a preset step size; the first threshold is less than the second threshold, and a steady-state maintenance interval is defined between the two. S333. Analyze the time series of glucose metabolism monitoring values, calculate the absolute value of its first difference as a measure of change trend, generate a sparse sampling label when the change trend measure is lower than the stationary threshold, otherwise generate a dense sampling label. S334. Read the signal strength indication of the end-side communication module to assess data transmission reliability, and read the individualized tissue fluid-blood glucose hysteresis phase to assess metabolic dynamic alignment requirements; combine the data transmission reliability and metabolic dynamic alignment requirements to generate a spatiotemporal sparse sampling strategy and a communication interruption tolerance sleep threshold: when the communication quality is lower than the interruption tolerance threshold, the local caching strategy is preferentially enabled and the sleep duration is extended, while a communication interruption tolerance sleep threshold is generated, which defines the maximum number of continuous acquisition cycles that the device can maintain autonomously in the absence of network coverage; when the hysteresis phase exceeds the dynamic threshold, the dense sampling identifier is preferentially used; combine the above sparse sampling identifier or dense sampling identifier, communication interruption tolerance sleep threshold and local caching strategy to generate the final spatiotemporal sparse sampling strategy and sleep control command.

[0020] A non-invasive glycemic monitoring system based on multi-source surface signal fusion is disclosed to implement a non-invasive glycemic monitoring method based on multi-source surface signal fusion. The system includes: an ESP32 main controller, a temperature and humidity sensor, skin impedance electrodes, a multi-wavelength optical sensor, an LCD, a buzzer, a 4G module, and power management. The multi-wavelength optical sensor comprises six channels with center wavelengths of 660nm, 905nm, 1200nm, 1300nm, 1450nm, and 1550nm. Specifically: The end-side multi-source signal acquisition and preprocessing module is implemented by the ESP32 main controller scheduling temperature and humidity sensor, skin impedance electrode and multi-wavelength optical sensor. It is configured to respond to monitoring trigger, pre-acquire temperature, humidity, multi-frequency impedance and 660nm and 905nm dual-wavelength reflectance to construct blood perfusion equivalent index. When it deviates, low-frequency microcurrent thermal stimulation is applied through the skin impedance electrode and the recovery curve is tracked until it falls into the ideal window. After stabilization, the six wavelength signals are collected by the multi-wavelength optical sensor and jointly pre-processed to generate a standardized feature vector, which is then uploaded to the cloud server through the 4G module. The cloud-based physical state decoupling and interlayer purification module is configured to establish a physical state decoupling model based on standardized feature vectors, extract the discrete step moments of sweat, the baseline of microvascular automatic fluctuations, and the baseline of glycosylation interference for compensation and elimination and background absorption subtraction, and extract metabolic thermal hysteresis features and ion migration features to generate physical state decoupling feature vectors; and establish an interlayer signal purification model based on the interlayer unmixing parameter matrix to extract the pure optical features of each cortical layer and splice the multi-source fusion feature vectors. The cloud-based metabolic inversion and calibration module is configured to perform cross-physical field metabolic inversion and hysteresis phase dynamic estimation, calculate the drift factor and state correction coefficient to generate a preliminary calibration blood glucose value; based on the user's historical incremental deviation mode and LED aging spectral drift compensation vector, the preliminary calibration blood glucose value is recalibrated to generate a glucose metabolism monitoring value, which is transmitted back to the ESP32 main controller through the 4G module. The ESP32 main controller controls the LCD to display the glucose metabolism monitoring value and controls the buzzer to output a prompt signal. The edge-side strategy control and low-power management module, implemented by the ESP32 main controller in conjunction with power management, is configured to adjust the low-frequency microcurrent thermal stimulation parameters based on the recovery time change rate, generate a spatiotemporal sparse sampling strategy and a sleep threshold based on the trend of changes in glucose metabolism monitoring values, and shut down the power of the temperature and humidity sensor, skin impedance electrode, multi-wavelength optical sensor and 4G module through the power management, and reduce the frequency of the ESP32 main controller to enter deep sleep and wait to be woken up.

[0021] The beneficial effects of this invention are: 1. A blood perfusion equivalent index is constructed by pre-collecting dual-frequency impedance and dual-wavelength optical reflectivity. Low-frequency microcurrent thermal stimulation is used to regulate skin microcirculation and track the recovery curve to the ideal window to ensure blood flow steady state before collection. A three-dimensional coupling interference field is established to decouple sweat pulse step, microvascular automatic fluctuation and glycosylation interference, and eliminate the coupling modulation of multi-source signals by the skin microenvironment.

[0022] 2. Based on the six-wavelength differentiated penetration depth, an interlayer aliasing deconvolution and memory effect decorrelation model is established. The unmixing matrix is ​​solved iteratively by non-negative matrix decomposition to purify the six-channel optical signals between layers, separating the pure optical characteristics of the epidermis, superficial dermis and deep dermis. This breaks through the limitation of existing technologies that simply splice multi-wavelength signals, and preserves layer-specific physiological information.

[0023] 3. A state-space model with tissue fluid glucose concentration as the hidden state is constructed, and an augmented state Kalman filter is used to jointly estimate the individualized lag phase. Secondary calibration is performed by combining the historical incremental bias of a single user with LED aging spectral drift compensation to suppress long-term drift and adapt to individual differences, so that the error of glucose metabolism monitoring value is stabilized within ±0.5 mmol / L, and the light source aging mismatch is eliminated.

[0024] 4. The device dynamically adjusts thermal stimulation parameters based on the recovery time change rate. It generates a spatiotemporal sparse sampling strategy and sleep threshold based on the trend of changes in glucose metabolism monitoring values ​​and communication quality. During the metabolic steady period, it performs sparse sampling and quickly enters deep sleep, which significantly reduces the power consumption of the whole machine and extends the battery life of the device to more than 15 days, realizing a closed loop of end-to-cloud collaboration. Attached Figure Description

[0025] Figure 1This is a flowchart illustrating the method described in Embodiments 1-4 of the present invention; Figure 2 This is a framework diagram of the system described in Embodiment 5 of the present invention; Figure 3 This is a schematic diagram of the electronic device in this invention; Figure 4 This is a circuit schematic diagram of the ESP32 main control module in Embodiment 5 of the present invention; Figure 5 This is a schematic diagram of the ESP32 extended I / O and power control circuit in Embodiment 5 of the present invention; Figure 6 This is a schematic diagram of the temperature and humidity sensor circuit in Embodiment 5 of the present invention; Figure 7 This is a schematic diagram of the sensor daughterboard interface circuit in Embodiment 5 of the present invention; Figure 8 This is a schematic diagram of the LCD display module interface circuit in Embodiment 5 of the present invention; Figure 9 This is a schematic diagram of the buzzer drive circuit in Embodiment 5 of the present invention; Figure 10 This is a circuit schematic diagram of the power management module in Embodiment 5 of the present invention. Detailed Implementation

[0026] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention. Example 1

[0027] like Figure 1 As shown, this invention provides a non-invasive method for monitoring glucose metabolism based on multi-source surface signal fusion, specifically including the following steps: S1. Response monitoring triggers, pre-collects temperature, humidity, multi-frequency impedance and dual-wavelength reflectivity to construct the blood perfusion equivalent index, when it deviates, low-frequency microcurrent thermal stimulation is applied through impedance electrodes and the recovery curve is tracked until it falls into the ideal window, after stabilization, multi-wavelength signals are collected and pre-processed to generate a standardized feature vector and uploaded. S2. The cloud-based system establishes a physical state decoupling model based on standardized feature vectors, extracts multidimensional interference baselines for compensation and elimination and background absorption subtraction, and extracts metabolic thermal hysteresis and ion migration features; it purifies the pure optical features of each cortex based on the interlayer unmixing parameter matrix and splices them to generate a multi-source fusion feature vector; it performs cross-physical field metabolic inversion and hysteresis phase dynamic estimation, calculates the drift factor and state correction coefficient to generate a preliminary calibrated blood glucose value. S3. The cloud performs secondary calibration of the initial blood glucose value based on the user's historical incremental deviation mode and LED aging spectral drift compensation vector to generate glucose metabolism monitoring values ​​and send them back; the terminal adjusts the thermal stimulation parameters based on the recovery time change rate, generates a spatiotemporal sparse sampling strategy and a dormancy threshold based on the trend of glucose metabolism monitoring values, and enters deep dormancy by reducing the frequency and waiting to be woken up. Example 2

[0028] Preferably, in Embodiment 1, step S1 is specifically implemented through the following steps: S11. Low-resolution pre-scanning and blood perfusion equivalent index construction are as follows: Upon responding to the monitoring trigger command, the embedded main control unit wakes up from deep sleep, increases the CPU frequency to 80MHz, initializes the analog-to-digital converter and direct memory access channel, and starts the first sampling phase. This phase operates at a sampling frequency lower than the formal acquisition frequency, with skin temperature and skin humidity signals acquired at 10 sampling points per second. Multi-frequency skin impedance signals are measured by selecting at least 10 frequency points at logarithmic intervals in the range of 10Hz to 100kHz using a frequency sweep method, with a sampling rate of not less than 200kHz. Optical reflectivity signals at two characteristic wavelengths, 660nm and 905nm, are acquired at 100 sampling points per second for 5 seconds.

[0029] After acquisition, the embedded main control unit performs dual-frequency ratio calculation on the multi-frequency skin impedance signal, extracts the real part of the impedance at the 10kHz frequency point as the real part of the low-frequency impedance, extracts the real part of the impedance at the 100kHz frequency point as the real part of the high-frequency impedance, and calculates the ratio of the two as the impedance dual-frequency ratio; at the same time, it performs dual-wavelength ratio calculation on the optical reflectivity signals at two wavelengths of 660nm and 905nm, extracts the reflectivity at the 905nm wavelength and the reflectivity at the 660nm wavelength, and calculates the ratio of the two as the optical dual-wavelength ratio.

[0030] The impedance dual-frequency ratio and the optical dual-wavelength ratio are weighted and fused according to preset weights, with the impedance dual-frequency ratio accounting for 40% of the weight and the optical dual-wavelength ratio accounting for 60% of the weight. The weighted sum is then used to construct the real-time blood perfusion equivalent index.

[0031] S12. Judgment of perfusion equivalent deviation and thermal stimulation feedback control, specifically: S121. Comparison of the infusion equivalent index with the ideal window, specifically: The embedded main control unit reads the currently calculated real-time blood perfusion equivalent index value and compares it with a preset ideal perfusion equivalent window. This ideal window has a lower threshold and an upper threshold, with the lower threshold set to 0.6 and the upper threshold set to 1.4. If the index value is greater than or equal to the lower threshold and less than or equal to the upper threshold, it is determined that the current microcirculation perfusion state of the skin tissue is within the ideal window.

[0032] S122. Deviation Judgment and Dynamic Output of Thermal Stimulation, specifically: When the real-time blood perfusion equivalent index falls below the lower threshold or rises above the upper threshold, the embedded main control unit determines that the current microcirculation perfusion state of the skin tissue deviates from the ideal window and immediately initiates low-frequency microcurrent thermal stimulation output. This thermal stimulation is applied to the detection site through skin impedance electrodes, with a current frequency of 10kHz, falling within the category of low-frequency AC microcurrent. The power of the thermal stimulation is dynamically adjusted within the range of 0.1mW to 5mW according to the degree of deviation; the greater the deviation, the higher the power. The duration of the thermal stimulation is also dynamically adjusted within the range of 3s to 15s according to the degree of deviation; the greater the deviation, the longer the duration. The degree of deviation is defined as the absolute difference between the real-time blood perfusion equivalent index and the center value of the ideal window. The embedded main control unit determines the specific power and duration parameters based on this absolute difference using linear interpolation and outputs the corresponding pulse width modulation control signal to drive the constant current source circuit.

[0033] S123. Cooldown tracking and recovery curve generation, specifically: After the low-frequency microcurrent thermal stimulation ends, the embedded main control unit immediately enters the cooling period. During the cooling period, the embedded main control unit performs complex impedance measurement by selecting at least 10 frequency points at logarithmic intervals in the range of 10Hz to 100kHz using a frequency sweep method. It collects skin impedance signals at a sampling rate of not less than 200kHz and collects dual-wavelength optical reflectivity signals of 660nm and 905nm at 100 sampling points per second. The acquisition time is equal to the preset maximum cooling time, with a value range of 10s to 120s and a classic value of 60s.

[0034] After each sampling cycle of signal acquisition is completed, the embedded main control unit recalculates the blood perfusion equivalent index, records the index value and its corresponding sampling time, and arranges the blood perfusion equivalent index values ​​of multiple consecutive sampling times in chronological order to generate a dynamic recovery curve.

[0035] S13. High-resolution formal acquisition and generation of multi-source body surface signal dataset, specifically: During the cooling period, the embedded main control unit continuously monitors the dynamic recovery curve. When the blood perfusion equivalent index calculated at five consecutive sampling times falls within the preset ideal perfusion equivalent window, and the change in index value between two adjacent sampling times does not exceed 10% of the center value of the ideal window, the perfusion equivalent index is considered to have reached a stable state. At this point, the embedded main control unit immediately terminates the cooling period, stops the low-resolution sampling mode in the pre-scanning stage, and seamlessly switches to the high-resolution formal acquisition stage.

[0036] During the switching process, the embedded main control unit maintains the sensor power supply status and reconfigures the sampling parameters of the analog-to-digital converter through the direct memory access channel to synchronize the sampling clocks of all signals. In the formal acquisition phase, skin temperature, skin humidity, multi-frequency skin impedance, and optical reflectivity signals at six characteristic wavelengths are acquired synchronously and in parallel. Skin temperature and skin humidity signals are acquired at 10 sampling points per second. The multi-frequency skin impedance signal is measured using a frequency sweep method, selecting at least 10 frequency points at logarithmic intervals within the range of 10Hz to 100kHz for complex impedance measurement at a sampling rate of no less than 200kHz. The optical reflectivity signals at six characteristic wavelengths (660nm, 905nm, 1200nm, 1300nm, 1450nm, and 1550nm) are acquired wavelength-by-wave at 100 sampling points per second, with each wavelength acquired for 10 seconds, for a total acquisition time of 60 seconds. All signals are aligned with a unified timestamp to form a multi-source skin surface signal dataset.

[0037] S14. Joint preprocessing and standardized feature vector generation and uploading, specifically as follows: The embedded main control unit performs joint preprocessing on the multi-source body surface signal dataset, and generates standardized feature vectors after preprocessing: First, a moving average filter is performed. A sliding window with a length of 5 sampling points is selected. The arithmetic mean of the sampling points of each signal within the window is calculated. The average value is used as the filtered output at the center of the window. The entire signal is traversed sequentially to remove high-frequency random spikes.

[0038] Next, wavelet denoising is performed. The db4 wavelet basis function is used to decompose each signal into 5 levels to obtain high-frequency detail coefficients and low-frequency approximation coefficients. The median of the absolute value of each high-frequency detail coefficient is calculated and divided by 0.6745 to estimate the noise standard deviation of that level. The noise standard deviation is multiplied by a correction factor of 2.0 to obtain a soft threshold. If the soft threshold is lower than 0.02, it is forced to be 0.02; if it is higher than 0.15, it is forced to be 0.15. The soft threshold function is applied to the high-frequency detail coefficients to shrink the coefficients. The coefficients with absolute values ​​less than the soft threshold are set to zero, and the coefficients with absolute values ​​greater than the soft threshold are retained after subtracting the soft threshold. After reconstruction, motion artifacts and ambient light interference are removed.

[0039] Next, outlier removal is performed. The mean and standard deviation of each signal are calculated within a sliding window of 30 sampling points. Sampling points that exceed the mean by plus or minus 3 times the standard deviation are identified as outliers and replaced with the mean of the two adjacent sampling points before and after the outlier.

[0040] Then, signal normalization is performed using the min-max normalization method. Each signal is subtracted from its minimum value and then divided by the difference between its maximum and minimum values, linearly mapped to the numerical range of 0 to 1, thus eliminating dimensional differences.

[0041] Finally, baseline correction is performed using a sliding window baseline fitting strategy. The high-pass cutoff frequency is configured to be 0.05Hz, and the sliding window duration is 20s. Polynomial fitting is performed on the signal within the window to obtain the baseline trend. The baseline trend is subtracted from the original signal, and the signal components of the preset physiological frequency band above 0.05Hz are retained.

[0042] Simultaneously, the embedded main control unit records the time elapsed from the end of thermal stimulation to the point where the perfusion equivalent index falls into the ideal window in the dynamic recovery curve as the recovery time, and also records the cumulative runtime of LED light sources at wavelengths of 660nm, 905nm, 1200nm, 1300nm, 1450nm, and 1550nm since the device was activated. Finally, the standardized feature vector, recovery time, and cumulative runtime are uploaded to the cloud server via an AES-256 encrypted link. Example 3

[0043] Preferably, in Embodiment 1, step S2 is specifically implemented through the following steps: S21. Skin physical state decoupling and multidimensional interference compensation, specifically: S211. Three-dimensional coupled interference field modeling and multi-dimensional baseline extraction, specifically: A three-dimensional coupled interference field model of skin stratum corneum glycosylation-hydration-impedance and sweat electrolyte pulse discrete step model was established on a cloud server. The three-dimensional coupled interference field consists of three orthogonal dimensions. The first dimension is the stratum corneum glycation interference dimension, which uses the low-frequency drift component of the skin impedance signal to characterize the accumulation degree of stratum corneum glycation end products. This low-frequency drift component is obtained by low-pass filtering the skin impedance signal with a cutoff frequency of 0.1 Hz. The second dimension is the skin hydration state dimension, which uses the steady-state component of the skin humidity signal to characterize the stratum corneum water content. This steady-state component is obtained by moving average of the skin humidity signal. The third dimension is the interface impedance dimension, which uses the high-frequency component of the skin impedance signal and the transient response of the skin humidity signal to jointly estimate the time-varying thickness of the electrolyte film. The high-frequency component is obtained by high-pass filtering the skin impedance signal with a cutoff frequency of 10 Hz. The glycation interference baseline, the steady-state component of the hydration state, and the time-varying thickness of the electrolyte film are constructed into a three-dimensional state vector. The covariance matrix of this vector is calculated to characterize the coupling strength between each dimension. The main interference direction is determined based on the eigenvalue decomposition of the covariance matrix, and the three-dimensional coupled interference field is established.

[0044] The sweat electrolyte pulse discrete step model consists of a differential detection module and a peak marking module. The differential detection module performs a first-order differential operation on the skin humidity signal, and the peak marking module detects the peak points in the differential result and marks them as the discrete step moments of the sweat electrolyte pulse secretion event.

[0045] Based on the above model structure and its functions, the discrete step moments of sweat electrolyte pulses are extracted from the skin humidity signal in the standardized feature vector; the power spectrum peaks within a preset frequency band are extracted as the baseline for microvascular automaticity fluctuations by performing spectral analysis on the skin temperature signal, where the preset frequency band is 0.1Hz to 0.5Hz, which covers the typical frequency range of skin microvascular automaticity fluctuations; the stratum corneum glycosylation interference baseline is extracted from the low-frequency drift component of the skin impedance signal; and the time-varying thickness of the electrolyte film is estimated by jointly using the high-frequency component of the skin impedance signal and the transient response of the skin humidity signal.

[0046] S212. Extraction of metabolic thermal hysteresis and ion migration features, specifically: The cloud server performs first-order difference operations on the skin temperature signal in the standardized feature vector to calculate the rate of temperature change between adjacent sampling points, thereby identifying the starting point of the decay segment after the thermal stimulation ends. The decay segment is then fitted with an exponential decay curve using least-squares fitting to obtain the measured thermal decay time constant. This measured thermal decay time constant is then weighted and fused with the baseline thermal conduction time constant to serve as a metabolic thermal hysteresis feature. The baseline thermal conduction time constant is equal to the square of the skin surface feature thickness divided by the thermal diffusivity. The skin surface feature thickness ranges from 0.1 mm to 0.5 mm, with a classic value of 0.2 mm, and the thermal diffusivity ranges from 0.5 × 10⁻⁶. -7 m 2 / s to 2.0×10 -7 m 2 / s, classic value is 1.0×10 -7 m 2 / s; the weighting coefficient of the measured thermal decay time constant ranges from 0.60 to 0.80, and the classical value is 0.70. The weighting coefficient of the basic thermal conduction time constant ranges from 0.20 to 0.40, and the classical value is 0.30. The sum of the two weighting coefficients is 1.0.

[0047] Frequency domain decomposition of skin impedance signals was performed. Based on complex impedance spectrum data obtained by frequency sweeping, multiple frequency points were selected in the range of 10Hz to 100kHz. The real and imaginary parts of impedance at each frequency point were extracted. The deviation of the imaginary part of impedance at each frequency point from the reference imaginary part of impedance was then weighted and multiplied by preset ion mobility and impedance imaginary part response coefficient to obtain the extracellular fluid ion migration characteristics. The reference imaginary part of impedance is the average value of the imaginary part of impedance at each frequency point under historical stable conditions, or it can be obtained by statistically measuring the impedance spectrum during periods without sweat interference. The weighting coefficients corresponding to the weighted sum of each frequency point range from 0.05 to 0.25, and the sum of the weighting coefficients of all frequency points is 1.0. The mid-frequency range has a higher weight, while the low-frequency and high-frequency ranges have a lower weight. The preset ion mobility ranges from 0.5 to 2.0, with a classical value of 1.0, and the impedance imaginary part response coefficient ranges from 0.1 to 5.0, with a classical value of 1.0.

[0048] S213. Multidimensional disturbance compensation and physical state decoupling vector generation, specifically: Based on the extraction results of the three-dimensional coupled interference field and the discrete step model of sweat electrolyte pulses, the cloud server determines the current dominant interference type according to the main interference direction, and performs the following compensation and elimination in order of priority of dominant interference type: Based on the discrete step time output by the sweat electrolyte pulse discrete step model, step response compensation is performed on the skin impedance signal. Independent compensation baselines are established before and after the discrete step time to eliminate impedance abrupt interference caused by sweat pulse secretion.

[0049] Based on the time-varying thickness of the electrolyte film output in the third dimension of the three-dimensional coupled interference field, the compensated skin impedance signal is normalized and corrected by interface impedance normalization. This proportionally subtracts the impedance drift caused by changes in the electrolyte film thickness. Specifically, the interface impedance normalization correction uses a preset electrolyte film thickness-interface impedance mapping function, obtained through offline calibration experiments, and has the form Z... norm = Z comp / (1 + k·Δd), where Z norm For the normalized impedance, Z comp To compensate for the impedance, Δd is the deviation of the time-varying thickness of the electrolyte film relative to the reference thickness, and k is the thickness sensitivity coefficient, ranging from 0.01 to 0.1 μm. -1 Based on the stratum corneum glycosylation interference baseline output from the first dimension of the three-dimensional coupled interference field, stratum corneum background absorption compensation is performed on the optical reflectivity signal to subtract the optical background absorption caused by the accumulation of stratum corneum glycosylation end products from the original signal. Based on the microvascular automaticity fluctuation baseline, microcirculation perfusion phase modulation elimination is performed on the optical reflectivity signal to subtract the periodic optical scattering baseline drift caused by microvascular automaticity fluctuations from the original signal.

[0050] After the above compensation and background absorption subtraction, the processed temperature signal, humidity signal, impedance signal and optical reflectivity signal are recombined with metabolic thermal hysteresis feature and ion migration feature to generate physical state decoupling feature vector.

[0051] S22. Inter-layer aliasing deconvolution and multi-source fusion feature vector construction are as follows: S221. Based on the differential penetration depth and multiple scattering path correlation of six characteristic wavelengths in skin tissue, the cloud server establishes an interlayer aliasing deconvolution and memory effect decorrelation model. This model consists of four functional substructures: aliasing matrix transformation module, memory effect statistical correlation estimation module, decorrelation constraint module, and iterative solution module.

[0052] S222. Interlayer signal purification and extraction of pure optical features of each cortical layer, specifically including: S222a. The aliasing matrix transformation module is a multi-input multi-output blind deconvolution model framework. In this framework, the input is the optical reflectivity measurement signal with six characteristic wavelengths, and the output is three cortical source signals, namely the epidermal source signal, the superficial dermal source signal, and the deep dermal source signal. The core of the framework is the aliasing matrix, which describes the mapping relationship between the three cortical source signals after linear transformation to generate six wavelength observations.

[0053] Each cortical source signal corresponds to a specific optical scattering or absorption characteristic. Epidermal source signals mainly contain scattering responses at a wavelength of 660 nm, superficial dermal source signals contain tissue fluid scattering responses at wavelengths of 905 nm and 1200 nm, and deep dermal source signals contain water absorption responses at wavelengths of 1300 nm, 1450 nm, and 1550 nm. Each row of the aliasing matrix corresponds to the generation equation of the observation value for a characteristic wavelength, and each column corresponds to the contribution weight of a cortical source signal to the observation value at that wavelength. Through this framework, the measured signals at each wavelength are defined as the observation values ​​after transformation by the aliasing matrix of different cortical source signals, and a forward aliasing model from cortical source signals to wavelength observation values ​​is established.

[0054] S222b. The memory effect statistical correlation estimation module estimates the statistical correlation of the memory effect between wavelengths by calculating the mutual information function between each wavelength channel. The calculation of the mutual information function is based on the probability distribution of the measurement signal at each wavelength. By estimating the ratio of the joint probability density of the two wavelength channels to their respective marginal probability densities, the statistical dependence between the two channels is quantified.

[0055] Based on this statistical correlation, a decorrelation constraint term is established. This constraint term requires minimizing the mutual information among the demixed cortical source signals, i.e., achieving statistical independence. The decorrelation constraint term is combined with the reconstruction error term of the blind deconvolution model. The reconstruction error term is the mean square error between the observed value and the reconstructed value of the aliasing matrix, forming the complete objective function of the interlayer aliasing deconvolution and memory effect decorrelation model.

[0056] S222c. The decorrelation constraint module uses a non-negative matrix factorization algorithm to iteratively solve the objective function of the interlayer aliasing deconvolution and memory effect decorrelation model. The non-negative matrix factorization algorithm decomposes the aliasing matrix into the product of two non-negative matrices. One matrix is ​​the basis matrix, which represents the intrinsic spectral mode of the cortical source signal, and the other matrix is ​​the coefficient matrix, which represents the mixing ratio of each source signal in the observed signal.

[0057] During the iteration process, the algorithm alternately updates the basis matrix and the coefficient matrix. The update rule is based on the multiplication update rule to ensure that the result of each iteration is non-negative. The iteration terminates when the rate of change of the reconstruction error term is lower than the preset convergence threshold or the maximum number of iterations is reached. The preset convergence threshold is 0.0001 and the maximum number of iterations is 1000. After the iteration converges, the unmixing matrix is ​​extracted from the coefficient matrix. This unmixing matrix is ​​the generalized inverse transform of the aliasing matrix and is used to map the observed signal back to the cortical source signal space.

[0058] The S222d iterative solution module uses the unmixing matrix to perform a linear transformation on the six-channel measurement signals. This linear transformation is achieved through matrix multiplication, multiplying the six-dimensional observation signal vector with the unmixing matrix to generate a three-dimensional cortical source signal vector. During the transformation, each row of the unmixing matrix corresponds to an extraction filter for the cortical source signal, and the observation signals of each wavelength are weighted and combined.

[0059] After linear transformation, waveform estimates of each dermal source signal are separated. The first output is the epidermal optical characteristic component, mainly reflecting the scattering characteristics of hemoglobin in the stratum corneum; the second output is the superficial dermal optical characteristic component, mainly reflecting the scattering characteristics of tissue fluid beneath the epidermis; and the third output is the deep dermal optical characteristic component, mainly reflecting the water absorption characteristics of the dermis. The waveform estimates of the three dermal source signals are used as the pure optical characteristic outputs of each dermal layer.

[0060] S223. The pure optical features of each cortex are spliced ​​together with the metabolic thermal hysteresis features, ion migration features, and microvascular automatic fluctuation baselines to form a multi-source fusion feature vector.

[0061] S23. Cross-physical field metabolic inversion and preliminary calibration of blood glucose value generation, specifically: S231. Generation of multi-source fusion feature vectors and initial blood glucose estimates, specifically: The cloud server inputs the multi-source fused feature vectors into a pre-trained multivariate linear regression and random forest fusion model. This fusion model consists of three functional structures: a linear regression sub-model, a random forest sub-model, and a fusion decision layer. The linear regression sub-model receives multi-source fused feature vectors and establishes a global linear mapping relationship between the pure optical features, metabolic thermal hysteresis features, ion migration features, and blood glucose concentration of each cortex through a preset regression coefficient matrix, outputting an estimated linear trend value. The value range of each element in the preset regression coefficient matrix varies according to the feature type and the strength of physiological correlation. Elements corresponding to feature components positively correlated with blood glucose concentration take positive values, and elements corresponding to feature components negatively correlated with blood glucose concentration take negative values. The absolute value of each element ranges from 0.01 to 2.00. The classical values ​​of the weight elements corresponding to the pure optical features of each cortex are 0.10 to 0.50, the classical values ​​of the weight elements corresponding to the metabolic thermal hysteresis features are 0.05 to 0.20, the classical values ​​of the weight elements corresponding to the ion migration features are 0.10 to 0.30, and the classical values ​​of the weight elements corresponding to the baseline of microvascular automaticity fluctuations are 0.01 to 0.10.

[0062] The random forest sub-model consists of twenty decision trees. Each decision tree makes hierarchical branching judgments based on different random subsets of feature vectors. The optimal splitting node is determined by minimizing the Gini coefficient, capturing nonlinear interactions and local anomaly patterns between features, and outputting nonlinear correction estimates.

[0063] The fusion decision layer weights and fuses linear trend estimates with nonlinear corrected estimates, with the linear regression sub-model accounting for 40% of the weight and the random forest sub-model accounting for 60%. The initial blood glucose estimate is then output after weighted summation. This fusion structure balances the interpretability of linear trends with the adaptability of nonlinear patterns.

[0064] S232. State-space model construction and microcirculation perfusion rate mapping, specifically: The cloud server uses the metabolic thermal hysteresis feature as the thermal diffusion state component in the state vector to characterize the influence of the skin tissue thermal diffusion process on the dynamics of blood glucose metabolism. The initial blood glucose estimate is used as the observation output and directly input into the observation equation of the state space model. The peak amplitude of the power spectral density of the microvascular automaticity fluctuation baseline is mapped to the microcirculation perfusion rate through a preset perfusion rate conversion coefficient. The value of this conversion coefficient ranges from 0.5 to 2.0, with a classic value of 1.0. The mapped microcirculation perfusion rate is used as the time-varying gain coefficient in the state transition equation to reflect the modulation effect of microvascular automaticity fluctuation on the glucose diffusion rate of tissue fluid.

[0065] A state-space model is constructed, where the hidden state is defined as the glucose concentration of the tissue fluid, and the observed output nominally corresponds to the glucose concentration of the blood, but in reality, it is characterized by the noisy approximate observation of the glucose concentration of the tissue fluid using the initial blood glucose estimate. The state transition equation describes the dynamic evolution of the glucose concentration of the tissue fluid under the driving force of the thermal diffusion state component and modulated by the time-varying gain coefficient. The observation equation describes the noisy approximate observation of the glucose concentration of the tissue fluid using the initial blood glucose estimate. Since the non-invasive signal on the body surface directly reflects the glucose concentration of the tissue fluid, the observation matrix H=[1,0] directly extracts the state component of the glucose concentration of the tissue fluid.

[0066] S233. Augmented-state Kalman filtering and dynamic estimation of hysteresis phase, specifically: The cloud server uses the augmented state Kalman filter algorithm to recursively estimate the state space model, adding the lag time constant as an augmented state component to the original state vector, thus expanding the state vector from a single tissue fluid glucose concentration to an augmented state vector containing both tissue fluid glucose concentration and lag time constant.

[0067] At each sampling time, the algorithm sequentially executes a prediction step and an update step: the prediction step, based on the state transition equation, thermal diffusion state components, and time-varying gain coefficients, predicts the prior estimates of the tissue fluid glucose concentration and lag time constant for the next time step; the update step calculates the innovation sequence based on the observation equation and the initial blood glucose estimate, calculates the Kalman gain using the innovation covariance matrix and the prior estimate covariance matrix, and uses the Kalman gain to correct the prior estimation bias to obtain the posterior estimate. Through continuous recursive estimation, the algorithm dynamically outputs the tissue fluid glucose concentration estimation sequence and the individualized tissue fluid-blood glucose lag phase, which reflects the response delay time of the current user's tissue fluid glucose relative to blood glucose. Specifically, the state transition equation is: x k+1 = F k ·x k + G k ·u k + w k , where x k = [G t (k), τ(k)] T For the augmented state vector, G t (k) is the glucose concentration in the tissue fluid, τ(k) is the lag time constant, and F k Let G be the state transition matrix. k Let u be the time-varying gain coefficient matrix. k For the thermal diffusion state component, w k For process noise; the observation equation is: y k = H·x k + v k , where yk Here are the initial blood glucose estimates, H = [1, 0] is the observation matrix, and v k For observation noise; assume process noise w k With observation noise v k Both are zero-mean Gaussian white noise, and their covariance matrices are Q. k With R k Q k This is a 2×2 diagonal matrix, where the diagonal elements correspond to the process noise variances of tissue fluid glucose concentration and lag time constant, respectively, with values ​​ranging from 0.01–1.0 and 0.001–0.1; R k The observation noise variance is taken in the range of 0.1–5.0; the above covariance matrix is ​​obtained offline by statistical variance analysis of the state prediction residuals and observation residuals in the historical training data.

[0068] S234. Drift factor calculation and preliminary calibration of blood glucose value generation, specifically: The cloud server reads the tissue fluid glucose concentration estimation sequence and the individualized tissue fluid-blood glucose lag phase output by the augmented state Kalman filter algorithm. It then converts the tissue fluid glucose concentration estimation sequence into a blood glucose concentration estimation sequence using a preset lag compensation model. This model establishes a first-order inertial delay mapping relationship based on the individualized tissue fluid-blood glucose lag phase, using the lag time constant as the time delay parameter. The tissue fluid glucose concentration estimation sequence is then shifted backward along the time axis and smoothed to obtain the blood glucose concentration estimation sequence. Combining the user's stored historical blood glucose data sequence with the currently received skin temperature and skin humidity data, an exponentially weighted moving average method is used to calculate the long-term physiological drift factor and short-term state correction coefficient. The forgetting factor in the exponentially weighted moving average method ranges from 0.8 to 0.99, with a classic value of 0.95. Recent historical data is given higher weights, while the weights of older historical data decay exponentially, making the calculation results closer to the current physiological state.

[0069] The long-term physiological drift factor reflects the long-term trend deviation of the user's blood glucose metabolism and is determined by the weighted mean deviation between historical blood glucose data sequences and initial blood glucose estimates. The short-term state correction factor reflects the short-term measurement deviation caused by current skin temperature and humidity and is determined by the weighted difference between current skin condition and historical average skin condition. The blood glucose concentration estimate sequence is multiplied by the long-term physiological drift factor and the short-term state correction factor to generate a preliminary calibrated blood glucose value. Example 4

[0070] Preferably, in Embodiment 1, step S3 is implemented through the following steps: S31. Calculation of LED aging compensation vector and incremental calibration factor, specifically as follows: The cloud server reads the cumulative runtime of each wavelength LED light source. This runtime is uploaded by the embedded main control unit after each acquisition cycle and stored in the cloud database sequentially. Based on a preset aging decay curve, an LED aging spectral drift compensation vector is calculated. The preset aging decay curve is an exponential decay function of the light power output of each wavelength LED light source with cumulative runtime. The function form is that the current light power equals the exponential power of the product of the initial light power multiplied by the negative decay coefficient of the natural constant and the cumulative runtime. The decay coefficients for each wavelength are independently preset according to the rated life and light decay characteristics of the LED: 0.0001 per hour for 660nm, 0.00008 per hour for 905nm, 0.00012 per hour for 1200nm, 0.00015 per hour for 1300nm, 0.00018 per hour for 1450nm, and 0.0002 per hour for 1550nm. Substitute the cumulative operating time of each wavelength into the corresponding attenuation function to calculate the attenuation ratio of the current optical power relative to the initial optical power. The attenuation ratios of the six wavelengths are arranged and combined in wavelength order to form a six-dimensional LED aging spectral drift compensation vector.

[0071] Meanwhile, the cloud server reads the user's stored historical data sequence, extracts the deviation sequence between the preliminary calibrated blood glucose value and the corresponding reference true value in the historical data sequence, performs a first-order difference operation on the deviation sequence to obtain the deviation change rate sequence, and uses a sliding window with a length of 10 data points to average the deviation change rate sequence to calculate the increment of the current deviation level relative to the historical average deviation level. This increment serves as an individualized incremental calibration factor, reflecting the degree of systematic drift of the user's current physiological state relative to the historical benchmark.

[0072] S32. Secondary calibration and monitoring value feedback, specifically: The cloud server uses a personalized incremental calibration factor to correct the deviation of the initial calibrated blood glucose value. The correction method is to subtract the personalized incremental calibration factor from the initial calibrated blood glucose value to eliminate the systematic deviation caused by the user's current physiological drift.

[0073] The six-dimensional LED aging spectral drift compensation vector is weighted and aggregated into a single scalar compensation coefficient based on the contribution weight of each wavelength to the optical blood glucose estimation component. This contribution weight is determined by normalizing the coefficients of the corresponding optical features at each wavelength in the regression coefficient matrix of the multivariate linear regression sub-model. The LED aging spectral drift compensation vector is used to perform spectral gain compensation on the corrected blood glucose value. The reciprocal of the attenuation ratio of each wavelength is used as the gain compensation coefficient of the corresponding wavelength. Based on the regression coefficient matrix of the multivariate linear regression sub-model, the components corresponding to the optical reflectance feature and whose absolute regression coefficient value is greater than a preset threshold are selected and weighted to obtain the blood glucose estimation component dependent on the optical feature. The weighted summation of the components corresponding to the remaining temperature, impedance, ion migration, and metabolic thermal hysteresis features is used as the non-optical feature estimation component.

[0074] The amplitude of the blood glucose estimation component dependent on optical features in the multi-source fusion feature vector is restored. The restored blood glucose estimation component is then fused with the non-optical feature estimation component using weighted methods to generate a glucose metabolism monitoring value. This glucose metabolism monitoring value eliminates the dual effects of user physiological drift and light source aging drift. The weight coefficient of the restored blood glucose estimation component ranges from 0.60 to 0.90, with a classic value of 0.75. The weight coefficient of the non-optical feature estimation component ranges from 0.10 to 0.40, with a classic value of 0.25.

[0075] The cloud server packages and transmits the glucose metabolism monitoring values, individualized tissue fluid-blood glucose hysteresis phase, and individualized incremental calibration factors back to the embedded main control unit via an encrypted link. The embedded main control unit then receives and displays and stores the data locally.

[0076] S33. End-side strategy generation and dynamic adjustment of thermal stimulation parameters, specifically: S331. Recovery time history sliding window update and rate of change calculation, specifically: After each acquisition cycle, the embedded main control unit stores the recorded recovery time into a fixed-length historical sliding window. The window length is preset to 10 acquisition cycles and is managed using a cyclic overwrite mechanism. When the window is full of 10 recovery time data, the newly stored recovery time automatically overwrites the earliest stored recovery time, ensuring that the window always retains the recovery time data of the most recent 10 cycles.

[0077] After the window is updated, the embedded main control unit calculates the arithmetic mean of all 10 recovery times within the window as the baseline mean of recovery time; at the same time, it calculates the standard deviation of the recovery times within the window as a measure of the dispersion of recovery times; the recovery time change rate is defined as the normalized ratio obtained by subtracting the above mean from the recovery time of the current acquisition cycle and then dividing by the standard deviation. This change rate reflects the degree of deviation of the current skin blood flow response speed from the recent historical average level. A positive value indicates that the recovery time is prolonged, i.e., the blood flow response is degraded, and a negative value indicates that the recovery time is shortened, i.e., the blood flow response is improved, providing a quantitative basis for adjusting the thermal stimulation parameters in the next cycle.

[0078] S332. Dynamic adjustment of thermal stimulation power and duration, specifically: When the recovery time change rate is lower than the first threshold, the embedded main control unit determines that the skin microcirculation blood flow response tends to be stable, the skin tissue's sensitivity to thermal stimulation is increased, and the blood perfusion recovery speed is accelerated. The first threshold is preset to -0.5. At this time, the power of the low-frequency microcurrent thermal stimulation in the next cycle is reduced and the duration is shortened by a preset step size. The power reduction step size is 0.2mW, the duration shortening step size is 2s, and the adjusted power is not less than 0.1mW and the duration is not less than 3s.

[0079] When the rate of change of recovery time is higher than the second threshold, it is determined that the skin blood flow response is degraded, the microcirculation perfusion level is reduced, and the tissue response to thermal stimulation is sluggish. The second threshold is preset to +0.5. At this time, the thermal stimulation power is increased and the duration is extended by a preset step size. The power increase step size is 0.2mW and the duration extension step size is 2s. The adjusted power is not higher than 5mW and the duration is not higher than 15s.

[0080] A steady-state maintenance interval is defined between the first threshold of -0.5 and the second threshold of +0.5. When the rate of change of recovery time falls into this interval, the blood flow state is considered to be stable, and the current thermal stimulation parameters are kept unchanged.

[0081] S333. Analysis of trends in glucose metabolism monitoring values ​​and generation of sampling identifiers, specifically: The embedded main control unit analyzes the time series of glucose metabolism monitoring values ​​transmitted from the cloud, extracts the glucose metabolism monitoring values ​​of the most recent 3 to 5 consecutive sampling periods and arranges them in chronological order, calculates the absolute value of the difference between two adjacent glucose metabolism monitoring values ​​as a measure of the current trend; the absolute value of this first-order difference directly reflects the fluctuation range of blood glucose concentration between adjacent sampling periods, and the larger the value, the more drastic the change in metabolic state.

[0082] When the trend measurement is below the stability threshold, the user is determined to be in a metabolically stable state with minimal blood glucose fluctuations and slow changes in physiological parameters. A sparse sampling indicator is generated, and the subsequent sampling frequency is reduced to once every 30 minutes to conserve device power and extend battery life. When the trend measurement is above or equal to the stability threshold, the user is determined to be in a metabolically active state with significant blood glucose fluctuations, possibly during postprandial absorption or exercise stress. A dense sampling indicator is generated, and the subsequent sampling frequency is increased to once every 5 minutes to capture detailed dynamic changes in blood glucose. This stability threshold is set based on the statistical mean of the difference between the peak postprandial blood glucose fluctuation and fasting blood glucose in normal individuals, ranging from 0.5 mmol / L to 1.0 mmol / L, with a classic value of 0.8 mmol / L.

[0083] S334. Comprehensive evaluation of communication quality and lag phase, and strategy generation, specifically The embedded main control unit reads the signal strength indication from the end-side communication module to assess data transmission reliability. The signal strength indication ranges from -120dBm to -50dBm; values ​​closer to -50dBm indicate better communication quality, while values ​​closer to -120dBm indicate poorer communication quality. Simultaneously, it reads the individualized tissue fluid-blood glucose lag phase transmitted from the cloud to assess metabolic dynamic alignment requirements; a larger lag phase indicates a more severe delay in tissue fluid glucose relative to blood glucose.

[0084] To address the combined requirements of data transmission reliability and metabolic dynamic alignment, a spatiotemporal sparse sampling strategy and a communication interruption tolerance sleep threshold are generated: When communication quality falls below the interruption tolerance threshold of -100dBm, insufficient network coverage or weak signal is identified, and a local caching strategy is prioritized. The standardized feature vector to be uploaded and the recovery time are temporarily stored in local memory, and the sleep duration is extended to reduce power consumption. Simultaneously, a communication interruption tolerance sleep threshold of 5 acquisition cycles is generated, defining the maximum number of consecutive acquisition cycles the device can maintain autonomously in the absence of network coverage. When the phase lag exceeds the dynamic threshold of 10 minutes, the synchronization deviation between tissue fluid and blood glucose concentrations is deemed too large, and a dense sampling indicator is prioritized. The spatiotemporal sparse sampling strategy is generated by integrating these indicators and thresholds.

[0085] S34. Low-power sleep and wake-up wait, specifically: The embedded main control unit sets a wake-up timer according to the spatiotemporal sparse sampling strategy. The timer duration is determined by the sampling interval in the spatiotemporal sparse sampling strategy: 30 minutes in sparse sampling mode and 5 minutes in dense sampling mode. The power supplies to the skin temperature sensor, skin humidity sensor, multi-frequency skin impedance sensor, multi-wavelength optical sensor, and communication module are turned off, and the CPU frequency is reduced from 80MHz to 10MHz, entering deep sleep mode. In deep sleep mode, only the wake-up timer and external interrupt wake-up functions are retained; the CPU stops executing the main program, only maintaining timer counting and interrupt listening. When the wake-up timer reaches the preset sampling interval or an external trigger signal is received, the embedded main control unit wakes up from deep sleep mode, restores the CPU frequency to 80MHz, and sequentially powers on the sensor and communication module to begin monitoring for the next acquisition cycle. Example 5

[0086] like Figure 2 As shown, a non-invasive glycemic monitoring system based on multi-source surface signal fusion is used to implement a non-invasive glycemic monitoring method based on multi-source surface signal fusion. It includes: an ESP32 main controller, a temperature and humidity sensor, skin impedance electrodes, a multi-wavelength optical sensor, an LCD, a buzzer, a 4G module, and power management. The multi-wavelength optical sensor includes six channels with center wavelengths of 660nm, 905nm, 1200nm, 1300nm, 1450nm, and 1550nm; specifically: The end-side multi-source signal acquisition and preprocessing module is implemented by the ESP32 main controller scheduling temperature and humidity sensors, skin impedance electrodes and multi-wavelength optical sensors. It is configured to respond to monitoring triggers, pre-acquire temperature, humidity, multi-frequency impedance and 660nm and 905nm dual-wavelength reflectance to construct the blood perfusion equivalent index. When it deviates, low-frequency microcurrent thermal stimulation is applied through the skin impedance electrodes and the recovery curve is tracked until it falls into the ideal window. After stabilization, the six-wavelength signals are acquired by the multi-wavelength optical sensor and jointly preprocessed to generate a standardized feature vector, which is then uploaded to the cloud server through the 4G module. The cloud-based physical state decoupling and interlayer purification module is configured to establish a physical state decoupling model based on standardized feature vectors, extract the discrete step moments of sweat, the baseline of microvascular automatic fluctuations, and the baseline of glycosylation interference for compensation and elimination and background absorption subtraction, and extract metabolic thermal hysteresis features and ion migration features to generate physical state decoupling feature vectors; and establish an interlayer signal purification model based on the interlayer unmixing parameter matrix to extract the pure optical features of each cortical layer and splice the multi-source fusion feature vectors. The cloud-based metabolic inversion and calibration module is configured to perform cross-physical field metabolic inversion and hysteresis phase dynamic estimation, calculate the drift factor and state correction coefficient to generate a preliminary calibration blood glucose value; based on the user's historical incremental deviation mode and LED aging spectral drift compensation vector, the preliminary calibration blood glucose value is recalibrated to generate a glucose metabolism monitoring value, which is transmitted back to the ESP32 main controller via the 4G module. The ESP32 main controller controls the LCD to display the glucose metabolism monitoring value and controls the buzzer to output an alert signal. The edge-side strategy control and low-power management module, implemented by the ESP32 main controller in conjunction with power management, is configured to adjust the low-frequency microcurrent thermal stimulation parameters based on the recovery time change rate, generate a spatiotemporal sparse sampling strategy and sleep threshold based on the trend of changes in glucose metabolism monitoring values, and shut down the power of the temperature and humidity sensor, skin impedance electrode, multi-wavelength optical sensor and 4G module through power management, and reduce the frequency of the ESP32 main controller to enter deep sleep and wait to be woken up.

[0087] At the hardware implementation level, the end-side multi-source signal acquisition and preprocessing module, and the end-side strategy control and low-power management module are implemented by an embedded main control unit based on the ESP32-S3-WROOM-1-N16R8 and various functional peripheral circuits. For example... Figure 4 As shown, the ESP32 main control chip uses its general-purpose input / output pins and internally integrated analog-to-digital converter, direct memory access channel, and pulse width modulation output unit to realize the acquisition, preprocessing, display interaction, communication scheduling, and low power management of multi-source body surface signals.

[0088] like Figure 4 As shown, pins 37 and 38 of the ESP32 main control chip serve as the asynchronous serial communication transmitter and receiver, respectively, connected to the data transceiver of the 4G module via a serial bus, forming an encrypted data transmission channel between the terminal and the cloud server. Pins 15 and 4 serve as general-purpose digital-to-analog converter outputs, connected to the analog signal terminals of the sensor daughterboard interface, used to output low-frequency micro-current thermal stimulation signals. Pins 28 and 27 are connected to the button circuit, used to receive monitoring trigger commands and external wake-up signals. Pin 3 is connected to the reset circuit, and a resistor-capacitor reset network ensures stable power-on timing of the main control chip.

[0089] like Figure 5As shown, the ESP32 main control chip expands its parallel control ports through an expansion I / O chip. The expansion I / O chip is connected to the ESP32 main control chip via an I2C bus, and its parallel outputs output sensor power enable signals, 4G module power enable signals, buzzer control signals, and multi-channel sensor chip select signals, respectively. The fourth parallel output outputs the 4G module power enable signal, used to control the on / off state of the 4G module power supply circuit; the fifth parallel output outputs the sensor power enable signal, used to control the on / off state of the sensor power supply circuit; and the zeroth parallel output outputs the buzzer control signal, which, via a driver circuit, controls the buzzer to start and stop.

[0090] like Figure 6 As shown, the temperature and humidity sensor includes a first temperature and humidity sensing chip and a second temperature and humidity sensing chip. Both are connected to the sensor I2C clock line and sensor I2C data line of the ESP32 main control chip via an I2C bus, and are each controlled by an independent chip select signal. The power supply terminals of both the first and second temperature and humidity sensing chips are connected to the sensor power supply. The communication line is connected to the sensor power supply via a pull-up resistor to ensure the stability of the I2C bus level in the high-impedance state.

[0091] like Figure 7 As shown, the sensor daughterboard interface includes eight pins. Pins one and two are grounded; pins three and four are analog signal terminals connected to the skin impedance electrode, used to output a low-frequency micro-current excitation signal during the thermal stimulation phase and to acquire the skin impedance response signal during the acquisition phase; pin five is the sensor chip select signal; pins six and seven are the sensor I2C data line and clock line; and pin eight is the sensor power supply. A multi-wavelength optical sensor is integrated on the sensor daughterboard and connects to the ESP32 main control chip via the serial bus of this interface. The main control chip reads the optical reflectivity data of six characteristic wavelengths.

[0092] like Figure 8 As shown, the LCD display module is connected to the LCD_SPI2 clock line, master output / slave input line, master input / slave output line, and chip select line of the ESP32 master control chip via the SPI bus, realizing the display output of sugar metabolism monitoring values, signal strength, and power information. The backlight anode and backlight cathode of the LCD display module are connected to a 3.3V power supply through current-limiting resistors. The backlight enable terminal is connected to the backlight control pin of the ESP32 master control chip via a transistor switching circuit, and the master control chip adjusts the backlight brightness through pulse width modulation.

[0093] like Figure 9As shown, the buzzer driver circuit includes the buzzer body, a transistor switch, and a current-limiting resistor. The buzzer control signal is output from the extended I / O chip, connected to the base of the transistor via a base resistor, and the collector of the transistor is connected to the negative terminal of the buzzer. The positive terminal of the buzzer is connected to a 3.3V power supply via the current-limiting resistor. When the buzzer control signal is high, the transistor conducts, the buzzer is powered on, and it sounds, providing audio prompts for measurement completion and abnormal status.

[0094] like Figure 10 As shown, the power management module includes an ESP32 main control power supply circuit, a sensor power supply circuit, a 4G module power supply circuit, a battery interface circuit, and a SIM card slot power protection circuit. The ESP32 main control power supply circuit uses a low-dropout linear regulator to convert the battery voltage to 3.3V main control power, supplying power to the main control chip and expansion I / O chips. The sensor power supply circuit consists of an independent low-dropout linear regulator, whose enable terminal is controlled by the sensor power enable signal, outputting 3.3V sensor power, which only powers the temperature and humidity sensor, skin impedance electrode, and multi-wavelength optical sensor during the data acquisition phase. The 4G module power supply circuit consists of an independent power switch circuit, whose enable terminal is controlled by the 4G module power enable signal, supplying power only to the 4G communication module and SIM card slot during the data upload phase. The battery interface circuit includes a battery positive input terminal and a battery negative ground terminal, providing energy input to the entire device. The SIM card slot power protection circuit includes a transient suppression diode array to suppress electrostatic discharge and voltage surges on the SIM card signal line.

[0095] In deep sleep mode, the ESP32 main control chip simultaneously disables the sensor power enable signal and the 4G module power enable signal by extending the I / O chip, cutting off the output of the sensor power supply circuit and the 4G module power supply circuit, so that the temperature and humidity sensor, skin impedance electrode, multi-wavelength optical sensor, LCD display module, buzzer and 4G communication module are all in a power-off state; at the same time, the ESP32 main control chip reduces the central processing unit frequency to 10MHz, retaining only the wake-up timer and external interrupt listening function, thereby achieving microampere-level static power consumption of the whole machine.

[0096] In some embodiments, the non-invasive glycemic monitoring system based on multi-source surface signal fusion of the present invention can be implemented in a combination of hardware and software. As an example, the non-invasive glycemic monitoring system based on multi-source surface signal fusion of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the non-invasive glycemic monitoring method based on multi-source surface signal fusion of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0097] The modules described in the embodiments of the present invention can be implemented in software or hardware. The names of the modules are not, in some cases, intended to limit the functionality of the module itself.

[0098] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned non-invasive glucose monitoring methods based on multi-source surface signal fusion. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the non-invasive glucose monitoring method based on multi-source surface signal fusion shown in any embodiment of the present invention by calling the computer program.

[0099] In one alternative embodiment, an electronic device is provided, such as Figure 3 As shown, Figure 3 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0100] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0101] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0102] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0103] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0104] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0105] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0106] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any one of the above-mentioned non-invasive glucose monitoring methods based on multi-source body surface signal fusion.

[0107] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0108] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned non-invasive glucose monitoring method and system based on multi-source body surface signal fusion.

[0109] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0110] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in the methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0111] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EEPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0112] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0113] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0114] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0115] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0116] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A non-invasive method for monitoring glucose metabolism based on multi-source surface signal fusion, characterized in that, include: The blood perfusion equivalent index is constructed by pre-collecting temperature, humidity, multi-frequency impedance and dual-wavelength reflectivity. When it deviates, low-frequency microcurrent thermal stimulation is applied through impedance electrodes and the recovery curve is tracked until it falls into the ideal window. After stabilization, multi-wavelength signals are collected, pre-processed together to generate standardized feature vectors and uploaded. The cloud-based physical state decoupling model is established based on standardized feature vectors, and multidimensional interference baselines are extracted for compensation and elimination and background absorption subtraction, and metabolic thermal hysteresis and ion migration features are extracted. Based on the interlayer unmixing parameter matrix, the pure optical features of each cortex are purified and spliced ​​to generate a multi-source fusion feature vector; cross-physical field metabolic inversion and hysteresis phase dynamic estimation are performed, and the drift factor and state correction coefficient are calculated to generate a preliminary calibrated blood glucose value. The cloud-based system performs secondary calibration of the initial blood glucose value based on the user's historical incremental deviation pattern and LED aging spectral drift compensation vector, generating glucose metabolism monitoring values ​​that are then transmitted back. The end-to-end system adjusts the thermal stimulation parameters based on the rate of change of recovery time, generates a spatiotemporal sparse sampling strategy and a dormancy threshold based on the trend of changes in glucose metabolism monitoring values, and enters deep dormancy by reducing the frequency and waiting to be awakened.

2. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 1, characterized in that, The pre-collection temperature, humidity, multi-frequency impedance, and dual-wavelength reflectivity are used to construct the blood perfusion equivalent index. When deviations occur, low-frequency microcurrent thermal stimulation is applied via impedance electrodes, and the recovery curve is tracked until it falls within the ideal window. After stabilization, multi-wavelength signals are acquired, jointly pre-processed to generate a standardized feature vector, and uploaded. Specifically, this includes: In response to monitoring triggers, a low-resolution pre-scan is initiated to collect skin temperature, skin humidity, multi-frequency skin impedance, and optical reflectivity signals of two characteristic wavelengths. Based on the weighted fusion of the impedance dual-frequency ratio and the optical dual-wavelength ratio, a real-time blood perfusion equivalent index is constructed. The real-time blood perfusion equivalent index is compared with the preset ideal perfusion equivalent window. If it deviates, a low-frequency microcurrent thermal stimulation is applied through the skin impedance electrode. After heating, a cooling period is entered, and the dynamic recovery curve of the blood perfusion equivalent index is continuously tracked. When multiple consecutive sampling points of the dynamic recovery curve fall into the ideal window of perfusion equivalent, the perfusion equivalent index is determined to be stable, the cooling period ends and the system seamlessly switches to high-resolution formal acquisition. Simultaneously and in parallel, optical reflectivity signals of six characteristic wavelengths, as well as temperature, humidity, and impedance signals, are acquired to generate a multi-source body surface signal dataset. The dataset is jointly preprocessed to generate a standardized feature vector; the recovery time in the dynamic recovery curve and the cumulative runtime of each wavelength LED light source are recorded, and the standardized feature vector, recovery time and cumulative runtime are uploaded to the cloud server.

3. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 2, characterized in that, The real-time blood perfusion equivalent index is compared with a preset ideal perfusion equivalent window. If there is a deviation, low-frequency microcurrent thermal stimulation is applied. After heating, a cooling period is initiated, and the dynamic recovery curve is continuously tracked. Specifically, this includes: The real-time blood perfusion equivalent index is compared with a preset ideal perfusion equivalent window, which includes an upper threshold and a lower threshold. When the real-time blood perfusion equivalent index is lower than the lower threshold or higher than the upper threshold, it is determined that the deviation is from the ideal window, and the embedded main control unit starts the low-frequency microcurrent thermal stimulation output. After the low-frequency microcurrent thermal stimulation ends, a cooling period is entered. Skin impedance signals and dual-wavelength optical reflectivity signals are collected, the blood perfusion equivalent index is recalculated, and a dynamic recovery curve is generated.

4. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 1, characterized in that, The cloud-based system establishes a physical state decoupling model based on standardized feature vectors, extracts multidimensional interference baselines for compensation and elimination and background absorption subtraction, and extracts metabolic thermal hysteresis and ion migration features, specifically including: A three-dimensional coupled interference field of skin stratum corneum glycosylation-hydration-impedance and a discrete step model of sweat electrolyte pulses were established in the cloud. The discrete step time of sweat electrolyte pulses of skin humidity signal, microvascular automatic fluctuation baseline of skin temperature signal, glycosylation interference baseline of skin impedance signal and time-varying thickness of electrolyte film were extracted from the standardized feature vector. The skin temperature signal is first-order differential and combined with the decay time constant after thermal stimulation and steady-state temperature difference to estimate the skin-tissue fluid heat conduction hysteresis time constant as a metabolic heat hysteresis feature. Ion migration features are extracted by frequency domain decomposition of impedance signal. Step response compensation is performed on the impedance signal based on discrete step moments, interface impedance normalization correction is performed on the compensated impedance signal based on the time-varying thickness of the electrolyte film, microcirculation perfusion phase modulation elimination is performed on the optical reflectivity signal based on the microvascular automatic fluctuation baseline, and stratum corneum background absorption compensation is performed on the optical reflectivity signal based on the glycosylation interference baseline, generating a physical state decoupling feature vector.

5. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 1, characterized in that, The process of refining the pure optical features of each cortical layer based on the interlayer unmixing parameter matrix and splicing them to generate a multi-source fusion feature vector specifically includes: Based on the correlation between the differential penetration depth and multiple scattering path of six characteristic wavelengths in skin tissue, a model for interlayer aliasing deconvolution and memory effect decorrelation is established. The optical reflectivity signal in the physical state decoupled feature vector is input into the interlayer aliasing deconvolution and memory effect decorrelation model. The interlayer signal is purified by using the statistical correlation of the wavelength memory effect to extract the pure optical features of each cortex. The pure optical features of each cortex are spliced ​​together with metabolic thermal hysteresis features, ion migration features, and microvascular automatic fluctuation baselines to form a multi-source fusion feature vector.

6. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 5, characterized in that, The process of inputting the optical reflectivity signal from the physical state decoupled feature vector into the interlayer aliasing deconvolution and memory effect decorrelation model, and using the statistical correlation of the interwavelength memory effect to purify the interlayer signal and extract the pure optical features of each cortical layer specifically includes: A blind deconvolution model framework with multiple inputs and multiple outputs is constructed, and the measurement signals at each wavelength are defined as the observation values ​​after aliasing matrix transformation of different cortical source signals. By calculating the mutual information function between each wavelength channel, estimating the statistical correlation of the memory effect between wavelengths, establishing decorrelation constraint terms, and forming a complete objective function for the interlayer aliasing deconvolution and memory effect decorrelation model; The nonnegative matrix factorization algorithm is used to iteratively solve the objective function of the interlayer aliasing deconvolution and memory effect decorrelation model to obtain the unmixing matrix; The six-channel measurement signal is linearly transformed using a demixing matrix to separate the waveform estimates of each cortical source signal and output the optical feature components of the epidermis, superficial dermis, and deep dermis.

7. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 1, characterized in that, The process of performing cross-physical field metabolic inversion and hysteresis phase dynamic estimation, calculating drift factors and state correction coefficients to generate preliminary calibrated blood glucose values, specifically includes: The multi-source fusion feature vector is input into a pre-trained multivariate linear regression and random forest fusion model to output an initial blood glucose estimate. The metabolic thermal hysteresis feature is used as the thermal diffusion state component in the state vector, the initial blood glucose estimate is used as the observation output, and the peak amplitude of the power spectral density of the microvascular automatic fluctuation baseline is mapped to the microcirculation perfusion rate through a preset perfusion rate conversion coefficient, which is used as the time-varying gain coefficient in the state transition equation. A noisy approximate observation output with tissue fluid glucose concentration as the hidden state and the initial blood glucose estimate as the tissue fluid glucose concentration is constructed. An augmented state Kalman filter algorithm is used to augment the lag time constant into the state vector and perform joint recursive estimation with the tissue fluid glucose concentration, dynamically outputting the tissue fluid glucose concentration estimation sequence and the individualized tissue fluid-blood glucose lag phase. The estimated sequence of tissue fluid glucose concentration is read, and combined with the individualized tissue fluid-blood glucose lag phase, the estimated sequence of blood glucose concentration is obtained by converting it through a preset lag compensation model. Then, combined with the user's historical blood glucose data sequence and current skin temperature and skin humidity, the long-term physiological drift factor and short-term state correction coefficient are calculated using the exponential weighted moving average method. The estimated sequence of blood glucose concentration is multiplied by the long-term physiological drift factor and the short-term state correction coefficient to generate a preliminary calibrated blood glucose value.

8. The non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 1, characterized in that, The cloud-based system performs secondary calibration of the initial blood glucose value based on the user's historical incremental deviation mode and the LED aging spectral drift compensation vector to generate glucose metabolism monitoring values, which are then transmitted back. The device adjusts thermal stimulation parameters based on the recovery time rate of change, generates a spatiotemporally sparse sampling strategy and a dormancy threshold based on the trend of changes in glucose metabolism monitoring values, and enters deep dormancy by reducing the frequency to await awakening. Specifically, this includes: The cloud server calculates the LED aging spectral drift compensation vector based on the cumulative running time of each wavelength LED light source and the preset aging decay curve; and calculates the individualized incremental calibration factor based on the incremental deviation pattern of the user's historical data sequence. The initial calibrated blood glucose value is corrected for deviation using an individualized incremental calibration factor. The corrected blood glucose value is then decomposed into an optical feature-dependent blood glucose estimation component and a non-optical feature-dependent estimation component. The amplitude of the optical feature-dependent blood glucose estimation component is restored using an LED aging spectral drift compensation vector. The restored optical feature-dependent estimation component and the non-optical feature-dependent estimation component are then weighted and fused to generate a glucose metabolism monitoring value. The glucose metabolism monitoring value, hysteresis phase, and individualized incremental calibration factor are then fed back to the embedded main control unit. The embedded main control unit updates the historical sliding window based on the recovery time, calculates the rate of change of the recovery time, and dynamically adjusts the low-frequency microcurrent thermal stimulation power and duration for the next cycle; based on the trend of changes in glucose metabolism monitoring values, hysteresis phase, and communication quality, it generates a spatiotemporal sparse sampling strategy and a communication interruption fault-tolerant sleep threshold. The wake-up timer is set according to the spatiotemporal sparse sampling strategy, the power supply of the sensor and communication module is turned off, the frequency is reduced to enter deep sleep, and the system waits for the next cycle to wake up.

9. A non-invasive glucose monitoring method based on multi-source surface signal fusion according to claim 8, characterized in that, The embedded main control unit calculates the rate of change of recovery time based on the historical sliding window updated by recovery time, and dynamically adjusts the low-frequency microcurrent thermal stimulation power and duration for the next cycle. Based on the changing trends of glucose metabolism monitoring values, hysteresis phase, and communication quality, a spatiotemporal sparse sampling strategy and a communication interruption fault-tolerant sleep threshold are generated, specifically including: The embedded main control unit stores the recovery time recorded in each acquisition cycle into a fixed-length historical sliding window. When the window is full, the latest value is used to overwrite the oldest value. The mean and standard deviation of the recovery time within the window are calculated, and the recovery time change rate is defined as the ratio of the deviation of the current recovery time from the mean to the standard deviation. When the rate of change of recovery time is lower than the first threshold, the blood flow response is determined to be stable, and the power of low-frequency microcurrent thermal stimulation in the next cycle is reduced and the duration is shortened by a preset step size; when the rate of change of recovery time is higher than the second threshold, the blood flow response is determined to be degraded, and the thermal stimulation power is increased and the duration is extended by a preset step size. The time series of glucose metabolism monitoring values ​​are analyzed, and the absolute value of their first difference is calculated as a measure of the trend of change. When the measure of the trend of change is lower than the stationary threshold, a sparse sampling label is generated; otherwise, a dense sampling label is generated. The strategy integrates communication quality with the spatiotemporal sparse sampling strategy for generating hysteresis phase and the communication interruption tolerance sleep threshold: when the communication quality is lower than the interruption tolerance threshold, local buffering is enabled and sleep is extended, while a communication interruption tolerance sleep threshold is generated; when the hysteresis phase exceeds the dynamic threshold, dense sampling is preferred.

10. A non-invasive glucose monitoring system based on multi-source surface signal fusion, characterized in that, A method for implementing a non-invasive glycemic monitoring method based on multi-source surface signal fusion as described in any one of claims 1-9, comprising: an ESP32 main controller, a temperature and humidity sensor, a skin impedance electrode, a multi-wavelength optical sensor, an LCD, a buzzer, a 4G module, and power management; the multi-wavelength optical sensor includes six channels with center wavelengths of 660nm, 905nm, 1200nm, 1300nm, 1450nm, and 1550nm; specifically: The end-side multi-source signal acquisition and preprocessing module is implemented by the ESP32 main controller scheduling temperature and humidity sensor, skin impedance electrode and multi-wavelength optical sensor. It is configured to respond to monitoring trigger, pre-acquire temperature, humidity, multi-frequency impedance and 660nm and 905nm dual-wavelength reflectance to construct blood perfusion equivalent index. When it deviates, low-frequency microcurrent thermal stimulation is applied through the skin impedance electrode and the recovery curve is tracked until it falls into the ideal window. After stabilization, the six wavelength signals are collected by the multi-wavelength optical sensor and jointly pre-processed to generate a standardized feature vector, which is then uploaded to the cloud server through the 4G module. The cloud-based physical state decoupling and interlayer purification module is configured to establish a physical state decoupling model based on standardized feature vectors, extract the discrete step moments of sweat, the baseline of microvascular automatic fluctuations, and the baseline of glycosylation interference for compensation and elimination and background absorption subtraction, and extract metabolic thermal hysteresis features and ion migration features to generate physical state decoupling feature vectors; and establish an interlayer signal purification model based on the interlayer unmixing parameter matrix to extract the pure optical features of each cortical layer and splice the multi-source fusion feature vectors. The cloud-based metabolic inversion and calibration module is configured to perform cross-physical field metabolic inversion and hysteresis phase dynamic estimation, calculate the drift factor and state correction coefficient to generate a preliminary calibration blood glucose value; based on the user's historical incremental deviation mode and LED aging spectral drift compensation vector, the preliminary calibration blood glucose value is recalibrated to generate a glucose metabolism monitoring value, which is transmitted back to the ESP32 main controller through the 4G module. The ESP32 main controller controls the LCD to display the glucose metabolism monitoring value and controls the buzzer to output a prompt signal. The edge-side strategy control and low-power management module, implemented by the ESP32 main controller in conjunction with power management, is configured to adjust the low-frequency microcurrent thermal stimulation parameters based on the recovery time change rate, generate a spatiotemporal sparse sampling strategy and a sleep threshold based on the trend of changes in glucose metabolism monitoring values, and shut down the power of the temperature and humidity sensor, skin impedance electrode, multi-wavelength optical sensor and 4G module through the power management, and reduce the frequency of the ESP32 main controller to enter deep sleep and wait to be woken up.

Images