PPG blood pressure detection method capable of reducing false alarm of smart bracelet

Abstract

The invention discloses a PPG blood pressure detection method capable of reducing false alarms of an intelligent bracelet, and relates to the technical field of blood pressure detection.The PPG blood pressure detection method comprises the following steps that firstly, a PPG signal collection unit is used for collecting PPG signals; step 2, establishing an attention model in the PPG signal preprocessing unit, and performing intelligent signal processing on the PPG signals acquired by the PPG signal acquisition unit by utilizing an attention mechanism: focusing key blood pressure information, and filtering redundant information. According to the method, automatic feature extraction is realized through an attention mechanism, false alarms introduced by irrelevant features can be reduced, a personalized nonlinear model is established through deep learning, false alarms caused by individual differences can be reduced, a sample set is continuously expanded through transfer learning, false alarms of different crowds can be dynamically reduced, and the accuracy of the method is improved. By constructing an end-to-end process from collection to prediction, all the modules cooperatively improve the performance, and false alarms are jointly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of blood pressure detection, and particularly to a PPG blood pressure detection method for reducing false alarms of smart bracelets. Background Art

[0002] Photoplethysmography (PPG) is a method for detecting blood volume changes in living tissues by means of optoelectronic means, and can be used for continuous long-term detection by wearing a wearable device. At present, PPG has been widely used in the detection and evaluation of various hemodynamic parameters such as blood oxygen and heart rate. Research shows that PPG and arterial blood pressure are morphologically similar and can be used for blood pressure detection. The heartbeat generates a continuous pressure wave that is transmitted through the blood vessels and slightly changes the diameter of the blood vessels. The change in blood volume in the blood vessels can be detected by PPG, and its signal has the characteristics of non-stationarity and pseudo-periodicity.

[0003] Currently, the PPG sensor in a smart bracelet can collect the blood circulation changes in the skin microvasculature, and the sampling frequency is about 100 Hz. However, to accurately obtain the real-time blood pressure data of the user, two key problems still need to be solved: extracting personalized features and establishing a personalized prediction model. It is known that the accuracy of a blood pressure prediction model depends on extracting effective features from complex physiological signals and establishing a reliable prediction model. The simple linear regression model in existing products only establishes a linear relationship between blood pressure and limited features, which is difficult to meet these two requirements and is prone to false alarms. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides a PPG blood pressure detection method for reducing false alarms of smart bracelets. Automatic feature extraction is achieved through the attention mechanism, which can reduce false alarms caused by the introduction of irrelevant features. A personalized non-linear model is established through deep learning, which can reduce false alarms caused by individual differences. The sample set is continuously expanded through transfer learning, which can dynamically reduce false alarms for different populations. By constructing an end-to-end process from collection to prediction, each module collaborates to improve performance and jointly reduce false alarms.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a PPG blood pressure detection method for reducing false alarms of smart bracelets, comprising the following steps:

[0006] Step 1: Collect PPG signals using a PPG signal collection unit;

[0007] Step 2: Establish an attention model in the PPG signal preprocessing unit, and use the attention mechanism to perform intelligent signal processing on the PPG signals collected by the PPG signal collection unit: focus on key blood pressure information and filter out redundant information;

[0008] The working content of the attention model is as follows: perform non-linear transformation on each band feature, learn to generate an attention weight score between 0 and 1, re-weight and modulate the bands according to the attention weight scores of the bands, keep or amplify the important bands, and shrink or filter out the redundant bands. After attention modulation, each band is linearly reconstructed to generate a PPG signal with enhanced key features. The parameters of the attention model are trained by the backpropagation algorithm to make the reconstructed PPG signal have the best effect on blood pressure prediction;

[0009] Step 3: Construct an LSTM module and a CNN module in the blood pressure detection model establishment unit, connect the LSTM module and the CNN module, and perform joint time-domain sequence and frequency-domain feature modeling on the PPG signal. The LSTM module and the CNN module respectively optimize the time-domain and frequency-domain feature expressions, and use them jointly as a sub-network to enable the network to comprehensively learn the time-frequency attributes of the PPG signal and perform better blood pressure state expression;

[0010] Through end-to-end training of the network, automatically learn the PPG features that express the blood pressure state. The end-to-end training method enables the network modules to work together to optimize the correlation and discriminability of the feature expressions;

[0011] The output of the network is connected to a regression layer for blood pressure prediction. The network parameters are adjusted through error backpropagation. The regression task drives the network to adjust the feature expression to adapt to the blood pressure monitoring requirements. After the network training converges, the intermediate layer features can accurately reflect the encoding of the PPG for the blood pressure state. The trained network has the ability to extract optimized features. Through the network, online PPG feature extraction and blood pressure prediction are performed. After the network is modeled, it is deployed to the edge device to monitor blood pressure in real time;

[0012] Step 4: In the individual blood pressure prediction unit, establish a personalized non-linear regression model by using the PPG training data of each individual and the corresponding reference blood pressure values, and update the model parameters in an end-to-end manner to make the blood pressure prediction value as close as possible to the real blood pressure, so as to obtain a non-linear regression structure driven by the individual's own data, where the complex mapping relationship from the feature expression to the blood pressure value has adapted to the physiological characteristics of the individual;

[0013] Step 5: Use the model optimization unit to continuously collect a large-scale PPG data set containing more subjects, which is used as new sample data after preprocessing. Under the condition that the original model parameters remain unchanged, use the new sample data for incremental training to achieve the effect of smoothly optimizing the original model.

[0014] Preferably, the PPG signal acquisition unit 1 in Step 1 includes multiple light-emitting diodes, multiple photodiodes, multiple PPG sensors, an accelerometer, a filter, an analog front end, and a micropump;

[0015] The wavelengths of multiple light-emitting diodes are different. The wavelengths of commonly used light-emitting diodes include 660nm, 805nm, and 940nm;

[0016] Multiple photodiodes adopt a diagonal or circular layout to make the light intensity distribution more uniform and improve the signal acquisition quality;

[0017] Multiple PPG sensors adopt a high-density layout of light-emitting diodes and photodiodes, and this layout includes but is not limited to one of the 16×16 layout and 32×32 layout;

[0018] The accelerometer is used to detect the movement of the wrist or fingertip, and the algorithm is used to eliminate the influence of movement on the PPG signal to avoid generating artifacts;

[0019] The filter and analog front end are used to improve the signal quality and eliminate light and power noise;

[0020] The micro pump is used to improve the skin blood flow distribution and improve the PPG signal quality.

[0021] Preferably, the sampling frequency in step one satisfies the Nyquist sampling theorem, is higher than twice the PPG signal frequency band, that is, above 200Hz, and the sampling accuracy is between 12 - 16 bits.

[0022] Preferably, the attention weight score in step two represents the contribution of each frequency band to blood pressure information. The important frequency band obtains a higher score, and the redundant frequency band obtains a lower score.

[0023] Preferably, the LSTM module in step three contains multiple LSTM layers, which are used to learn the timing law in the PPG time-frequency diagram, capture the timing features of long-term dependencies through the gated recurrent unit, stack multiple LSTM layers, extract the timing features of different abstraction levels in the PPG signal, and model its curve shape;

[0024] The CNN module learns to extract local features in the PPG frequency domain through multi-layer convolution and pooling operations. The convolution filter in the CNN module can capture local features in the input data, and pooling improves the robustness of the features. Applied to the frequency domain expression of the PPG signal, it can learn the combination pattern of frequency components.

[0025] Preferably, in step four, in the detection stage, by directly inputting the new PPG data of this individual into its exclusive personalized non-linear model, the blood pressure can be quickly predicted without new model training.

[0026] Preferably, the large-scale PPG dataset collected in step five covers the demographic characteristics of different ages, genders, and body types.

[0027] Preferably, the incremental training in step five will retain the general knowledge of extracting time-frequency features and expressing the PPG shape in the original model. At the same time, by adjusting some parameters, the model can adapt to the feature patterns in the new samples, reducing the error caused by the difference in sample distribution.

[0028] Compared with the prior art, the present invention provides a PPG blood pressure detection method for reducing false alarms of smart bracelets, having the following beneficial effects:

[0029] 1. By applying the attention mechanism to analyze the original PPG signal in the signal acquisition and preprocessing stage, the contribution of different frequency components to blood pressure information can be automatically learned, so as to weight and strengthen important frequency features, reduce the interference of irrelevant or redundant features on the results, and avoid false alarms caused by these features. This mechanism that mimics the attention distribution of the human brain can make the model pay more attention to key features, effectively extract the frequency that has the greatest impact on blood pressure, and reduce the false alarm risk introduced by the attention mechanism itself.

[0030] 2. By using deep learning technology to establish a personalized and complex mapping relationship between PPG signal features and blood pressure status in the feature expression and regression prediction stage, false alarms caused by differences in human constitutions can be reduced. Compared with simple linear regression, deep learning can learn the internal laws of data through layer-by-layer abstraction, establish a non-linear model suitable for each individual, greatly improve the accuracy of the results, and reduce the false alarm rate caused by individual differences.

[0031] 3. By continuously collecting new samples containing more subjects and dynamically optimizing the model in a transfer learning manner, the coverage range of samples can be continuously expanded, enabling the model to adapt to the characteristics of more people, thereby reducing the false alarm rate when used by different people. Compared with static models, this optimization method enables the model to dynamically adapt to changes in individual states, continuously improve, and reduce false alarms from various sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a schematic diagram of the overall structure of the present invention;

[0033] Figure 2 is a schematic diagram of the structure of the PPG signal acquisition unit of the present invention;

[0034] Figure 3 is a schematic diagram of the structure of the PPG signal preprocessing unit of the present invention;

[0035] Figure 4 is a schematic diagram of the structure of the blood pressure detection model establishment unit of the present invention;

[0036] Figure 5 is a schematic diagram of the structure of the individual blood pressure prediction unit of the present invention.

[0037] In the figure:

[0038] 1. PPG signal acquisition unit; 2. PPG signal preprocessing unit; 3. Blood pressure detection model establishment unit; 4. Individual blood pressure prediction unit; 5. Model optimization unit. Specific implementation manners

[0039] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0040] As Figure 1 shown, the present invention provides a PPG blood pressure detection method for reducing false alarms of smart bracelets, including the following steps:

[0041] Step 1: Use the PPG signal acquisition unit 1 to collect PPG signals. In order to obtain high-quality PPG signals, the PPG signal acquisition unit 1 includes a plurality of light-emitting diodes, a plurality of photodiodes, a plurality of PPG sensors, an accelerometer, a filter, an analog front end, and a micropump;

[0042] It is necessary to select an appropriate working wavelength for the light-emitting diodes, generally in the range of infrared light and near-infrared light, which can better penetrate the skin and tissues to detect the pulsation information in the blood. Therefore, the wavelengths of the plurality of light-emitting diodes in the PPG signal acquisition unit 1 are different. The commonly used wavelengths of the light-emitting diodes include 660nm, 805nm, and 940nm, and different wavelengths are suitable for test subjects with different skin colors;

[0043] Since the distance between the light-emitting diode and the photodiode directly affects the quality of the PPG signal, if the distance is too small, it is easy to cause mutual interference of the optical paths; if the distance is too large, the light intensity attenuation is serious. Generally, the distance is preferably 5-15mm. Therefore, in the present invention, a plurality of photodiodes are arranged diagonally or in a ring shape, which can make the light intensity distribution more uniform and improve the signal acquisition quality;

[0044] A plurality of PPG sensors adopt a layout of high-density light-emitting diodes and photodiodes. This layout includes, but is not limited to, one of the 16×16 layout and the 32×32 layout. By adopting this layout method, more redundant signals can be obtained, and noise suppression and signal enhancement can also be achieved through algorithm processing;

[0045] In addition, it is necessary to reasonably design and arrange the directions of the light-emitting diodes and photodiodes to avoid direct light pollution of the receiving end. The scattering and diffraction of light also need to be considered, and light-shielding and collimation structures are adopted to improve the signal-to-noise ratio of the optical path. The gap between the light-emitting diode and the photodiode needs to be reduced to avoid interference from external light. At the same time, the heat dissipation requirements need to be considered to prevent the light-emitting diode from overheating and affecting the wavelength. The working current of the light-emitting diode also needs to be monitored and adjusted in real time to ensure the stability of its emission wavelength and intensity. Meanwhile, dynamic calibration can be performed online according to the skin color and tissue thickness of the tested person to improve the robustness of the acquisition.

[0046] Since both too low sampling rate and precision will lead to signal distortion and errors, in the present invention, the sampling frequency of the PPG signal acquisition unit 1 satisfies the Nyquist sampling theorem, being higher than twice the frequency band of the PPG signal, that is, above 200 Hz, and the sampling precision is between 12 - 16 bits.

[0047] The accelerometer is used to detect the movement of the wrist or fingertip, and the influence of the movement on the PPG signal is eliminated through an algorithm to avoid generating artifacts; the filter and analog front end are used to improve the signal quality and eliminate optical and power supply noise; the micropump is used to improve the skin blood flow distribution and enhance the PPG signal quality;

[0048] Through optimizing technical means in multiple aspects such as sensor parameter configuration, optical path design, data acquisition scheme, and signal processing algorithm, the present invention can enhance the accuracy and reliability of the acquisition from the source of the PPG signal, laying a foundation for improving the accuracy of subsequent heart rate detection and physiological parameter calculation.

[0049] Step 2: Establish an attention model in the PPG signal preprocessing unit 2, and use the attention mechanism to perform intelligent signal processing on the PPG signal collected by the PPG signal acquisition unit 1: focus on key blood pressure information and filter out redundant information;

[0050] The working content of the attention model is as follows: perform non-linear transformation on each frequency band feature to learn and generate an attention weight score between 0 and 1. The attention weight score represents the contribution of each frequency band to the blood pressure information. The important frequency bands obtain higher scores, and the redundant frequency bands obtain lower scores. According to the attention weight scores of the frequency bands, the frequency bands are re-weighted and modulated. The important frequency bands are maintained or amplified, and the redundant frequency bands are reduced or filtered out. After attention modulation, each frequency band is linearly reconstructed to generate a PPG signal with enhanced key features. The parameters of the attention model are trained through the backpropagation algorithm to make the effect of the reconstructed PPG signal on blood pressure prediction optimal;

[0051] In this way, the attention model can automatically learn which frequency bands are the most critical without manual setting of the frequency weight function. The entire attention allocation and re-weighting process mimics the way the human brain processes information and efficiently extracts key features.

[0052] Compared with the frequency weight scheme customized by experience, the automatic evaluation and weighting of frequency bands based on the attention mechanism can more intelligently focus on key blood pressure information, effectively filter out redundant components, reduce the risk of false alarms, and improve the accuracy and reliability of PPG measurement.

[0053] Step 3: Construct an LSTM module and a CNN module in the blood pressure detection model establishment unit 3 to jointly model the time-domain shape and frequency-domain energy, and learn the intrinsic timing law and frequency characteristics of the PPG signal. By obtaining the PPG time-frequency diagram containing time-domain shape information and frequency-domain energy information. By selecting an appropriate wavelet basis, the PPG signal can be decomposed into sub-band signals in several different time and frequency ranges, reflecting the time-domain waveform characteristics and frequency-domain energy distribution of the PPG signal. This provides input data for subsequent time-frequency joint modeling.

[0054] Among them, the LSTM module contains multiple LSTM layers, which are used to learn the timing law in the PPG time-frequency diagram, capture the timing characteristics of long-term dependencies through gated recurrent units, stack multiple LSTM layers, extract the timing characteristics at different abstraction levels in the PPG signal, and model its curve shape;

[0055] The CNN module learns to extract local features in the PPG frequency domain through multiple convolutional and pooling operations. The convolutional filters in the CNN module can capture local features in the input data, and pooling improves the robustness of the features. Applied to the frequency-domain expression of the PPG signal, it can learn the combination pattern of frequency components.

[0056] Then, connect the LSTM module and the CNN module to jointly model the time-domain sequence and frequency-domain characteristics of the PPG signal. The LSTM module and the CNN module respectively optimize the feature expressions in the time domain and the frequency domain, and use them jointly as a sub-network, so that the network comprehensively learns the time-frequency attributes of the PPG signal and performs a better blood pressure state expression;

[0057] Through end-to-end training of the network, automatically learn the PPG features expressing the blood pressure state. The end-to-end training method enables the network modules to work together and optimize the correlation and discriminability of the feature expressions;

[0058] The network output is connected to a regression layer for blood pressure prediction. Adjust the network parameters through error backpropagation. The regression task drives the network to adjust the feature expression to adapt to the blood pressure monitoring requirements; after the network training converges, the intermediate layer features can accurately reflect the encoding of the PPG for the blood pressure state. The trained network has the ability to extract optimized features. Through the network, online PPG feature extraction and blood pressure prediction are performed. After the network is modeled, it is deployed to the edge device to monitor blood pressure in real time;

[0059] Step 4: In the individual blood pressure prediction unit 4, a personalized non-linear regression model is established by using each individual's own PPG training data and the corresponding reference blood pressure values, and the model parameters are updated in an end-to-end manner to make the blood pressure prediction value as close as possible to the true blood pressure, so as to obtain a non-linear regression structure driven by the individual's own data, where the complex mapping relationship from the feature expression to the blood pressure value has adapted to the physiological characteristics of the individual;

[0060] In the detection stage, by directly inputting the individual's new PPG data into its exclusive personalized non-linear model, the blood pressure can be quickly predicted without new model training. Compared with the general model, this personalized modeling strategy can greatly reduce the blood pressure prediction error caused by individual differences and improve the accuracy and reliability of non-invasive monitoring.

[0061] Step 5: Use the model optimization unit 5 to continuously collect a large-scale PPG data set containing more subjects. The collected large-scale PPG data set covers demographic characteristics such as different age groups, genders, body types, etc. The collected data is preprocessed, including denoising and normalization, and the cleaned high-quality PPG signal is extracted. The feature expression is carried out in the same way as the original model, such as time-frequency feature extraction based on wavelet transform, etc. At the same time, the corresponding reference blood pressure value needs to be obtained as the supervision label for these new samples.

[0062] After preprocessing, it is used as new sample data, and under the condition that the original model parameters remain unchanged, the new sample data is used for incremental training. This transfer learning can avoid losing the knowledge extracted by the original model and at the same time adapt to the feature distribution of the new samples. Compared with training from scratch, transfer learning is more efficient. Incremental training will retain the general knowledge such as time-frequency features extracted and PPG shape expressed in the original model. At the same time, by adjusting some parameters, the model is adapted to the feature patterns in the new samples and the error caused by the difference in sample distribution is reduced. Progressive end-to-end fine-tuning is used for incremental learning. Compared with large-scale parameter adjustment, this fine-tuning method can effectively prevent the model from deviating too much from the original knowledge and achieve the effect of smoothly optimizing the original model.

[0063] There may be feature patterns in the new samples that are not covered by the original model. Incremental learning can enable the model to come into contact with these new patterns, correct the original biases, and enhance the generalization ability of the model. A regular or dynamic data set update strategy can be designed to avoid incremental training being too frequent or too little and control the rhythm of model optimization. The enhanced model will be applied to the prediction of new samples. If the error still exists, a new round of incremental learning can be triggered to achieve continuous improvement. Through this incremental learning mechanism, the applicable range of the model can be dynamically and progressively expanded, the blood pressure prediction error of different populations can be reduced, and the model can be made more robust.

[0064] Through this systematic and process-oriented optimization approach, each module collaborates in innovation to jointly enhance the accuracy of the results, comprehensively reducing false alarms from different sources in different links overall, and achieving the goal of improving the user experience and promoting product progress. This multi-pronged solution can effectively address the issue of false alarms in blood pressure prediction.

[0065] The above has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and all these changes and improvements fall within the scope of the present invention claimed.

Claims

1. A PPG blood pressure detection method for reducing false alarms of smart bracelets, characterized in that: It includes the following steps: Step 1: Use the PPG signal acquisition unit (1) to collect PPG signals; Step 2: Establish an attention model in the PPG signal preprocessing unit (2), and use the attention mechanism to perform intelligent signal processing on the PPG signals collected by the PPG signal acquisition unit (1): focus on key blood pressure information and filter out redundant information; The working content of the attention model is as follows: perform non-linear transformation on each frequency band feature, learn to generate an attention weight score between 0 and 1, and re-weight and modulate the frequency bands according to the attention weight scores of the frequency bands. Important frequency bands are maintained or amplified, and redundant frequency bands are reduced or filtered out. After attention modulation, each frequency band is linearly reconstructed to generate a PPG signal with enhanced key features. The parameters of the attention model are trained by the backpropagation algorithm to make the reconstructed PPG signal have the best effect on blood pressure prediction; Step 3: Construct an LSTM module and a CNN module in the blood pressure detection model establishment unit (3), connect the LSTM module and the CNN module, and perform joint time-domain sequence and frequency-domain feature modeling on the PPG signal. The LSTM module and the CNN module respectively optimize the time-domain and frequency-domain feature expressions, and use them jointly as a sub-network to enable the network to comprehensively learn the time-frequency attributes of the PPG signal and perform better blood pressure state expression; Through end-to-end training of the network, automatically learn the PPG features that express the blood pressure state. The end-to-end training method enables the network modules to work together to optimize the correlation and discriminability of the feature expressions; The output of the network is connected to the regression layer for blood pressure prediction. The network parameters are adjusted through error backpropagation. The regression task drives the network to adjust the feature expression to adapt to the blood pressure monitoring requirements; after the network training converges, the intermediate layer features can accurately reflect the encoding of the PPG for the blood pressure state. The trained network has the ability to extract optimized features. Online PPG feature extraction and blood pressure prediction are performed through the network. After the network is modeled, it is deployed to the edge device to monitor blood pressure in real time; Step 4: In the individual blood pressure prediction unit (4), establish a personalized non-linear regression model by using the PPG training data of each individual and the corresponding reference blood pressure values, and update the model parameters in an end-to-end manner to make the blood pressure prediction value approach the real blood pressure, so as to obtain a non-linear regression structure driven by the individual's own data, where the complex mapping relationship from the feature expression to the blood pressure value has adapted to the physiological characteristics of the individual; Step 5: Use the model optimization unit (5) to continuously collect a large-scale PPG data set containing more subjects, which is used as new sample data after preprocessing, and perform incremental training using the new sample data while keeping the original model parameters unchanged to achieve the effect of smoothly optimizing the original model.

2. The PPG blood pressure detection method for reducing false alarms of smart bracelets according to claim 1, characterized in that: The PPG signal acquisition unit (1) in Step 1 includes multiple light-emitting diodes, multiple photodiodes, multiple PPG sensors, an accelerometer, a filter, an analog front end, and a micropump; The wavelengths of the multiple light-emitting diodes are different, and the commonly used wavelengths of the light-emitting diodes include 660 nm, 805 nm, and 940 nm; The multiple photodiodes adopt a diagonal or annular layout to make the light intensity distribution more uniform and improve the signal acquisition quality; The multiple PPG sensors adopt a high-density layout of light-emitting diodes and photodiodes, and this layout includes but is not limited to one of the 16×16 layout and 32×32 layout; The accelerometer is used to detect the movement of the wrist or fingertip, and the influence of the movement on the PPG signal is eliminated through an algorithm to avoid generating artifacts; The filter and the analog front end are used to improve the signal quality and eliminate light and power supply noise; The micropump is used to improve the skin blood flow distribution and improve the PPG signal quality.

3. A PPG blood pressure detection method for reducing false alarms of a smart bracelet according to claim 1, characterized in that: The sampling frequency in Step 1 satisfies the Nyquist sampling theorem, is higher than twice the PPG signal frequency band, that is, above 200 Hz, and the sampling accuracy is between 12 - 16 bits.

4. A PPG blood pressure detection method for reducing false alarms of a smart bracelet according to claim 1, characterized in that: The attention weight score in Step 2 represents the contribution of each frequency band to blood pressure information. The important frequency band obtains a higher score, and the redundant frequency band obtains a lower score.

5. A PPG blood pressure detection method for reducing false alarms of a smart bracelet according to claim 1, characterized in that: The LSTM module in Step 3 contains multiple LSTM layers, which are used to learn the timing rules in the PPG time-frequency diagram, capture the timing features of long-term dependencies through a gated recurrent unit, stack multiple LSTM layers, extract the timing features of different abstraction levels in the PPG signal, and model its curve shape; The CNN module learns to extract local features in the PPG frequency domain through multiple convolutional and pooling operations. The convolutional filter in the CNN module can capture local features in the input data, and pooling improves the robustness of the features. Applied to the frequency domain expression of the PPG signal, it can learn the combination pattern of frequency components.

6. A PPG blood pressure detection method for reducing false alarms of a smart bracelet according to claim 1, characterized in that: In Step 4, in the detection stage, by directly inputting the new PPG data of this individual into its exclusive personalized non-linear model, the blood pressure can be quickly predicted without new model training.

7. A PPG blood pressure detection method for reducing false alarms of a smart bracelet according to claim 1, characterized in that: The large-scale PPG dataset collected in Step 5 covers the demographic characteristics of different ages, genders, and body types.

8. A PPG blood pressure detection method for reducing false alarms of a smart bracelet according to claim 1, characterized in that: The incremental training in step five will retain the general knowledge of extracting time-frequency features and expressing the PPG shape in the original model. At the same time, by adjusting some parameters, the model can be adapted to the feature patterns in the new samples, reducing the error caused by the difference in sample distribution.

Images