Blood oxygen estimation method and system based on multi-band periodic waveform driving

Abstract

The present invention discloses a blood oxygen estimation method and system based on multi-band periodic waveform driving, belonging to the technical field of blood oxygen monitoring in plateau environments. It solves the problems existing in the prior devices when determining the blood oxygen saturation of the area to be detected, such as insufficient measurement accuracy, large influence of motion interference, and unstable measurement results caused by individual differences in special environments such as plateaus, which in turn affect the reduction of the blood oxygen detection inference speed. The method includes real-time collection of PPG physiological signal data by a wristband-type physiological data collection module, preprocessing of the PPG physiological signal data, execution of an end-to-end E2E model, and output of the SpO2 value under the periodic waveform. The end-to-end E2E model of the present invention can learn and compensate for the physiological characteristics among different users, and achieve personalized adjustment for the influence of motion interference, so as to provide more accurate blood oxygen level measurement results in special environments such as plateaus, and use the CFCs algorithm to reduce the calculation amount and accelerate the inference speed.

Description

Technical Field

[0001] The present invention belongs to the technical field of blood oxygen monitoring in plateau environments, and particularly relates to a blood oxygen estimation method and system based on multi-band periodic waveform driving. Background Art

[0002] Blood oxygen saturation (SpO2) is a key physiological parameter that measures the degree of oxygen binding to hemoglobin in the blood and is crucial for evaluating an individual's health status. In the field of health monitoring, the monitoring of blood oxygen saturation is not only of great significance for the diagnosis and treatment of respiratory diseases, cardiovascular diseases, etc., but also in special environments such as plateaus, due to the thin air, the monitoring of blood oxygen saturation is particularly critical for the prevention and treatment of altitude diseases (such as high altitude pulmonary edema and high altitude cerebral edema).

[0003] In the plateau hypoxic environment, blood oxygen saturation is a key indicator to measure the available oxygen level in the human body and is also one of the important physiological parameters to evaluate an individual's adaptability to the plateau environment. Under normal circumstances, the arterial blood oxygen saturation of the human body in plain areas is 97%-100%, and when the saturation is below 90%, it is considered hypoxemia. In plateau areas, the partial pressure of oxygen decreases, the oxygen intake of the human body decreases, and the blood oxygen saturation will decrease significantly. Persistent low blood oxygen saturation may cause a series of discomfort symptoms and even endanger life.

[0004] Traditional blood oxygen monitoring devices, such as finger clip oximeters, although widely used in medical environments, have obvious limitations in extreme environments such as plateaus. These devices usually adopt the 1-photodetector (1PD) technology and calculate blood oxygen saturation by measuring the absorption differences of light with different wavelengths. Although this technology is relatively accurate in medical monitoring, it has the following deficiencies:

[0005] The finger clip oximeter requires the user to clip it on the finger regularly for measurement, and it cannot achieve 24-hour continuous monitoring, which limits its application in scenarios such as plateaus that require continuous monitoring; and due to the design of the finger clip oximeter, long-term wearing may cause finger discomfort and affect the user's daily life and sleep.

[0006] Meanwhile, Chinese Patent CN112472079B discloses a blood oxygen saturation detection device, equipment and storage medium. When the determination module receives a blood oxygen saturation detection instruction, it determines the area to be detected according to the blood oxygen saturation detection instruction; the extraction module acquires the initial optical signal of the area to be detected and performs feature extraction on the initial optical signal to obtain signal feature information; the detection module determines the blood oxygen saturation of the area to be detected according to the signal feature information. However, when the existing device's detection module determines the blood oxygen saturation of the area to be detected according to the signal feature information, there are problems such as insufficient measurement accuracy, large influence of motion interference, and unstable measurement results caused by individual differences in special environments such as plateaus, which in turn affect the reduction of the blood oxygen detection inference speed. To address the above problems, we propose a blood oxygen estimation method and system based on multi-band periodic waveform driving. Summary of the Invention

[0007] The purpose of the present invention is to provide a blood oxygen estimation method and system based on multi-band periodic waveform driving for the deficiencies of the prior art, and to solve the problems of insufficient measurement accuracy, large influence of motion interference, and unstable measurement results caused by individual differences in special environments such as plateaus when the existing device's detection module determines the blood oxygen saturation of the area to be detected according to the signal feature information, which in turn affects the reduction of the blood oxygen detection inference speed.

[0008] The present invention is implemented as follows. The blood oxygen estimation method based on multi-band periodic waveform driving includes:

[0009] Real-time collection of PPG physiological signal data based on a wristband-type physiological data collection module;

[0010] Load the PPG physiological signal data, preprocess the PPG physiological signal data, intercept the periodic band time segment signal according to the number of peaks, and fuse the periodic band time segment signals at the decision-making level based on the decision rule to construct a PPG physiological signal feature vector.

[0011] Pre-construct an end-to-end E2E model (End to end, E2E), use the PPG physiological signal feature vector as the input, execute the end-to-end E2E model, and the end-to-end E2E model performs signal feature extraction on the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combines a multi-layer perceptron to identify and analyze the signal feature set, and outputs the SpO2 value under the periodic waveform.

[0012] The wristband-type physiological data collection module includes:

[0013] Four photoelectric sensors, which are used to collect PPG physiological signal data in real time, and the photoelectric sensors are symmetrically distributed at 4 points on the inner side of the wearable device;

[0014] Multi - band LED lamp, the multi - band LED lamp is deployed in the middle area of the photoelectric sensor and adopts a recessed design. The multi - band LED lamp is used to assist the photoelectric sensor to work;

[0015] Microcontroller MCU, the microcontroller MCU is used to control the multi - band LED lamp to collect PPG signals of different bands and different positions in a time - sharing manner, and convert the PPG analog signals into PPG digital signals in different bands through a high - sampling - rate analog - to - digital converter ADC.

[0016] The method for pre - processing PPG physiological signal data includes:

[0017] Load the PPG physiological signal data, perform spectral analysis on the PPG physiological signal data, observe the signal time - domain waveform, and reconstruct and denoise the PPG physiological signal data based on the Daubechies wavelet basis function to obtain the reconstructed and denoised AC and DC signals;

[0018] Load the reconstructed and denoised AC and DC signals, and perform filtering processing on the AC and DC signals. Among them, when performing filtering processing on the AC and DC signals, pre - process the original signals of different wavelengths collected by each group of photoelectric sensors. Extract the DC signal through a low - pass filter with a cut - off frequency of 0.1Hz, calculate the average value of the PPG physiological signal within a predetermined time window as an estimate of the DC component. As the window moves, continuously update the average value to track the change of the DC signal in real time. Use a high - pass filter with a cut - off frequency of 0.5Hz to 10Hz to allow the AC component to pass through, filter out the DC component and low - frequency interference to obtain the AC signal. The AC and DC signals under each waveform are expressed as: Obtain the PPG physiological signal data after filtering the AC and DC signals, perform PPG signal detection on the PPG physiological signal data, and output the AC and DC of the best PPG signals of different bands identified from each group of photoelectric sensors;

[0019] Calculate different peaks based on the waveform of the entire PPG physiological signal, intercept time periods according to the number of peaks in different bands, and obtain the DC and AC time lengths of each band of the PPG physiological signals of different photoelectric sensors as band signals ;

[0020] Decision-level fusion of periodic band time-segment signals based on decision rules to construct a PPG physiological signal feature vector. When performing decision-level fusion of periodic band time-segment signals based on decision rules, the three-band signals of each optoelectronic sensor are analyzed independently, and then the analysis results are combined according to the decision rules. For the effective signals of each band, if two or more band signals under each optoelectronic sensor are considered valid PPG signals, then the photodiode PD in that optoelectronic sensor is considered. According to the decision result, the most effective PPG signal in each optoelectronic sensor is selected to construct the PPG physiological signal feature vector. The PPG physiological signal feature vector is expressed as:

[0021] .

[0022] A method for reconstructing and denoising PPG physiological signal data based on Daubechies wavelet basis functions, including:

[0023] Obtain PPG physiological signal data, observe the time-domain waveform of the signal, determine the distribution of the main frequency components of the signal through spectral analysis, and judge the position and bandwidth of the noise in the frequency domain;

[0024] Preset the low-order Daubechies wavelet db2 and the high-order Daubechies wavelet db6, and judge whether the PPG physiological signal is smooth;

[0025] If the PPG physiological signal is a smooth signal, perform discrete wavelet decomposition on the DC and AC signals of the PPG physiological signal based on the preset Daubechies wavelet db2 and the decomposition level. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer;

[0026] If the PPG physiological signal is a non-smooth signal, perform discrete wavelet decomposition on the DC and AC signals of the PPG physiological signal based on the preset Daubechies wavelet db6 and the decomposition level. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer;

[0027] Integrate at least one set of high-frequency coefficients, process the high-frequency coefficients based on an adaptive threshold, use the minimum Stein unbiased risk estimation criterion to determine the threshold, and set the threshold T;

[0028] Judge whether the absolute value of the high-frequency coefficient is greater than the set threshold T;

[0029] If the absolute value of the high-frequency coefficient is less than or equal to the set threshold T, set the amplitude of the high-frequency coefficient to 0;

[0030] If the absolute value of the high-frequency coefficient is greater than the set threshold T, subtract the sign function multiple of the threshold T from the absolute value of the high-frequency coefficient to shrink the high-frequency coefficient, that is ;

[0031] Load the high-frequency coefficients and low-frequency coefficients after threshold processing, and use the inverse discrete wavelet transform algorithm to perform wavelet reconstruction processing on the high-frequency coefficients and low-frequency coefficients to obtain the reconstructed and denoised AC and DC signals.

[0032] When performing PPG signal detection on the PPG physiological signal data, preprocess the original signals collected in 3 bands under each photoelectric sensor, use template matching technology to identify whether there is a PPG signal in each photoelectric sensor, and select the signal with the highest similarity as the effective PPG signal. First, a data set containing PPG signals of different individuals and under different conditions is known. Then, high-quality PPG signal samples are selected from the data set as PPG signal sample templates. The PPG signal samples have clear waveform features, with obvious peaks and valleys.

[0033] Extract the same features as the PPG signal sample template from the original signals collected by each photoelectric sensor, at the same positions of the peaks and valleys, and use the correlation coefficient to calculate the similarity between the AC and DC of the PPG signal of each photoelectric sensor and the AC and DC of the PPG template.

[0034] The correlation coefficient calculation formula is as follows:

[0035] Select the signal with a similarity higher than the threshold as the effective PPG signal according to the similarity calculation result. The threshold is determined by evaluating the classification accuracy through classification experiments, and output the AC and DC of the best PPG signals in different bands identified from each group of photoelectric sensors.

[0036] The method of intercepting the time period according to the number of peaks in different bands includes:

[0037] Obtain the PPG physiological signal and record the positions of the signal peak points.

[0038] Perform band periodic detection on the PPG physiological signal to identify the bands in the PPG physiological signal, and use the peak detection algorithms of the first derivative and the second derivative to identify the peaks of the bands, expressed as:

[0039] Among them, Peaks1 represents the first derivative, and Peaks2 represents the second derivative.

[0040] Load the band periodic detection results, and perform time period interception on the PPG physiological signal based on the band periodic detection results.

[0041] Among them, when intercepting the time period, determine the number of peaks according to the peak solution result of the band , and perform periodic time period interception , a peak point before the first peak is used as the start time t start , the peak point corresponding to the start time after the peak point is used as the end time t end , so the total intercepted time period is represented by M as ;

[0042] After intercepting the PPG physiological signal in the time period, a scaling method is used for time period alignment processing to ensure that the time period lengths of different waveforms are the same, and the fixed time period length is calculated as follows:

[0043] Using the scaling method, a scaling ratio is added after each waveform of the fixed length as a mark, and the scaling ratio is calculated based on the ratio of the actual length of the time period to the fixed length ;

[0044] The interception of the time periods of all bands is performed synchronously, the physiological information at the same time is extracted, and the DC and AC time lengths of each band of the PPG physiological signals of different optoelectronic sensors are obtained as the signal band numbers .

[0045] The end-to-end E2E model includes an input layer, a CFCs module, a residual connection module, a multi-layer perceptron, and an output layer. The CFCs module is used to extract features from the PPG physiological signal feature vector to obtain a signal feature set, and the multi-layer perceptron is used to identify and analyze the signal feature set and output the SpO2 value under the periodic waveform;

[0046] The CFCs module includes a first-layer CFCs, a second-layer CFCs, and a third-layer CFCs. The first-layer CFCs is connected to the input layer, and the first-layer CFCs, the second-layer CFCs, and the third-layer CFCs are connected through the residual connection module. The residual connection module includes a first-layer residual connection, a second-layer residual connection, and a third-layer residual connection. The first-layer residual connection, the second-layer residual connection, and the third-layer residual connection are respectively arranged after the first-layer CFCs, the second-layer CFCs, and the third-layer CFCs.

[0047] The method for extracting signal features from the PPG physiological signal feature vector by the end-to-end E2E model based on the closed-form solution feature extraction algorithm of explicit time dependence includes:

[0048] Load the PPG physiological signal feature vector and input the PPG physiological signal feature vector into the input layer of the end-to-end E2E model;

[0049] The PPG physiological signal feature vector input by the input layer First, feature extraction is performed through the first-layer CFCs, and through , and A neural network is used to simulate the dynamic behavior of neurons, and the output is expressed as ;

[0050] The output of the first - layer CFCs is added to the input by the first - layer residual connection to form the first - layer residual connection, and the output is expressed as ;

[0051] The result of the first - layer residual connection serves as the input of the second - layer CFCs. The second - layer CFCs apply the neural network again , and for feature extraction, which is expressed as ;

[0052] The output of the second - layer CFCs is added to the result of the first - layer residual connection by the second - layer residual connection to form the second - layer residual connection, and the output is expressed as ;

[0053] The result of the second - layer residual connection serves as the input of the third - layer CFCs. The third - layer CFCs apply the neural network again , and for feature extraction, which is expressed as ;

[0054] The output of the third - layer CFCs is added to the result of the second - layer residual connection by the third - layer residual connection to form the third - layer residual connection, and the output is expressed as ;

[0055] The output results of the third - layer residual connection are integrated into a signal feature set.

[0056] The multi - layer perceptron includes three fully - connected layers. The input dimension of the multi - layer perceptron matches the output dimension of the CFCs as , and dimensionality reduction is performed through three fully - connected layers. The final output dimension is 1. The fully - connected layers are used to map the input features to the output space through linear transformation.

[0057] On the other hand, the present invention also provides a blood oxygen estimation system based on a multi - band periodic waveform drive. The blood oxygen estimation system based on a multi - band periodic waveform drive includes:

[0058] A wrist - band type physiological data acquisition module, which is used to collect PPG physiological signal data in real time;

[0059] A data preprocessing module for preprocessing PPG physiological signal data, intercepting the periodic band time segment signals according to the number of peaks, making decision-level fusion of the periodic band time segment signals based on decision rules, and constructing a PPG physiological signal feature vector;

[0060] An oxygen saturation detection module for pre-constructing an end-to-end E2E model, taking the PPG physiological signal feature vector as input, executing the end-to-end E2E model, and the end-to-end E2E model extracting signal features from the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combining a multi-layer perceptron to identify and analyze the signal feature set to output the SpO2 value under the periodic waveform.

[0061] Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects:

[0062] In the embodiments of the present invention, the periodic time segment is dynamically intercepted through band peak detection, and a feature extraction algorithm based on CFCs is used to efficiently extract signal features directly from the original data, improving the processing efficiency. The end-to-end E2E model can adaptively learn the physiological characteristics of different users, and perform real-time compensation for motion interference in the plateau environment, improving the measurement robustness. By optimizing the calculation amount through the CFCs algorithm, the inference speed is accelerated, meeting the low-power and real-time monitoring requirements in special scenarios such as the plateau, so as to more accurately detect abnormal changes in blood oxygen and provide reliable support for health early warning.

[0063] In the embodiments of the present invention, the wristband-type physiological data acquisition module allows users to collect physiological data by wearing the device when performing blood oxygen monitoring, and is specifically used for real-time and continuous collection of users' physiological parameters, especially non-invasive measurement through PPG. And in order to enhance the stability of the device during exercise and reduce the influence of motion artifacts, the number of photoelectric sensors is four groups, and the photoelectric sensors are symmetrically arranged outside the multi-band LED lights, which is suitable for long-term wearing by users. Its low-power characteristic makes it very suitable for integration into wearable devices such as smart bracelets, with minimal interference to users' daily lives.

[0064] In the embodiments of the present invention, through wavelet transform and reconstruction, high-frequency noise and interference in the signal can be effectively removed, the main features of the signal can be retained, the quality of the signal can be improved. By calculating the average value through a moving window, the change of the DC signal can be tracked in real time, ensuring the accuracy and timeliness of the DC signal. Through PPG signal detection, the best PPG signals in different bands can be accurately identified, thereby improving the representativeness of the signal. The decision rule can be adjusted according to different users and environments, improving the adaptability and flexibility of the model. By selecting the most effective PPG signals, high-quality feature vectors can be constructed, providing accurate inputs for subsequent blood oxygen saturation calculation. The preprocessing method not only removes noise and interference, but also enhances the useful components of the signal, improving the measurement accuracy and reliability. Through these preprocessing methods, high-quality input data can be provided for subsequent blood oxygen saturation calculation, thereby improving the performance and accuracy of the entire system.

[0065] In the embodiments of the present invention, the end-to-end E2E model uses the CFCs module to reduce the computational amount and accelerate the training and inference speed, approximating the analytical solution of the neural dynamics system. Through the layer-by-layer training and feature extraction of three layers of CFCs, and through residual connections between CFCs, important parameters affecting blood oxygen changes are captured, and then through a multi-layer perceptron, accurate SpO2 values under the periodic waveform are output. The end-to-end E2E model reduces the need for feature engineering, simplifies the development process, makes model training and prediction more efficient, and fully utilizes the advantages of the end-to-end E2E model in the training and prediction processes, that is, without complex feature engineering, it can directly learn the prediction results of blood oxygen saturation from the original data. In the end-to-end E2E model, the explicit time-dependent closed-form solution (CFCs) can improve the accuracy of feature extraction and the efficiency of the entire training process.

[0066] In the embodiments of the present invention, an end-to-end E2E model is provided. The end-to-end E2E model is driven by a multi-band periodic waveform, supports real-time and continuous blood oxygen saturation monitoring, helps to detect abnormal changes in blood oxygen levels in a timely manner, and provides early health warnings for individuals in high-altitude environments. The E2E model algorithm can learn and compensate between different users, thereby achieving personalized adjustment and being able to adapt to the physiological characteristics of different users. Description of the Drawings

[0067] Figure 1 It is a schematic flowchart of the implementation of the blood oxygen estimation method based on multi-band periodic waveform driving provided by the present invention.

[0068] Figure 2 It shows a schematic installation diagram of a photoelectric sensor and a multi-band LED lamp in the embodiments of the present invention.

[0069] Figure 3Shows a schematic diagram of the implementation process of the PPG physiological signal data preprocessing method.

[0070] Figure 4 Shows a schematic diagram of the implementation process of the method for extracting signal features of the PPG physiological signal feature vector by the end-to-end E2E model based on the closed-form solution feature extraction algorithm for explicit time dependence.

[0071] Figure 5 Shows a schematic diagram of the structure of the blood oxygen estimation system driven by multi-band periodic waveforms. Detailed implementation manners

[0072] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the specification and claims of this application or the above drawings are used to distinguish different objects and not to describe a specific order.

[0073] Referring to "embodiments" herein means that the specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of this application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0074] When the existing device detection module determines the blood oxygen saturation of the area to be detected according to the signal characteristic information, there are problems such as insufficient measurement accuracy, large influence of motion interference, and unstable measurement results caused by individual differences in special environments such as plateaus, which in turn affect the reduction of the blood oxygen detection inference speed. To address the above problems, we propose a blood oxygen estimation method and system based on multi-band periodic waveform driving. Briefly, when the method is implemented, it first collects PPG physiological signal data in real time based on the wristband-type physiological data acquisition module 100, then preprocesses the PPG physiological signal data, fuses the periodic band time segment signals at the decision-making level based on decision rules, constructs a PPG physiological signal feature vector, and pre-constructs an end-to-end E2E model (End to end, E2E). Finally, the end-to-end E2E model is executed. The end-to-end E2E model extracts signal features from the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combines a multi-layer perceptron to identify and analyze the signal feature set, and outputs the SpO2 value under the periodic waveform. In the embodiment of the present invention, the periodic time segment can be intercepted according to the number of band peaks, and the signal features can be directly extracted from the original data through the explicit time-dependent closed-form solution feature extraction algorithm for easy feature extraction. The end-to-end E2E model can learn and compensate for the physiological characteristics between different users, and achieve personalized adjustment for the influence of motion interference, so as to provide more accurate blood oxygen level measurement results in special environments such as plateaus. Moreover, the end-to-end E2E model uses the CFCs (Closed Form Continuous-time Neural Networks) algorithm to reduce the calculation amount and speed up the inference speed, which is particularly important for health monitoring in special environments such as plateaus, and can timely detect abnormal changes in blood oxygen levels, thus helping to provide early health warnings.

[0075] An embodiment of the present invention provides a blood oxygen estimation method based on multi-band periodic waveform driving. Figure 1 The implementation flow schematic diagram of the blood oxygen estimation method based on multi-band periodic waveform driving is shown. The blood oxygen estimation method based on multi-band periodic waveform driving specifically includes:

[0076] S10, collecting PPG physiological signal data in real time based on the wristband-type physiological data acquisition module 100;

[0077] S20, loading the PPG physiological signal data, preprocessing the PPG physiological signal data, intercepting the periodic band time segment signal according to the number of peaks, fusing the periodic band time segment signals at the decision-making level based on decision rules, and constructing a PPG physiological signal feature vector;

[0078] S30. Pre-build an end-to-end E2E model. Using the PPG physiological signal feature vector as input, execute the end-to-end E2E model. The end-to-end E2E model performs signal feature extraction on the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combines a multi-layer perceptron to identify and analyze the signal feature set, and outputs the SpO2 value under the periodic waveform.

[0079] In the embodiments of the present invention, a periodic time period is dynamically intercepted through band peak detection, and a feature extraction algorithm based on CFCs is used to efficiently extract signal features directly from the original data, improving the processing efficiency. The end-to-end E2E model can adaptively learn the physiological characteristics of different users and perform real-time compensation for motion interference in the plateau environment, improving the measurement robustness. By optimizing the calculation amount through the CFCs algorithm, the inference speed is accelerated, meeting the low-power and real-time monitoring requirements in special scenarios such as the plateau, so as to more accurately detect abnormal changes in blood oxygen and provide reliable support for health warning.

[0080] In the embodiments of the present invention, the wristband-type physiological data acquisition module 100 includes:

[0081] Four photoelectric sensors. The photoelectric sensors are used to collect PPG (photoplethysmogram) physiological signal data in real time, and the photoelectric sensors are symmetrically distributed at 4 points on the inner side of the wearable device;

[0082] A multi-band LED lamp. The multi-band LED lamp is deployed in the middle area of the photoelectric sensors and adopts a concave design. The multi-band LED lamp is used to assist the photoelectric sensors to work;

[0083] In this embodiment, the wristband-type physiological data acquisition module 100 allows users to collect physiological data by wearing the device when performing blood oxygen monitoring. It is specifically used to collect the physiological parameters of users in real time and continuously, especially to achieve non-invasive measurement through PPG. In order to enhance the stability of the device during exercise and reduce the influence of motion artifacts, the number of photoelectric sensors can be four groups, and the photoelectric sensors are symmetrically arranged outside the multi-band LED lamp. Among them, Figure 2 The installation schematic diagram of the photoelectric sensor and the multi-band LED lamp in the embodiments of the present invention is shown. Among them, the photoelectric sensor adopts a convex design to ensure more stable contact with the skin; while the multi-band LED lamp is designed to be concave. Such a design helps to expand the radiation area of light, thereby effectively suppressing the artifacts caused by motion. The physiological signal data is extracted through the photoplethysmography (PPG) of the wearable device, and the pulsation signal is obtained by using the spectral absorption ability of hemoglobin in blood vessels for different wavelengths (green light 550nm, red light 660nm, and infrared light 940nm) and the difference in blood flow volume during heart beats.

[0084] A microcontroller MCU is used to time - share control a multi - band LED lamp to collect PPG signals at different bands and different positions, and convert the PPG analog signals into PPG digital signals at different bands through a high - sampling - rate analog - to - digital converter ADC.

[0085] It should be noted that in terms of hardware configuration, the module irradiates four specific positions on the wrist through the embedded LED light source. The light emitted by these light sources will be absorbed by the tissues under the skin to different degrees, and the reflected light intensity signal will then be received by the photoelectric sensor and converted into an electrical signal, so as to obtain PPG signals at different positions under the same wavelength. Among them, the direct - current component (DC) is mainly caused by relatively stable factors such as ambient light and the dark current of the photoelectric sensor, reflecting a baseline signal with a relatively flat waveform, and its basic output is generated by its own characteristics; while the alternating - current component (AC) is a signal component related to pulse beating. When the heart beats, the volume of peripheral blood vessels will change periodically. For red light (r, 660nm), infrared light (ir, 940nm) and green light (g, 550nm), although their absorption and scattering characteristics in human tissues are different, they will all change periodically with the change of blood vessel volume. The microcontroller (MCU) is used to time - share control the LED light source, and then collect PPG signals at different bands and different positions. Through a high - sampling - rate analog - to - digital converter (ADC), the analog signal is converted into a digital signal at different bands.

[0086] The embodiment of the present invention provides a method for pre - processing PPG physiological signal data. Figure 3 The figure shows a schematic implementation flow chart of the method for pre - processing PPG physiological signal data. The method for pre - processing PPG physiological signal data specifically includes:

[0087] S101, load the PPG physiological signal data, perform spectrum analysis on the PPG physiological signal data, observe the signal time - domain waveform, and reconstruct and denoise the PPG physiological signal data based on the Daubechies wavelet basis function to obtain the AC and DC signals after reconstruction and denoising;

[0088] S102, load the AC and DC signals after reconstruction and denoising, and perform filtering processing on the AC and DC signals. Among them, when performing filtering processing on the AC and DC signals, pre - process the original signals collected by each group of photoelectric sensors at different wavelengths. Extract the DC signal through a low - pass filter with a cut - off frequency of 0.1Hz, calculate the average value of the PPG physiological signal within a predetermined time window as the estimated value of the direct - current component. As the window moves, continuously update the average value to track the change of the DC signal in real - time. Use a high - pass filter with a cut - off frequency of 0.5Hz to 10Hz to allow the alternating - current component to pass through, filter out the direct - current component and low - frequency interference, and obtain the AC signal. , the AC and DC signals under each waveform are expressed as:

[0089]

[0090] S103. Obtain the PPG physiological signal data after filtering the AC and DC signals, perform PPG signal detection on the PPG physiological signal data, and output the AC and DC of the best PPG signals in different bands identified from each group of optoelectronic sensors;

[0091] In this embodiment, when performing PPG signal detection on the PPG physiological signal data, preprocess the original signals collected in 3 bands under each optoelectronic sensor, use template matching technology to identify whether there is a PPG signal in each optoelectronic sensor, and select the signal with the highest similarity as the effective PPG signal. First, a dataset containing PPG signals of different individuals and under different conditions is known. Then, high-quality PPG signal samples are selected from the dataset as the PPG signal sample templates. The PPG signal samples have clear waveform features, obvious peaks and valleys;

[0092] Extract the same features as the PPG signal sample template from the original signals collected by each optoelectronic sensor, the same positions of the peaks and valleys, and use the correlation coefficient to calculate the similarity between the AC and DC of the PPG signals of each optoelectronic sensor and the AC and DC of the PPG template;

[0093] Among them, the correlation coefficient calculation formula:

[0094]

[0095] Select the signal with a similarity higher than the threshold as the effective PPG signal according to the similarity calculation result. The threshold is determined by evaluating the classification accuracy through classification experiments, and output the AC and DC of the best PPG signals in different bands identified from each group of optoelectronic sensors.

[0096] S104. Solve the concentration ratio k of oxyhemoglobin HbO2 and deoxyhemoglobin Hb that is directly related to blood oxygen saturation based on the AC and DC of the best PPG signal;

[0097] It should be noted that the measurement of blood oxygen saturation generally adopts the dual-wavelength transmission method, which usually involves two main light absorption characteristics: HbO2 and Hb. These characteristics can be measured by light of different wavelengths, and its theoretical basis stems from the Lambert-Beer law. The light source propagates in the medium, and the expression relationship is as follows:

[0098]

[0099] Among them, is the transmitted (received) light intensity, is the intensity of incident light, is the reflection coefficient of different molecules, is the molecular concentration, is the incident diameter;

[0100] During human breathing, the dilation and constriction of blood vessels will affect the propagation path of light in tissues. HbO2 and Hb in the blood absorb the AC light intensity, and this change is caused by which increases , that is, the change amount of the light propagation path in tissues. In the PPG signal, due to the change in light intensity caused by and the reference light intensity can be expressed as:

[0101]

[0102] Among them, under the same waveform, let , take the DC component of the waveform as , then there is a difference between any point on the waveform and . For convenience, take the difference between the maximum value in the waveform and the DC component, and calculate . The DC under different waveforms is equal.

[0103] First, for the waveform signals under green light, red light, and infrared light in a fixed time period, there is the following formula for each peak point:

[0104]

[0105] Then, take the logarithm of both sides of the equation to get and , and through a certain ratio calculation, the expression is as follows:

[0106]

[0107] Then, according to the subsequent equation derivation:

[0108]

[0109] Among them, k is a constant, and then the relationship between the concentration ratio of HbO2 and Hb and k is expressed as follows:

[0110]

[0111] In the band signals within the same time period, for the calculation of SpO2, it is crucial to solve the concentration ratio of oxyhemoglobin (HbO2) and deoxyhemoglobin (Hb) that is directly related to blood oxygen saturation based on different waveform signals. Due to the existence of different light sources (green light, red light, and infrared light), directly using different light sources for signal comparison is unreliable. As can be seen from the derived formula, the blood oxygen value is mainly related to the value of p of two band light beams (the commonly used method is to use infrared and red light), and the p value is the ratio of the AC and DC components in the PPG data of this band. The DC component generally uses the baseline of the signal, and the difference between any point on the waveform and the baseline can be used as the AC. The commonly used method is to select the maximum value point, and then fit a linear or polynomial equation. However, since the PPG signal is easily interfered by various noises, the maximum value point is prone to distortion, which in turn affects the accuracy of blood oxygen.

[0112] Therefore, the present invention obtains PPG data under 3 bands (green light, red light, and infrared light), intercepts samples according to the waveform, and obtains the value of "p" by using the complete cycle as a sample, and then obtains the final blood oxygen value.

[0113] S105, calculate different peaks based on the waveform of the entire PPG physiological signal, intercept the time period according to the number of peaks in different bands, and obtain the DC and AC time lengths of each band of the PPG physiological signal of different photoelectric sensors as band signals ;

[0114] S106, perform decision-level fusion of periodic band time period signals based on decision rules to construct a PPG physiological signal feature vector. When performing decision-level fusion of periodic band time period signals based on decision rules, independently analyze the 3 band signals of each photoelectric sensor, and then merge the analysis results according to the decision rules. For the valid signals of each band, if two or more band signals under each photoelectric sensor are considered valid PPG signals, then the photodiode PD in this photoelectric sensor is considered. Select the most effective PPG signal in each photoelectric sensor according to the decision result. For example, the band signal with the highest similarity can be selected as the final PPG signal, and then construct a PPG physiological signal feature vector. The PPG physiological signal feature vector is expressed as:

[0115]

[0116] In the embodiments of the present invention, through wavelet transform and reconstruction, high-frequency noise and interference in the signal can be effectively removed, the main features of the signal can be retained, the quality of the signal can be improved. By calculating the average value through a moving window, the change of the DC signal can be tracked in real time to ensure the accuracy and timeliness of the DC signal. Through PPG signal detection, the optimal PPG signals in different bands can be accurately identified, thereby improving the representativeness of the signal. The decision rule can be adjusted according to different users and environments to improve the adaptability and flexibility of the model. By selecting the most effective PPG signal, a high-quality feature vector can be constructed, providing accurate input for subsequent blood oxygen saturation calculation. The preprocessing method not only removes noise and interference, but also enhances the useful components of the signal, improving the measurement accuracy and reliability. Through these preprocessing methods, high-quality input data can be provided for subsequent blood oxygen saturation calculation, thereby improving the performance and accuracy of the entire system.

[0117] The embodiments of the present invention provide a method for reconstructing and denoising PPG physiological signal data based on Daubechies wavelet basis functions. The method for reconstructing and denoising PPG physiological signal data based on Daubechies wavelet basis functions specifically includes:

[0118] S1011, obtain PPG physiological signal data, observe the time-domain waveform of the signal, and through spectral analysis, determine the distribution of the main frequency components of the signal and judge the position and bandwidth of the noise in the frequency domain;

[0119] It should be noted that when reconstructing and denoising PPG physiological signal data based on Daubechies wavelet basis functions, a comprehensive analysis of the original signal is required. This includes observing the time-domain waveform of the signal to understand its general change trend, amplitude range, and whether there are obvious abnormal fluctuations or mutation points. Through spectral analysis, determine the distribution of the main frequency components of the signal and judge the approximate position and bandwidth of the noise in the frequency domain.

[0120] S1012, preset the low-order Daubechies wavelet db2 and the high-order Daubechies wavelet db6, and judge whether the PPG physiological signal is smooth; according to whether the signal is relatively smooth and changes slowly, set the low-order Daubechies wavelet db2, which can effectively remove noise without overly complex calculations; if the signal has more complex details and higher frequency components, then set the high-order Daubechies wavelet db6. By calculating the decomposition level , decompose the signal, where is the signal length.

[0121] S1013, If the PPG physiological signal is a smooth signal, perform discrete wavelet decomposition on the DC and AC signals of the PPG physiological signal based on the preset Daubechies wavelet db2 and the decomposition level. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer;

[0122] S1014, If the PPG physiological signal is a non-smooth signal, perform discrete wavelet decomposition on the DC and AC signals of the PPG physiological signal based on the preset Daubechies wavelet db6 and the decomposition level. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer;

[0123] It should be noted that discrete wavelet decomposition (DWT) is performed on the DC and AC signals using the selected Daubechies wavelet basis function and the determined decomposition level. The original signal is used as the approximation signal of the bottom layer (the first layer), and through convolution operations with the low-pass filter and high-pass filter corresponding to the Daubechies wavelet, and downsampling operations, it is decomposed into the low-frequency coefficients (approximation coefficients) and high-frequency coefficients (detail coefficients) of the second layer. Then, the low-frequency coefficients of the second layer are continued to be decomposed in the same way, and the low-frequency coefficients and high-frequency coefficients of subsequent layers are obtained in turn until the predetermined decomposition level is reached. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer.

[0124] S1015, Integrate at least one set of high-frequency coefficients, process the high-frequency coefficients based on an adaptive threshold, use the minimum Stein unbiased risk estimation criterion to determine the threshold, and set the threshold T;

[0125] S1016, Determine whether the absolute value of the high-frequency coefficient is greater than the set threshold T;

[0126] S1017, If the absolute value of the high-frequency coefficient is less than or equal to the set threshold T, set the amplitude of the high-frequency coefficient to 0;

[0127] S1018, If the absolute value of the high-frequency coefficient is greater than the set threshold T, subtract the sign function multiple of the threshold T from the absolute value of the high-frequency coefficient to shrink the high-frequency coefficient, that is ;

[0128] In the embodiments of the present invention, the high-frequency coefficients are processed by an adaptive threshold method, and the Stein unbiased risk estimate criterion is used to minimize to determine the threshold, which better adapts to the noise of different layers and different signal localities. Set the threshold T. For the elements in the high-frequency coefficients of each layer, if the absolute value is greater than the threshold T, it indicates that the coefficient may contain more signal information but is also mixed with noise. Then subtract the sign function multiple of the threshold T from it, which shrinks the coefficient to a certain extent. If the absolute value is less than or equal to the threshold T, when the coefficient amplitude is small and lower than the set threshold, it is very likely to belong to noise, and it is set to 0, which can effectively remove this part of the suspected noise signal and reduce the interference of noise on the reconstructed signal.

[0129] S1019, Load the high-frequency coefficients and low-frequency coefficients after threshold processing, and use the inverse discrete wavelet transform algorithm to perform wavelet reconstruction processing on the high-frequency coefficients and low-frequency coefficients to obtain the reconstructed and denoised AC and DC signals.

[0130] In this embodiment, the high-frequency coefficients of each layer after threshold processing and the low-frequency coefficients of the last layer are used to perform wavelet reconstruction according to the inverse discrete wavelet transform (IDWT) algorithm. Starting from the top layer (the last layer of decomposition), the processed high-frequency coefficients and low-frequency coefficients are convolved through upsampling and the reconstruction filter corresponding to the Daubechies wavelet to gradually synthesize the coefficients of the next layer, and so on until the entire denoised DC and AC signals are reconstructed.

[0131] The embodiments of the present invention provide a method for intercepting time periods according to the number of peaks in different bands. The method for intercepting time periods according to the number of peaks in different bands specifically includes:

[0132] S1051, Obtain the PPG physiological signal and record the positions of the signal peak points;

[0133] S1052, Perform band periodicity detection on the PPG physiological signal to identify the bands in the PPG physiological signal, and use the peak detection algorithms of the first derivative and the second derivative to identify the peaks of the bands, expressed as:

[0134]

[0135] Among them, Peaks1 represents the first derivative, and the peak appears at the point where it changes from positive to negative. Peaks2 represents the second derivative, and the peak appears where the second derivative is negative;

[0136] S1053, Load the band periodicity detection results, and intercept the time period of the PPG physiological signal based on the band periodicity detection results;

[0137] Among them, when intercepting the time period, determine the number of peaks according to the peak solution result of the band , Periodic time segment extraction , A certain peak point before the first peak is used as the start time t start , The peak point corresponding to the start time after the peak point is used as the end time t end , So the total intercepted time period is represented by M as ;

[0138] S1054, The PPG physiological signal after time segment extraction is processed by a scaling method for time segment alignment to ensure that the time segment lengths of different waveforms are the same, and the fixed time segment length is calculated as follows:

[0139]

[0140] During the time segment extraction process, it is necessary to determine the time segment alignment method. The scaling method is adopted, and a scaling ratio is added after each waveform of a fixed length as a mark. The scaling ratio is calculated based on the ratio of the actual length of the time segment to the fixed length ;

[0141] S1055, The time segment extraction of all bands is carried out synchronously, and the physiological information at the same time is extracted to obtain the band signals with the DC and AC time lengths of for each band of the PPG physiological signals of different optoelectronic sensors .

[0142] In the embodiment of the present invention, the end-to-end E2E model includes an input layer, a CFCs module, a residual connection module, a multi-layer perceptron, and an output layer. The CFCs module is used to extract features from the PPG physiological signal feature vector to obtain a signal feature set, and the multi-layer perceptron is used to identify and analyze the signal feature set and output the SpO2 value under the periodic waveform;

[0143] It should be noted that in the design and construction of the end-to-end (E2E) model, an input layer is designed to receive the preprocessed PPG signal and other relevant parameters to form a feature vector. The CFCs module is used to reduce the computational load and accelerate the training and inference speed, approximating the analytical solution of the neural dynamics system. Through the layer-by-layer training and feature extraction of three layers of CFCs, and residual connections between the CFCs, important parameters affecting blood oxygen changes are captured. Then, through a multi-layer perceptron, an accurate SpO2 value under the periodic waveform is output. The end-to-end (E2E) model reduces the need for feature engineering, simplifies the development process, makes model training and prediction more efficient, and fully utilizes the advantages of the end-to-end (E2E) model in the training and prediction processes, that is, without complex feature engineering, it can directly learn the prediction results of blood oxygen saturation from the original data. In the end-to-end (E2E) model, the closed-form solution with explicit time dependence (CFCs) can improve the accuracy of feature extraction and the efficiency of the entire training process.

[0144] In the embodiment of the present invention, the designed input layer receives the preprocessed PPG signal as the input signal. The length of the time series is T. There are a total of 4 PDs in the PPG acquisition, each PD has 3 wavelengths, and the AC component of each wavelength has t. The time period is intercepted through the number of periodic waveforms, and there are l sub-time periods. These features serve as components of the input signal of.

[0145] The CFCs module includes the first layer of CFCs, the second layer of CFCs, and the third layer of CFCs. The first layer of CFCs is connected to the input layer, and the first layer of CFCs, the second layer of CFCs, and the third layer of CFCs are connected through a residual connection module. The residual connection module includes the first layer of residual connection, the second layer of residual connection, and the third layer of residual connection. The first layer of residual connection, the second layer of residual connection, and the third layer of residual connection are respectively arranged after the first layer of CFCs, the second layer of CFCs, and the third layer of CFCs.

[0146] The multi-layer perceptron (MLP) includes three fully connected layers. The input dimension of the multi-layer perceptron matches the output dimension of the CFCs to be , and dimensionality reduction is performed through three fully connected layers. The final output dimension is 1. The fully connected layer is used to map the input features to the output space through a linear transformation.

[0147] The input dimension of the multi-layer perceptron matches the output dimension of the CFCs to be , and dimensionality reduction is performed through three fully connected layers. The final output dimension is 1. The purpose of the fully connected layer (linear layer) is to map the input features to the output space through a linear transformation. First, the weights of the fully connected layer are initialized, and they are initialized through a normal distribution. Assume the input feature is , and the input feature dimension is , the output dimension of the first fully connected layer is , and the output dimension of the second fully connected layer is , the output dimension of the third layer is 1, and the corresponding weight matrices for each layer are , .

[0148] Among them, in the forward propagation process, the input features are linearly transformed through the weight matrix and then non-linearly transformed through the activation function to obtain the input of the next layer. Then the output of the first fully connected layer can be expressed as:

[0149]

[0150] Among them, is the bias vector of the first layer, is the Sigmoid activation function. Similarly, the outputs of the second fully connected layer and the third fully connected layer are respectively expressed as follows:

[0151]

[0152] In the backpropagation process, the loss function is calculated at the last layer of the network output layer, and regularization techniques are used to prevent overfitting. The loss between the predicted value and the true value is calculated at the network output layer, and the loss function is expressed as follows:

[0153]

[0154] Among them, is the true value, is the output of the fully connected layer (predicted value). The gradient of the loss function will be backpropagated to the first two layers at the output layer to update the weights of these layers.

[0155] In this embodiment, a neuron is set in the output layer, and the Sigmoid function is used as the activation function of the output layer to calculate the predicted value of the output layer , and then the output value is multiplied by 100, and the result is clipped to the interval [0, 100]:

[0156]

[0157] It should be noted that when building the end-to-end E2E model, a training data set marked with the true blood oxygen saturation value is used to train the end-to-end model, optimize the model parameters, test the model performance on an independent validation set to ensure the accuracy and generalization ability of the model, deploy a wristband-type blood oxygen detection device in the plateau area, continuously monitor the blood oxygen level, evaluate the adaptability of individuals to the plateau environment, and provide important health assessment data for the residents and travelers in the plateau area.

[0158] In an embodiment of the present invention, an end-to-end (E2E) model is provided. The end-to-end (E2E) model is driven by a multi-band periodic waveform and supports real-time and continuous blood oxygen saturation monitoring, which helps to promptly detect abnormal changes in blood oxygen levels and provide early health warnings for individuals in high-altitude environments. The E2E model algorithm can learn and compensate among different users, thereby achieving personalized adjustment and being able to adapt to the physiological characteristics of different users.

[0159] An embodiment of the present invention provides a method for extracting signal features of a PPG physiological signal feature vector by the end-to-end (E2E) model based on an explicit time-dependent closed-form solution feature extraction algorithm. Figure 4 The following shows a schematic implementation flowchart of the method for extracting signal features of a PPG physiological signal feature vector by the end-to-end (E2E) model based on an explicit time-dependent closed-form solution feature extraction algorithm. The method for extracting signal features of a PPG physiological signal feature vector by the end-to-end (E2E) model based on an explicit time-dependent closed-form solution feature extraction algorithm specifically includes:

[0160] S201, load the PPG physiological signal feature vector and input the PPG physiological signal feature vector into the input layer of the end-to-end (E2E) model;

[0161] S202, the PPG physiological signal feature vector input by the input layer first performs feature extraction through the first layer of CFCs, and through , and a neural network to simulate the dynamic behavior of neurons, and the output is expressed as ;

[0162] S203, the first-layer residual connection adds the output of the first layer of CFCs to the input to form the first-layer residual connection, and the output is expressed as ;

[0163] S204, the result of the first-layer residual connection is used as the input of the second layer of CFCs, and the second layer of CFCs applies the neural network , and again for feature extraction, expressed as ;

[0164] S205, the second-layer residual connection adds the output of the second layer of CFCs to the result of the first-layer residual connection to form the second-layer residual connection, and the output is expressed as ;

[0165] S206. The result of the second - layer residual connection serves as the input to the third - layer CFCs. The third - layer CFCs apply the neural network again , and perform feature extraction, denoted as ;

[0166] S207. The third - layer residual connection adds the output of the third - layer CFCs to the result of the second - layer residual connection to form the third - layer residual connection, and the output is denoted as ;

[0167] S208. Integrate the output result of the third - layer residual connection into a signal feature set.

[0168] In this embodiment, to design the entire process of feature extraction and training, CFCs are used to significantly reduce the computational amount and accelerate the training and inference speed. The basic idea of CFCs is to approximate the analytical solution of the neural dynamics system through a network - parameterized function. Based on the simplified LiquidNN model, each node in the network represents a neuron, and the connections between nodes represent synapses. The approximate solution of CFCs can be expressed as:

[0169]

[0170] where is the potential of the neuron at the initial time. is its synaptic reversal potential, serving as a bias. is the input signal, After passing through the synaptic non - linear transformation function, the neural network is used to replace in the formula, and is expressed as follows:

[0171]

[0172] where is a neural network parameterized by the parameter parameters.

[0173] Since will rapidly approach 0 as the time t increases, resulting in the gradients received by the subsequent layers of the network becoming smaller and smaller when passing through the backpropagation algorithm. The hyperbolic tangent (tanh) function σ is used to replace the exponential decay term, and is expressed as follows:

[0174]

[0175] Use the neural network to approximate ; Similarly, use the neural network to approximate , the update of the state can be expressed as:

[0176] Thus, the final .

[0177] The embodiment of the present invention provides a blood oxygen estimation system based on multi-band periodic waveform driving. Figure 5 FIG. shows a schematic structural diagram of a blood oxygen estimation system based on multi-band periodic waveform driving. The blood oxygen estimation system based on multi-band periodic waveform driving specifically includes:

[0178] A wristband-type physiological data acquisition module 100, which is used to collect PPG physiological signal data in real time;

[0179] A data preprocessing module 200, which is used to preprocess the PPG physiological signal data, intercept the periodic band time period signal according to the number of peaks, fuse the periodic band time period signals at the decision-making level based on decision rules, and construct a PPG physiological signal feature vector;

[0180] A blood oxygen detection module 300, which is used to pre-construct an end-to-end E2E model, take the PPG physiological signal feature vector as input, execute the end-to-end E2E model. The end-to-end E2E model performs signal feature extraction on the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combines a multi-layer perceptron to identify and analyze the signal feature set, and outputs the SpO2 value under the periodic waveform.

[0181] In summary, the present invention provides a blood oxygen estimation method and system based on multi-band periodic waveform driving. In the embodiment of the present invention, the periodic time period is dynamically intercepted through band peak detection, and the signal features are efficiently extracted directly from the original data by using a feature extraction algorithm based on CFCs, which improves the processing efficiency. The end-to-end E2E model can adaptively learn the physiological characteristics of different users, and perform real-time compensation for the motion interference in the plateau environment, improve the measurement robustness, optimize the calculation amount through the CFCs algorithm, speed up the inference speed, meet the low-power and real-time monitoring requirements in special scenarios such as the plateau, and thus more accurately detect the abnormal changes in blood oxygen, providing reliable support for health early warning.

[0182] It should be noted that for the foregoing embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps may be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0183] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting the protection scope of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict and without creative efforts, combine, add or delete the features in the embodiments of the present invention according to the situation or make other adjustments, so as to obtain different technical solutions that essentially do not deviate from the concept of the present invention, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A blood oxygen estimation method based on multi-band periodic waveform driving, characterized in that, Including: Real-time collection of PPG physiological signal data based on a wristband-type physiological data collection module; Load the PPG physiological signal data, preprocess the PPG physiological signal data, intercept the periodic band time segment signal according to the number of peaks, perform decision-level fusion on the periodic band time segment signal based on decision rules, and construct a PPG physiological signal feature vector; Pre-construct an end-to-end E2E model, use the PPG physiological signal feature vector as the input, execute the end-to-end E2E model, and the end-to-end E2E model performs signal feature extraction on the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combines a multi-layer perceptron to identify and analyze the signal feature set, and outputs the SpO2 value under the periodic waveform; The method for the end-to-end E2E model to perform signal feature extraction on the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm includes: Load the PPG physiological signal feature vector and input the PPG physiological signal feature vector into the input layer of the end-to-end E2E model; The PPG physiological signal feature vector input by the input layer First, feature extraction is performed through the first layer of CFCs. Through , and The neural network simulates the dynamic behavior of neurons, and the output is expressed as ; The first-layer residual connection adds the output of the first-layer CFCs to the input to form the first-layer residual connection, and the output is denoted as ; The result of the first-layer residual connection is used as the input of the second-layer CFCs, and the second-layer CFCs apply the neural network again , and perform feature extraction, denoted as ; The second-layer residual connection adds the output of the second-layer CFCs to the result of the residual connection of the first layer to form the residual connection of the second layer, and the output is expressed as ; The result of the second - layer residual connection is used as the input of the third - layer CFCs, and the third - layer CFCs apply the neural network again , and for feature extraction, denoted as ; The third-layer residual connection adds the output of the third-layer CFCs to the result of the residual connection of the second layer to form the residual connection of the third layer, and the output is expressed as ; Integrate the output result of the third-layer residual connection into a signal feature set.

2. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 1, wherein: The wristband-type physiological data collection module includes: Four photoelectric sensors, which are used to collect PPG physiological signal data in real time, and the photoelectric sensors are symmetrically distributed at 4 points on the inner side of the wearable device; A multi-band LED lamp, which is deployed in the middle area of the photoelectric sensor and adopts a concave design, and the multi-band LED lamp is used to assist the photoelectric sensor to work; A microcontroller MCU, which is used to control the multi-band LED lamp to collect PPG signals of different bands and different points in a time-sharing manner, and converts the PPG analog signal into a PPG digital signal under different bands through a high-sampling-rate analog-to-digital converter ADC.

3. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 1, wherein: The method for preprocessing the PPG physiological signal data includes: Load the PPG physiological signal data, perform spectral analysis on the PPG physiological signal data, observe the signal time-domain waveform, and reconstruct and denoise the PPG physiological signal data based on the Daubechies wavelet basis function to obtain the reconstructed and denoised AC and DC signals; Load the reconstructed and denoised AC and DC signals, and perform filtering processing on the AC and DC signals. Among them, when performing filtering processing on the AC and DC signals, preprocess the original signals of different wavelengths collected by each group of photoelectric sensors, extract the DC signal through a low-pass filter with a cut-off frequency of 0.1 Hz, calculate the average value of the PPG physiological signal within a predetermined time window as an estimate of the DC component, and continuously update the average value as the window moves to track the change of the DC signal in real time. Use a high-pass filter with a cut-off frequency of 0.5 Hz to 10 Hz to allow the AC component to pass through and filter out the DC component and low-frequency interference to obtain the AC signal. The AC and DC signals under each waveform are expressed as: Obtain the PPG physiological signal data after AC and DC signal filtering processing, perform PPG signal detection on the PPG physiological signal data, and output the AC and DC of the best PPG signals in different bands identified from each group of optoelectronic sensors; Calculate different peaks based on the waveforms of the entire PPG physiological signal, intercept time periods according to the number of peaks in different bands, and obtain the DC and AC time lengths of each band of the PPG physiological signal of different optoelectronic sensors as band signals ; Decision-level fusion of periodic band time-segment signals based on decision rules to construct a PPG physiological signal feature vector. When performing decision-level fusion of periodic band time-segment signals based on decision rules, the three-band signals of each optoelectronic sensor are analyzed independently, and then the analysis results are combined according to the decision rules. For the valid signals of each band, if two or more band signals under each optoelectronic sensor are considered valid PPG signals, then the photodiode PD in that optoelectronic sensor is considered. According to the decision results, the most effective PPG signal in each optoelectronic sensor is selected to construct the PPG physiological signal feature vector. The PPG physiological signal feature vector is expressed as: .

4. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 3, wherein: The method for reconstructing and denoising the PPG physiological signal data based on the Daubechies wavelet basis function includes: Obtain PPG physiological signal data, observe the time-domain waveform of the signal, through spectral analysis, determine the distribution of the main frequency components of the signal, and judge the position and bandwidth of the noise in the frequency domain; Preset the low-order Daubechies wavelet db2 and the high-order Daubechies wavelet db6, and judge whether the PPG physiological signal is smooth; If the PPG physiological signal is a smooth signal, perform discrete wavelet decomposition on the DC and AC signals of the PPG physiological signal based on the preset Daubechies wavelet db2 and the decomposition level. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer; If the PPG physiological signal is a non-smooth signal, perform discrete wavelet decomposition on the DC and AC signals of the PPG physiological signal based on the preset Daubechies wavelet db6 and the decomposition level. During the decomposition process, record the low-frequency coefficients and high-frequency coefficients of each layer; Integrate at least one set of high-frequency coefficients, process the high-frequency coefficients based on an adaptive threshold, use the minimum Stein unbiased risk estimation criterion to determine the threshold, and set the threshold T; Judge whether the absolute value of the high-frequency coefficient is greater than the set threshold T; If the absolute value of the high-frequency coefficient is less than or equal to the set threshold T, set the amplitude of the high-frequency coefficient to 0; If the absolute value of the high-frequency coefficient is greater than the set threshold T, subtract the sign function multiple of the threshold T from the absolute value of the high-frequency coefficient to shrink the high-frequency coefficient, that is Load the high-frequency and low-frequency coefficients after threshold processing, and use the inverse discrete wavelet transform algorithm to perform wavelet reconstruction processing on the high-frequency and low-frequency coefficients to obtain the reconstructed and denoised AC and DC signals.

5. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 3, wherein: When performing PPG signal detection on the PPG physiological signal data, preprocess the original signals collected in 3 bands under each photoelectric sensor, use template matching technology to identify whether there is a PPG signal in each photoelectric sensor, and select the signal with the highest similarity as the effective PPG signal. First, a dataset containing PPG signals of different individuals and under different conditions is known. Then, high-quality PPG signal samples are selected from the dataset as PPG signal sample templates. The PPG signal samples have clear waveform features, obvious peaks and valleys; Extract the same features as the PPG signal sample template from the original signals collected by each photoelectric sensor, at the same positions of the peaks and valleys, and use the correlation coefficient to calculate the similarity between the AC and DC of the PPG signal of each photoelectric sensor and the AC and DC of the PPG template; Where The formula for calculating the correlation coefficient: Select the signals with similarity higher than the threshold as valid PPG signals according to the similarity calculation results. The threshold is determined by evaluating the classification accuracy through classification experiments, and output the AC and DC of the best PPG signals in different bands identified from each group of optoelectronic sensors.

6. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 4, characterized in that: The method of intercepting time periods according to the number of peaks in different bands includes: Obtain the PPG physiological signal and record the positions of the signal peak points; Perform band periodicity detection on the PPG physiological signal, identify the bands in the PPG physiological signal, and use the peak detection algorithms of the first derivative and the second derivative to identify the peaks of the bands, expressed as: Among them, Peaks1 represents the first derivative, and Peaks2 represents the second derivative; Load the band periodicity detection results, and intercept the time periods of the PPG physiological signal based on the band periodicity detection results; Among them, when intercepting the time period, the number of peak values is determined according to the peak value solving result of the wave band , and the periodic time period is intercepted , a certain wave peak point before the first peak value is used as the start time t start , the wave peak point corresponding to the start time after the wave peak point is used as the end time t end , so the total intercepted time period is represented by M as ; After intercepting the PPG physiological signal in the time period, a scaling method is used for time period alignment processing to ensure that the time period lengths of different waveforms are the same, and the fixed time period length is calculated, which is expressed as follows: Using the scaling method, a scaling ratio is added after each waveform of a fixed length as a mark, and the scaling ratio is calculated based on the ratio of the actual length of the time period to the fixed length ; Synchronously perform time period truncation for all bands, extract physiological information at the same time, and obtain the DC and AC time lengths of each band of different optoelectronic sensor PPG physiological signals as band signals .

7. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 2, wherein: The end-to-end E2E model includes an input layer, a CFCs module, a residual connection module, a multi-layer perceptron, and an output layer. The CFCs module is used to extract features from the PPG physiological signal feature vector to obtain a signal feature set, and the multi-layer perceptron is used to identify and analyze the signal feature set and output the SpO2 value under the periodic waveform; The CFCs module includes the first-layer CFCs, the second-layer CFCs, and the third-layer CFCs. The first-layer CFCs are connected to the input layer, and the first-layer CFCs, the second-layer CFCs, and the third-layer CFCs are connected through a residual connection module. The residual connection module includes the first-layer residual connection, the second-layer residual connection, and the third-layer residual connection, and the first-layer residual connection, the second-layer residual connection, and the third-layer residual connection are respectively arranged after the first-layer CFCs, the second-layer CFCs, and the third-layer CFCs.

8. The blood oxygen estimation method based on multi-band periodic waveform driving according to claim 7, characterized in that: The multi-layer perceptron includes three fully-connected layers, and the input dimension of the multi-layer perceptron matches the output dimension of the CFCs as , and dimensionality reduction is performed through three fully-connected layers. The final output dimension is 1. The fully-connected layer is used to map the input features to the output space through a linear transformation.

9. A blood oxygen estimation system driven by a multi-band periodic waveform, for implementing the blood oxygen estimation method driven by a multi-band periodic waveform according to any one of claims 1-8, characterized in that: The blood oxygen estimation system based on multi-band periodic waveform driving includes: A wristband-type physiological data acquisition module, which is used to collect PPG physiological signal data in real time; A data preprocessing module, which is used to preprocess the PPG physiological signal data, intercept the periodic band time segment signal according to the number of peaks, make a decision-level fusion of the periodic band time segment signal based on decision rules, and construct a PPG physiological signal feature vector; A blood oxygen detection module, which is used to pre-build an end-to-end E2E model, use the PPG physiological signal feature vector as the input, execute the end-to-end E2E model, and the end-to-end E2E model performs signal feature extraction on the PPG physiological signal feature vector based on an explicit time-dependent closed-form solution feature extraction algorithm to obtain a signal feature set, and combines a multi-layer perceptron to identify and analyze the signal feature set, and outputs the SpO2 value under the periodic waveform.

Images