A blood oxygen detection calculation method based on infrared light and visible light

Abstract

The present invention discloses a blood oxygen detection and calculation method based on infrared and visible light, relating to the technical field of blood characteristic measurement. The method comprises: obtaining physiological indicator information and pulse wave detection information of each wristband wearer, wherein the physiological indicator information includes at least one of body fat information and age information, and the pulse wave detection information includes infrared light pulse wave signals and visible light pulse wave signals; clustering each pulse wave detection information based on the physiological indicator information to obtain multiple pulse wave detection groups; determining the light absorption anomaly of each pulse wave detection information based on the infrared light pressure value and visible light pressure value of each pulse wave detection information in the pulse wave detection group; screening a target training data set from each pulse wave detection group based on each light absorption anomaly; and training a target model based on the target training data set to obtain a blood oxygen detection model, so as to implement blood oxygen detection and calculation through the blood oxygen detection model. The present invention can improve the accuracy of blood oxygen detection and calculation.

Description

Technical Field

[0001] The present invention relates to the technical field of blood characteristic measurement, and in particular to a blood oxygen detection and calculation method based on infrared light and visible light. Background Art

[0002] Fitness trackers are commonly equipped with a blood oxygen monitoring function, which focuses on measuring the levels of oxyhemoglobin and deoxyhemoglobin. This function works because different types of hemoglobin have different absorption characteristics under infrared and visible light. By measuring and analyzing these absorption differences, the levels of oxyhemoglobin and deoxyhemoglobin can be estimated.

[0003] In practical applications, a fitness tracker collects pulse wave signals from the human wrist under infrared and visible light. These signals and the resulting blood oxygen saturation are combined into a training dataset. This dataset is then used to train a neural network model, enabling more efficient and accurate blood oxygen detection.

[0004] However, a major problem currently faced is that the overall accuracy of the pulse wave signal measured by the sports bracelet is poor, which directly leads to an increase in the analysis error of the trained blood oxygen detection model, thereby affecting the accuracy of the blood oxygen detection calculation. Summary of the Invention

[0005] The embodiment of the present invention provides a blood oxygen detection and calculation method based on infrared light and visible light, which can reduce the error of blood oxygen detection calculation and improve the accuracy of blood oxygen detection calculation.

[0006] According to a first aspect of an embodiment of the present invention, a blood oxygen detection and calculation method based on infrared light and visible light is provided, the method comprising:

[0007] Acquiring physiological indicator information and pulse wave detection information of each wristband wearer, where the physiological indicator information includes at least one of body fat information and age information, and the pulse wave detection information includes infrared light pulse wave signals and visible light pulse wave signals;

[0008] Based on the physiological index information, each pulse wave detection information is clustered to obtain multiple pulse wave detection groups;

[0009] Determining the light absorption abnormality of each pulse wave detection information based on the infrared light pressure value and the visible light pressure value of each pulse wave detection information in the pulse wave detection group, wherein the light absorption abnormality is used to represent the degree of deviation abnormality between the pulse wave detection information and the corresponding pulse wave detection group;

[0010] Based on the degree of abnormal light absorption, a target training data set is selected from each pulse wave detection group;

[0011] Based on the target training data set, the target model is trained to obtain a blood oxygen detection model, so as to realize blood oxygen detection calculation through the blood oxygen detection model.

[0012] Furthermore, the present invention also proposes that, based on the infrared light pressure value and the visible light pressure value of each pulse wave detection information in the pulse wave detection group, the light absorption abnormality of each pulse wave detection information is determined, which may specifically include:

[0013] Performing mean processing on each infrared light pressure value and each visible light pressure value of each pulse wave detection information in the pulse wave detection group to obtain an infrared light pressure mean value and a visible light pressure mean value;

[0014] The light absorption abnormality of the target pulse wave detection information is determined by using the difference between the infrared light pressure value and the infrared light pressure mean of the target pulse wave detection information, and the difference between the visible light pressure value and the visible light pressure mean. The target pulse wave detection information is any pulse wave detection information.

[0015] Furthermore, the present invention also proposes that, before selecting a target training data set from each pulse wave detection group based on each light absorption abnormality, the method may further include:

[0016] For the infrared light pulse wave signal and the visible light pulse wave signal of the pulse wave detection information, the area between two adjacent valley points is determined as a segmented area, thereby obtaining a plurality of first segmented areas of the infrared light pulse wave signal and a plurality of second segmented areas of the visible light pulse wave signal;

[0017] For a target segmented area, determining a motion interference parameter of the target segmented area based on a time span and a pulse pressure of the target segmented area, where the target segmented area is any one of the first segmented area and the second segmented area;

[0018] Determining the pulse wave error of the pulse wave detection information by using the motion interference parameters of the pulse wave detection information and the light absorption abnormality of the pulse wave detection information;

[0019] Determining a training priority coefficient of the pulse wave detection information by utilizing the pulse wave error of the pulse wave detection information;

[0020] Based on the degree of light absorption anomaly, a target training data set is selected from each pulse wave detection group, which may include:

[0021] Based on each training priority coefficient, a target training data set is screened from each pulse wave detection group.

[0022] Furthermore, the present invention also proposes that the pulse pressure includes a pulse DC component and a pulse AC component;

[0023] For the target segmented area, based on the time span and pulse pressure of the target segmented area, the motion interference parameter of the target segmented area is determined, which may specifically include:

[0024] Determine the periodic regularity of the target segmented area based on the time span of the target segmented area;

[0025] determining the motion interference tendency of the target segmented area based on the pulse DC component and the pulse AC component at each moment in the target segmented area;

[0026] The motion interference parameters of the target segmented area are determined by using the periodic regularity presentation degree and motion interference tendency degree of the target segmented area.

[0027] Furthermore, the present invention also proposes that, based on the time span of the target segmented area, determining the periodic regularity of the target segmented area may specifically include:

[0028] Obtaining a first time span mean value corresponding to a segmented region in a target pulse wave signal and a second time span mean value corresponding to a segmented region in a reference pulse wave signal, where the target pulse wave signal is the pulse wave signal belonging to the target segmented region, and the reference pulse wave signal is the pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal;

[0029] Determining the first period feature conformity of the target segmented area by using the time span of the target segmented area and the first time span mean;

[0030] Determining the second period feature conformity of the reference segmented region using the time span of the reference segmented region and the second time span mean, where the reference segmented region is a segmented region in the reference pulse wave signal corresponding to the target segmented region;

[0031] The periodic regularity presentation degree of the target segmented area is determined based on the first periodic feature conformity and the second periodic feature conformity.

[0032] Furthermore, the present invention also proposes that, based on the first periodic feature conformity and the second periodic feature conformity, determining the periodic regularity presentation degree of the target segmented area may specifically include:

[0033] The first calculated value is obtained by taking the absolute value of the difference between the first period characteristic conformity and the second period characteristic conformity;

[0034] Normalizing the reciprocal of the first calculated value to obtain a second calculated value;

[0035] The periodic regularity presentation degree of the target segmented area is determined using the second calculated value and the first periodic feature conformity.

[0036] Furthermore, the present invention also proposes determining the motion interference tendency of the target segmented area based on the pulse DC component and pulse AC component at each moment in the target segmented area, which may specifically include:

[0037] Based on the pulse DC component and the pulse AC component at each moment in the target segmented area, respectively determining the component transfer interference degree at each moment in the target segmented area;

[0038] Obtaining the Pearson correlation coefficient of pulse pressure at each moment in the target segmented region and the reference segmented region, where the reference segmented region is the segmented region in the reference pulse wave signal corresponding to the target segmented region, the reference pulse wave signal is the pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal, and the target pulse wave signal is the pulse wave signal belonging to the target segmented region;

[0039] The motion interference tendency of the target segmented area is determined using the transfer interference degree of each component and the Pearson correlation coefficient.

[0040] Furthermore, the present invention also proposes to determine the component transfer interference degree at each moment in the target segmented area based on the pulse DC component and the pulse AC component at each moment in the target segmented area, which may specifically include:

[0041] Obtaining a first slope of a pulse DC component and a second slope of a pulse AC component at a target time in a target segmented area, where the target time is any time in the target segmented area;

[0042] The ratio of the first slope to the second slope is normalized to obtain the component transfer interference degree at the target time in the target segment area.

[0043] Furthermore, the present invention also proposes that obtaining the Pearson correlation coefficient of the pulse pressure in the target segmented area and the reference segmented area at each moment may specifically include:

[0044] Taking the target moment in the target segmented area as the center, extend the target duration to both sides to obtain the first analysis period, where the target moment is any moment in the target segmented area;

[0045] Taking the target moment in the reference segment area as the center, extend the target duration to both sides to obtain the second analysis period;

[0046] The first analysis period is compared with the second analysis period to obtain the Pearson correlation coefficient of the pulse pressure in the target segmented area and the reference segmented area at the target time.

[0047] Furthermore, the present invention also proposes that, based on the target training data set, a target model is trained to obtain a blood oxygen detection model, which may specifically include:

[0048] Divide the target training dataset into multiple training batches;

[0049] Performing data augmentation and data preprocessing on the first training data samples in each training batch to obtain second training data samples;

[0050] According to the priority order of the training batches, the model parameters of the target model are adjusted based on the second training data samples in the training batches in sequence until the preset training end conditions are met to obtain a blood oxygen detection model.

[0051] The present invention has the following beneficial effects:

[0052] In the infrared and visible light-based blood oxygen measurement calculation method provided in an embodiment of the present invention, physiological indicators (such as body fat and age) and pulse wave detection information (infrared and visible light pulse wave signals) of the wristband wearer are first obtained. Physiological indicators can reflect the impact of individual differences on the pulse wave signal. Next, the pulse wave detection information is clustered based on the physiological indicators, grouping similar individual data into the same group to form multiple pulse wave detection groups, which facilitates subsequent analysis of different group characteristics. The light absorption anomaly degree is then determined based on the infrared and visible light pressure values ​​of each pulse wave detection information within the group. The light absorption anomaly degree indicates the degree of deviation from the pulse wave detection information within the group and can effectively identify abnormal signals caused by factors such as wristband measurement errors and individual physiological fluctuations. The target training data set is then filtered based on the light absorption anomaly degree to eliminate abnormal signals and ensure training data quality. Finally, the target model is trained with a high-quality target training data set to obtain a blood oxygen detection model. Since the training data is accurate and reliable, the trained model is more accurate, which can reduce the error in blood oxygen detection calculation and improve the accuracy of blood oxygen detection calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 A schematic flow chart of a first blood oxygen detection and calculation method based on infrared light and visible light provided by an embodiment of the present invention;

[0055] Figure 2 A schematic flow chart of a second blood oxygen detection calculation method based on infrared light and visible light provided by an embodiment of the present invention;

[0056] Figure 3A schematic diagram of the division of segmented areas provided by one embodiment of the present invention;

[0057] Figure 4 A schematic diagram of the process of S202 provided in one embodiment of the present invention;

[0058] Figure 5 This is a flow chart of S402 provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0059] To further illustrate the technical means and effects of the present invention to achieve the intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effects of a blood oxygen detection and calculation method based on infrared and visible light proposed by the present invention. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable form.

[0060] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0061] It should be noted that the acquisition, storage, use, and processing of data in the technical solution of the present invention comply with the relevant provisions of laws and regulations.

[0062] It should be noted that in the embodiments of the present invention, certain software, components, models and other existing solutions in the industry may be mentioned. They should be regarded as exemplary and their purpose is only to illustrate the feasibility of implementing the technical solution of the present invention, but it does not mean that the applicant has or will necessarily use the solution.

[0063] Traditional blood oxygen detection methods rely on dual-channel infrared and visible light pulse wave signals to construct training datasets. Individual physiological differences and motion interference during signal acquisition can lead to non-uniform sample distribution in the original dataset. Physiological indicators such as body fat percentage and age affect the optical properties of subcutaneous tissue, resulting in differences in optical signal absorption intensity between individuals. Furthermore, limb movement in dynamic environments distorts the morphology of the photoplethysmography signal, reducing signal periodicity and amplitude stability. These factors combine to make it difficult for conventional methods to distinguish valid physiological features from noise when screening training data, resulting in non-ideal feature mapping relationships learned during model training.

[0064] For example, when a user wears a wristband during an indoor fitness session, the myoelectric interference and limb displacement generated by the exercise cause the baseline of the visible light channel signal to drift, while the pulse wave peak of the infrared light channel experiences abnormal attenuation. When directly using pulse wave signals containing such interference to generate training samples, individuals with a body fat percentage above 30% experience a deviation from the linear relationship between the infrared light absorption intensity and the standard model due to the scattering effect of the subcutaneous fat layer. In elderly users, changes in vascular elasticity cause the rising slope of the pulse wave to decrease. In this case, the training dataset contains both distorted and normal samples, but traditional clustering methods cannot effectively identify subsets of samples with similar physiological characteristics but different signal quality, resulting in gradient direction conflicts during the model weight update process.

[0065] Faced with the above problems, the present invention first considers optimizing the training dataset construction process through data stratification and quality screening. When traditional methods directly use raw signals to construct training sets, they do not fully analyze the impact of physiological differences on optical signal absorption, and ignore the signal morphological distortion caused by motion interference. The present invention recognizes that there are systematic differences in the optical properties of subcutaneous tissue of wearers with different body fat percentages and ages, and it is necessary to establish a correlation mechanism between physiological characteristics and signal quality in the data preprocessing stage. To this end, attempts are made to use individual physiological indicators as the basis for data grouping, and to form a subset of samples with similar optical absorption characteristics through clustering, thereby reducing the interference of intra-group signal intensity differences on model training. At the same time, a dynamic quality assessment indicator is introduced to address motion interference, and the degree of signal deviation is quantified by optical absorption anomalies. High-quality samples are screened out in combination with pulse wave period stability analysis. Compared with the existing technology that directly uses full data training, the present invention chooses to construct a hierarchical and multi-dimensional screening mechanism to improve the quality of the dataset through the dual strategy of physiological feature clustering and abnormal signal elimination.

[0066] In this regard, Figure 1 As shown, a flow chart of a blood oxygen detection and calculation method based on infrared light and visible light is provided. The blood oxygen detection and calculation method based on infrared light and visible light can be applied to a server. The blood oxygen detection and calculation method based on infrared light and visible light can include the following steps S101 to S105:

[0067] S101, obtaining physiological indicator information and pulse wave detection information of each wristband wearer, where the physiological indicator information includes at least one of body fat information and age information, and the pulse wave detection information includes an infrared light pulse wave signal and a visible light pulse wave signal;

[0068] S102, clustering each pulse wave detection information based on the physiological indicator information to obtain multiple pulse wave detection groups;

[0069] S103, determining a light absorption abnormality of each pulse wave detection information based on the infrared light pressure value and the visible light pressure value of each pulse wave detection information in the pulse wave detection group, where the light absorption abnormality is used to represent the degree of deviation abnormality between the pulse wave detection information and the corresponding pulse wave detection group;

[0070] S104, based on each light absorption abnormality, screening target training data sets from each pulse wave detection group;

[0071] S105: Based on the target training data set, the target model is trained to obtain a blood oxygen detection model, so as to implement blood oxygen detection calculation through the blood oxygen detection model.

[0072] In this embodiment, physiological indicator information refers to the wearer's body fat percentage or age data obtained through the wristband. This can be implemented using a bioelectrical impedance sensor or a user input interface. Physiological indicator information is used to reflect individual physiological differences, thereby more accurately dividing similar groups during the subsequent clustering process. For example, people with high body fat have higher fat content in wrist tissue. This causes excessive absorption and consumption of infrared and visible light in fat, leading to normal physiological differences in blood oxygen measurement data between different wearers. Older wearers also have slower blood circulation rates, which can also cause differences in blood oxygen measurement information. Therefore, to reduce the analytical errors caused by normal physiological differences in blood oxygen detection, clustering is performed based on the wearer's body fat, age, and other physiological indicator information to obtain multiple pulse wave detection groups.

[0073] Pulse wave detection information refers to the periodic pulse fluctuation signal collected by infrared and visible light sensors. It can be specifically achieved using photoelectric plethysmography technology. Pulse wave detection information is used to reflect the absorption characteristics of hemoglobin in the blood to light of different wavelengths.

[0074] Clustering refers to grouping and classifying pulse wave data according to physiological indicators. It can be implemented using the K-means algorithm or the hierarchical clustering algorithm. The role of clustering is to reduce the interference of individual physiological differences on pulse wave analysis and improve the consistency of data within the group.

[0075] Light absorption anomaly refers to the degree of deviation calculated by comparing the difference between infrared light pressure values ​​and visible light pressure values ​​within the same group. This can be achieved using the Euclidean distance or standardized residual method. Light absorption anomaly is used to identify abnormal pulse wave signals and eliminate low-quality data caused by measurement errors.

[0076] The target training dataset refers to the high-quality pulse wave data set retained after anomaly screening. This can be achieved by setting an anomaly threshold or sorting and selecting the best method. The role of the target training dataset is to provide reliable samples for model training and reduce the impact of noise data on model accuracy.

[0077] The blood oxygen detection model refers to a mapping model between pulse wave signals and blood oxygen saturation established through a machine learning algorithm. It can be implemented using a convolutional neural network or a support vector machine. The role of the blood oxygen detection model is to improve the accuracy of blood oxygen detection calculations through optimized training data.

[0078] The working process and principle of the present invention are as follows: first, the wristband wearer's physiological indicator information and pulse wave detection information are obtained. The physiological indicator information includes at least one of body fat information and age information, and the pulse wave detection information includes infrared light pulse wave signals and visible light pulse wave signals. This information serves as the basis for subsequent processing.

[0079] Next, the pulse wave detection information is clustered based on the physiological index information to obtain multiple pulse wave detection groups. This step groups data with similar physiological characteristics, providing a more accurate basis for subsequent analysis.

[0080] Next, the pulse wave detection information from each pulse wave detection group is analyzed. Based on the infrared and visible light pressure values, the light absorption anomaly of each pulse wave detection information is determined. This light absorption anomaly indicates the degree of deviation from the corresponding pulse wave detection group. This step identifies abnormal data within the pulse wave detection group and improves the accuracy of subsequent analysis.

[0081] Based on the calculated light absorption anomaly, the target training dataset is selected from each pulse wave detection group. This step improves the quality of the training dataset by eliminating abnormal data in the pulse wave detection group.

[0082] Finally, the target model is trained using the selected target training data set to obtain a blood oxygen detection model. The trained blood oxygen detection model can achieve more accurate blood oxygen detection calculations.

[0083] As an example, first obtain the wristband wearer's physiological indicator information and pulse wave detection information. Physiological indicator information includes body fat percentage and age, and pulse wave detection information includes infrared light pulse wave signals and visible light pulse wave signals. Using the K-means clustering algorithm, the pulse wave detection information is clustered based on body fat percentage and age, resulting in multiple pulse wave detection groups. For example, the data can be divided into a low body fat young group, a high body fat young group, a low body fat middle-aged group, a high body fat middle-aged group, and so on.

[0084] For each pulse wave detection group, the average of the infrared and visible light pressure values ​​is calculated. The difference between the pressure value and the average for each pulse wave detection is then calculated to determine the degree of light absorption anomaly. A light absorption anomaly threshold is set to filter out data with a light absorption anomaly below the threshold, forming the target training dataset.

[0085] A deep learning model, such as a long short-term memory network, is then used as the target model and trained using the target training dataset. Batch gradient descent is used to optimize model parameters until the model converges or reaches a preset number of training rounds. The resulting blood oxygen detection model can accept new pulse wave signals as input and output the corresponding blood oxygen saturation estimate.

[0086] This embodiment first obtains the wristband wearer's physiological indicators (such as body fat and age) and pulse wave detection information (infrared and visible light pulse wave signals). Physiological indicators can reflect the impact of individual differences on the pulse wave signal. Next, the pulse wave detection information is clustered based on the physiological indicators, grouping similar individual data into the same group to form multiple pulse wave detection groups. This facilitates subsequent analysis of different group characteristics. The light absorption anomaly is then determined based on the infrared and visible light pressure values ​​of each pulse wave detection information within the group. The light absorption anomaly indicates the degree of deviation from the pulse wave detection information within the group, effectively identifying abnormal signals caused by factors such as wristband measurement errors and individual physiological fluctuations. A target training dataset is then selected based on the light absorption anomaly, and abnormal signals are eliminated to ensure training data quality. Finally, a target model is trained using the high-quality target training dataset to obtain a blood oxygen detection model. Because the training data is accurate and reliable, the trained model is more accurate, thereby reducing errors in blood oxygen detection calculations and improving the accuracy of blood oxygen detection calculations.

[0087] In some of the aforementioned embodiments of the present invention, after clustering pulse wave detection information based on physiological indicator information, the light absorption abnormality of each pulse wave detection information needs to be determined. However, when determining the light absorption abnormality, if a unified benchmark for the pressure value distribution within the same detection group is not established, the abnormality calculation result may be affected by the discreteness within the group, and may not accurately reflect the degree of deviation between individual data and the overall status of the group.

[0088] In this regard, the present invention further proposes that S103 may specifically include:

[0089] Performing mean processing on each infrared light pressure value and each visible light pressure value of each pulse wave detection information in the pulse wave detection group to obtain an infrared light pressure mean value and a visible light pressure mean value;

[0090] The light absorption abnormality of the target pulse wave detection information is determined by using the difference between the infrared light pressure value and the infrared light pressure mean of the target pulse wave detection information, and the difference between the visible light pressure value and the visible light pressure mean. The target pulse wave detection information is any pulse wave detection information.

[0091] In this embodiment, the mean processing can adopt the arithmetic mean or weighted mean method, wherein the weighting coefficient can be dynamically adjusted according to the difference in physiological indicators corresponding to each pulse wave detection information. The difference calculation can be in the form of absolute value difference or square difference. When the square difference is used, normalization processing is required to eliminate the dimension difference. The determination of the light absorption abnormality can be combined with the linear combination of the infrared light pressure difference and the visible light pressure difference, wherein the combination weight is set according to the measurement stability of the two light sources. For example, in a motion interference environment, the visible light signal fluctuates greatly. At this time, the weight ratio of the visible light pressure difference can be reduced to improve the robustness of the abnormality assessment.

[0092] Specifically, all infrared light pressure values ​​within the same pulse wave detection group are first averaged to obtain the infrared light pressure mean value that reflects the typical state within the group. Similarly, the visible light pressure mean value is calculated in the same way. Subsequently, for the target pulse wave detection information, the absolute deviations of its infrared light pressure values ​​from the group mean are calculated and summed to obtain the total infrared light deviation. The visible light deviation is obtained using the same method. Finally, the deviations of the two light sources are weighted and summed or multiplied according to a preset ratio to obtain the light absorption anomaly that comprehensively reflects the degree of signal deviation.

[0093] As an example, the light absorption abnormality of the target pulse wave detection information can be determined by the following formula 1:

[0094] Formula 1

[0095] In formula 1, It is used to characterize the light absorption abnormality of the i-th pulse wave detection information in the pulse wave detection group. The infrared light pressure value used to represent the jth data in the i-th pulse wave detection information, Used to represent the visible light pressure value of the jth data in the i-th pulse wave detection information. It is used to represent the average infrared light pressure value of the jth data of all pulse wave detection information in the pulse wave detection group corresponding to the i-th pulse wave detection information. It is used to represent the mean value of the visible light pressure of all pulse wave detection information in the jth data of the pulse wave detection group corresponding to the i-th pulse wave detection information. M is used to represent the amount of data in the i-th pulse wave detection information. Indicates taking the absolute value.

[0096] in, The larger it is, the greater the deviation between the infrared light pressure value in the i-th pulse wave detection information and the overall data in the similar pulse wave detection group; The larger the value is, the greater the deviation between the visible light pressure value in the i-th pulse wave detection information and the overall data in the similar pulse wave detection group is.

[0097] This embodiment effectively identifies and quantifies abnormal data in pulse wave detection information. By calculating the degree of light absorption anomaly, the degree of deviation of each pulse wave detection piece from the overall data can be accurately assessed. This method enables targeted screening of high-quality training data, thereby improving the training effectiveness and accuracy of subsequent blood oxygen detection models. Furthermore, this method provides a quantitative basis for the identification and processing of abnormal data, helping to improve the overall accuracy and reliability of blood oxygen detection.

[0098] In some of the above-mentioned schemes of the present invention, pulse wave detection information is clustered using physiological indicator information, and the target training data set is screened based on the degree of light absorption abnormality. However, the impact of motion interference on the pulse wave signal is not considered during the screening process, resulting in some pulse wave detection information containing motion noise being selected into the training data set, affecting the training effect of the blood oxygen detection model.

[0099] In this regard, Figure 2 As shown, the present invention further proposes that before S104, the blood oxygen detection calculation method based on infrared light and visible light may also include the following S201 to S204:

[0100] S201, for the infrared light pulse wave signal and the visible light pulse wave signal of the pulse wave detection information, determining the area between two adjacent valley points as a segmented region, thereby obtaining a plurality of first segmented regions of the infrared light pulse wave signal and a plurality of second segmented regions of the visible light pulse wave signal;

[0101] S202, determining a motion interference parameter of a target segmented region based on a time span and a pulse pressure of the target segmented region, where the target segmented region is any one of the first segmented region and the second segmented region;

[0102] S203, determining the pulse wave error of the pulse wave detection information using the motion interference parameters of the pulse wave detection information and the light absorption abnormality of the pulse wave detection information;

[0103] S204, determining a training priority coefficient for the pulse wave detection information using the pulse wave error of the pulse wave detection information;

[0104] S104 may specifically include:

[0105] Based on each training priority coefficient, a target training data set is screened from each pulse wave detection group.

[0106] In this embodiment, the segmentation of the region is achieved by detecting valley points. Figure 3As shown, a schematic diagram of the division of segmented regions is provided, wherein the maximum pulse pressure within a pulse cycle is a peak point 310, and the minimum pulse pressure within a pulse cycle is a valley point 320, and the area between two adjacent valley points 320 constitutes a single segmented region.

[0107] The motion interference parameter is used to measure the degree of motion interference on the pulse wave signal within the target segment. Because human motion can interfere with the pulse wave signal, causing signal distortion, determining the motion interference parameter helps evaluate signal quality.

[0108] Pulse wave error is a metric used to measure the overall error in pulse wave detection information, taking into account various motion interference parameters and light absorption anomalies. A greater pulse wave error indicates a greater deviation from the actual physiological condition and poorer signal quality.

[0109] The training priority coefficient is a parameter determined based on the pulse wave error. It is used to prioritize each pulse wave detection group in subsequent model training. A higher training priority coefficient indicates that the data quality of that pulse wave detection group is relatively good, and it should be given a higher weight or priority in model training.

[0110] As an example, a signal processing algorithm (such as a derivative-based extreme point detection algorithm) is first used to detect valley points in infrared and visible light pulse wave signals. Specifically, the first-order derivative of the signal is calculated. Points where the derivative is zero and the second-order derivative is greater than zero are considered valley points. Then, based on the order of the detected valley points, the signal segment between two adjacent valley points is defined as a segmented region.

[0111] Then, the start and end times of the target segmented area are recorded to determine the time span of the target segmented area; at the same time, the pulse pressure value corresponding to the target segmented area is measured in real time. The motion interference parameter D can be calculated using a preset function or model that comprehensively considers the influence of the time span T and the pulse pressure P. For example, D = f(T, P), where f can be a linear function, a nonlinear function, or a model trained based on a machine learning algorithm. In practical applications, a large amount of experimental data is needed to determine the specific form of the function f so that the motion interference parameter can accurately reflect the degree to which the pulse wave signal in the target segmented area is interfered with by motion.

[0112] Then, using the motion interference parameters of the pulse wave detection information and the light absorption abnormality of the pulse wave detection information, the pulse wave error of the pulse wave detection information is determined by the following formula 2:

[0113] Formula 2

[0114] In formula 2, The pulse wave error used to characterize the i-th pulse wave detection information, The light absorption abnormality used to characterize the i-th pulse wave detection information, The average value of each motion interference parameter used to characterize the i-th pulse wave detection information.

[0115] Then, the pulse wave error of the pulse wave detection information is used to determine the training priority coefficient of the pulse wave detection information through the following formula 3:

[0116] Formula 3

[0117] In formula 3, The training priority coefficient used to characterize the i-th pulse wave detection information, The pulse wave error used to characterize the i-th pulse wave detection information, and norm is used to characterize normalization processing. It should be noted that to ensure meaningful calculation results, when performing fractional operations, in the embodiment of the present invention, when the denominator is zero, a correction factor of substantially zero is added to the denominator to prevent the denominator from being zero. The value of the correction factor is set by the implementer based on actual conditions and is not specifically limited by the present invention.

[0118] This embodiment effectively screens high-quality pulse wave detection data, reducing noise and interference in the training dataset. By accounting for motion interference, the robustness of the blood oxygen detection model is improved. Furthermore, by introducing a training priority coefficient, full utilization of high-quality data is ensured during model training, thereby enhancing the overall performance and accuracy of the blood oxygen detection model.

[0119] Some of the aforementioned embodiments of the present invention propose filtering training data by segmenting the pulse wave signal into segments and calculating motion interference parameters. However, due to physiological differences or motion interference among wearers, the time spans of the segmented regions may exhibit irregular periodic variations. Furthermore, the DC and AC components of the pulse wave may fluctuate abnormally due to motion, thus affecting the accuracy of the motion interference parameters.

[0120] In this regard, Figure 4 As shown, the present invention further proposes that pulse pressure includes a pulse DC component and a pulse AC component;

[0121] S202 may specifically include the following S401 to S403:

[0122] S401, determining the periodic regularity presentation degree of the target segmented area based on the time span of the target segmented area;

[0123] S402, determining the motion interference tendency of the target segmented area based on the pulse DC component and the pulse AC component at each moment in the target segmented area;

[0124] S403: Determine the motion interference parameter of the target segmented region by using the periodic regularity presentation degree and the motion interference tendency degree of the target segmented region.

[0125] In this embodiment, the periodicity degree measures the degree of periodicity within the target segment. Normally, pulse wave signals exhibit a certain degree of periodicity. A higher periodicity degree indicates that the signal segment conforms more closely to the normal periodicity and is likely to be less susceptible to interference. Conversely, a lower periodicity degree indicates that the signal is subject to significant interference, disrupting its periodicity.

[0126] The pulse DC component is the relatively stable portion of the pulse wave signal that does not change rapidly with the pulse beat. It generally reflects the average level of the pulse wave signal or the background signal strength. The pulse AC component is the portion of the pulse wave signal that changes rapidly with the pulse beat. It directly reflects the pulse's beating condition and contains important information such as pulse frequency and amplitude.

[0127] The motion interference tendency describes the degree to which the pulse wave signal is affected by motion interference within the target segment. By analyzing the relationship between the DC and AC components of the pulse at each moment, it can be determined whether the effect of motion interference on the pulse wave signal is gradually increasing, decreasing, or remaining relatively stable.

[0128] As an example, first, the relative deviation between the time span of the target segmented region and the average time span of each segmented region of the pulse wave signal to which the target segmented region belongs, i.e., the periodic coefficient of variation, is calculated. The periodic coefficient of variation reflects the degree of dispersion of the periodic values; a smaller coefficient of variation indicates a more stable period and a higher degree of periodic regularity. The degree of periodic regularity is then determined based on the magnitude of the periodic coefficient of variation of the target segmented region. For example, a pre-set mapping relationship can be used to assign different periodic regularity values ​​to different periodic coefficient of variation intervals.

[0129] Then, for each moment within the target segmented area, filtering and other methods can be used to extract the DC and AC components of the pulse. For example, a low-pass filter can be used to extract the DC component, while a high-pass filter can be used to extract the AC component. The cutoff frequency of the low-pass filter is typically set at a lower frequency band to filter out high-frequency interference in the pulse wave signal while retaining the relatively stable DC component. The cutoff frequency of the high-pass filter is determined based on the main frequency range of the pulse wave signal to extract the AC component that reflects the pulse beat.

[0130] Next, calculate the DC-to-AC ratio at each moment and the rate of change of this ratio between adjacent moments. This rate of change is then analyzed to determine its trend. If the rate of change shows a clear increasing or decreasing trend over a period of time, it indicates that motion interference has affected the ratio of the DC to AC components of the pulse, indicating a high degree of motion interference. If the rate of change is relatively stable and small, it indicates a low degree of motion interference. This trend can be quantified using metrics such as the variance and slope of the rate of change.

[0131] Finally, the motion interference parameter is determined by the following formula 4:

[0132] Formula 4

[0133] In formula 4, The motion interference parameter used to characterize the rth segment area in the i-th pulse wave detection information, It is used to characterize the motion interference tendency of the rth segment area in the i-th pulse wave detection information. Used to characterize the motion interference tendency of the rth segment area in the i-th pulse wave detection information.

[0134] in, The larger it is, the greater the impact of motion interference on the proportional relationship between the pulse DC component and the AC component in the rth segment area of ​​the i-th pulse wave detection information, that is, the larger the motion interference parameter is; The larger the value is, the more stable the period of the rth segment area in the i-th pulse wave detection information is, that is, the smaller the motion interference parameter is.

[0135] This embodiment effectively identifies and quantifies motion interference in pulse wave signals, thereby improving the accuracy and reliability of subsequent blood oxygen detection calculations. Specifically, by analyzing the periodic regularity of the pulse wave signal and the changing characteristics of the DC / AC components, it is possible to more accurately determine whether abnormal fluctuations in the signal are caused by physiological changes or motion interference. By introducing motion interference parameters, a more reliable basis is provided for subsequent data screening and model training, thereby building a more robust blood oxygen detection model.

[0136] Some of the aforementioned solutions proposed evaluating motion interference parameters by determining the degree of periodic regularity within segmented regions. However, relying solely on the time span of a single pulse wave signal in this process cannot accurately reflect the true physiological cycle characteristics. Individual differences or transient interference can easily lead to deviations in periodic regularity determination, affecting the reliability of subsequent motion interference parameters.

[0137] In this regard, the present invention further proposes that S401 may specifically include:

[0138] Obtaining a first time span mean value corresponding to a segmented region in a target pulse wave signal and a second time span mean value corresponding to a segmented region in a reference pulse wave signal, where the target pulse wave signal is the pulse wave signal belonging to the target segmented region, and the reference pulse wave signal is the pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal;

[0139] Determining the first period feature conformity of the target segmented area by using the time span of the target segmented area and the first time span mean;

[0140] Determining the second period feature conformity of the reference segmented region using the time span of the reference segmented region and the second time span mean, where the reference segmented region is a segmented region in the reference pulse wave signal corresponding to the target segmented region;

[0141] The periodic regularity presentation degree of the target segmented area is determined based on the first periodic feature conformity and the second periodic feature conformity.

[0142] In this embodiment, it should be noted that the periodicity of the corresponding pulse wave signals for both visible and infrared light is caused by the contraction and relaxation of arterial vessels, reflected in the regular changes in the pressure on the corresponding hemoglobin. That is, although visible light is more sensitive to oxyhemoglobin and infrared light is more sensitive to deoxyhemoglobin, both are affected by the contraction and relaxation of arterial vessels in arterial blood in the same way. In other words, for a single blood oxygen measurement, the periodicity of visible and infrared light should be consistent. Therefore, the degree of periodicity of the target segmented region can be determined based on the consistency of the first periodic feature of the target segmented region with the consistency of the second periodic feature of the reference segmented region.

[0143] The target pulse wave signal and the reference pulse wave signal may be an infrared pulse wave signal and a visible light pulse wave signal, respectively, and both have a synchronous acquisition characteristic. The first time span mean is the average of the time spans of all segmented regions in the target pulse wave signal, and the second time span mean is the average of the time spans of the corresponding segmented regions in the reference pulse wave signal.

[0144] The first-period feature conformance is determined by calculating the deviation ratio between the target segment's time span and the mean of the first time span. The second-period feature conformance is determined using the same method, calculating the deviation ratio between the reference segment's time span and the mean of the second time span. The degree of periodic regularity is determined by comparing the differences between the two periodic feature conformances.

[0145] As an example, first obtain the first time span mean value corresponding to the segmented region in the target pulse wave signal and the second time span mean value corresponding to the segmented region in the reference pulse wave signal. The target pulse wave signal is the pulse wave signal belonging to the target segmented region, and the reference pulse wave signal is the pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal.

[0146] Then, the time span of the target segmented area and the first time span mean are used to determine the first period feature conformity of the target segmented area. Specifically, the first period feature conformity can be determined by the following formula 5:

[0147] Formula 5

[0148] In formula 5, It is used to characterize the periodic feature conformity of the rth segment area in the i-th pulse wave detection information. Used to represent the time span of the rth segment area in the i-th pulse wave detection information, Used to represent the time span mean of the pulse wave signal belonging to the rth segmented area in the i-th pulse wave detection information.

[0149] Referring again to the above formula 5, the second period feature conformity of the reference segmented region is determined using the time span of the reference segmented region and the second time span mean.

[0150] Finally, the periodic regularity of the target segmented area is determined based on the first periodic feature conformity and the second periodic feature conformity. For example, the periodic regularity of the target segmented area can be obtained by taking a weighted average of the first periodic feature conformity and the second periodic feature conformity.

[0151] This embodiment effectively assesses the periodic regularity of the target segmented region, providing a more accurate data basis for subsequent blood oxygen detection calculations. By considering the time span characteristics of the target and reference pulse wave signals, this method can more comprehensively reflect the periodic changes in the pulse wave signal, improving the accuracy and reliability of blood oxygen detection.

[0152] In some of the above-mentioned schemes of the present invention, it is proposed to evaluate the periodic characteristics of the target segmented area by the periodic regularity presentation degree. However, directly comparing the first periodic feature conformity with the second periodic feature conformity to determine the periodic regularity presentation degree of the target segmented area has low accuracy.

[0153] In this regard, the present invention further proposes determining the periodic regularity presentation degree of the target segmented area based on the first periodic feature conformity and the second periodic feature conformity, which may specifically include:

[0154] The first calculated value is obtained by taking the absolute value of the difference between the first period characteristic conformity and the second period characteristic conformity;

[0155] Normalizing the reciprocal of the first calculated value to obtain a second calculated value;

[0156] The periodic regularity presentation degree of the target segmented area is determined using the second calculated value and the first periodic feature conformity.

[0157] In this embodiment, the first calculated value is obtained by taking the absolute value of the difference between the first period feature conformity and the second period feature conformity, and the second calculated value is obtained by normalizing the inverse of the first calculated value. The periodic regularity presentation degree is further determined in combination with the second calculated value and the first period feature conformity.

[0158] As an example, the periodic regularity presentation degree of the target segmented area can be specifically determined by the following formula 6:

[0159] Formula 6

[0160] In formula 6, It is used to characterize the periodic regularity of the rth segment area in the i-th pulse wave detection information. The absolute value of the difference between the periodic feature conformity of the rth segmented area in the i-th pulse wave detection information and the periodic feature conformity of the corresponding reference segmented area is used to represent the periodic feature conformity of the rth segmented area in the i-th pulse wave detection information. It is used to characterize the periodic feature conformity of the rth segment area in the i-th pulse wave detection information, and norm is used to characterize the normalization processing.

[0161] Among them, if The smaller it is, the better the periodic consistency of the segmented area in the infrared light pulse wave signal and the visible light pulse wave signal, which means the higher the degree of periodic regularity.

[0162] This embodiment enables a more accurate assessment of the degree of periodicity within a target segmented region. This effectively identifies pulse wave signal segments with good periodicity, thereby improving the accuracy of subsequent blood oxygen detection calculations. Furthermore, by introducing a second calculated value, this solution simultaneously considers the periodic characteristics of both the target and reference pulse wave signals, making the assessment of periodicity more comprehensive and objective.

[0163] In some of the above-mentioned schemes of the present invention, the method for determining the motion interference tendency based on the pulse DC component and the pulse AC component can reflect the influence of motion interference to a certain extent. However, in actual applications, since the signal interference caused by wrist movement is instantaneous and random, it is difficult to accurately quantify the degree of motion interference tendency by relying solely on the component changes of a single pulse wave signal. In particular, when the synchronization differences between different optical signals are not fully considered, there may be deviations in the assessment of the motion interference tendency.

[0164] In this regard, Figure 5 As shown, S402 may specifically include the following S501 to S503:

[0165] S501, determining the component transfer interference degree at each moment in the target segmented area based on the pulse DC component and the pulse AC component at each moment;

[0166] S502, obtaining the Pearson correlation coefficient of the pulse pressure at each time point between the target segmented region and the reference segmented region, where the reference segmented region is the segmented region in the reference pulse wave signal corresponding to the target segmented region, the reference pulse wave signal is the pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal, and the target pulse wave signal is the pulse wave signal belonging to the target segmented region;

[0167] S503: Determine the motion interference tendency of the target segmented area by using the transfer interference degree of each component and each Pearson correlation coefficient.

[0168] In this embodiment, the component transfer interference degree is determined by normalizing the ratio of the first slope of the pulse DC component to the second slope of the pulse AC component at the target moment. The Pearson correlation coefficient is then compared by extending the target duration from the target moment to form a first analysis period and a second analysis period. For example, the first slope is the linear regression coefficient of the DC component within 0.5 seconds before and after the target moment, and the second slope is the rate of change of the AC component within the same time window. The ratio is normalized and constrained to the interval [0, 1]. The target duration can be set to 0.3 seconds. In this case, the first analysis period covers three sampling points before and after the target moment, and the second analysis period corresponds to the same time span of the reference pulse wave signal.

[0169] Specifically, when the wrist moves, the DC component and AC component in the infrared and visible light signals will produce asynchronous fluctuations. By calculating the ratio of the DC component slope to the AC component slope at the target moment, the coordination of the component changes at that moment can be quantified. The higher the degree of deviation of the ratio from 1, the greater the signal distortion caused by motion interference. At the same time, adjacent time periods are intercepted with the target moment as the center, and the Pearson correlation coefficients of the two light signals in the corresponding time periods are calculated respectively. When the correlation coefficient is lower than 0.8, it indicates that there is significant desynchronization between the signals. After weighted summation of the component transfer interference degree and the Pearson correlation coefficient, when the weighted value exceeds the threshold of 0.6, it is determined that high-intensity motion interference exists at that moment.

[0170] As an example, the component transfer interference degree at each moment in the target segmented region is determined based on the pulse DC component and pulse AC component at each moment. Specifically, data for the pulse DC component and pulse AC component at each moment in the target segmented region are first acquired. Then, the data at each moment is processed using a predetermined component transfer interference degree algorithm to calculate the component transfer interference degree at that moment.

[0171] Furthermore, the Pearson correlation coefficient of the pulse pressure in the target segmented region and the reference segmented region at each moment is obtained. The reference segmented region is the segmented region in the reference pulse wave signal that corresponds to the target segmented region, and the reference pulse wave signal is the pulse wave signal in the pulse wave detection information that corresponds to the target pulse wave signal, excluding the target pulse wave signal. In a specific implementation, an infrared pulse wave signal can be selected as the target pulse wave signal, and a visible light pulse wave signal can be selected as the reference pulse wave signal. At each moment, the Pearson correlation coefficient of the pulse pressure data at that moment is calculated for the target segmented region and the reference segmented region.

[0172] Thus, the motion interference tendency of the target segmented area is determined by using the transfer interference degree of each component and the Pearson correlation coefficient. Specifically, the motion interference tendency of the target segmented area can be determined by the following formula 7:

[0173] Formula 7

[0174] In formula 7, It is used to characterize the motion interference tendency of the rth segment area in the i-th pulse wave detection information. It is used to represent the average value of the Pearson correlation coefficients in the rth segment area in the i-th pulse wave detection information. Used to represent the average value of the transfer interference degree of each component in the rth segment area in the i-th pulse wave detection information.

[0175] This embodiment effectively assesses the degree of motion interference in pulse wave signals. By analyzing the changing characteristics of the pulse's DC and AC components and comparing the correlations between pulse wave signals under different light sources, signal anomalies caused by motion can be accurately identified. This method not only considers the internal characteristics of a single signal but also exploits the relationships between multiple light source signals, thereby improving the accuracy and reliability of motion interference identification.

[0176] In some of the aforementioned solutions, the method for determining the component transfer interference degree during the calculation of the motion interference tendency for the target segmented region suffers from insufficient accuracy. Specifically, this is because traditional methods fail to fully consider the dynamic relationship between the DC and AC components of the pulse, resulting in a biased assessment of the sensitivity to motion interference in the signal.

[0177] In this regard, the present invention further proposes that S501 may specifically include:

[0178] Obtaining a first slope of a pulse DC component and a second slope of a pulse AC component at a target time in a target segmented area, where the target time is any time in the target segmented area;

[0179] The ratio of the first slope to the second slope is normalized to obtain the component transfer interference degree at the target time in the target segment area.

[0180] In this embodiment, the first slope is calculated using linear regression within a sliding time window, with the window length set as an adjustable parameter ranging from 0.2 seconds to 0.5 seconds. The second slope is calculated using a synchronous difference method based on the rate of change between adjacent sampling points. Normalization is achieved using the z-score method, specifically subtracting the group mean from the ratio and dividing it by the standard deviation to ensure data comparability across different signal channels. The time window configuration is linked to the sampling frequency of the pulse wave signal. When the sampling frequency is 100 Hz, the window length preferably includes 20 to 50 data points.

[0181] Specifically, a time window of 0.1 seconds before and after the target moment is selected, and a linear fit is performed on the DC component data within the window to obtain the first slope to characterize the steady-state change trend of the DC component. The difference between adjacent sampling points of the AC component within the window is synchronously extracted, and the average change rate is calculated as the second slope to reflect the instantaneous fluctuation characteristics of the pulse wave. The dynamic relationship coefficient of the two components is obtained by ratio calculation, and then standardized to eliminate the influence of individual differences. This method effectively captures the signal distortion characteristics caused by motion interference. When motion interference causes a sudden change in the DC component, the first slope increases significantly, while the regular fluctuation of the AC component is destroyed, resulting in an abnormal decrease in the second slope. At this time, the abnormal change in the ratio can accurately reflect the degree of component transfer interference. By quantifying the coordinated change pattern of the two components, the sensitivity of motion interference detection can be improved, providing a more reliable evaluation basis for subsequent training data screening.

[0182] As an example, when a wrist-worn wearable device collects visible light pulse wave signals, the time series data for a target segmented region of the signal is differentiated for the DC and AC components of the pulse at each sampling moment within that region. Taking a DC component data point at a specific moment as an example, the DC component data segments 50 milliseconds before and after that moment are taken, and the DC component slope corresponding to that moment is obtained through least squares fitting. Simultaneously, the AC component data segments with the same time span are taken and the AC component slope is calculated. The absolute ratio of the DC component slope to the AC component slope is input into a sigmoid function for normalization, and the resulting component transfer interference value at that moment is output, constrained to a range between 0 and 1.

[0183] This embodiment effectively addresses the difficulty in identifying interference caused by signal component offsets during motion. By quantitatively evaluating the dynamic relationship between the DC and AC components at each moment, it accurately captures periods of signal distortion caused by limb movement, providing a reliable basis for interference assessment for subsequent screening of high-quality training data and ultimately improving the blood oxygen detection model's ability to resist motion interference.

[0184] In some of the above solutions of the present invention, when determining the Pearson correlation coefficient, only the pulse pressure at a single moment in the target segmented region is compared with the pulse pressure in the reference segmented region, which may easily result in a low accuracy of the Pearson correlation coefficient.

[0185] In this regard, the present invention further proposes that S502 may specifically include:

[0186] Taking the target moment in the target segmented area as the center, extend the target duration to both sides to obtain the first analysis period, where the target moment is any moment in the target segmented area;

[0187] Taking the target moment in the reference segment area as the center, extend the target duration to both sides to obtain the second analysis period;

[0188] The first analysis period is compared with the second analysis period to obtain the Pearson correlation coefficient of the pulse pressure in the target segmented area and the reference segmented area at the target time.

[0189] In this embodiment, the target moment is an arbitrarily selected time point in the target segmented area. Subsequent analysis will be carried out around this moment. By constructing an analysis period centered on this moment, changes in characteristics such as pulse pressure are studied.

[0190] The target duration is a fixed value used to determine the length of the analysis period. The target duration is extended in both directions, centered on the target moment, to construct the time window for analysis.

[0191] The first analysis period is the time period obtained by extending the target duration to both sides of the target moment in the target segmented area. The pulse pressure and other data within this period will be used for subsequent comparative analysis with the data of the corresponding period in the reference segmented area.

[0192] The second analysis period is a time period obtained by extending the target duration to both sides with the target moment in the reference segmented area as the center. It has the same time span as the first analysis period and is used for data comparison with the first analysis period.

[0193] Pearson's correlation coefficient: A statistical indicator used to measure the degree of linear correlation between two variables, with a value range of -1 to 1. In this embodiment, it measures the linear correlation between the pulse pressures of the target segmented area and the reference segmented area around the target time (i.e., within the first and second analysis periods). When the Pearson's correlation coefficient is close to 1, it indicates a strong positive correlation between the two variables; when it is close to -1, it indicates a strong negative correlation; and when it is close to 0, it indicates a weak linear correlation.

[0194] As an example, an arbitrary moment in the target segmented area is selected as the target moment. With this target moment as the reference point, the preset time lengths are extended in the forward and backward directions of the time axis to form the first analysis period. A time point corresponding to the target moment is selected in the reference segmented area, and the preset time lengths are extended in the same manner to form the second analysis period. By calculating the data sequences of the pulse pressure waveforms in the first and second analysis periods and comparing their similarity using the Pearson correlation coefficient formula, the Pearson correlation coefficient value for the target moment is finally obtained. This value is used to quantify the consistency of pressure changes in the time dimension of the two segmented areas.

[0195] This embodiment effectively identifies the correlations between pressure changes across different optical bands within the same pulse wave signal, eliminating interference from segmented waveform distortion caused by limb movement. By dynamically analyzing pressure correlations within local time periods, the accuracy of noise signal recognition is improved, providing a reliable basis for subsequent screening of high-quality training data and ultimately reducing errors introduced by motion artifacts in the blood oxygen detection model.

[0196] In some of the above-mentioned solutions of the present invention, there may be quality differences or uneven distribution of samples in the training data set. Directly using unprocessed training data may reduce the efficiency of model training, resulting in slow convergence or falling into local optimality during the model parameter adjustment process, affecting the final performance of the blood oxygen detection model.

[0197] In this regard, the present invention further proposes that S105 may specifically include:

[0198] Divide the target training dataset into multiple training batches;

[0199] Performing data augmentation and data preprocessing on the first training data samples in each training batch to obtain second training data samples;

[0200] According to the priority order of the training batches, the model parameters of the target model are adjusted based on the second training data samples in the training batches in sequence until the preset training end conditions are met to obtain a blood oxygen detection model.

[0201] In this embodiment, the division of the training data set can be based on the number of samples or the time sequence, and each training batch contains the same or different number of samples. Data enhancement processing may include amplitude scaling, time stretching or adding noise to the pulse wave signal, and data preprocessing may include normalization, filtering or outlier removal. The priority order can be determined based on the light absorption anomaly, pulse wave error or average value of the training priority coefficient of the samples in the training batch, and batches with higher priority will participate in the training first. The preset training end condition can be set as the loss function dropping to a threshold or the number of training rounds reaching an upper limit. The modularization of the training process is achieved by dividing the batches, data enhancement and preprocessing improve data quality, and the priority order is adjusted to optimize the parameter update path.

[0202] Specifically, after the training data set is divided into multiple batches, each batch is independently enhanced and preprocessed. For example, random amplitude changes are applied to the infrared and visible light pulse wave signals to simulate the effects of different wearing tightness, or high-frequency motion interference is eliminated through sliding window filtering. The preprocessed second training data samples are input into the model in sequence according to batch priority. Batches with higher priority may contain samples with lower light absorption abnormalities. Prioritized training can accelerate the model's capture of core features. The model parameters are updated after each batch of training and are gradually optimized through multiple rounds of iterations until the loss function stabilizes or reaches the preset round number. The final blood oxygen detection model can more accurately process pulse wave signals of different qualities and improve the robustness of blood oxygen detection calculations.

[0203] As an example, the training dataset is divided into six training batches containing 300 samples each. The original samples in each training batch are first augmented by adding Gaussian noise, keeping the noise amplitude within 5% of the peak-to-peak value of the pulse wave signal. Normalization is also performed to map the signal amplitude to the range [-1, 1]. During the preprocessing phase, the augmented data undergoes Butterworth bandpass filtering with a passband range of 0.5Hz-8Hz, and baseline drift removal is performed. Training batches are prioritized by calculating the average pulse wave signal-to-noise ratio (SNR) of each batch. Batches with SNRs above 35dB are designated as high priority. During model training, high-priority batches are loaded first in each training cycle. An adaptive learning rate adjustment strategy is employed, with an initial learning rate of 0.001. If the validation set loss does not decrease for three consecutive cycles, the learning rate is decayed to 0.5. Gradient clipping is performed after every three training cycles, with a threshold of 1.0. Training is terminated when the validation set accuracy remains stable for ten consecutive training cycles.

[0204] This embodiment effectively addresses the slow model convergence problem caused by uneven distribution of training data, optimizing the model parameter update path by dynamically adjusting the training sequence. Data augmentation and preprocessing significantly reduce the interference of motion artifacts on pulse wave feature extraction, and the adaptive learning mechanism enables the model to automatically adapt to differences in feature distributions of data of varying quality during training. The introduction of a gradient clipping strategy effectively avoids the gradient explosion phenomenon that can occur during training, thereby ensuring that the blood oxygen detection model ultimately possesses stable computational accuracy and generalization performance.

[0205] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0206] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A blood oxygen detection calculation method based on infrared light and visible light, characterized in that: The method comprises: Acquiring physiological indicator information and pulse wave detection information of each wristband wearer, wherein the physiological indicator information includes at least one of body fat information and age information, and the pulse wave detection information includes infrared light pulse wave signals and visible light pulse wave signals; Clustering the pulse wave detection information based on the physiological indicator information to obtain multiple pulse wave detection groups; determining a light absorption abnormality of each pulse wave detection information based on the infrared light pressure value and the visible light pressure value of each pulse wave detection information in the pulse wave detection group, wherein the light absorption abnormality is used to represent the degree of deviation abnormality between the pulse wave detection information and the corresponding pulse wave detection group; Based on each of the light absorption abnormalities, a target training data set is selected from each of the pulse wave detection groups; Based on the target training data set, training the target model to obtain a blood oxygen detection model, so as to implement blood oxygen detection calculation through the blood oxygen detection model; Before selecting a target training data set from each of the pulse wave detection groups based on each of the light absorption abnormalities, the method further includes: For the infrared light pulse wave signal and the visible light pulse wave signal of the pulse wave detection information, determining a segmented region between two adjacent valley points to obtain a plurality of first segmented regions of the infrared light pulse wave signal and a plurality of second segmented regions of the visible light pulse wave signal; For a target segmented area, determining a motion interference parameter of the target segmented area based on a time span and a pulse pressure of the target segmented area, the target segmented area being any one of the first segmented area and the second segmented area; determining a pulse wave error of the pulse wave detection information by using the motion interference parameters of the pulse wave detection information and the light absorption abnormality of the pulse wave detection information; Determining a training priority coefficient of the pulse wave detection information by utilizing a pulse wave error degree of the pulse wave detection information; The step of selecting a target training data set from each of the pulse wave detection groups based on each of the light absorption abnormalities comprises: Based on each of the training priority coefficients, a target training data set is selected from each of the pulse wave detection groups; The pulse pressure includes a pulse DC component and a pulse AC component; The step of determining the motion interference parameter of the target segmented region based on the time span and pulse pressure of the target segmented region includes: Determining a periodic regularity presentation degree of the target segmented area based on a time span of the target segmented area; Determining a motion interference tendency of the target segmented area based on the pulse DC component and the pulse AC component at each moment in the target segmented area; the motion interference tendency is used to describe the changing trend or degree of motion interference of the pulse wave signal in the target segmented area; comparing the correlation of the pulse wave signals under different light sources by analyzing the changing characteristics of the pulse DC component and the pulse AC component; The motion interference parameter of the target segmented area is determined by using the periodic regularity presentation degree and the motion interference tendency degree of the target segmented area.

2. The blood oxygen detection calculation method based on infrared light and visible light according to claim 1, characterized in that: The determining of the light absorption abnormality of each pulse wave detection information based on the infrared light pressure value and the visible light pressure value of each pulse wave detection information in the pulse wave detection group includes: performing mean processing on each of the infrared light pressure values ​​and each of the visible light pressure values ​​of each of the pulse wave detection information in the pulse wave detection group to obtain an infrared light pressure mean value and a visible light pressure mean value; The light absorption abnormality of the target pulse wave detection information is determined by using the difference between the infrared light pressure value of the target pulse wave detection information and the infrared light pressure mean, as well as the difference between the visible light pressure value and the visible light pressure mean. The target pulse wave detection information is any one of the pulse wave detection information.

3. The blood oxygen detection calculation method based on infrared light and visible light according to claim 1, characterized in that: The determining of the periodic regularity presentation degree of the target segmented area based on the time span of the target segmented area includes: Obtaining a first time span mean value corresponding to a segmented region in a target pulse wave signal and a second time span mean value corresponding to a segmented region in a reference pulse wave signal, wherein the target pulse wave signal is a pulse wave signal belonging to the target segmented region, and the reference pulse wave signal is a pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal; Determining a first period feature conformity of the target segmented region by using the time span of the target segmented region and the first time span mean; Determining the second period feature conformity of the reference segmented region using the time span of the reference segmented region and the second time span mean, the reference segmented region being the segmented region in the reference pulse wave signal corresponding to the target segmented region; The periodic regularity presentation degree of the target segmented area is determined based on the first periodic feature conformity and the second periodic feature conformity.

4. The blood oxygen detection calculation method based on infrared light and visible light according to claim 3, characterized in that: The determining the periodic regularity presentation degree of the target segmented area based on the first periodic feature conformity and the second periodic feature conformity includes: Taking the absolute value of the difference between the first period feature conformity and the second period feature conformity to obtain a first calculated value; performing normalization processing on the reciprocal of the first calculated value to obtain a second calculated value; The periodic regularity presentation degree of the target segmented area is determined by using the second calculated value and the first periodic feature conformity.

5. The blood oxygen detection and calculation method based on infrared light and visible light according to claim 1, characterized in that: The determining of the motion interference tendency of the target segmented area based on the pulse DC component and the pulse AC component at each moment in the target segmented area includes: determining the component transfer interference degree at each moment in the target segmented area based on the pulse DC component and the pulse AC component at each moment; Obtaining a Pearson correlation coefficient of the pulse pressure at each moment in the target segmented region and a reference segmented region, wherein the reference segmented region is a segmented region in the reference pulse wave signal corresponding to the target segmented region, the reference pulse wave signal is a pulse wave signal corresponding to the pulse wave detection information other than the target pulse wave signal, and the target pulse wave signal is the pulse wave signal to which the target segmented region belongs; The motion interference tendency of the target segmented area is determined by using the component transfer interference degree and the Pearson correlation coefficient.

6. The blood oxygen detection calculation method based on infrared light and visible light according to claim 5, characterized in that: The determining of the component transfer interference degree at each moment in the target segmented area based on the pulse DC component and the pulse AC component at each moment includes: Acquire a first slope of the pulse DC component and a second slope of the pulse AC component at a target time in the target segmented area, wherein the target time is any time in the target segmented area; The ratio of the first slope to the second slope is normalized to obtain the component transfer interference degree at the target time in the target segmented area.

7. The blood oxygen detection and calculation method based on infrared light and visible light according to claim 5, characterized in that: The obtaining of the Pearson correlation coefficient of the pulse pressure at each moment in the target segmented region and the reference segmented region includes: Taking the target moment in the target segmented area as the center, extending the target duration to both sides to obtain a first analysis period, wherein the target moment is any moment in the target segmented area; Taking the target moment in the reference segmented area as the center, extending the target duration to both sides to obtain a second analysis period; The first analysis period is compared with the second analysis period to obtain a Pearson correlation coefficient of the pulse pressure in the target segmented area and the pulse pressure in the reference segmented area at the target time.

8. The blood oxygen detection and calculation method based on infrared light and visible light according to any one of claims 1 to 7, characterized in that: The step of training a target model based on the target training data set to obtain a blood oxygen detection model includes: Dividing the target training data set into multiple training batches; Performing data augmentation and data preprocessing on the first training data samples in each of the training batches to obtain second training data samples; According to the priority order of the training batches, the model parameters of the target model are adjusted based on the second training data samples in the training batches in sequence until the preset training end conditions are met to obtain the blood oxygen detection model.

Images