A contactless blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture

Abstract

The application discloses a non-contact blood pressure measurement method based on a millimeter wave radar and a hierarchical multi-model architecture, which comprises data acquisition and preprocessing; PWV estimation based on a single radar; classification layer processing: target classification based on pulse wave explicit features; multi-model layer processing: blood pressure regression detection based on target sub-classes and a model pool; statistical layer processing: statistical filtering of detection results and output of final results; the application utilizes a single radar and a hierarchical multi-model architecture to realize detection of a pulse wave transmission speed, improves blood pressure detection precision, simplifies a system and operation, and makes detection more accurate and highly adaptive due to the hierarchical multi-model structure.

Description

Technical Field

[0001] This invention relates to the field of non-contact human feature detection radar technology, specifically to a non-contact blood pressure measurement method based on millimeter-wave radar and a hierarchical multi-model architecture. Background Technology

[0002] Cardiovascular disease poses a significant threat to human health and safety. Blood pressure, as a crucial physiological parameter, is vital for the long-term monitoring of cardiovascular diseases, particularly in prevention and diagnosis. While traditional contact blood pressure measurement techniques are relatively mature, they impose significant limitations on patients and hinder long-term continuous monitoring. With increasing health demands, non-contact blood pressure measurement technologies have emerged, among which millimeter-wave-based blood pressure measurement solutions have seen rapid development due to their low cost and high convenience.

[0003] The most common method for measuring blood pressure using millimeter waves is as follows: First, a millimeter-wave radar device is used to detect the human pulse wave. Then, a conversion model between the pulse wave and blood pressure (systolic blood pressure SBP and diastolic blood pressure DBP) is established to measure blood pressure. Establishing this blood pressure conversion model is crucial. There are two main approaches: one is "modeling based on explicit pulse wave features," which involves establishing a conversion model from explicit pulse wave features to blood pressure. Explicit pulse wave features include pulse wave velocity (PWV), pulse wave frequency (PWF), and dominant wave height. The other is "modeling based on deep neural networks," which does not focus on specific pulse wave features but directly establishes a neural network conversion model from pulse wave waveform to blood pressure. The neural network automatically extracts features and calculates blood pressure.

[0004] (1) For blood pressure measurement schemes based on explicit features, there are two main problems.

[0005] PWV measurement is relatively cumbersome. In millimeter-wave non-contact blood pressure measurement systems, the conventional approach to measuring PWV involves deploying dual radar measurements at two points based on the same pulse wave conduction tree. PWV is calculated by simultaneously monitoring the pulse wave relationship between the two measurement points. This method is cumbersome to implement, thus limiting its application scope; the feature set is incomplete, limiting accuracy. Although feature engineering can select the feature set most correlated with blood pressure, the complexity of blood pressure models currently prevents the extraction of complete features that perfectly match blood pressure. Therefore, the conversion model from pulse wave features to blood pressure is theoretically inaccurate, limiting the accuracy of blood pressure measurement.

[0006] (2) The "blood pressure measurement scheme based on deep neural networks" has the following three main problems.

[0007] The complexity and computational cost are significant. The performance of a neural network is highly correlated with the number of neurons. To adapt to different populations and blood pressure ranges, the final model typically has a high number of layers and neurons per layer, posing a considerable challenge to the computational power of small, embedded devices. Furthermore, the "utilization" of known important features is insufficient. Deep neural networks rely on automatic feature acquisition during model training, lacking effective means to transfer prior information into the model. Therefore, the "utilization" of known important features (such as PWV) is usually lower than that of dedicated models for those features. Overfitting or underfitting issues also exist. While increased model complexity can improve performance in deep neural networks, it also increases the demands on the dataset. Generally, more complex models are more prone to overfitting or underfitting due to dataset bias. Summary of the Invention

[0008] Technical objective: To address the shortcomings of existing blood pressure measurement methods, this invention discloses a non-contact blood pressure measurement method based on millimeter-wave radar and a hierarchical multi-model architecture, which improves blood pressure detection accuracy while reducing system computing power.

[0009] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0010] A non-contact blood pressure measurement method based on millimeter-wave radar and a hierarchical multi-model architecture includes:

[0011] The raw ADC data of the target area is collected using millimeter-wave radar equipment, and FFT is performed on the ADC data to generate a one-dimensional FFT result. Based on the result, a range azimuth map is constructed. Target detection is performed based on the range azimuth map to identify the AOI area. Pitch DBF angle measurement is performed within the AOI area to generate a radar point cloud.

[0012] Based on radar point cloud data, the pulse wave propagation path of the target area is estimated. Multiple key nodes are selected on the propagation path, and the pulse wave phase information of multiple key nodes is extracted frame by frame to form a phase matrix. Each row in the phase matrix represents the pulse wave sequence of a key node.

[0013] The pulse wave transmission time between key nodes is calculated by cross-correlation, generating a transmission time matrix. The relative position matrix between key nodes is also calculated. Based on the transmission time matrix and the relative position matrix between key nodes, the pulse wave transmission velocity is calculated.

[0014] Several classification models are constructed based on the explicit features of the pulse wave, and these models are trained based on the corresponding feature subsets and sample sets.

[0015] The pulse wave sequence of key nodes is segmented, each pulse wave is extracted separately, and pulse waves with obvious features are selected to form a pulse wave set. Based on the features in the feature subset, feature extraction is performed on each pulse wave in the pulse wave set to obtain a feature matrix, which is then input into the classification model to obtain the classification result of each pulse wave.

[0016] Based on the classification results of each pulse wave, the regression model that best matches each pulse wave is selected from the pre-trained model pool. After processing the pulse wave, it is input into the regression model to obtain the blood pressure detection result.

[0017] Perform sliding window histogram statistics on the blood pressure measurement results for each pulse wave, and output the final blood pressure measurement result based on the statistical results.

[0018] Beneficial Effects: The non-contact blood pressure measurement method based on millimeter-wave radar and a hierarchical multi-model architecture provided by this invention has the following beneficial effects:

[0019] 1. This invention constructs multiple classification and regression models to classify and regress different types of pulse wave signals. It can select the most suitable model for blood pressure detection based on the actual situation. This multi-model architecture improves the adaptability of pulse wave signals under different physiological states, thereby significantly improving the accuracy and reliability of blood pressure detection. This invention extracts explicit features of the pulse wave, such as pulse wave transmission velocity and pulse wave frequency, and constructs classification models based on these explicit features. Since these features are directly related to blood pressure, the classification model can screen out more representative pulse wave signals, enabling more accurate blood pressure prediction.

[0020] 2. This invention designs a pulse wave transmission velocity detection scheme based on a single radar. Compared with the conventional single radar detection scheme that does not have the ability to directly detect PWV, this scheme introduces the characteristic PWV that is highly correlated with blood pressure, thereby improving the detection accuracy. Compared with the conventional dual radar scheme that can directly measure PWV, this scheme simplifies the system and operation.

[0021] 3. This invention processes blood pressure detection results from pulse wave signals using sliding window histogram statistics, reducing fluctuations in detection results caused by individual abnormal pulse waves and enhancing the stability and reliability of blood pressure detection results. When processing pulse wave signals, the scheme employs techniques such as phase unwinding and cross-correlation calculation, effectively reducing phase jumps and noise interference caused by signal periodicity, and ensuring the accuracy of signal processing. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0023] Figure 1 This is a flowchart illustrating the overall processing of the present invention;

[0024] Figure 2 This is a schematic diagram illustrating the detection process of the arm according to the present invention;

[0025] Figure 3 This is a schematic diagram of the RAMap of the present invention;

[0026] Figure 4 This is a schematic diagram of the convolution process of the RAMap of the present invention after convolution processing;

[0027] Figure 5 This is a schematic diagram of the model pool structure of the multi-model layer of the present invention;

[0028] Figure 6 This is a histogram showing the detection of systolic and diastolic blood pressure according to the present invention. Detailed Implementation

[0029] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0030] In one specific embodiment, the frame rate f of the millimeter-wave radar device sz The frequency is 1000Hz, one frame contains one pulse, the number of sampling points per frame is 256, the bandwidth is 2.5GHz, the antenna has 3 transmit and 4 receive, and a total of 12 virtual receiving channels. However, in this embodiment, only 7 of the azimuth channels are used to measure the azimuth angle, and 3 of the elevation channels are used to measure the elevation angle.

[0031] like Figure 1 As shown, the specific steps of a non-contact blood pressure measurement method based on millimeter-wave radar and a hierarchical multi-model architecture provided in this embodiment are as follows:

[0032] S101. Use millimeter-wave radar equipment to collect raw ADC data of the target area, perform FFT on the ADC data to generate a one-dimensional FFT result, construct a range azimuth map based on the result, perform target detection based on the range azimuth map, identify the AOI area, perform pitch DBF angle measurement within the AOI area and generate a radar point cloud.

[0033] This step involves data acquisition and preprocessing, encompassing the process from acquiring ADC data to generating AOI (Area of ​​Interest) point clouds. The specific steps are as follows:

[0034] The S110 millimeter-wave radar equipment contains several virtual receiving channels configured for transmitting and receiving pulse signals.

[0035] In one embodiment, the millimeter-wave radar device includes a radar device including N ch Each frame contains one virtual receive channel and one Chirp of N. ch Channel data, with a distance of N0 from the sampling points, in this embodiment N ch The value is 12, N0 is 256, according to... Figure 2 The arm data was collected for blood pressure monitoring in the manner shown. The arm area was chosen because the arm is relatively long, which makes it easier to estimate the pulse wave propagation path and calculate the pulse wave velocity (PWV). The arm was kept as still as possible during the process, and the distance between the radar device and the center of the arm was preferably within the range of [0.3, 0.8] meters.

[0036] The S120 millimeter-wave radar equipment transmits pulse signals and receives signals reflected from the target area, generating raw ADC data containing one pulse per frame. The received raw ADC data is stored in a matrix according to several virtual receiving channels. Each row of the matrix represents a virtual receiving channel, and the number of columns represents the number of distance sampling points.

[0037] In one embodiment, the received raw ADC data is saved to matrix D according to the channels. adc In the middle, D adc The size is N ch ×N0.

[0038] S130. Window the saved original ADC data, perform a one-dimensional FFT on the windowed data to convert the time domain signal into a frequency domain signal, and generate a one-dimensional FFT result.

[0039] For N ch The raw ADC data D of each receiving channel adc Perform an N1-point windowed FFT along the distance dimension to obtain a one-dimensional FFT result D. 1d D 1d It is of size N ch The matrix of size N1 is calculated as follows:

[0040] D 1d =FFT(win(D) adc ),N1)

[0041] The win(·) operator represents windowing. In this case, a 256-point Hanning window is used, and the number of FFT points N1 is 256. Windowing can reduce spectral leakage and enhance signal quality.

[0042] S140. Combining the one-dimensional FFT results from multiple virtual receiving channels with the pre-calibrated azimuth steering vector table, perform azimuth DBF processing on all range cells to generate a two-dimensional range-azimuth map.

[0043] In D1d Extracting the submatrix D corresponding to the azimuth channel from the matrix 1d,az D 1d,az It is of size N ch,az A matrix of ×N1, based on the orientation guidance vector table SVT az For matrix D 1d,az A range-azimuth map (RAMap) can be obtained by performing a digital beamforming (DBF) scan, as shown in the diagram. Figure 3 As shown, the specific calculation formula is as follows:

[0044] RAMap = SVT az *D 1d,az

[0045] SVT az This is a azimuth guidance vector table with a size of N. az ×N ch,az SVT az N needs to be pre-calibrated and obtained. az The number of angles used to indicate the azimuth.

[0046] In one embodiment, the number of azimuth channels N used is... ch,az The value is 7. A total of 121 angles were calibrated in the azimuth direction, that is, a guide vector was calibrated for each degree in the angle range [-60, 60] degrees, resulting in the azimuth guide vector SVT. az The size is 121×7.

[0047] S150. Based on the generated range-azimuth map, the CFAR algorithm is applied for target detection, and the AOI region is determined based on the CFAR detection results.

[0048] Target detection is performed using Constant False Alarm Rate (CFAR) based on a RAMap 2D graph. In one embodiment, the CFAR strategy can be CA-CFAR, OS-CFAR, etc., and the detected targets are stored in a target array TAR1. The i-th target point in TAR1 is a quadruple (r) consisting of the target's distance, azimuth, amplitude, and noise. i ,az i ,amp i ,ns i ).

[0049] The goal of AOI region selection is to retain targets within the arm area as much as possible while eliminating targets outside the arm area. The basic method is: using W... r ×W a A convolution map CMap is obtained by convolving a rectangular window with the absolute value of RAMap, such as... Figure 4As shown, the peak position (r0, a0) of CMap is taken as the energy center, where r0 represents the distance and a0 represents the azimuth angle;

[0050] In TAR1, all targets within the distance range [r0-Δr, r0+Δr] are selected. Then, the selected targets are sorted by energy, and targets with higher energy are selected proportionally, with a threshold value of G. rt ;

[0051] Density clustering is performed on the selected targets to obtain multiple distance regions A0, A1, ..., A i The number of points to be removed is less than the threshold G. A The region is defined as A, and the remaining regions are merged into a single region, which is the AOI region. aoi and A aoi The included points are saved to the target set TAR2.

[0052] S160. For targets in the AOI area, perform pitch DBF processing based on the pre-calibrated pitch steering vector table to obtain target pitch angle information.

[0053] The basic method for pitch DBF angle measurement and point cloud construction is: in the one-dimensional FFT result matrix D 1d Select N from ch,el The data corresponding to each pitch channel are organized into array D. 1d,el D 1d,el It is of size N ch,el A matrix of size N1; for the i-th target point in TAR2, extract the value in D based on its distance cell. 1d,el Extract the corresponding complex vector V after the one-dimensional FFT transformation. i For vector V i The elevation DBF spectrum can be obtained by performing an elevation DBF scan. The specific calculation formula is as follows:

[0054] S i,el =SVT el *V i

[0055] SVT el This is a pitch guidance vector table with a size of N. el ×N ch,el SVT el N needs to be pre-calibrated and obtained. el This refers to the number of angles used for pitch calibration.

[0056] In the pitch DBF spectrum i,el In the middle, find the peak point by absolute value, and take its corresponding angle as the pitch angle el. i The horizontal axis in the pitch DBF spectrum is the pitch angle el. i.

[0057] S170. Combining range, azimuth, and elevation information, a radar point cloud of the AOI area is generated.

[0058] For the i-th target point in TAR2, based on its distance r i Azimuth angle (az) i And the previously calculated pitch angle el i The coordinates (x, y) of the target point in the three-dimensional Cartesian coordinate system are calculated. i ,y i ,z i ), combined with the known amplitude amp i and noise ns i To form the target quintuple (x i ,y i ,z i ,amp i ,ns i Repeat this operation until all target points in TAR2 are transformed into a quintuple representation as described above; then convert N... frm All the aforementioned quintuples obtained from the frame are cached in the detection point set PC, which is the radar point cloud of the AOI region, where N frm This is the number of frames required to complete one blood pressure measurement.

[0059] In one embodiment, the number of pitch channels N is used. ch,el The value is 3. A total of 21 angles were calibrated in the pitch direction, that is, a guide vector was calibrated for each degree in the angle range [-10, 10] degrees, resulting in the pitch guide vector SVT. el The size is 21×7. Additionally, the data accumulation window in this case is 20 seconds, and the frame rate is 1000Hz, so N... frm It is 20,000.

[0060] S102. Estimate the pulse wave propagation path of the target area based on radar point cloud data. Select multiple key nodes on the propagation path and extract the pulse wave phase information of multiple key nodes frame by frame to form a phase matrix. Each row in the phase matrix represents the pulse wave sequence of a key node.

[0061] This step calculates the pulse wave propagation velocity (PWV) based on single radar point cloud information and pulse wave characteristics of key nodes. Since the human arm posture is not fixed during detection (different degrees of bending and different directions), and the pulse wave propagation direction in the arm area can be considered to be basically consistent with the arm shape, the pulse wave propagation path needs to be estimated by estimating the arm shape before calculating the PWV. The detailed steps are as follows:

[0062] S210. Using radar point cloud data of the AOI area, the point cloud area is divided into multiple closed areas of equal area according to the extension direction of the point cloud.

[0063] In one embodiment, the point cloud region is divided into N according to the point cloud extension direction. part A series of equal-area enclosed regions, Part1, Part2, ..., Part... n This ensures coverage of the entire arm surface and facilitates subsequent path estimation.

[0064] S220. Extract all points in each closed region and calculate its geometric center point. Combine the center points of all regions into a center point sequence. Perform linear interpolation on the center point sequence to generate a denser path point sequence, which is the estimated pulse wave propagation path.

[0065] In one embodiment, for the i-th closed region Part i Extract all points in the region and calculate their geometric centers to obtain the center point CP. i After calculating the center points of all regions, a sequence of center points is obtained. Linear interpolation is performed on the center point sequence CPL to obtain the path point sequence PPL, which is the estimated pulse wave propagation path.

[0066] In one embodiment, the number of segments N part The value is 10, and the length of the interpolated PPL is 100.

[0067] S230. Select several key nodes from the point cloud data sequentially along the estimated pulse wave propagation path as key nodes of the pulse wave. Except for the first key node, the Euclidean distance between each key node and the previous key node is greater than the set threshold.

[0068] Along the PPL direction, select N points sequentially in the point cloud. keyp Each detection point serves as a key node in the pulse wave. In one embodiment, the key node KP is selected. i At this time, there are two requirements: 1) Except for KP1, any other critical node must have a Euclidean distance from the previous critical node that is greater than the threshold G. d ;2)KP i It needs to be close enough to PPL, that is, to ensure KP i It is the point among all detection points in the point cloud that is closest to a certain path point in the PPL. In this embodiment, the number of key nodes is N. keyp The value is 5.

[0069] S240. Extract the phase information of the pulse wave frame by frame at each key node position, and perform phase unwinding processing on the extracted phase information.

[0070] Due to the periodicity of the pulse wave, phase jumps may occur. Therefore, the extracted phase information undergoes unwinding processing to ensure phase continuity. keyp Phase information is extracted frame by frame from each key node location, and unwinding is performed.

[0071] S250. The phase information extracted from all key nodes frame by frame is used to construct a phase matrix.

[0072] When the total number of frames reaches N frm When, take out the nearest N. frm The frame phase data constitutes the phase matrix PHA, and the size of PHA is N. keyp ×N frm Each row of the phase matrix represents the pulse wave of a key node.

[0073] S103. Calculate the pulse wave transmission time between key nodes by cross-correlation, generate a transmission time matrix, and calculate the relative position matrix between key nodes. Based on the transmission time matrix and the relative position matrix between key nodes, calculate the pulse wave transmission speed.

[0074] Different rows of the phase matrix PHA represent samples of the same pulse wave at different locations. Theoretically, there is a translational matching relationship between the rows, and the propagation time of the pulse wave between two key nodes can be determined by detecting the translational amount. However, due to factors such as noise, clutter, and location differences, the pulse waves in each row will not be exactly the same, and may even have significant differences. If only pulse wave feature points are used for matching (such as matching based on the pulse wave peak point), large errors may occur. Therefore, this embodiment uses a multi-pulse wave cross-correlation strategy to calculate the translational amount to improve measurement accuracy. The specific steps are as follows:

[0075] S310. For each key node, calculate the cross-correlation function between its pulse wave sequence and the pulse wave sequences of other key nodes, and denote the pulse wave sequence of the i-th key node as PW. i PW i =PHA(i,:), the cross-correlation function R between the pulse wave sequences of the i-th key node and the j-th key node. ij as follows:

[0076]

[0077] Among them, PW i PW represents the pulse wave sequence of the i-th key node. j Let N represent the pulse wave sequence of the j-th key node; k is an integer representing the offset of the pulse wave sampling points. frmThis represents the total number of frames, which is also the total number of sampling points in the pulse wave sequence, where n represents the index of the pulse wave sampling point;

[0078] S320. Calculate the pulse wave transmission time between each key node using the cross-correlation function, specifically the transmission time PWTT between the pulse wave of the i-th node and the pulse wave of the j-th node. ij The calculation formula is as follows:

[0079]

[0080] Among them, PWTT ij T represents the pulse wave transmission time between the i-th critical node and the j-th critical node. frm The frame period, also known as the pulse wave sampling period, is represented in one embodiment as T. frm It takes 1ms. Represents the cross-correlation function R ij (k) The k value corresponding to the peak position;

[0081] S330, construct a pulse wave transmission time matrix from all key nodes. Where N keyp The number of critical nodes is represented by PWTT(i,j), which represents the pulse wave transmission time between the i-th and j-th nodes. Therefore, PWTT is a symmetric matrix. Since the diagonal elements represent the propagation time from the critical node to itself, they are 0.

[0082] S340. Use radar point cloud data to determine the specific coordinates of each key node in three-dimensional space, and calculate the Euclidean distance between each key node. The calculation formula is as follows:

[0083]

[0084] Among them, LOC ij Let x represent the Euclidean distance from the i-th critical node to the j-th critical node. i y i and z i Let x and x represent the three-dimensional coordinates of the i-th key node, respectively. j y j and z j These represent the three-dimensional coordinates of the j-th key node;

[0085] S350, construct a relative position matrix from the Euclidean distances between all key nodes.

[0086] S360. Calculate the pulse wave propagation velocity matrix PWVM between each key node based on the propagation time matrix and the relative position matrix. The formula for calculating the pulse wave propagation velocity PWVM(i,j) between the i-th key node and the j-th key node is as follows:

[0087]

[0088] S370. Extract the upper or lower triangular portion of the transmission velocity matrix PWVM, excluding diagonal elements, and calculate its mean to obtain the pulse wave transmission velocity PWV. The calculation method is as follows:

[0089]

[0090] S104. Construct several classification models based on the explicit features of the pulse wave, and train the several classification models based on the corresponding feature subsets and sample sets.

[0091] This step classifies the total space of pulse wave samples based on explicit pulse wave features, dividing the complex total sample space into multiple relatively simple sample subspaces. Samples within each subspace have high consistency. Modeling and detecting based on subspaces offers two advantages:

[0092] (1) The subspace structure is relatively simple, so the complexity of the regression model is also relatively low, which can improve the detection accuracy while reducing computing power.

[0093] (2) The most suitable detection model can be independently established for different subspaces to achieve the optimization of detection performance.

[0094] S410. Extract explicit features from the pulse wave signal, perform correlation analysis based on the explicit features, and select a feature subset that is highly correlated with the target classification task. The feature subset contains multiple explicit features.

[0095] Many features can be used for pulse wave classification, including pulse wave velocity (PWV), pulse wave frequency (PWF), main wave height (h1), descending isthmus height (h2), dicrotic wave height (h3), rapid ejection phase (t1), left ventricular ejection phase (t2), diastolic phase (t3), cardiac cycle (t4), pulse wave area (S1), systolic area (S2), and diastolic area (S3). In practice, derived features, such as the relative height of the descending isthmus (h), are often used. 21 The relative height h of the diphtheria wave 31 Relative area S during the contraction period 21 and diastolic relative area S 31 These features are all based on consensus. Different features have different correlations with blood pressure, and a subset of features with higher correlations needs to be analyzed in advance for classification.

[0096] The classification layer can have one or two levels of classification. Whether it is one or two levels depends on factors such as storage space, computing power requirements, accuracy requirements, blood pressure range, and the target population. Under normal circumstances, only one level of classification is needed. However, for some special cases, such as the pulse wave characteristics of atrial fibrillation patients are significantly different from those of the general population, if it is necessary to cover these special cases, targeted classification processing is required, which is where two levels of classification come in.

[0097] Here are typical feature subsets used for primary and secondary classification, respectively. This subset configuration is optional and not the only one. If primary classification is used, the feature subset can be configured as follows:

[0098] FS1 = [PWV, PWF, S 21 ]

[0099] If a two-level classification is used, the corresponding feature subset 1 and feature subset 2 can be configured as follows:

[0100]

[0101] S420. Construct several classification models based on feature subsets, input the feature subsets in the sample set into the classification models for training, adjust the model parameters to minimize the classification error, and obtain the trained classification model. The sample set includes sample data and target classification labels corresponding to the feature subsets.

[0102] Whether it's primary or secondary classification, the model building and usage are similar; only the feature subset and model differ. Classification modeling can use neural network models or multinomial models. Here, taking the feature set FS1 for primary classification as an example, we present a modeling method based on multinomial fitting as follows:

[0103] The sample set X is constructed as follows, where PWVL, PWFL, and SL are used. 21 These are pulse wave velocity (PWV), pulse wave frequency (PWF), and systolic relative area (S), respectively. 21 Let L be a column vector, where all elements are 1.

[0104] X = [PWVL] 5 PWVL 4 PWVL 3 PWVL 2 PWVL, PWFL 2 ,PWFL,SL 21 ,L]

[0105] The label set Y corresponding to the sample set is as follows, where SBPL is the systolic blood pressure column vector and DBPL is the diastolic blood pressure column vector:

[0106] Y = [SBPL, DBPL]

[0107] The polynomial coefficient matrix FP can then be obtained as follows:

[0108] FP = (X T X) -1 X T Y

[0109] For a sample x in the feature subset, the approximate values ​​of systolic and diastolic blood pressure can be calculated based on the coefficient matrix FP. Based on these approximate values, a linear classifier CF can be used to calculate its classification C, as follows:

[0110] C = PWC(x)

[0111] =CF(x×FP)

[0112] In the above formula, PWC is the final classifier model, while CF is a simple classifier based on systolic and diastolic blood pressure. It directly outputs different classification values ​​for different combinations of systolic and diastolic blood pressure intervals, resulting in a total of N classifications. cf kind.

[0113] After completing the PWC modeling, we can then perform classification calculations based on the actual collected pulse waves. The pre-trained first-level classification model is denoted as PWC1. If second-level classification is used, the second-level classification model is denoted as PWC2. The specific classification steps are as follows:

[0114] S105. Segment the pulse wave sequence of key nodes, extract each pulse wave separately, and select pulse waves with obvious features to form a pulse wave set. Based on the features in the feature subset, extract features from each pulse wave in the pulse wave set to obtain a feature matrix, and input it into the classification model to obtain the classification result of each pulse wave.

[0115] S510. The pulse wave sequence of key nodes is segmented by setting a fixed length to divide the pulse wave sequence into individual pulse waves, and the segmented individual pulse waves are extracted to form multiple independent pulse waves. Pulse waves with obvious characteristics are selected to form a pulse wave set.

[0116] The obtained pulse wave sequences of each key node are segmented, and the pulse wave sequences of each key node are PW. i Length N frm This yields individual pulse waves, each with a fixed length of N. pw The system identifies points and filters out pulse wave waveforms with more prominent characteristics, saving them to a pulse wave set PWS. PWS contains N... pws A pulse wave.

[0117] In one embodiment, N pw If the original pulse wave length is 1280 points, it will be less than 1280 points. In this case, zeros need to be added to both ends of the original pulse wave to make its length 1280 points. The number of zeros added to both ends should be equal or differ by 1.

[0118] S520. Extract corresponding features for each pulse wave in the pulse wave set based on a pre-selected feature subset, arrange all extracted features into a feature matrix, and then input the feature matrix into a pre-trained classification model.

[0119] For each pulse wave in the PWS, the corresponding results are extracted according to the features in feature subset 1 and saved into the feature matrix FSM. If there is a secondary classification, the corresponding results are extracted according to the features in feature subset 2 and saved into the feature matrix FSM. Each row of FSM corresponds to a feature group of one pulse wave in the PWS.

[0120] S530: The classification model classifies each pulse wave based on the input feature matrix and outputs the classification result for each pulse wave.

[0121] In the feature matrix FSM, feature groups corresponding to feature subset 1 are extracted row by row, and input into the first-level classification model PWC1 to obtain the first-level classification result. The classification output corresponding to the i-th row of input is denoted as C1. i The complete classification result corresponding to all rows is denoted as C1; the feature groups corresponding to feature subset 2 are extracted row by row in the feature matrix FSM, and input into the secondary classification model PWC2 to obtain the secondary classification result. The classification output corresponding to the i-th row of input is denoted as C2. i The complete classification result corresponding to all rows is denoted as C2;

[0122] If only primary classification is used, the final classification result is C = C1. If secondary classification is used, the final classification result is...

[0123] C = Comb(C1, C2)

[0124] Here, Comb(·) represents a linear combination of C1 and C2, if C1 contains N. c1 There are N categories, C2 contains N c2 If there are 1 category, then the final number of categories is N. cf =N c1 ×N c2 .

[0125] S106. Based on the classification results of each pulse wave, select the regression model that best matches each pulse wave from the pre-trained model pool, process the pulse wave, and input it into the regression model to obtain the blood pressure detection result.

[0126] This embodiment employs a multi-model layer, supporting the selection of different models for each pulse wave subclass to achieve optimal matching, thereby maximizing the overall detection performance of the system. Each subclass corresponds to one model, and the models of all subclasses together constitute a model pool. The models in the model pool can be:

[0127] (1) The same structure of the same type of neural network, such as CNN, and the same structure and number of layers;

[0128] (2) Different structures of the same type of neural network, such as CNN, but with different layer structures or number of layers;

[0129] (3) Different types of neural networks, such as a combination of CNN, LSTM, MultiResUNet and other neural network types;

[0130] The overall structure of the multi-model layer is as follows Figure 5 As shown, this structure selects the best model from the model pool based on the subclass for blood pressure detection. All models in the model pool need to be built before the formal implementation of blood pressure regression detection. In order to ensure that each subclass can obtain its own best matching model, multiple network models and multiple network structures must be selected for each subclass. Through iterative debugging and verification, the model with the highest final accuracy is selected as the final adapted model.

[0131] In step S510, a pulse wave set PWS is obtained, which contains all pulse waves within 20 seconds at 5 key nodes. The typical number of pulse waves is around 100. Blood pressure measurement is performed on each pulse wave in the PWS, as follows:

[0132] S610. Pre-build a model pool containing multiple regression models. Each regression model is optimized for a pulse wave category. Based on the pulse wave classification results, select a regression model from the model pool that matches the pulse wave category.

[0133] For the i-th pulse wave in PWS i According to its classification result C i Select the corresponding regression model NN i And set it as the current blood pressure detection model, i.e., NN act =NN i .

[0134] S620. The pulse wave is downsampled and normalized, and the processed pulse wave is input into the selected regression model for calculation to obtain the blood pressure detection result of the pulse wave.

[0135] The original pulse wave was downsampled by 1 / 5 sampling, resulting in the pulse wave pulse waveform (PWS). i The result after downsampling is denoted as PWD. iThe number of individual pulse wave points was reduced from 1280 to 256, and then a normalization operation was performed, i.e., the pulse wave PWD was reduced. i The values ​​are normalized to the range [0,1] to obtain the normalized pulse wave PWN. i The method is as follows:

[0136]

[0137] Normalized pulse wave waveform PWN i Input to the currently activated neural network model NN act The single pulse wave blood pressure (SBP) result can be obtained from this. i and DBP i The test results SBP i and DBP i Buffered into the SBPL and DBPL queues.

[0138] After all pulse waves in the PWS have been measured according to the above steps, a single pulse wave blood pressure measurement result with the same number of pulse waves as in the PWS can be obtained. This number is denoted as N. nndet These results are stored in queues SBPL and DBPL according to systolic and diastolic blood pressure, respectively, for subsequent statistical filtering.

[0139] S107. Perform sliding window histogram statistics on the blood pressure detection results for each pulse wave, and output the final blood pressure detection result based on the statistical results.

[0140] Because non-contact measurements are inevitably affected by factors such as noise, clutter, and body movement, single-pulse wave detection results may exhibit significant deviations. This solution, however, outputs N signals for each measurement. nndet Since there are multiple blood pressure measurement results, it is possible to combine multiple sets of results to improve the accuracy of the measurement. Therefore, a sliding window histogram statistical method is used to analyze N... nndet The results are statistically filtered, and the filtered result is used as the final blood pressure measurement result, thereby improving the accuracy of the final detection. The steps are as follows:

[0141] S710. Set the size and step size of the sliding window. The sliding window slides at the set step size, and the blood pressure monitoring results of each pulse wave in the window are statistically analyzed to generate a histogram.

[0142] (1) Perform sliding window histogram statistics on SBPL and DBPL respectively.

[0143] SBPL and DBPL each contain N nndet systolic blood pressure results and N nndetFor each diastolic blood pressure result, a sliding window histogram (HIST) was performed on both. This involved counting the number of times a test result fell within a window of length Wlen, and the resulting histograms were then used to obtain the systolic blood pressure histograms. sbp Histogram of diastolic and systolic blood pressure (HIST) dbp ,like Figure 6 As shown, in one embodiment, Wlen is taken as twice the sample standard deviation and is an integer value. Furthermore, the minimum value of Wlen is 1, i.e.:

[0144] For systolic blood pressure:

[0145] Wlen=max(round(2×std(SBPL)),1)

[0146] HIST sbp Each bin unit height represents the bin located in [x] in the SBPL. sbp -wlen,x sbp The number of detection points in the interval, where x sbp The x-axis of the bin unit represents the systolic pressure.

[0147] Regarding diastolic blood pressure:

[0148] Wlen=max(round(2×std(DBPL)),1)

[0149] HIST dbp Each bin unit height represents the height of the bin located in [x] in DBPL. dbp -wlen,x dbp The number of detection points in the interval, where x dbp The x-axis of the bin unit represents diastolic blood pressure.

[0150] S720. For all window cells in the histogram, identify the blood pressure interval with the highest frequency, calculate the mean of the blood pressure detection results within the sliding window in the blood pressure interval with the highest frequency, and output the mean as the final blood pressure detection result.

[0151] The window corresponding to the largest number of values ​​is selected as the target window. Values ​​outside the window are removed as outliers, and the average value of the test results within the window is used as the final blood pressure test result. The procedure is as follows:

[0152] 1) In the systolic blood pressure detection histogram HIST sbp Search for the peak value and determine the detection window corresponding to the peak value. Use the average of all systolic blood pressure detection results within the detection window as the final systolic blood pressure detection result SBP_OUT.

[0153] 2) In the diastolic blood pressure detection histogram (HIST) dbpSearch for peak values ​​and determine the detection window corresponding to the peak value. Use the average of all diastolic blood pressure detection results within the detection window as the final diastolic blood pressure detection result DBP_OUT.

[0154] The invention can be summarized into the following five steps: (1) data acquisition and preprocessing; (2) PWV estimation based on a single radar; (3) classification layer processing: target classification based on explicit pulse wave features; (4) multi-model layer processing: blood pressure regression detection based on target subclasses and model pools; (5) statistical layer processing: statistical filtering of detection results and output of final results. The invention designs a hierarchical multi-model detection architecture. In the "classification layer", first-level or second-level classification is used to make full use of prior information to realize the division of detection scenarios, thereby reducing the complexity of the problem and reducing the computational cost of the system. In the "multi-model layer", the best matching model is established for each subclass to improve detection accuracy and achieve the overall performance optimization of the system. Finally, in the "statistical layer", the results are filtered based on statistical distribution to remove abnormal samples, further improving measurement accuracy and detection stability.

[0155] The non-contact blood pressure measurement method based on millimeter-wave radar and hierarchical multi-model architecture can be applied to a smart card device based on an integrated millimeter-wave radar module. The radar module contains 3 transmit channels and 4 receive channels, which are configured as a virtual array of 1 transmit and 12 receive channels through MIMO.

[0156] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A contactless blood pressure measurement method based on millimeter wave radar and a hierarchical multi-model architecture, characterized by, The method comprises the following steps: Collecting original ADC data of a target area by using a millimeter wave radar device, and performing FFT on the ADC data to generate a one-dimensional FFT result, constructing a range-azimuth map based on the result, performing target detection based on the range-azimuth map, identifying an AOI area, performing elevation DBF angle measurement in the AOI area, and generating a radar point cloud; Estimating the pulse wave propagation path of the target area based on the radar point cloud data, selecting multiple key nodes on the propagation path, and extracting the pulse wave phase information of the multiple key nodes frame by frame to form a phase matrix, wherein each row in the phase matrix represents a pulse wave sequence of a key node; Calculating the pulse wave transmission time between the key nodes by cross-correlation, generating a transmission time matrix, and calculating the relative position matrix between the key nodes, and calculating the pulse wave transmission speed based on the transmission time matrix and the relative position matrix between the key nodes; Constructing a plurality of classification models based on the explicit features of the pulse waves, and training the plurality of classification models based on the corresponding feature subsets and sample sets; Segmenting the pulse wave sequence of the key nodes, extracting each pulse wave separately, and selecting pulse waves with obvious features to form a pulse wave set, extracting features of each pulse wave in the pulse wave set based on the features in the feature subset to obtain a feature matrix, and inputting the feature matrix into the classification model to obtain a classification result of each pulse wave; Selecting a regression model that best matches each pulse wave from a pre-trained model pool based on the classification result of each pulse wave, inputting the processed pulse wave into the regression model to obtain a blood pressure detection result; Performing sliding window histogram statistics on the blood pressure detection result of each pulse wave, and outputting the final blood pressure detection result based on the statistical result.

2. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, The method comprises the following steps: Collecting original ADC data of a target area by using a millimeter wave radar device, and performing FFT on the ADC data to generate a one-dimensional FFT result, constructing a range-azimuth map based on the result, performing target detection based on the range-azimuth map, identifying an AOI area, performing elevation DBF angle measurement in the AOI area, and generating a radar point cloud; The millimeter wave radar device comprises a plurality of virtual receiving channels configured for transmitting and receiving pulse signals; The millimeter wave radar device transmits pulse signals and receives signals reflected from the target area, generates original ADC data of each frame containing a channel data of a pulse, and saves the received original ADC data into a matrix according to the plurality of virtual receiving channels, wherein each row of the matrix represents all distance sampling points of a virtual receiving channel, and each column represents data of a plurality of virtual channels at a sampling time, and the number of columns represents the number of distance sampling points; Performing windowing on the saved original ADC data, performing one-dimensional FFT on the windowed data, and converting the time domain signal into a frequency domain signal to generate a one-dimensional FFT result; Performing azimuth DBF processing on all distance units by combining the one-dimensional FFT results of the plurality of virtual receiving channels and a pre-calibrated azimuth steering vector table to generate a two-dimensional range-azimuth map; Based on the generated range-azimuth map, applying a CFAR algorithm to perform target detection to obtain target distance and azimuth angle information, and determining an AOI area based on the CFAR detection result. For the target in the AOI region, a pitch DBF processing is performed based on a pre-calibrated pitch guide vector table to obtain target pitch angle information; Combined with the distance, azimuth angle and pitch angle information, a radar point cloud of the AOI region is generated.

3. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, Based on the radar point cloud data, a pulse wave propagation path of the target region is estimated, and a plurality of key nodes are selected on the propagation path, and pulse wave phase information of the plurality of key nodes is extracted frame by frame to form a phase matrix, including: Using the radar point cloud data of the AOI region, the point cloud region is divided into a plurality of closed regions of equal area according to the extension direction of the point cloud; For each closed region, all points therein are extracted and the geometric center point thereof is calculated, and the center points of all regions form a center point sequence, and the center point sequence is linearly interpolated to generate a denser path point sequence, i.e. an estimated pulse wave propagation path; Along the estimated pulse wave propagation path, a plurality of detection points are sequentially selected from the point cloud data as pulse wave key nodes, and the Euclidean distance between each key node and the previous key node is greater than a set threshold, except for the first key node; The phase information of the pulse wave is extracted frame by frame at each key node position, and the extracted phase information is subjected to phase unwrapping processing. The phase information extracted frame by frame at all key nodes forms a phase matrix, each row of the phase matrix corresponding to a pulse wave sequence of a key node, and each column corresponding to a sampling point of the pulse wave.

4. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, The pulse wave transmission time between the key nodes is calculated by cross-correlation to generate a transmission time matrix, and a relative position matrix between the key nodes is calculated, based on the transmission time matrix and the relative position matrix between the key nodes, the pulse wave transmission velocity is calculated, including: For each key node, the cross-correlation function between its pulse wave sequence and the pulse wave sequences of other key nodes is calculated as follows wherein, PW i denotes the pulse wave sequence of the i-th key node, PW j denotes the pulse wave sequence of the j-th key node, k is an integer, denotes the pulse wave sampling point offset number, N frm denotes the total frame number, i.e. the total pulse wave sampling point number, and n denotes the pulse wave sampling point index; The pulse wave transmission time between each key node is calculated using the cross-correlation function, and the calculation formula is as follows: wherein, PWTT ij denotes the pulse wave transmission time between the i-th key node and the j-th key node, T frm denotes the frame period, i.e., the pulse wave sampling period, denotes the cross-correlation function R ij (k) the k value corresponding to the peak position; The pulse wave transmission time between all key nodes forms a pulse wave transmission time matrix where N keyp represents the number of key nodes; The specific coordinates of each key node in the three-dimensional space are determined using the radar point cloud data, and the Euclidean distance between each key node is calculated, and the calculation formula is as follows: wherein LOC ij represents the Euclidean distance from the ith key node to the jth key node, x i , y i , and z i represent the three-dimensional coordinates of the ith key node, x j , y j , and z j represent the three-dimensional coordinates of the jth key node; The Euclidean distances between all key nodes form the relative position matrix Based on the transmission time matrix and the relative position matrix, a transmission velocity matrix PWVM of the pulse wave between the key nodes is calculated, and the calculation formula of the transmission velocity PWVM(i,j) between each key node is as follows: The upper triangular or lower triangular part of the transmission velocity matrix PWVM excluding the diagonal elements is taken out, and the mean value is calculated to obtain the pulse wave transmission velocity PWV, and the calculation method is as follows: 。 5. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, Based on the explicit features of the pulse wave, a plurality of classification models are constructed, and the plurality of classification models are trained based on the corresponding feature subsets and sample sets, including: Explicit features are extracted from the pulse wave signal, and a correlation analysis is performed based on the explicit features, and a feature subset with high correlation to the target classification task is selected, and the feature subset includes a plurality of explicit features; Based on the feature subset, a plurality of classification models are constructed, the feature subset in the sample set is input into the classification model for training, the model parameters are adjusted to minimize the classification error, and a trained classification model is obtained, and the sample set includes sample data and target classification labels corresponding to the feature subset.

6. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, The pulse wave sequence of the key node is segmented, each pulse wave is extracted separately, and the pulse wave set is formed by screening the pulse wave with obvious features. The feature matrix is obtained by extracting the features of each pulse wave in the pulse wave set based on the features in the feature subset, and inputting the feature matrix into the classification model to obtain the classification result of each pulse wave, including: The pulse wave sequence of the key node is segmented, and the pulse wave sequence is segmented into single pulse wave by setting a fixed length, and the segmented single pulse wave is extracted to form multiple independent pulse waves, and the pulse wave with obvious features is screened to form a pulse wave set; Based on the pre-selected feature subset, the corresponding features of each pulse wave in the pulse wave set are extracted, all extracted features are arranged into a feature matrix, and the feature matrix is input into the pre-trained classification model; The classification model classifies each pulse wave based on the input feature matrix and outputs the classification result of each pulse wave.

7. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, According to the classification result of each pulse wave, the regression model most matched with each pulse wave is selected from the model pool pre-trained, and the blood pressure detection result is obtained by inputting the processed pulse wave into the regression model, including: A model pool containing multiple regression models is constructed in advance, and a regression model is optimized for a pulse wave category. According to the classification result of the pulse wave, the regression model matched with the pulse wave category is selected from the model pool; The pulse wave is down-sampled and normalized, and the processed pulse wave is input into the selected regression model for calculation to obtain the blood pressure detection result of the pulse wave.

8. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1, characterized in that, The blood pressure detection result of each pulse wave is subjected to sliding window histogram statistics, and the final blood pressure detection result is output based on the statistical result, including: The size and step length of the sliding window are set, the sliding window slides at the set step length, and the blood pressure monitoring result of each pulse wave in the window is counted to generate a histogram; For all window units in the histogram, the blood pressure interval with the highest frequency is identified, the mean value of the blood pressure detection result in the sliding window is calculated in the blood pressure interval with the highest frequency, and the mean value is output as the final blood pressure detection result.

Images