CN115607126A - A non-contact blood pressure measurement method based on pulsed ultra-wideband radar - Google Patents

Abstract

The invention provides a non-contact blood pressure measuring method based on a pulse ultra-wideband radar, belonging to the field of non-contact vital sign detection; the method specifically comprises the following steps: firstly, an ultra-wideband radar continuously transmits pulse signals, and after the pulse signals are reflected by a human body to be detected, a wall body and the ground in an effective detection area, received signals are accumulated line by line to form a two-dimensional radar signal matrix M1; after direct current components are removed, band-pass filtering and noise waves are carried out, a row of signals with the largest energy are selected and used as vital sign signals S1; then, by eliminating the movement and the respiratory interference of the vital sign signal S1, a heartbeat signal S2 and a pulse wave signal S3 are obtained; extracting a single-beat pulse wave signal S4 according to the period and the correlation of the two signals, and training a blood pressure prediction neural network after standardization; and finally, aiming at a new subject, acquiring a single-beat pulse wave signal S4 of the new subject, standardizing the signal, and inputting the standardized signal into a trained blood pressure prediction neural network to predict the blood pressure. The invention does not need to contact the skin and can conveniently and accurately measure the blood pressure.

Description

A non-contact blood pressure measurement method based on pulsed ultra-wideband radar

technical field

The invention belongs to the field of non-contact vital sign detection, and in particular relates to a non-contact blood pressure measurement method based on pulse ultra-wideband radar.

Background technique

Hypertension is one of the most common chronic diseases. According to statistics, about a quarter of the people in China suffer from hypertension, which can induce many cardiovascular and cerebrovascular diseases. Therefore, blood pressure monitoring is of great significance for disease prevention. Most of the commonly used sphygmomanometers are cuff-type and need to be tightly tied to the arm of the subject. This is not suitable for patients whose arms are injured and are not suitable for skin-to-skin contact. The cuff also adds inconvenience to the measurement.

Non-contact measurement provides convenience for long-term blood pressure monitoring. Ultra-wideband radar has the advantages of low power consumption, strong anti-interference ability, and strong penetration. There have been many researches on non-contact vital sign monitoring using ultra-wideband radar. With the development of technology, the radar-based non-contact blood pressure measurement method has gradually appeared and has a good development prospect.

Currently, radar-based blood pressure measurement methods combine electrocardiography or photoplethysmography with radar. This type of method also requires the sensor to be attached to the body, which brings inconvenience. The method of only using radar is generally: by obtaining the beating waveform of the heart or carotid artery and wrist radial artery, according to certain characteristics of the waveform, such as systolic period, amplitude, slope and other parameters, calculate the pulse wave transit time, according to linear regression Fit an expression for blood pressure. However, these methods usually require the subject to lie flat, remain completely still or even hold his breath during the measurement. However, when the body moves slightly, the noise interference is large, the pulse wave will be distorted, and it is difficult to directly extract blood pressure-related features. The fit is poor.

In the prior art, document 1: CN202111244372.5 provides a human blood pressure detection method based on millimeter-wave radar signals, collecting millimeter-wave radar signals at a preset height from the human wrist, and generating predictions through a three-part neural network blood pressure. This method needs to train the three parts of the neural network separately, and the operation is cumbersome, which is not conducive to long-term real-time monitoring in the future, and the generalization ability of the neural network is weak, and all rely on the neural network for data processing and prediction. The robustness may also be insufficient. Document 2: CN202210010289.X also provides a blood pressure detection method based on millimeter-wave radar, using wavelet packet decomposition to reconstruct the pulse wave signal, and performing feature point detection on it, and performing regression analysis based on the feature point data to obtain blood pressure detection results , but in the actual scene, the waveform is often distorted, and the feature point data will be blurred or even abnormal, which cannot fit the relationship with blood pressure well.

Contents of the invention

In view of this, the present invention proposes a non-contact blood pressure measurement method based on pulse ultra-wideband radar, which can automatically capture the deep temporal characteristics of the pulse wave signal by using a neural network, and accurately and conveniently monitor blood pressure.

The non-contact blood pressure measurement method based on pulsed ultra-wideband radar, the specific steps are as follows:

Step 1. The ultra-wideband radar continuously transmits pulse signals, which are received by the receiving antenna after being reflected by the human body, wall and ground in the effective detection area, and stacked and accumulated row by row to form a two-dimensional radar signal matrix M1;

The two-dimensional radar echo signal matrix is recorded as:

The row vector represents the fast time dimension, which is positively correlated with the detection distance; the column vector represents the slow time dimension, which is positively correlated with the data accumulation time, and each element x _ij of the matrix represents the radar signal sampling value.

Step 2: Preprocessing the radar signal matrix M1, removing the DC component, performing band-pass filtering and removing clutter, to obtain the preprocessed radar signal matrix M.

Step 3. Calculate the energy value of each column from the radar signal matrix M by column, and select the largest column signal as the vital sign signal S1;

Step 4: Eliminate movement and respiration interference on the vital sign signal S1 through a variational mode decomposition algorithm to obtain a heartbeat signal S2 and a pulse wave signal S3.

First, for the vital sign signal S1, obtain modal signals with different center frequencies through variational mode decomposition, perform fast Fourier transform to calculate its frequency, and select a standard heart rate value;

The formula for variational modal decomposition is:

Where {μ _i } and {ω _i } correspond to the i-th modal component and center frequency after decomposition respectively, δ(t) is the Dirac function, β is the quadratic penalty factor, γ is the Lagrangian multiplier sub, f(t) is the original signal.

Then, a modal signal whose center frequency is a standard heart rate value is selected as the heartbeat signal S2.

Finally, except for the modal signal with the smallest energy, the heartbeat signal and the modal signals with energy smaller than the heartbeat signal are selected to be superimposed, and regarded as the pulse wave signal S3.

Step 5, according to the cycle and correlation of the heartbeat signal S2 and the pulse wave signal S3, extract the single beat pulse wave signal S4;

Single-beat pulse wave extraction includes:

Firstly, the single-beat heartbeat signal is segmented by taking the minimum value point of the heartbeat signal as the segmentation point.

Then, the single-beat pulse wave signal is segmented by taking the minimum value point of the pulse wave signal closest to the minimum value point of the heartbeat signal as the segmentation point.

Finally, calculate the correlation coefficient between the single-beat heartbeat signal and the single-beat pulse wave signal, and select the single-beat pulse wave signal S4 with a correlation coefficient greater than 0.6;

The formula for calculating the correlation coefficient is:

Where g _p is a single-beat pulse wave signal, g _h is a single-beat heartbeat signal, Cov() is a covariance formula, and σ ² () is a variance formula.

Step 6, setting up the blood pressure prediction neural network, and inputting the blood pressure prediction neural network into the blood pressure prediction neural network after normalizing the single-shot pulse wave signal S4 for training.

Step 7. For new subjects, repeat the above steps 1 to 5 to obtain the single-shot pulse wave signal S4, and input it into the trained blood pressure prediction neural network after normalization to directly obtain the blood pressure result.

The advantages of the present invention are:

The invention provides a non-contact blood pressure measurement method based on pulse ultra-wideband radar, which can conveniently and accurately measure blood pressure without touching the skin. Different from methods based on signal processing, the present invention not only uses ultra-wideband radar to extract pulse wave signals related to blood pressure, but also adaptively analyzes the timing characteristics of signals through neural networks, and outputs blood pressure results according to feature mapping.

Description of drawings

FIG. 1 is a flowchart of a non-contact blood pressure measurement method based on pulsed ultra-wideband radar according to the present invention.

Fig. 2 is a schematic diagram of the radar signal matrix after preprocessing according to the present invention.

Fig. 3 is a schematic diagram of vital sign signals extracted by the present invention.

Fig. 4 is a schematic diagram of the pulse wave signal extracted by the present invention.

Fig. 5 is a schematic diagram of a single beat pulse wave extracted by the present invention.

Fig. 6 is a schematic diagram of the structure of the blood pressure prediction neural network of the present invention.

detailed description

The present invention will be further explained in detail below in conjunction with the embodiments and the accompanying drawings. The embodiments described below are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

A non-contact blood pressure measurement method based on pulsed ultra-wideband radar in the present invention can automatically capture the deep time sequence characteristics of pulse wave signals by using neural network, and accurately and conveniently monitor blood pressure; it is specifically divided into the following steps:

Step 1: Radar signal acquisition. The ultra-wideband radar continuously transmits pulse signals. After the signals are reflected by objects, walls and the ground in the effective detection area, they are received by the receiving antenna and stacked and accumulated row by row to form a two-dimensional radar signal matrix M1;

Step 2: Radar signal preprocessing. Remove the DC component, perform bandpass filtering and clutter removal.

Step 3: Vital sign signal extraction. For the preprocessed radar signal matrix M, select a column of signals with the largest energy in the fast time dimension as the vital sign signal S1, and the energy of each column of signals can be expressed by calculating its variance.

Step 4: Movement and Breathing Interference Elimination. Through variational mode decomposition (Variational ModeDecomposition, VMD), the vital sign signal S1 is eliminated from movement and breathing interference, and a heartbeat signal S2 and a pulse wave signal S3 are obtained.

Step 5: Single beat pulse wave extraction. According to the period and correlation between the heartbeat signal S2 and the pulse wave signal S3, the single-beat pulse wave signal S4 is extracted, and standardized as the input of the neural network for blood pressure prediction.

Step 6: Set up the blood pressure prediction neural network.

Step 7: Divide the training set, verification set and test set, and train the neural network.

Step 8: Predict the subject's blood pressure through the trained blood pressure prediction neural network.

As shown in Figure 1, the specific steps are as follows:

Step 1. The ultra-wideband radar continuously transmits pulse signals, which are received by the receiving antenna after being reflected by the human body, wall and ground in the effective detection area, and stacked and accumulated row by row to form a two-dimensional radar signal matrix M1;

The two-dimensional radar echo signal matrix is recorded as:

The row vector represents the fast time dimension, which is positively correlated with the detection distance; the column vector represents the slow time dimension, which is positively correlated with the data accumulation time, and each element x _ij of the matrix represents the radar signal sampling value.

Step 2: Preprocessing the radar signal matrix M1, removing the DC component, performing band-pass filtering and removing clutter, to obtain the preprocessed radar signal matrix M.

The preprocessing operation includes: firstly, segmenting each row of the radar signal matrix M1 according to a specific application scenario, taking an average value for each segment, subtracting the corresponding average value from each segment, and removing the DC component.

In this embodiment, each row is selected to be segmented with 156 columns of data;

Then, for the radar signal matrix from which the DC component has been removed, a bandpass filter with a passband of 6.5GHz-8GHz is used to filter each row of data.

Finally, for the radar signal matrix after bandpass filtering, the static clutter is removed by the moving average algorithm, and the radar signal matrix M is obtained;

The static clutter of each row is the weighted sum of the static clutter of the previous row and the data of this row, expressed as:

C(t,τ)=a C(t-1,τ)+(1-a)x(t,τ)

Among them, C(t,τ) is the static clutter of this row, a is the weight, which is set to 0.9 in this embodiment, C(t-1,τ) is the static clutter of the previous row, and x(t,τ) is For the data in this row, the static clutter of the first row is 0.1·x(1,τ) during initialization.

Step 3. Calculate the energy value of each column from the radar signal matrix M by column, and select the largest column signal as the vital sign signal S1;

The energy of each column signal is expressed by calculating the sum of its squares, and the energy calculation formula of the jth column signal is:

Where m is the signal length.

Step 4: Eliminate movement and respiration interference on the vital sign signal S1 through a variational mode decomposition algorithm to obtain a heartbeat signal S2 and a pulse wave signal S3.

First, for the vital sign signal S1, obtain modal signals with different center frequencies through variational mode decomposition, perform fast Fourier transform to calculate its frequency, and select a standard heart rate value;

The normal human heart rate range is 0.8-2Hz, that is, 50-120 beats per minute, so the frequency falling within 0.8-2Hz is selected as the heart rate value.

The formula for variational modal decomposition is:

Where {μ _i } and {ω _i } correspond to the i-th modal component and center frequency after decomposition respectively, δ(t) is the Dirac function, β is the quadratic penalty factor, γ is the Lagrangian multiplier sub, f(t) is the original signal.

Then, a modal signal whose center frequency is a standard heart rate value is selected as the heartbeat signal S2.

Finally, except for the modal signal with the smallest energy, the heartbeat signal and the modal signals with energy smaller than the heartbeat signal are selected to be superimposed, and regarded as the pulse wave signal S3.

Step 5, according to the cycle and correlation of the heartbeat signal S2 and the pulse wave signal S3, extract the single beat pulse wave signal S4;

Single-beat pulse wave extraction includes:

Firstly, the single-beat heartbeat signal is segmented by taking the minimum value point of the heartbeat signal as the segmentation point.

Then, the single-beat pulse wave signal is segmented by taking the minimum value point of the pulse wave signal closest to the minimum value point of the heartbeat signal as the segmentation point.

Finally, calculate the correlation coefficient between the single-beat heartbeat signal and the single-beat pulse wave signal, and select the single-beat pulse wave signal S4 with a correlation coefficient greater than 0.6;

The formula for calculating the correlation coefficient is:

Where g _p is a single-beat pulse wave signal, g _h is a single-beat heartbeat signal, Cov() is a covariance formula, and σ ² () is a variance formula.

Step 6, setting up the blood pressure prediction neural network, and inputting the blood pressure prediction neural network into the blood pressure prediction neural network after normalizing the single-shot pulse wave signal S4 for training.

The blood pressure prediction neural network includes the following structures:

First, the standardized single-beat pulse wave signal S4 is used as a sample, and input to two branches for convolution of different scales;

The structure of the first branch is 5×1 one-dimensional convolution layer, BN layer, maximum pooling layer, 5×1 one-dimensional convolution layer, BN layer, 5×1 one-dimensional convolution layer, BN layer, 1× 1 one-dimensional convolution layer, maximum pooling layer; the structure of the second branch is 3×1 one-dimensional convolution layer, BN layer, maximum pooling layer, 3×1 one-dimensional convolution layer, BN layer, 3× 1 one-dimensional convolutional layer, BN layer, 1×1 one-dimensional convolutional layer, maximum pooling layer.

Then, the feature maps obtained by the convolution of the two branches are spliced, and then pass through the Dropout layer, two layers of gated recurrent units (Gated Recurrent Unit, GRU), multi-head attention module, dropout layer and two layers of fully connected layers, and output the prediction results ;

Among them, the Dropout layer is used to prevent overfitting, the gated recurrent unit and the multi-head attention module can capture the deep temporal features of the signal, and the fully connected layer can map the features to the output blood pressure value.

Step 7. For new subjects, repeat the above steps 1 to 5 to obtain the single-shot pulse wave signal S4, and input it into the trained blood pressure prediction neural network after normalization to directly obtain the blood pressure result.

Example:

In this example, the pulse ultra-wideband radar is installed on a bracket about 60 cm away from the subject's chest. The person to be detected is an adult male, aged 23, weighing 65 kg, and 180 cm tall.

The steps of the processing flow are as follows:

Step 1: Use the pulse ultra-wideband radar to obtain radar signals. The pulse ultra-wideband radar continuously sends pulse signals to the test site. After being reflected by the human body to be tested and the ground in the test site, it is received by the radar receiving antenna. matrix data form for storage and further analysis;

代表实数域。Among them, the row vector of the two-dimensional matrix represents the fast time dimension of the pulse ultra-wideband radar, which is positively correlated with the detection distance. The detection distance is set to 3 meters in this example, that is, each row vector of the two-dimensional matrix in this example has 437 columns The fixed length of ; the column vector represents the slow time dimension, which is positively correlated with the data accumulation time. The radar sampling frame rate is set to 20 frames per second, that is, the length of the column vector of the two-dimensional matrix in this example increases by 20 rows per second. Each group collects data for 10 seconds to form a 200*437 two-dimensional matrix M1,

represents the field of real numbers.

Step 2: Preprocess the radar signal matrix M1, remove the DC component, perform band-pass filtering and remove clutter; obtain the preprocessed radar signal matrix M

The preprocessed radar signal matrix is shown in Figure 2;

在快时间维度挑选能量最大的一列信号，视为生命体征信号S1，

每列信号的能量可以通过计算其方差的形式来表示，生命体征信号如图3所示。Step 3: Vital sign signal extraction. In a static environment, the ups and downs caused by the chest vibration occupy the maximum energy, so the preprocessed radar signal matrix

Select a series of signals with the largest energy in the fast time dimension as the vital sign signal S1,

The energy of each column signal can be expressed by calculating its variance, and the vital sign signal is shown in Figure 3.

和脉搏波信号

脉搏波信号如图4所示；Step 4: Movement and Breathing Interference Elimination. Through variational mode decomposition, the vital sign signal S1 eliminates motion and respiration interference, and obtains the heartbeat signal

and pulse wave signal

The pulse wave signal is shown in Figure 4;

Specific steps include:

Step 4-1: For the vital sign signal S1, obtain modal signals with different center frequencies through variational mode decomposition, and perform Fast Fourier Transform (FFT) to calculate its frequency. The heart rate range of normal people is 0.8 -2Hz, so select the frequency that falls within 0.8-2Hz as the heart rate value.

Step 4-2: Based on the heart rate value as a standard, select the modal signal whose center frequency is closest to the heart rate as the heartbeat signal S2.

Step 4-3: Because environmental noise usually occupies the smallest energy, in order to restore the waveform with more blood pressure related information, in addition to the modal signal with the smallest energy, select the heartbeat signal and the modal signal with energy less than the heartbeat signal to superimpose, It is regarded as the pulse wave signal S3.

标准化后作为神经网络的输入，如图5所示，具体步骤包括：Step 5: Single beat pulse wave extraction. According to the cycle and correlation of the heartbeat signal S2 and the pulse wave signal S3, extract the single beat pulse wave signal

After normalization, it is used as the input of the neural network, as shown in Figure 5, and the specific steps include:

Step 5-1: Segment the single-beat heartbeat signal with the minimum point of the heartbeat signal as the segmentation point.

Step 5-2: Segment the single-beat pulse wave signal with the minimum value point of the pulse wave signal closest to the minimum value point of the heartbeat signal as the segmentation point.

Step 5-3: Calculate the correlation coefficient between the single-beat heartbeat signal and the single-beat pulse wave signal, and select the single-beat pulse wave signal with a correlation coefficient greater than 0.6, because the period of the pulse wave signal is generally not greater than 1.5 seconds, so the single-beat The pulse wave is intercepted or zero-filled at 30 sampling points in 1.5 seconds, and then standardized as the input of the neural network.

Step 6: Set up the neural network for blood pressure prediction. The network structure is shown in Figure 6, including:

Firstly, convolution of different scales is performed through two branches, and the structure is sequentially one-dimensional convolutional layer, BN layer, maximum pooling layer, one-dimensional convolutional layer, BN layer, one-dimensional convolutional layer, BN layer, 1×1 One-dimensional convolution layer, maximum pooling layer, the convolution kernel sizes of the first three convolution layers of the two branches are different, 5 and 3 respectively, and the rest of the layers are the same.

Convolutions of different scales can extract different features of the waveform. The 1×1 convolution kernel is used to increase the dimension, the BN layer is used to prevent gradient explosion, and the maximum pooling layer is used to reduce the amount of calculation and increase robustness.

Then, the feature maps obtained by the two branches are spliced, and the prediction results are output through the dropout layer, two layers of gated recurrent units, multi-head attention module, dropout layer, and two layers of fully connected layers. The dropout layer is used to prevent over-fitting, the GRU network can obtain the sequence characteristics of the sequence, and the multi-head attention module can capture the deep sequence relationship of the sequence.

Step 7: Train the neural network based on the single-beat pulse wave signal and its corresponding label, where the label is the real blood pressure value measured by the sphygmomanometer. The loss function uses Mean Square Error (Mean Square Error, MSE), the activation function is Relu, the optimizer is Adam, the batch size is 32, and a total of 50 rounds of training are performed.

Step 8: Predict the blood pressure of the subjects through the trained blood pressure prediction neural network.

Obviously, those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. In this way, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims (6)

1. A non-contact blood pressure measuring method based on pulse ultra-wideband radar is characterized by comprising the following specific steps:

firstly, an ultra-wideband radar continuously transmits pulse signals, the pulse signals are reflected by a human body, a wall body and the ground in an effective detection area, and then are received by a receiving antenna, and are stacked and accumulated line by line to form a two-dimensional radar signal matrix M1; removing direct current components, performing band-pass filtering and clutter removal to obtain a preprocessed radar signal matrix M;

then, calculating the energy value of each column from the radar signal matrix M according to the column, selecting the largest column of signals as the vital sign signals S1; eliminating motion and respiration interference on the vital sign signal S1 through a variational modal decomposition algorithm to obtain a heartbeat signal S2 and a pulse wave signal S3; further extracting a single-beat pulse wave signal S4 according to the period and the correlation of the heartbeat signal S2 and the pulse wave signal S3;

the single-beat pulse wave extraction process comprises the following steps:

firstly, taking a heartbeat signal minimum value point as a segmentation point, and segmenting a single-beat heartbeat signal;

then, taking the minimum value point of the pulse wave signal closest to the minimum value point of the heartbeat signal as a segmentation point, and segmenting the single-beat pulse wave signal;

finally, calculating a correlation coefficient between the single-beat heartbeat signal and the single-beat pulse wave signal, and selecting a single-beat pulse wave signal S4 with the correlation coefficient larger than 0.6;

the correlation coefficient calculation formula is:

wherein g is _p For a single beat of pulse wave signal, g _h For a single beat heartbeat signal, cov () is the covariance formula, σ ² () Is a variance formula;

finally, setting a blood pressure prediction neural network, standardizing the single-beat pulse wave signal S4, and inputting the standardized single-beat pulse wave signal into the blood pressure prediction neural network for training; and inputting a new single-beat pulse wave signal S4 extracted by the subject after the training is finished, and inputting after standardization to directly obtain a blood pressure result.

2. The method for non-contact blood pressure measurement based on the pulse ultra-wideband radar as claimed in claim 1, wherein the two-dimensional radar echo signal matrix

Wherein the row vector represents a fast time dimension and is positively correlated with the detection distance; the column vector represents the slow time dimension and is positively correlated with the data accumulation time, and each element x of the matrix _ij Representing radar signal sample values.

3. The method for measuring the blood pressure in the non-contact mode based on the pulse ultra-wideband radar as claimed in claim 1, wherein the pre-processing of the two-dimensional radar signal matrix M1 comprises:

firstly, segmenting each row of a radar signal matrix M1 according to a specific application scene, taking an average value of each segment, subtracting the corresponding average value of each segment, and removing a direct current component;

then, filtering each line of data of the radar signal matrix with the direct-current component removed by using a band-pass filter;

finally, removing static clutter from the filtered radar signal matrix through a moving average algorithm to obtain a radar signal matrix M;

the static clutter of each row is the weighted sum of the static clutter of the previous row and the data of the current row, and is expressed as:

C(t,τ)＝a·C(t-1,τ)+(1-a)·x(t,τ)

wherein C (t, tau) is the static clutter of the current row, a is the weight, C (t-1, tau) is the static clutter of the previous row, x (t, tau) is the data of the current row, and the static clutter of the first row is 0.1. X (1, tau) during initialization.

4. The method for non-contact blood pressure measurement based on the pulse ultra-wideband radar is characterized in that the energy value of each column is calculated from the radar signal matrix M according to the column, and the energy calculation formula of the jth column signal is as follows:

where m is the signal length.

5. The non-contact blood pressure measuring method based on the pulse ultra-wideband radar according to claim 1, wherein the specific process of obtaining the heartbeat signal S2 and the pulse wave signal S3 is as follows:

firstly, obtaining modal signals with different center frequencies from a vital sign signal S1 through variational modal decomposition, carrying out fast Fourier transform to calculate the frequencies of the modal signals, and selecting a standard heart rate value;

the formula of the variation modal decomposition is as follows:

wherein [ mu ] _i And { omega } and _i the i-th modal component and the center frequency after decomposition are respectively corresponding to the f-th modal component and the center frequency, delta (t) is a Dirac function, beta is a secondary penalty factor, gamma is a Lagrangian multiplier operator, and f (t) is an original signal;

then, selecting a modal signal with the center frequency being a standard heart rate value as a heartbeat signal S2;

and finally, except the modal signal with the minimum energy, selecting a heartbeat signal and a modal signal with the energy smaller than the heartbeat signal for superposition, and regarding the superposition as a pulse wave signal S3.

6. The method for non-contact blood pressure measurement based on the pulse ultra-wideband radar is characterized in that the blood pressure prediction neural network comprises the following structures:

firstly, taking a single beat pulse wave signal S4 after standardization as a sample, and inputting two branches to carry out convolution with different scales;

the structure of the first branch is sequentially a 5 multiplied by 1 one-dimensional convolution layer, a BN layer, a maximum pooling layer, a 5 multiplied by 1 one-dimensional convolution layer, a BN layer, a 1 multiplied by 1 one-dimensional convolution layer and a maximum pooling layer; the structure of the second branch is sequentially a 3 multiplied by 1 one-dimensional convolution layer, a BN layer, a maximum pooling layer, a 3 multiplied by 1 one-dimensional convolution layer, a BN layer, a 1 multiplied by 1 one-dimensional convolution layer and a maximum pooling layer;

then, splicing the feature maps obtained by convolution of the two branches, sequentially passing through a Dropot layer, two layers of gating circulation units, a multi-head attention module, a Dropot layer and two layers of full connection layers, and outputting a prediction result;

the Dropout layer is used for preventing overfitting, the gating circulation unit and the multi-head attention module can capture deep timing characteristics of signals, and the full connection layer can map the characteristics to output blood pressure values.

Images