Blood pressure measurement system and method based on collaborative learning of oscillometric wave and PPG signal

Abstract

The present invention provides a blood pressure measurement system and method based on collaborative learning of oscillation waves and PPG signals, and relates to the field of wearable medical health monitoring technology. The present invention gives full play to the advantages of CNN in local feature extraction, Transformer in time series modeling, and PPG signals in assisting blood pressure prediction, which can significantly improve the accuracy of blood pressure estimation. It uses deep learning to automatically extract complex features and reduce dependence on manual feature engineering, and uses traditional machine learning to enhance model interpretability. Combining the advantages of both, it improves the accuracy and robustness of blood pressure estimation. The present invention has a wide range of applicability and can be applied to various scenarios such as home health monitoring, clinical monitoring, and wearable devices. At the same time, it uses multi-task learning to simultaneously predict systolic and diastolic blood pressure and share feature extraction networks. Moreover, the algorithm is a lightweight algorithm that can be deployed on wearable devices to achieve low-power, high-efficiency real-time blood pressure estimation, providing strong support for health monitoring.

Description

Technical Field

[0001] The present invention relates to the field of wearable medical health monitoring technology, and in particular to a blood pressure measurement system and method based on collaborative learning of oscillation waves and PPG signals. Background Art

[0002] Hypertension is one of the most common chronic diseases worldwide, affecting approximately 1.8 billion adults worldwide, according to the World Health Organization. Accurate and convenient blood pressure monitoring is key to early detection of hypertension, evaluating treatment effectiveness, and preventing complications. While the traditional mercury sphygmomanometer is the gold standard, it relies on manual operation and cannot provide real-time monitoring. Electronic sphygmomanometers utilize the oscillometric method for non-invasive measurement, but their accuracy is significantly affected by individual differences and measurement posture. With the widespread adoption of wearable devices, continuous blood pressure monitoring technology based on photoplethysmography (PPG) has become a research hotspot. However, existing methods generally suffer from issues such as insufficient measurement accuracy and weak anti-interference capabilities.

[0003] Traditional oscillometric blood pressure monitors rely on Korotkoff sound measurements as a calibration standard. However, the Korotkoff sound method itself is subject to significant subjectivity and is susceptible to environmental noise. Studies have shown that different doctors' judgments on the onset of Korotkoff sounds can vary by as much as 5-10 mmHg. Existing machine learning models are often trained based on the mapping between oscillometric waves and Korotkoff sound blood pressure readings, but their accuracy is limited by the reliability of the reference standard.

[0004] Existing methods primarily rely on handcrafted features (such as oscillation wave peak and area) or single deep learning models (such as CNNs) for feature extraction. Handcrafted features struggle to capture complex hemodynamic changes, while pure deep learning models suffer from poor interpretability and overfitting to small sample sizes. Furthermore, existing research rarely integrates the multimodal features of oscillation waves and PPG signals, failing to fully utilize the complementary information between different signal modalities.

[0005] Existing models are typically trained on specific populations (such as healthy adults) and are less adaptable to specific groups such as those with obesity and arteriosclerosis. Clinical data show that traditional methods can have an error of more than 15 mmHg in systolic blood pressure measurement in the elderly.

[0006] Patent CN117678987A proposes a blood pressure measurement method that acquires the PPG and accelerometer (ACC) signals of the person being tested, processes them to obtain R-wave peak data to calculate PTT data, and simultaneously extracts PPG signal features. Both are then input into a pre-trained model to produce blood pressure measurement results. This method utilizes signal processing to avoid manual measurement and reduce costs. However, this method has limitations. Relying solely on PPG and ACC signals, it is prone to domain shift, resulting in large measurement errors in specific populations and difficulty ensuring high accuracy and reliability. Furthermore, the lack of clarity regarding the source and characteristics of the training data leads to insufficient model generalization and inability to adapt to diverse scenarios and populations.

[0007] Therefore, there is an urgent need for a joint measurement method of oscillation waves and PPG signals suitable for home health management that can solve the problem of small samples and enhance clinical credibility. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a blood pressure measurement system and method based on collaborative learning of oscillation waves and PPG signals to achieve accurate estimation of blood pressure.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0010] In one aspect, the present invention provides a blood pressure measurement system based on collaborative learning of oscillatory waves and PPG signals, comprising an oscillatory wave and PPG signal database construction module, a data preprocessing module, a feature extraction module, and a blood pressure estimation module;

[0011] The oscillatory wave and PPG signal database stores oscillatory wave and PPG signals and corresponding diastolic and systolic blood pressures of 3,000 subjects;

[0012] The data preprocessing module is used to filter the acquired oscillation wave signal and PPG signal to remove common noise;

[0013] The feature extraction module constructs a CNN-Transformer network, including a CNN feature extractor, a position encoder, a multi-layer Transformer encoder, a global average pooling layer, a fully connected layer, and a feature fusion module. The CNN feature extractor is used to extract local features of the oscillatory wave signal and the PPG signal, and the multi-layer Transformer encoder is used to extract the temporal features of the oscillatory wave signal. On this basis, the position encoder is used to add position encoding to the extracted features to retain sequence information and enhance the model's perception of temporal features. The global average pooling layer performs global average pooling on the output of the multi-layer Transformer encoder and then inputs it into the fully connected layer. The feature fusion module combines time-frequency analysis and statistical features to extract manual features, fuses them with the features extracted by deep learning, and divides the training set into the test set.

[0014] The blood pressure estimation module uses the XGBoost model to estimate blood pressure based on the extracted features, achieving accurate prediction of systolic and diastolic blood pressure.

[0015] Furthermore, the oscillation wave and PPG signal database construction module includes an oscillation wave signal device, a cuff, a PPG sensor, a blood pressure measuring device, and a database; first, the oscillation wave signal device is correctly placed on the subject's upper arm and fixed with the cuff; the cuff will gradually increase the pressurization and record the subject's oscillation wave data under different pressures, and store it in the database; then, the PPG sensor is correctly placed at the fingertips. During the collection process, the subject is required to remain still to reduce the impact of motion artifacts on the PPG signal, and the subject's systolic and diastolic blood pressures are measured by the blood pressure measuring device and stored in the database.

[0016] Furthermore, the data preprocessing module includes an oscillation wave data preprocessing module and a PPG data preprocessing module;

[0017] The oscillation wave data preprocessing module first uses a low-pass filter to reduce noise on the original oscillation wave signal to remove high-frequency noise in the signal, thereby retaining the low-frequency components in the oscillation wave signal; then performs a high-pass filter to remove the low-frequency components in the signal, thereby retaining the high-frequency components in the oscillation wave signal;

[0018] The PPG data preprocessing module first uses low-pass filtering to remove high-frequency noise in the PPG signal and retains the useful components in the PPG signal; then uses high-pass filtering to remove low-frequency drift or baseline drift to retain important information in the PPG signal; and uses a waveform quality assessment method based on wavelet transform to detect and remove motion artifacts or poor signal segments. The calculation formula is as follows:

[0019] (1);

[0020] in, a and b are scale and translation parameters, ψ is the mother wavelet function; f(t) is the original signal, which is a function of t.

[0021] The Savitzky-Golay filter technology is used to smooth the signal and reduce high-frequency fluctuations. The calculation formula is as follows:

[0022] (2);

[0023] in, c j is the Savitzky-Golay coefficient, m is half the window length, x i+j is the original data value of the input signal at position i+j, where j is the offset index within the window.

[0024] Furthermore, the specific implementation method of the blood pressure estimation module is:

[0025] First, set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth;

[0026] Then use the training set to train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean square error, as shown below:

[0027] (9);

[0028] in, θ is the parameter of the XGBoost model, y i It is the real blood pressure value. is the blood pressure value predicted by the model;

[0029] Use the trained XGBoost model to predict the test set and get the predicted blood pressure value ;

[0030] Then use the evaluation indicators to measure the performance of the model, including mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE):

[0031] (10);

[0032] (11);

[0033] (12);

[0034] Among them, ntest is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0035] On the other hand, the present invention also provides a blood pressure measurement method based on collaborative learning of oscillation waves and PPG signals, which is implemented by the above-mentioned blood pressure measurement system based on collaborative learning of oscillation waves and PPG signals, and includes the following steps:

[0036] Step 1: Construct an oscillatory wave signal and a PPG signal database, including oscillatory wave signals, PPG signals and their corresponding systolic and diastolic pressures;

[0037] Step 2: Preprocess the oscillation wave signals in the database, including low-pass filtering and high-pass filtering, to optimize the noise, baseline drift and amplitude differences between different waveforms in the oscillation wave signals;

[0038] Step 3: Preprocess the PPG signals in the database, including low-pass filtering, high-pass filtering, artifact removal, and smoothing to remove noise, baseline drift, and amplitude differences between different waveforms in the PPG signals;

[0039] Step 4: Perform feature extraction and blood pressure estimation modeling; establish a CNN-Transformer network, including a CNN feature extractor, positional encoding, a multi-layer Transformer encoder, global average pooling, and a fully connected layer; use CNN to extract local features of the oscillatory wave signal and PPG signal and concatenate them to form a joint feature; then use the multi-layer Transformer encoder to extract the temporal features of the oscillatory wave signal; add positional encoding to the extracted features to preserve sequence information and enhance the model's perception of temporal features; perform global average pooling on the output of the multi-layer Transformer encoder and input it into the fully connected layer;

[0040] Step 5: Predict blood pressure based on the newly collected oscillation wave signal data and PPG signal, including preprocessing of the oscillation wave signal and PPG signal and blood pressure estimation based on machine learning and deep learning models; fuse the features extracted by deep learning with the manual features extracted by time-frequency analysis and statistical features, and then input them into the XGBoost model to learn the mapping relationship between the oscillation wave signal, PPG signal and systolic and diastolic blood pressure, so as to achieve accurate estimation of diastolic and systolic blood pressure.

[0041] Furthermore, the specific method of step 1 is:

[0042] Step 1.1: Place the oscillatory wave signal device correctly on the subject's upper arm and secure it with a cuff;

[0043] Step 1.2: The cuff is gradually pressurized, and the oscillation wave data of the subject at different pressures is recorded and stored;

[0044] Step 1.3: Place the PPG sensor correctly on the fingertip; during the acquisition process, ask the subject to remain still to reduce the impact of motion artifacts on the PPG signal;

[0045] Step 1.4: Measure the subject's systolic and diastolic blood pressure using a blood pressure measuring device.

[0046] Furthermore, the specific method of step 2 is:

[0047] Step 2.1: Use a low-pass filter to remove high-frequency noise from the original oscillation wave signal, thereby retaining the low-frequency components of the oscillation wave signal;

[0048] Step 2.2: Use high-pass filtering to remove the influence of baseline drift on the oscillation wave, thereby retaining the high-frequency components in the oscillation wave signal.

[0049] Furthermore, the specific method of step 3 is:

[0050] Step 3.1: Use low-pass filtering to remove high-frequency noise and retain the useful components in the PPG signal;

[0051] Step 3.2: Use high-pass filtering to remove low-frequency drift or baseline drift and retain important information in the PPG signal;

[0052] Step 3.3: Use wavelet transform waveform quality assessment method to detect and remove motion artifacts or bad signal segments;

[0053] Step 3.4: Smooth the signal using Savitzky-Golay filter technique to reduce high frequency fluctuations.

[0054] Furthermore, the specific method of step 4 is:

[0055] Step 4.1: Prepare the data by loading the oscillatory wave signal, PPG signal and its corresponding diastolic and systolic blood pressure values, and dividing the data into training and test sets.

[0056] Step 4.2: Build a deep learning model, that is, build a CNN-Transformer network, including a CNN feature extractor, positional encoding, a multi-layer Transformer encoder, global average pooling, and a fully connected layer;

[0057] Step 4.2.1: Input oscillatory wave signal and PPG signal;

[0058] Step 4.2.2: Construct the first CNN network to extract features from the oscillation wave signal. The specific method is as follows:

[0059] Construct convolution layer 1 with a kernel size of 5, a number of kernels of 16, and a ReLU activation function.

[0060] Pooling layer 1, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0061] Construct the second convolution layer with a kernel size of 5, a number of kernels of 32, and a ReLU activation function.

[0062] Pooling layer 2, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0063] Output layer 1, outputs features extracted from the oscillation wave signal;

[0064] Step 4.2.3: Construct the second CNN network to extract features of the PPG signal. The specific method is as follows:

[0065] Construct convolution layer 3 with a kernel size of 3, a number of kernels of 16, and a ReLU activation function.

[0066] Pooling layer 3, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0067] Construct convolutional layer 4 with a kernel size of 5, a number of kernels of 32, and a ReLU activation function.

[0068] Pooling layer 4, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0069] Construct convolution layer 5 with a kernel size of 7 and a number of 64 kernels, using the ReLU activation function.

[0070] Pooling layer 5, the pooling method is average pooling, and the pooling kernel size is 2;

[0071] Output layer 2, outputs features extracted from PPG signals;

[0072] Step 4.2.4: Concatenate the feature vectors extracted from the oscillatory wave signal and the PPG signal to form a joint feature, then transform the input shape and convert the output of the CNN part into the shape required by the Transformer network;

[0073] Step 4.2.5: Build a Transformer network to capture the timing dependency of the oscillatory wave signal. The specific method is as follows:

[0074] Step 4.2.5.1: Input embedding, converting the input features into high-dimensional feature vectors;

[0075] Step 4.2.5.2: Position coding, add position coding, retain the timing information, introduce the timing information into the feature vector, and calculate the formula as shown in formula (5):

[0076] (5);

[0077] in, POS For location, i For the dimension, d model is the dimension of the model;

[0078] Step 4.2.5.3: Multi-head self-attention mechanism, calculate attention scores, and perform weighted summation based on these scores to capture the dependencies between different positions in the input sequence; the calculation process of self-attention is as shown in formula (6):

[0079] (6);

[0080] in, Q 、 K and V Obtained by linear transformation of input features, d k for K Dimensions;

[0081] Step 4.2.5.4: Layer normalization to stabilize and accelerate the training process;

[0082] Step 4.2.5.5: Feedforward neural network, further process features through nonlinear activation function (ReLU) to enhance the expressive power of the model. The calculation process is as shown in formula (7):

[0083] (7);

[0084] Among them, max(0,xW1+b1) is a nonlinear activation function, W 1 and W 2 is the weight matrix, b 1 and b 2 is bias;

[0085] Step 4.2.5.6: Layer normalization to stabilize and accelerate the training process;

[0086] Step 4.2.5.7: Repeat steps 4.2.5.3 to 4.2.5.6 six times;

[0087] Step 4.2.6: Perform global average pooling on the output of the Transformer network;

[0088] Step 4.2.7: Input the result of global average pooling into the fully connected layer.

[0089] Furthermore, the specific method of step 5 is:

[0090] Step 5.1: Preprocess the newly collected oscillatory wave signal and PPG signal data, and filter the oscillatory wave signal and PPG signal;

[0091] Step 5.2: Extract features from the processed oscillatory wave signal and PPG signal;

[0092] Step 5.2.1: Use CNN to extract time-frequency features of oscillatory wave signals and PPG signals to capture local pattern information;

[0093] Step 5.2.2: Add positional encoding to the extracted features to preserve sequence information and enhance the model's ability to model temporal dependencies.

[0094] Step 5.2.3: Use the multi-head self-attention mechanism and feedforward neural network to capture the long-range dependencies of the signal and improve the feature expression capability;

[0095] Step 5.2.4: Perform average pooling on the extracted feature vectors to obtain a fixed-length feature vector, reduce redundant information, and improve computational efficiency.

[0096] Step 5.3: Fuse the pooled feature vector with the manually extracted time-frequency features and statistical features, and input them into the XGBoost model to learn the mapping relationship between the oscillation wave signal, PPG signal, and blood pressure value, and achieve accurate estimation of diastolic and systolic blood pressure. The specific method is as follows:

[0097] Step 5.3.1: Input fusion features;

[0098] Step 5.3.2: Data preprocessing, including data cleaning and data standardization; the data standardization formula is as follows:

[0099] (8);

[0100] Among them, x ij is the jth eigenvalue of the i-th sample, x ij * is the jth eigenvalue of the i-th sample after standardization, μ j is the mean of the jth feature, σ j is the standard deviation of the jth feature;

[0101] Step 5.3.3: Divide the preprocessed dataset into training sets Dtrain and test set Dtest ;

[0102] Step 5.3.4: Set the XGBoost hyperparameters, including the number of trees, learning rate, and maximum depth.

[0103] Step 5.3.5: Using the training set Dtrain Train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean square error:

[0104] (9);

[0105] in, θ is the parameter of the XGBoost model, y i It is the real blood pressure value. is the blood pressure value predicted by the model;

[0106] Step 5.3.6: Use the trained XGBoost model to test the set Dtest Make a prediction and get the predicted blood pressure value ;

[0107] Step 5.3.7: Use evaluation metrics to measure the performance of the model. Evaluation metrics include mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE):

[0108] (10);

[0109] (11);

[0110] (12);

[0111] Among them, n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0112] The beneficial effect of adopting the above technical solution is that the blood pressure measurement system and method based on collaborative learning of oscillation wave and PPG signal provided by the present invention can effectively extract local features in oscillation wave signal and PPG signal and capture the temporal dependency of the signal by combining the advantages of CNN and Transformer. In the feature extraction process, CNN is used to extract local features of oscillation wave signal and PPG signal, and Transformer is used to capture the temporal features of oscillation wave signal, so as to obtain high-dimensional and highly expressive deep features. These deep features are combined with manual features extracted by time-frequency analysis and statistical features, and trained by XGBoost model to learn the mapping relationship between oscillation wave signal, PPG signal and systolic pressure and diastolic pressure, thereby improving the accuracy of blood pressure prediction. The training data of this model comes from different individuals, which effectively reduces the impact of individual differences on the blood pressure estimation results and enhances the applicability and versatility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 A structural block diagram of a blood pressure measurement system based on collaborative learning of oscillation waves and PPG signals provided by an embodiment of the present invention;

[0114] Figure 2 Implementation block diagram of the oscillation wave data preprocessing module provided in an embodiment of the present invention;

[0115] Figure 3 Implementation block diagram of the PPG data preprocessing module provided in an embodiment of the present invention;

[0116] Figure 4 Implementation block diagram of the CNN-Transformer network module provided by an embodiment of the present invention;

[0117] Figure 5 A block diagram of a CNN structure for extracting oscillation wave signal features provided by an embodiment of the present invention;

[0118] Figure 6 A block diagram of the CNN structure for extracting PPG signal features provided by an embodiment of the present invention;

[0119] Figure 7 A block diagram of the Transformer network model implementation provided by an embodiment of the present invention;

[0120] Figure 8 This is a structural block diagram of the XGBoost model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0121] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0122] In this embodiment, a blood pressure measurement system based on collaborative learning of oscillation waves and PPG signals is provided. Figure 1 As shown, it includes the construction of oscillation wave database, data preprocessing module, feature extraction and fusion module, and blood pressure estimation module.

[0123] The oscillation wave database stores the oscillation wave signals and corresponding diastolic and systolic blood pressures of 3,000 subjects; the data preprocessing module is used to remove common noise in the oscillation wave signals through filtering. First, the oscillation wave signal device is correctly placed on the subject's upper arm and fixed with a cuff to ensure accuracy and comfort. The cuff will gradually increase the pressure and record the subject's oscillation wave data under different pressures and store it. Then, the PPG sensor is correctly placed at the fingertips. During the acquisition process, the subject is required to remain still, avoid large movements or posture changes, reduce the impact of motion artifacts on the PPG signal, and measure the subject's systolic and diastolic blood pressures using a blood pressure measurement device.

[0124] The data preprocessing module is used to filter the acquired oscillation wave signal and PPG signal to remove common noise. Figure 2 As shown in the figure, the original oscillation wave signal is firstly subjected to noise reduction using a low-pass filter. The main purpose of low-pass filtering is to remove high-frequency noise in the signal, such as power supply interference and high-frequency random noise, thereby retaining the low-frequency components in the oscillation wave signal; then high-pass filtering is performed. The main purpose of high-pass filtering is to remove low-frequency components in the signal, such as low-frequency noise such as baseline drift and respiratory artifacts, thereby retaining the high-frequency components in the oscillation wave signal. Figure 3 As shown in the figure, firstly, low-pass filtering is used to remove high-frequency noise of the PPG signal, such as light source changes or motion artifacts, and retain the useful components in the PPG signal; then high-pass filtering is used to remove low-frequency drift or baseline drift (such as slow changes caused by breathing and body position changes) to retain important information in the PPG signal; a waveform quality assessment method based on wavelet transform is used to detect and remove motion artifacts or bad signal segments, and its calculation formula is as shown in (1); the signal is smoothed by Savitzky-Golay filter technology to reduce high-frequency fluctuations, and its calculation formula is as shown in formula (2).

[0125] (1);

[0126] in, a and b are scale and translation parameters, ψ is the mother wavelet function, f(t) is the original signal, and is a function of t.

[0127] (2);

[0128] in,c j is the Savitzky-Golay coefficient, m is half the window length, x i+j is the original data value of the input signal at position i+j, where j is the offset index within the window.

[0129] The feature extraction and fusion module uses deep learning (CNN and Transformer networks) to extract high-level features. It also combines time-frequency analysis and statistical features, two traditional machine learning methods, to obtain handcrafted features and perform feature fusion to enhance the model's predictive capabilities. First, data preparation is performed by loading oscillatory wave signals, PPG signals, and their corresponding diastolic and systolic blood pressure values, and dividing them into training and test sets. Next, a feature extraction module is established, consisting of a CNN feature extractor, a positional encoder, a multi-layer Transformer encoder, a global average pooling layer, and a fully connected layer, to construct a CNN-Transformer network, such as Figure 4 As shown in the figure. Furthermore, we combined two traditional machine learning methods, time-frequency analysis and statistical features, to extract handcrafted features. These features were then fused with features extracted through deep learning to enhance the model's predictive capabilities. Finally, we evaluated the model using the XGBoost model, validated its performance using a test set, detected overfitting, and optimized the model to ensure accuracy and generalization. The detailed process for building the CNN-Transformer network is as follows:

[0130] Step 1: Input layer: Input oscillating wave signal.

[0131] Step 2: The first CNN network: extract features from the oscillation wave signal. The network structure is as follows: Figure 5 The specific steps are as follows:

[0132] Construct the first convolution layer with a kernel size of 5 and a number of 16 kernels. Use the ReLU activation function. The output of the first convolution layer is as shown in formula (3):

[0133] (3);

[0134] in, x i+m,j+n is the input signal, w m,n is the convolution kernel, b For bias.

[0135] In the first pooling layer, the pooling method is maximum pooling, the pooling kernel size is 2, and the maximum pooling process is as shown in formula (4):

[0136] (4);

[0137] Construct the second convolution layer with a convolution kernel size of 5, a number of convolution kernels of 32, and a ReLU activation function.

[0138] In the second pooling layer, the pooling method is maximum pooling and the pooling kernel size is 2.

[0139] Output layer 1 outputs the features extracted from the oscillation wave signal.

[0140] Step 3: Second CNN network: Extract features from PPG signals. The network structure is as follows: Figure 6 The specific steps are as follows:

[0141] Construct convolution layer 3 with a kernel size of 5, a number of kernels of 16, and a ReLU activation function.

[0142] In the third pooling layer, the pooling method is maximum pooling and the pooling kernel size is 2.

[0143] Construct convolutional layer 4 with a kernel size of 5, a number of kernels of 32, and a ReLU activation function.

[0144] In the fourth pooling layer, the pooling method is maximum pooling and the pooling kernel size is 2.

[0145] Construct convolution layer 5 with a kernel size of 7 and a number of 64 kernels, using the ReLU activation function.

[0146] Pooling layer 5, the pooling method is average pooling, and the pooling kernel size is 2;

[0147] Output layer 2 outputs the features extracted from the PPG signal.

[0148] Step 4: Feature vector concatenation and input reshape transformation: The feature vectors extracted from the oscillatory wave signal and the PPG signal are concatenated to form a joint feature. The input shape is then transformed to convert the output of the CNN part into the shape required by the Transformer network.

[0149] Step 5: Transformer network: Capture the temporal dependency of the oscillating wave signal. The network structure is as follows: Figure 7 The specific steps are as follows:

[0150] Step 5.1: Input embedding, converting the input features into high-dimensional feature vectors.

[0151] Step 5.2: Position coding, add position coding, retain the timing information, introduce the timing information into the feature vector, and calculate the formula as (5).

[0152] (5);

[0153] in, POS For location, i For the dimension, d model The dimensions of the model.

[0154] Step 5.3: Multi-head self-attention mechanism, calculates attention scores and performs weighted summation based on these scores to capture the dependencies between different positions in the input sequence. The calculation process of self-attention is as shown in formula (6):

[0155] (6);

[0156] in, Q 、 K and V Obtained by linear transformation of input features, d k for K dimension.

[0157] Step 5.4: Layer normalization to stabilize and accelerate the training process.

[0158] Step 5.5: Feedforward neural network, further process features through nonlinear activation function to enhance the expressive power of the model. The calculation process is as shown in formula (7):

[0159] (7);

[0160] Among them, max(0,xW1+b1) nonlinear activation function, W 1 and W 2 is the weight matrix, b 1 and b 2 is the bias.

[0161] Step 5.6: Layer normalization to stabilize and speed up the training process.

[0162] Step 5.7: Repeat steps 5.3 to 5.6 six times.

[0163] Step 6: Global average pooling: Perform global average pooling on the output of Transformer.

[0164] Step 7: Fully connected layer: Input the result of global average pooling into the fully connected layer.

[0165] The blood pressure estimation module adopts the XGBoost machine learning model to learn the mapping relationship between oscillation waves, PPG signals and blood pressure values based on training data, so as to achieve accurate estimation of diastolic and systolic blood pressure. First, the newly collected oscillation wave signal data is preprocessed. Then, the processed oscillation wave data is input into the machine learning and deep learning model. First, CNN and Transformer are used to extract features, and manual features are extracted in combination with time-frequency analysis and statistical features. Position encoding is added to the features to retain sequence information. Then, the multi-head self-attention mechanism and feedforward neural network are used to capture long-range dependencies and further enhance the expressive power of features. Next, the feature vector is compressed into a vector of fixed length through global average pooling. Finally, the pooled feature vector is trained and mapped in combination with the traditional machine learning model XGBoost to achieve estimation of diastolic and systolic blood pressure.

[0166] The specific steps of building the XGBoost network are as follows:

[0167] Step 1: Input fusion features.

[0168] Step 2: Data preprocessing, including data cleaning and data standardization. The data standardization formula is as follows:

[0169] (8);

[0170] Among them, x ij is the jth eigenvalue of the i-th sample, is the jth eigenvalue of the i-th sample after standardization, μ j is the mean of the jth feature, σ j is the standard deviation of the j-th feature.

[0171] Step 3: Divide the preprocessed dataset into training sets Dtrain and test set Dtest , usually divided into 80% training set and 20% test set.

[0172] Step 4: Set the hyperparameters of XGBoost, the number of trees (n_estimators) equal to 100, the learning rate (learning_rate) equal to 0.01, and the maximum depth (max_depth) equal to 5.

[0173] Step 5: Train the model using the training set Dtrain Train the XGBoost model. The objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error (MSE):

[0174] (9);

[0175] in, θ are the parameters of the model, y i It is the real blood pressure value. is the blood pressure value predicted by the model.

[0176] Step 6: Use the trained XGBoost model to test the set Dtest Make a prediction and get the predicted blood pressure value .

[0177] Step 7: Use evaluation metrics to measure the performance of the model. The evaluation metrics include mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE):

[0178] (10);

[0179] (11);

[0180] (12);

[0181] Among them, n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0182] Through the above steps, the blood pressure estimation module can effectively extract features from the oscillation wave and PPG signals and predict blood pressure values.

[0183] A blood pressure measurement method based on collaborative learning of oscillation waves and PPG signals is also provided, which is implemented by the above-mentioned blood pressure measurement system based on collaborative learning of oscillation waves and PPG signals, and includes the following steps:

[0184] Step 1: Construct an oscillation wave signal and PPG signal database, including oscillation wave signals, PPG signals and their corresponding systolic and diastolic pressures. The specific method is:

[0185] Step 1.1: Place the oscillatory wave signal device correctly on the subject's upper arm and secure it with the cuff to ensure accuracy and comfort.

[0186] Step 1.2: The cuff is gradually pressurized, and the oscillation wave data of the subject at different pressures is recorded and stored.

[0187] Step 1.3: Place the PPG sensor correctly on the fingertip. During the acquisition process, the subject is required to remain still and avoid large movements or posture changes to reduce the impact of motion artifacts on the PPG signal.

[0188] Step 2: Preprocess the oscillation wave signals in the database, including low-pass filtering and high-pass filtering, to optimize the noise, baseline drift and amplitude differences between different waveforms in the oscillation wave signals. The specific method is:

[0189] Step 2.1: Use a low-pass filter to remove high-frequency noise from the original oscillation wave signal, thereby retaining the low-frequency components in the oscillation wave signal.

[0190] Step 2.2: Use high-pass filtering to remove the influence of baseline drift on the oscillation wave, thereby retaining the high-frequency components in the oscillation wave signal.

[0191] Step 3: Preprocess the PPG signals in the database, including low-pass filtering, high-pass filtering, artifact removal, and smoothing to remove noise, baseline drift, and amplitude differences between different waveforms in the PPG signals. The specific methods are as follows:

[0192] Step 3.1: Use low-pass filtering to remove high-frequency noise, such as light source changes or motion artifacts, and retain the useful components in the PPG signal.

[0193] Step 3.2: Use high-pass filtering to remove low-frequency drift or baseline drift, such as slow changes caused by breathing and body position changes, to retain important information in the PPG signal.

[0194] Step 3.3: Use wavelet transform waveform quality assessment method to detect and remove motion artifacts or bad signal segments.

[0195] Step 3.4: Smooth the signal using Savitzky-Golay filter technique to reduce high frequency fluctuations.

[0196] Step 4: Perform feature extraction and blood pressure estimation modeling; use CNN to extract local features of the oscillatory wave signal and PPG signal, and use Transformer to extract the temporal features of the oscillatory wave signal to obtain high-dimensional, highly expressive deep features. The specific method is as follows:

[0197] Step 4.1: Prepare the data by loading the oscillatory wave signal, PPG signal, and their corresponding diastolic and systolic blood pressure values, and dividing the data into training and test sets.

[0198] Step 4.2: Build a deep learning model, that is, build a CNN-Transformer network, including a CNN feature extractor, positional encoding, a multi-layer Transformer encoder, global average pooling, and a fully connected layer.

[0199] Step 4.2.1: Input the oscillatory wave signal and PPG signal.

[0200] Step 4.2.2: Construct the first CNN network to extract features from the oscillation wave signal. The specific method is as follows:

[0201] Construct convolution layer 1 with a kernel size of 5, a number of kernels of 16, and a ReLU activation function.

[0202] Pooling layer 1, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0203] Construct the second convolution layer with a kernel size of 5, a number of kernels of 32, and a ReLU activation function.

[0204] Pooling layer 2, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0205] Output layer 1 outputs the features extracted from the oscillation wave signal.

[0206] Step 4.2.3: Construct the second CNN network to extract features of the PPG signal. The specific method is as follows:

[0207] Construct convolution layer 3 with a kernel size of 3, a number of kernels of 16, and a ReLU activation function.

[0208] Pooling layer 3, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0209] Construct convolutional layer 4 with a kernel size of 5, a number of kernels of 32, and a ReLU activation function.

[0210] Pooling layer 4, the pooling method is maximum pooling, and the pooling kernel size is 2;

[0211] Construct convolution layer 5 with a kernel size of 7 and a number of 64 kernels, using the ReLU activation function.

[0212] Pooling layer 5, the pooling method is average pooling, and the pooling kernel size is 2;

[0213] Output layer 2 outputs the features extracted from the PPG signal.

[0214] Step 4.2.4: Concatenate the feature vectors extracted from the oscillatory wave signal and the PPG signal to form a joint feature, then transform the input shape and convert the output of the CNN part into the shape required by the Transformer network.

[0215] Step 4.2.5: Build a Transformer network to capture the timing dependency of the oscillatory wave signal. The specific method is as follows:

[0216] Step 4.2.5.1: Input embedding, converting the input features into high-dimensional feature vectors;

[0217] Step 4.2.5.2: Position coding, add position coding, retain the timing information, introduce the timing information into the feature vector, and calculate the formula as shown in formula (5):

[0218] (5);

[0219] in, POS For location, i For the dimension, d model is the dimension of the model;

[0220] Step 4.2.5.3: Multi-head self-attention mechanism, calculate attention scores, and perform weighted summation based on these scores to capture the dependencies between different positions in the input sequence; the calculation process of self-attention is as shown in formula (6):

[0221] (6);

[0222] in, Q 、 K and V Obtained by linear transformation of input features, d k for K Dimensions;

[0223] Step 4.2.5.4: Layer normalization to stabilize and accelerate the training process;

[0224] Step 4.2.5.5: Feedforward neural network, further process features through nonlinear activation function to enhance the expressive power of the model. The calculation process is as shown in formula (7):

[0225] (7);

[0226] Among them, max(0,xW1+b1) nonlinear activation function, W 1 and W 2 is the weight matrix, b 1 and b 2 is bias;

[0227] Step 4.2.5.6: Layer normalization to stabilize and accelerate the training process;

[0228] Step 4.2.5.7: Repeat steps 4.2.5.3 to 4.2.5.6 six times.

[0229] Step 4.2.6: Perform global average pooling on the output of the Transformer network.

[0230] Step 4.2.7: Input the result of global average pooling into the fully connected layer.

[0231] Step 5: Predict blood pressure based on the newly collected oscillatory wave signal data and PPG signal, including preprocessing of the oscillatory wave signal and PPG signal and blood pressure estimation based on machine learning and deep learning models; the features extracted by deep learning are integrated with the manual features extracted by time-frequency analysis and statistical features, and then input into the XGBoost model to learn the mapping relationship between the oscillatory wave signal, PPG signal and systolic and diastolic blood pressure, thus achieving accurate estimation of diastolic and systolic blood pressure. The specific method is as follows:

[0232] Step 5.1: Preprocess the newly collected oscillatory wave signal and PPG signal data, and perform filtering on the oscillatory wave signal and PPG signal.

[0233] Step 5.2: Extract features from the processed oscillation wave signal and PPG signal. The specific method is:

[0234] Step 5.2.1: Use CNN to extract time-frequency features from oscillatory wave signals and PPG signals to capture local pattern information.

[0235] Step 5.2.2: Add positional encoding to the extracted features to preserve sequence information and enhance the model's ability to model temporal dependencies.

[0236] Step 5.2.3: Utilize a multi-head self-attention mechanism and a feedforward neural network to capture long-range signal dependencies and improve feature representation. Long-range dependencies refer to the correlations between signal elements separated by long time steps or spatial locations. For example, over a long period of time, the oscillation waves generated by each heartbeat are linked to the waveform characteristics of the PPG signal. The amplitude of the oscillation waves at the current moment is related to the strength of the heart's contraction and the state of the blood vessels several seconds ago. This correlation, which spans a long time interval, is a long-range dependency.

[0237] Step 5.2.4: Perform average pooling on the extracted feature vectors to obtain a fixed-length feature vector, reduce redundant information, and improve computational efficiency.

[0238] Step 5.3: Fuse the pooled feature vector with the manually extracted time-frequency features and statistical features, and input them into the XGBoost model to learn the mapping relationship between the oscillation wave signal, PPG signal and blood pressure value, and achieve accurate estimation of diastolic and systolic blood pressure. Figure 8 As shown, the specific method is:

[0239] Step 5.3.1: Input fusion features.

[0240] Step 5.3.2: Data preprocessing, including data cleaning and data standardization. The data standardization formula is as follows:

[0241] (8);

[0242] Among them, x ij is the jth eigenvalue of the i-th sample, x ij * is the jth eigenvalue of the i-th sample after standardization, μ j is the mean of the jth feature, σ j is the standard deviation of the j-th feature.

[0243] Step 5.3.3: Divide the preprocessed dataset into training sets Dtrain and test set Dtest .

[0244] Step 5.3.4: Set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth.

[0245] Step 5.3.5: Using the training set Dtrain Train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean square error:

[0246] (9);

[0247] in, θ is the parameter of the XGBoost model, y i It is the real blood pressure value. is the blood pressure value predicted by the model.

[0248] Step 5.3.6: Use the trained XGBoost model to test the set Dtest Make a prediction and get the predicted blood pressure value .

[0249] Step 5.3.7: Use evaluation metrics to measure the performance of the model. Evaluation metrics include mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE):

[0250] (10);

[0251] (11);

[0252] (12);

[0253] Among them, n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0254] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A blood pressure measurement system based on collaborative learning of oscillatory waves and PPG signals, characterized by: It includes oscillation wave and PPG signal database construction module, data preprocessing module, feature extraction module and blood pressure estimation module; The oscillatory wave and PPG signal database stores oscillatory wave and PPG signals and corresponding diastolic and systolic blood pressures of 3,000 subjects; The data preprocessing module is used to filter the acquired oscillation wave signal and PPG signal to remove common noise; The feature extraction module constructs a CNN-Transformer network, including a CNN feature extractor, a position encoder, a multi-layer Transformer encoder, a global average pooling layer, a fully connected layer, and a feature fusion module. The CNN feature extractor is used to extract local features of the oscillatory wave signal and the PPG signal, and the multi-layer Transformer encoder is used to extract the temporal features of the oscillatory wave signal. On this basis, the position encoder is used to add position encoding to the extracted features to retain sequence information and enhance the model's perception of temporal features. The global average pooling layer performs global average pooling on the output of the multi-layer Transformer encoder and then inputs it into the fully connected layer. The feature fusion module combines time-frequency analysis and statistical features to extract manual features, fuses them with the features extracted by deep learning, and divides the training set into the test set. The blood pressure estimation module uses the XGBoost model to estimate blood pressure based on the extracted features, achieving accurate prediction of systolic and diastolic blood pressure.

2. The blood pressure measurement system based on collaborative learning of oscillatory waves and PPG signals according to claim 1, characterized in that: The oscillation wave and PPG signal database construction module includes an oscillation wave signal device, a cuff, a PPG sensor, a blood pressure measurement device, and a database. First, the oscillation wave signal device is correctly placed on the subject's upper arm and fixed with the cuff. The cuff will gradually increase the pressure and record the subject's oscillation wave data under different pressures and store it in the database. Then, the PPG sensor is correctly placed at the fingertips. During the collection process, the subject is required to remain still to reduce the impact of motion artifacts on the PPG signal. The subject's systolic and diastolic blood pressures are measured by the blood pressure measurement device and stored in the database.

3. The blood pressure measurement system based on collaborative learning of oscillatory waves and PPG signals according to claim 1, characterized in that: The data preprocessing module includes an oscillation wave data preprocessing module and a PPG data preprocessing module; The oscillation wave data preprocessing module first uses a low-pass filter to reduce noise on the original oscillation wave signal to remove high-frequency noise in the signal, thereby retaining the low-frequency components in the oscillation wave signal; then performs a high-pass filter to remove the low-frequency components in the signal, thereby retaining the high-frequency components in the oscillation wave signal; The PPG data preprocessing module first uses low-pass filtering to remove high-frequency noise in the PPG signal and retains the useful components in the PPG signal; then uses high-pass filtering to remove low-frequency drift or baseline drift to retain important information in the PPG signal; and uses a waveform quality assessment method based on wavelet transform to detect and remove motion artifacts or poor signal segments. The calculation formula is as follows: （1）； in, a and b are scale and translation parameters, ψ is the mother wavelet function; f(t) is the original signal, which is a function of t; The Savitzky-Golay filter technology is used to smooth the signal and reduce high-frequency fluctuations. The calculation formula is as follows: （2）； in, c j is the Savitzky-Golay coefficient, m is half the window length, x i+j is the original data value of the input signal at position i+j, where j is the offset index within the window.

4. The blood pressure measurement system based on collaborative learning of oscillatory waves and PPG signals according to claim 1, characterized in that: The specific implementation method of the blood pressure estimation module is: First, set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth; Then use the training set to train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean square error, as shown below: （9）； in, θ are the parameters of the XGBoost model, y i It is the real blood pressure value. is the blood pressure value predicted by the model; Use the trained XGBoost model to predict the test set and get the predicted blood pressure value ; Then use the evaluation indicators to measure the performance of the model, including mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE): （10）； （11）； （12）； Among them, n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

5. A blood pressure measurement method based on collaborative learning of oscillatory waves and PPG signals, characterized by: The blood pressure measurement system based on collaborative learning of oscillation waves and PPG signals according to claim 1 is implemented, comprising the following steps: Step 1: Construct an oscillatory wave signal and a PPG signal database, including oscillatory wave signals, PPG signals and their corresponding systolic and diastolic pressures; Step 2: Preprocess the oscillation wave signals in the database, including low-pass filtering and high-pass filtering, to optimize the noise, baseline drift and amplitude differences between different waveforms in the oscillation wave signals; Step 3: Preprocess the PPG signals in the database, including low-pass filtering, high-pass filtering, artifact removal, and smoothing to remove noise, baseline drift, and amplitude differences between different waveforms in the PPG signals; Step 4: Perform feature extraction and blood pressure estimation modeling; establish a CNN-Transformer network, including a CNN feature extractor, positional encoding, a multi-layer Transformer encoder, global average pooling, and a fully connected layer; use CNN to extract local features of the oscillatory wave signal and PPG signal and concatenate them to form a joint feature; then use the multi-layer Transformer encoder to extract the temporal features of the oscillatory wave signal; add positional encoding to the extracted features to preserve sequence information and enhance the model's perception of temporal features; perform global average pooling on the output of the multi-layer Transformer encoder and input it into the fully connected layer; Step 5: Predict blood pressure based on the newly collected oscillation wave signal data and PPG signal, including preprocessing of the oscillation wave signal and PPG signal and blood pressure estimation based on machine learning and deep learning models; fuse the features extracted by deep learning with the manual features extracted by time-frequency analysis and statistical features, and then input them into the XGBoost model to learn the mapping relationship between the oscillation wave signal, PPG signal and systolic and diastolic blood pressure, so as to achieve accurate estimation of diastolic and systolic blood pressure.

6. The blood pressure measurement method based on collaborative learning of oscillatory waves and PPG signals according to claim 5, characterized in that: The specific method of step 1 is: Step 1.1: Place the oscillatory wave signal device correctly on the subject's upper arm and secure it with a cuff; Step 1.2: The cuff is gradually pressurized, and the oscillation wave data of the subject at different pressures is recorded and stored; Step 1.3: Place the PPG sensor correctly on the fingertip; during the acquisition process, ask the subject to remain still to reduce the impact of motion artifacts on the PPG signal; Step 1.4: Measure the subject's systolic and diastolic blood pressure using a blood pressure measuring device.

7. The blood pressure measurement method based on collaborative learning of oscillatory waves and PPG signals according to claim 5, characterized in that: The specific method of step 2 is: Step 2.1: Use a low-pass filter to remove high-frequency noise from the original oscillation wave signal, thereby retaining the low-frequency components of the oscillation wave signal; Step 2.2: Use high-pass filtering to remove the influence of baseline drift on the oscillation wave, thereby retaining the high-frequency components in the oscillation wave signal.

8. The blood pressure measurement method based on collaborative learning of oscillatory waves and PPG signals according to claim 5, characterized in that: The specific method of step 3 is: Step 3.1: Use low-pass filtering to remove high-frequency noise and retain the useful components in the PPG signal; Step 3.2: Use high-pass filtering to remove low-frequency drift or baseline drift and retain important information in the PPG signal; Step 3.3: Use wavelet transform waveform quality assessment method to detect and remove motion artifacts or bad signal segments; Step 3.4: Smooth the signal using Savitzky-Golay filter technique to reduce high frequency fluctuations.

9. The blood pressure measurement method based on collaborative learning of oscillatory waves and PPG signals according to claim 5, characterized in that: The specific method of step 4 is: Step 4.1: Prepare the data by loading the oscillatory wave signal, PPG signal and its corresponding diastolic and systolic blood pressure values, and dividing the data into training and test sets. Step 4.2: Build a deep learning model, that is, build a CNN-Transformer network, including a CNN feature extractor, positional encoding, a multi-layer Transformer encoder, global average pooling, and a fully connected layer; Step 4.2.1: Input oscillatory wave signal and PPG signal; Step 4.2.2: Construct the first CNN network to extract features from the oscillation wave signal. The specific method is as follows: Construct convolution layer 1 with a kernel size of 5, a number of kernels of 16, and a ReLU activation function. Pooling layer 1, the pooling method is maximum pooling, and the pooling kernel size is 2; Construct the second convolution layer with a kernel size of 5, a number of kernels of 32, and a ReLU activation function. Pooling layer 2, the pooling method is maximum pooling, and the pooling kernel size is 2; Output layer 1, outputs features extracted from the oscillation wave signal; Step 4.2.3: Construct the second CNN network to extract features of the PPG signal. The specific method is as follows: Construct convolution layer 3 with a kernel size of 3, a number of kernels of 16, and a ReLU activation function. Pooling layer 3, the pooling method is maximum pooling, and the pooling kernel size is 2; Construct convolutional layer 4 with a kernel size of 5, a number of kernels of 32, and a ReLU activation function. Pooling layer 4, the pooling method is maximum pooling, and the pooling kernel size is 2; Construct convolution layer 5 with a kernel size of 7 and a number of 64 kernels, using the ReLU activation function. Pooling layer 5, the pooling method is average pooling, and the pooling kernel size is 2; Output layer 2, outputs features extracted from PPG signals; Step 4.2.4: Concatenate the feature vectors extracted from the oscillatory wave signal and the PPG signal to form a joint feature, then transform the input shape and convert the output of the CNN part into the shape required by the Transformer network; Step 4.2.5: Build a Transformer network to capture the timing dependency of the oscillatory wave signal. The specific method is as follows: Step 4.2.5.1: Input embedding, converting the input features into high-dimensional feature vectors; Step 4.2.5.2: Position coding, add position coding, retain the timing information, introduce the timing information into the feature vector, and calculate the formula as shown in formula (5): （5）； in, POS For location, i For the dimension, d model is the dimension of the model; Step 4.2.5.3: Multi-head self-attention mechanism, calculate attention scores, and perform weighted summation based on these scores to capture the dependencies between different positions in the input sequence; the calculation process of self-attention is as shown in formula (6): （6）； in, Q 、 K and V Obtained by linear transformation of input features, d k for K Dimensions; Step 4.2.5.4: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.5: Feedforward neural network, further process features through nonlinear activation function to enhance the expressive power of the model. The calculation process is as shown in formula (7): （7）； Among them, max(0,xW1+b1) is a nonlinear activation function, W 1 and W 2 is the weight matrix, b 1 and b 2 is bias; Step 4.2.5.6: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.7: Repeat steps 4.2.5.3 to 4.2.5.6 six times; Step 4.2.6: Perform global average pooling on the output of the Transformer network; Step 4.2.7: Input the result of global average pooling into the fully connected layer.

10. The blood pressure measurement method based on collaborative learning of oscillatory waves and PPG signals according to claim 9, characterized in that: The specific method of step 5 is: Step 5.1: Preprocess the newly collected oscillatory wave signal and PPG signal data, and filter the oscillatory wave signal and PPG signal; Step 5.2: Extract features from the processed oscillatory wave signal and PPG signal; Step 5.2.1: Use CNN to extract time-frequency features of oscillatory wave signals and PPG signals to capture local pattern information; Step 5.2.2: Add positional encoding to the extracted features to preserve sequence information and enhance the model's ability to model temporal dependencies. Step 5.2.3: Use the multi-head self-attention mechanism and feedforward neural network to capture the long-range dependencies of the signal and improve the feature expression capability; Step 5.2.4: Perform average pooling on the extracted feature vectors to obtain a fixed-length feature vector, reduce redundant information, and improve computational efficiency. Step 5.3: Fuse the pooled feature vector with the manually extracted time-frequency features and statistical features, and input them into the XGBoost model to learn the mapping relationship between the oscillation wave signal, PPG signal, and blood pressure value, and achieve accurate estimation of diastolic and systolic blood pressure. The specific method is as follows: Step 5.3.1: Input fusion features; Step 5.3.2: Data preprocessing, including data cleaning and data standardization; the data standardization formula is as follows: （8）； Among them, x ij is the jth eigenvalue of the i-th sample, x ij * is the jth eigenvalue of the i-th sample after standardization, μ j is the mean of the jth feature, σ j is the standard deviation of the jth feature; Step 5.3.3: Divide the preprocessed dataset into training sets Dtrain and test set Dtest ; Step 5.3.4: Set the XGBoost hyperparameters, including the number of trees, learning rate, and maximum depth. Step 5.3.5: Using the training set Dtrain Train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean square error: （9）； in, θ is the parameter of the XGBoost model, y i It is the real blood pressure value. is the blood pressure value predicted by the model; Step 5.3.6: Use the trained XGBoost model to test the set Dtest Make a prediction and get the predicted blood pressure value ; Step 5.3.7: Use evaluation metrics to measure the performance of the model. Evaluation metrics include mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE): （10）； （11）； （12）； Among them, n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

Images