Non-contact blood pressure measurement method based on millimeter wave radar and layered multi-model architecture

Abstract

The invention discloses a non-contact blood pressure measurement method based on a millimeter wave radar and a layered multi-model architecture. The non-contact blood pressure measurement method comprises the steps of data acquisition and preprocessing; pWV estimation is carried out based on a single radar; classification layer processing: performing target classification based on pulse wave explicit features; performing multi-model layer processing: performing blood pressure regression detection based on the target subclass and the model pool; statistical layer processing: performing statistical filtering on a detection result and outputting a final result; a single radar and a layered multi-model architecture are utilized, detection of the pulse wave transmission speed is achieved, the blood pressure detection precision is improved, and meanwhile the system and operation are simplified; the layered multi-model structure enables the detection to be more accurate and has high adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of non-contact human feature detection radar, and in particular to a non-contact blood pressure measurement method based on millimeter wave radar and a hierarchical multi-model architecture. Background Art

[0002] Cardiovascular disease is a major threat to human health and life safety. As one of the important physiological parameters of the human body, long-term monitoring of blood pressure is of great significance for the prevention and diagnosis of cardiovascular disease. Although traditional contact blood pressure measurement technology is relatively mature, it has great constraints on patients and it is difficult to achieve long-term continuous monitoring. As people's health needs increase, non-contact blood pressure measurement technology has emerged. Among them, millimeter-wave-based blood pressure measurement solutions have developed rapidly due to their low cost and high convenience.

[0003] The most common method of using millimeter waves to measure blood pressure is: first use millimeter wave radar equipment to detect the human pulse wave, and then measure blood pressure by establishing a conversion model between the pulse wave and blood pressure (systolic pressure SBP and diastolic pressure DBP). Among them, establishing a blood pressure conversion model is the key to the solution. There are two mainstream solutions for establishing a blood pressure conversion model. One is "modeling based on explicit features of pulse waves", that is, establishing a conversion model from explicit features of pulse waves to blood pressure. The explicit features of pulse waves include pulse wave transmission velocity (PWV, Pulse Wave Velocity), pulse wave frequency (PWF, Pulse Wave Frequency), main wave height, etc. The second is "modeling based on deep neural network", that is, not deliberately paying attention to the specific features of the pulse wave, but directly establishing a neural network conversion model from pulse wave waveform to blood pressure, and the neural network automatically extracts features and calculates blood pressure.

[0004] (1) For the blood pressure measurement scheme based on explicit features, there are mainly two problems:

[0005] PWV measurement is cumbersome. In the millimeter wave non-contact blood pressure measurement system, the conventional method of measuring PWV is to deploy dual radar measurements at two points based on the same pulse wave conduction tree, and calculate PWV by synchronously monitoring the pulse wave relationship between the two measuring points. This method is cumbersome to deploy and implement, so its application scope is also limited; the feature set is incomplete and the accuracy is limited. Although feature engineering can select the feature set with the highest correlation with blood pressure, due to the complexity of the blood pressure model, it is currently impossible to extract complete features that match blood pressure. Therefore, the conversion model from pulse wave features to blood pressure is not accurate in theory, which limits the accuracy of blood pressure measurement.

[0006] (2) Regarding the “blood pressure measurement solution based on deep neural network”, there are mainly three problems:

[0007] High complexity and high computing power consumption. The ability of a neural network is highly correlated to the number of neurons. In order to adapt to different populations and different blood pressure ranges, the number of layers of the final model and the number of neurons in each layer are usually not low, which poses a great challenge to the computing power of small and micro embedded devices; insufficient "utilization" of known important features. The features of deep neural networks are automatically acquired through model training, and there is a lack of effective means to pass prior information to the model. The "utilization" of known important features (such as PWV) is usually lower than that of dedicated models for this feature; overfitting or underfitting problems. For deep neural networks, although the increased complexity of the model can improve performance, the requirements for the data set will also increase. Generally, the more complex the model, the higher the possibility of overfitting or underfitting problems due to data set bias. Summary of the invention

[0008] Technical purpose: In view of the shortcomings of existing blood pressure measurement, the present invention discloses a non-contact blood pressure measurement method based on millimeter-wave radar and hierarchical multi-model architecture, which improves the blood pressure detection accuracy while reducing the 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 hierarchical multi-model architecture, comprising:

[0011] 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, build 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 in the AOI area and generate a radar point cloud;

[0012] The pulse wave propagation path of the target area is estimated based on the radar point cloud data, 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 a pulse wave sequence of a key node.

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

[0014] Building several classification models based on the explicit features of the pulse wave, and training the several classification models based on the corresponding feature subsets and sample sets;

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

[0016] According to the classification result of each pulse wave, a regression model that best matches each pulse wave is selected from the pre-trained model pool, and the pulse wave is processed and input into the regression model to obtain the blood pressure detection result;

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

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

[0019] 1. The present invention constructs multiple classification models and regression models to perform classification and regression analysis on different types of pulse wave signals, and can select the most matching model for blood pressure detection according to actual conditions. This multi-model architecture improves the adaptability to pulse wave signals under different physiological states, thereby significantly improving the accuracy and reliability of blood pressure detection; the present invention extracts explicit features of the pulse wave, such as pulse wave transmission speed, pulse wave frequency, etc., and constructs a classification model based on these explicit features. Since these features are directly related to blood pressure, more representative pulse wave signals can be screened out through the classification model, so that blood pressure prediction can be performed more accurately.

[0020] 2. The present 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 has a high correlation 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. The present invention processes the blood pressure detection results of the pulse wave signal through sliding window histogram statistics, thereby reducing the fluctuation of the detection results caused by individual abnormal pulse waves and enhancing the stability and reliability of the blood pressure detection results; when processing the pulse wave signal, the scheme adopts phase unwrapping and cross-correlation calculation and other technologies, which effectively reduces the phase jump and noise interference caused by the periodicity of the signal, and ensures the accuracy of signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0023] Figure 1 It is the overall processing flow chart of the present invention;

[0024] Figure 2 A schematic diagram of detecting an arm according to the present invention;

[0025] Figure 3 A RAMap schematic diagram of the present invention;

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

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

[0028] Figure 6 It is the detection histogram of systolic pressure and diastolic pressure of the present invention. DETAILED DESCRIPTION

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

[0030] In a specific embodiment, the frame rate f of the millimeter wave radar device sz The frequency is 1000Hz, 1 frame contains 1 pulse, the number of sampling points in a single frame is 256 points, the bandwidth is 2.5GHz, the antenna has 3 transmit and 4 receive, and there are a total of 12 virtual receiving channels. However, in this embodiment, only 7 azimuth channels are used to measure the azimuth angle, and 3 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 hierarchical multi-model architecture provided in this embodiment are as follows:

[0032] S101. Use a millimeter wave radar device to collect original 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 in the AOI area and generate a radar point cloud.

[0033] This step is data collection and preprocessing, which includes the process from collecting ADC data to generating AOI (Area Of Interest) area point cloud. The specific steps are as follows:

[0034] S110. The millimeter wave radar device includes a plurality of virtual receiving channels configured to transmit and receive pulse signals.

[0035] In one embodiment, the millimeter wave radar device includes a radar device including N ch Virtual receiving channels, each frame contains 1 Chirp N ch Channel data, the number of distance sampling points is N0, in this embodiment N ch is 12, N0 is 256, according to Figure 2 The arm data is collected for blood pressure monitoring in the manner shown. The arm area data is collected because the arm is relatively long, which facilitates the estimation of the pulse wave conduction path and the calculation of the pulse wave velocity PWV. During the process, the arm is kept still as much as possible, and the distance between the radar device and the center of the arm is preferably within the range of [0.3, 0.8] meters.

[0036] S120, the millimeter wave radar device transmits a pulse signal and receives a signal reflected from the target area, generating raw ADC data containing one pulse per frame, and the received raw ADC data is saved in a matrix according to a number of 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 into the matrix D according to the channel. adc In, D adc The size is N ch ×N0.

[0038] S130 , performing windowing processing on the saved original ADC data, performing one-dimensional FFT on the windowed data, converting the time domain signal into a frequency domain signal, and generating a one-dimensional FFT result.

[0039] To N ch The raw ADC data D of the receiving channels adc Perform N1-point windowed FFT operation along the distance dimension to obtain a one-dimensional FFT result D 1d , D 1d is of size N ch ×N1 matrix, the specific calculation is 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 spectrum leakage and enhance signal quality.

[0042] S140, combining the one-dimensional FFT results of multiple virtual receiving channels and the pre-calibrated azimuth steering vector table, performing azimuth DBF processing on all range units to generate a two-dimensional range-azimuth map;

[0043] In D1d Extract the sub-matrix D corresponding to the orientation channel from the matrix 1d,az , D 1d,az is of size N ch,az ×N1 matrix, according to the azimuth steering vector table SVT az For the matrix D 1d,az By performing azimuth digital beamforming (DBF) scanning, the range-azimuth map RAMap can be obtained, as shown in the schematic diagram. Figure 3 As shown, the specific calculation formula is

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

[0045] Among them SVT az is the azimuth guidance vector table, size is N az ×N ch,az , SVT az Need to be calibrated in advance, N az The number of angles for azimuth calibration.

[0046] In one embodiment, the number of azimuth channels used is N ch,az =7, a total of 121 angles are calibrated in the azimuth direction, that is, a steering vector is calibrated for each degree in the angle range [-60,60] degrees, and the azimuth steering vector SVT is obtained. az The size is 121×7.

[0047] S150, based on the generated range-azimuth map, applying the CFAR algorithm to perform target detection, and determining the AOI area based on the CFAR detection result;

[0048] Based on the RAMap two-dimensional graph, constant false alarm rate (CFAR) is used to detect the target. In one embodiment, the CFAR strategy can be CA-CFAR, OS-CFAR, etc., and the detected target is saved in the target array TAR1. The i-th target point in TAR1 is a four-tuple consisting of the target's distance, azimuth, amplitude and noise (r i ,az i ,amp i ,ns i ).

[0049] The purpose of AOI area selection is to keep the targets in the arm area as much as possible and remove the targets outside the arm range. The basic method is: use W r ×W a The convolution map CMap is obtained by convolving the 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;

[0050] In TAR1, all targets whose distance is in the interval [r0-Δr, r0+Δr] are screened out, and then the screened targets are sorted according to energy, and the target points with higher energy are screened out in proportion. The proportion threshold is G rt ;

[0051] Density clustering is performed on the selected targets to obtain multiple distance regions A0, A1, ..., A i , remove points less than the threshold G A The remaining areas are merged into one area, which is the AOI area, denoted as A aoi , and A aoi The included points are saved to the target set TAR2.

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

[0053] The basic method of pitch DBF angle measurement and point cloud construction is: in the one-dimensional FFT result matrix D 1d Filter out N ch,el The data corresponding to the pitch channels form an array D 1d,el ,D 1d,el is of size N ch,el ×N1 matrix; for the i-th target point in TAR2, take out the 1d,el Take out the corresponding one-dimensional FFT transformed complex vector V i ; for vector V i The pitch DBF spectrum can be obtained by performing a pitch DBF scan. The specific calculation formula is:

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

[0055] Among them SVT el is the pitch guidance vector table, size is N el ×N ch,el , SVT el Need to be calibrated in advance, N el The number of angles for pitch calibration.

[0056] In the pitch DBF spectrum S i,el In the example, find the peak point by absolute value and use its corresponding angle as the pitch angle el i , the horizontal coordinate in the pitch DBF spectrum is the pitch angle el i.

[0057] S170 , combining the distance, azimuth, and elevation information to generate a radar point cloud of the AOI area.

[0058] For the i-th target point in TAR2, according to its distance r i 、azimuth az i And the pitch angle el calculated previously i Calculate the coordinates of the target point in the three-dimensional Cartesian coordinate system (x i ,y i ,z i ), combined with the known amplitude amp i and noise ns i , forming the target quintuple (x i ,y i ,z i ,amp i ,ns i ), repeat this operation until all target points in TAR2 are converted into quintuple representation according to the above method; frm All the above five-tuples obtained from the frame are cached in the detection point set PC, which is the radar point cloud of the AOI area, where N frm It is the number of frames required to complete a blood pressure test.

[0059] In one embodiment, the number of pitch channels used is N ch,el =3, a total of 21 angles are calibrated in the pitch direction, that is, a steering vector is calibrated for each degree in the angle range [-10,10] degrees, and the pitch steering vector SVT is obtained. el The size of N is 21 × 7. In addition, the data accumulation window in this case is 20 seconds and the frame rate is 1000 Hz, so N frm is 20000.

[0060] S102, 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, 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.

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

[0062] S210 , using the radar point cloud data in the AOI area, dividing the point cloud area into a plurality of 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 part closed regions of equal area Part1, Part2, ..., Part n , to ensure coverage of the entire arm surface and facilitate subsequent path estimation.

[0064] S220, extract all points in each closed area and calculate its geometric center point, form the center points of all areas into a center point sequence, perform linear interpolation on the center point sequence, and generate a denser path point sequence, that is, the estimated pulse wave propagation path.

[0065] In one embodiment, for the i-th closed area Part i , take out all the points in the area and calculate their geometric center to get the center point CP i , after completing the calculation of the center points of all regions, we get the center point sequence 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 partitions N part The value is 10, and the length of the interpolated PPL is 100.

[0067] S230, selecting a number of key nodes from the point cloud data in sequence along the estimated pulse wave propagation path as pulse wave key nodes, wherein the Euclidean distance between each key node except the first key node and the previous key node is greater than a set threshold.

[0068] Along the PPL direction, select N points in the point cloud in turn. keyp Detection points as key nodes of pulse wave In one embodiment, a key node KP is selected i There are two requirements: 1) Except for KP1, the Euclidean distance between any other key node and the previous key node must be higher than the threshold G d 2) KP i Need to be close enough to PPL, that is, to ensure KP i It is the point among all the detection points in the point cloud that is closest to a path point in the PPL. In this embodiment, the number of key nodes N keyp The value is 5.

[0069] S240, extracting phase information of the pulse wave frame by frame at each key node position, and performing phase unwrapping processing on the extracted phase information.

[0070] Due to the periodicity of the pulse wave, phase jumps may occur. Therefore, the extracted phase information is unwrapped to ensure the continuity of the phase. keyp Phase information is extracted frame by frame at each key node position and de-wrapping is performed.

[0071] S250, constructing a phase matrix from the phase information extracted frame by frame from all key nodes.

[0072] When the total number of frames reaches N frm When 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, calculating the pulse wave transmission time between key nodes by cross-correlation, generating a transmission time matrix, and calculating the relative position matrix between key nodes, and calculating the pulse wave transmission speed based on the transmission time matrix and the relative position matrix between key nodes.

[0074] Different rows of the phase matrix PHA represent the sampling of the same pulse wave at different positions. Theoretically, there is a translation matching relationship between rows, and the propagation time of the pulse wave between two key nodes can be determined by detecting the translation 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 there may even be large differences. If only the pulse wave feature points are used for matching (such as matching based on the pulse wave peak point), there may be large errors. Therefore, in this embodiment, a multi-pulse wave cross-correlation strategy is used to calculate the translation amount to improve the 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 record the pulse wave sequence of the i-th key node as PW i , that is, PW i =PHA(i,:), the cross-correlation function R between the pulse wave sequence of the i-th key node and the pulse wave sequence of the j-th key node ij as follows:

[0076]

[0077] Among them, PW i represents the pulse wave sequence of the i-th key node, PW j represents the pulse wave sequence of the jth key node; k is an integer, representing the pulse wave sampling point offset number, N frmrepresents the total number of frames, that is, the total number of sampling points of the pulse wave sequence, and n represents the pulse wave sampling point index;

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

[0079]

[0080] Among them, PWTT ij represents the pulse wave transmission time between the i-th key node and the j-th key node, T frm represents the frame period, that is, the pulse wave sampling period. In one embodiment, the frame period T frm 1ms, represents the cross-correlation function R ij (k) k value corresponding to the peak position;

[0081] S330, 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. Since PWTT(i,j) represents the pulse wave transmission time between the i-th and j-th nodes, it can be seen that PWTT is a symmetric matrix. Since the diagonal elements represent the propagation time from the key node to itself, it is 0;

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

[0083]

[0084] Among them, LOC ij represents the Euclidean distance from the i-th key node to the j-th key node, x i ,y i and z i Represent the three-dimensional coordinates of the i-th key node, x j ,y j and z j Respectively represent the three-dimensional coordinates of the j-th key node;

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

[0086] S360, calculating the transmission speed matrix PWVM of the pulse wave between each key node based on the transmission time matrix and the relative position matrix, wherein the calculation formula of the pulse wave transmission speed PWVM(i,j) between the i-th key node and the j-th key node is:

[0087]

[0088] S370, extract the upper triangular or lower triangular part of the transmission velocity matrix PWVM and exclude the diagonal elements, and calculate the mean thereof to obtain the pulse wave transmission velocity PWV, and the calculation method is as follows:

[0089]

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

[0091] This step classifies the total space of pulse wave samples based on the explicit features of the pulse wave, and divides the complex total space of samples into multiple relatively simple sample subspaces. The samples within the subspaces have high consistency. Modeling and detection of subspaces can obtain two advantages:

[0092] (1) The subspace structure is relatively simple, so the complexity of the established regression model is also relatively low, which can reduce computing power overhead while improving detection accuracy;

[0093] (2) The most matching detection model can be independently established for different subspaces to optimize the detection performance

[0094] S410, extracting explicit features from the pulse wave signal, performing correlation analysis based on the explicit features, and selecting a feature subset with high correlation with the target classification task, wherein the feature subset includes multiple explicit features;

[0095] There are many features that can be used for pulse wave classification. The main features include pulse wave transmission velocity PWV, pulse wave frequency PWF, main wave height h1, descending isthmus height h2, dicrotic wave height h3, rapid ejection period t1, left ventricular ejection period t2, diastolic period t3, cardiac cycle t4, pulse wave area S1, systolic area S2, and diastolic area S3. In addition, in actual use, some derived features are preferred, such as the relative height of descending isthmus h 21 、Relative height of dicrotic wave h 31 , relative area during systole S 21 and diastolic relative area S 31 Etc. These features are given based on consensus. Different features have different correlations with blood pressure. It is necessary to pre-analyze and select feature subsets with higher correlation for classification.

[0096] The classification layer can have one or two levels of classification. Whether it is the first or second level needs to be determined by comprehensive consideration of factors such as storage space, computing power requirements, accuracy requirements, blood pressure range and applicable population. Under normal circumstances, only one level of classification is required, but for some special cases, such as the pulse wave characteristics of patients with atrial fibrillation are quite different from those of the general population. If such special cases need to be covered, they must be classified and processed in a targeted manner. At this time, secondary classification is needed.

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

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

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

[0100]

[0101] S420, constructing several classification models based on feature subsets, inputting the feature subsets in the sample set into the classification model for training, adjusting the model parameters to minimize the classification error, and obtaining a trained classification model, wherein the sample set includes sample data corresponding to the feature subsets and target classification labels.

[0102] Whether it is primary classification or secondary classification, the establishment and use of the model are similar, except that the feature subsets and models that the classification depends on are different. The classification modeling can be a neural network model or a polynomial model. Here, taking the feature set FS1 of the primary classification as an example, the modeling method based on polynomial fitting is given as follows:

[0103] The sample set X is established as follows, where PWVL, PWFL and SL 21 They are pulse wave velocity PWV, pulse wave frequency PWF and relative systolic area S 21 A column vector of , where L is a column vector whose elements are all 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, SBPL is the systolic pressure column vector, and DBPL is the diastolic pressure column vector:

[0106] Y=[SBPL,DBPL]

[0107] Then the polynomial coefficient matrix FP can be obtained as follows:

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

[0109] For a sample x in the feature subset, the rough values ​​of the systolic and diastolic blood pressures corresponding to the sample can be calculated according to the coefficient matrix FP. Based on the rough values, the linear classifier CF can be further 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, and CF is a simple classifier based on systolic and diastolic blood pressure, that is, it directly outputs different classification values ​​for different combinations of systolic and diastolic blood pressure intervals, which can be divided into N cf kind.

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

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

[0115] S510, segmenting the pulse wave sequence of the key node, segmenting the pulse wave sequence into individual pulse waves by setting a fixed length, extracting the segmented individual pulse waves to form multiple independent pulse waves, and screening out pulse waves with obvious characteristics to form a pulse wave set;

[0116] The pulse wave sequence of each key node is segmented, and the pulse wave sequence PW of each key node is i The length is N frm , get a separate pulse wave, the length of a single pulse wave is fixed to N pw points, and select the pulse wave waveforms with more obvious features, and save them into the pulse wave set PWS, which contains N pws A pulse wave.

[0117] In one embodiment, N pw The length of the original pulse wave directly segmented is necessarily less than 1280. In this case, it is necessary to fill zeros at both ends of the original pulse wave to make its length 1280 points. The number of zeros filled at both ends is equal or differs by 1.

[0118] S520, extracting corresponding features from each pulse wave in the pulse wave set based on a pre-selected feature subset, arranging all extracted features into a feature matrix, and then inputting the feature matrix into a pre-trained classification model;

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

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

[0121] Extract the feature group corresponding to feature subset 1 row by row in the feature matrix FSM, and input it into the primary classification model PWC1 to obtain the primary classification result. The classification output corresponding to the i-th row input is recorded as C1 i , the complete classification results corresponding to all rows are recorded as C1; extract the feature groups corresponding to feature subset 2 row by row in the feature matrix FSM, and input them into the secondary classification model PWC2 to obtain the secondary classification results. The classification output corresponding to the i-th row input is recorded as C2 i , the complete classification results corresponding to all rows are recorded as C2;

[0122] If only the first-level classification is used, the final classification result is C = C1. If the second-level classification is used, the final classification result is

[0123] C=Comb(C1,C2)

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

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

[0126] This embodiment uses a multi-model layer to support the selection of different models for each pulse wave subclass to achieve the best match, so that the overall detection performance of the system is optimized. Each subclass corresponds to a 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 neural network type, such as CNN, with the same structure and number of layers;

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

[0129] (3) Different neural network types, such as mixed use 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 in the figure, this structure selects the best model in the model pool for blood pressure detection according to the subclass. All models in the model pool need to be established before the formal implementation of blood pressure regression detection. In order for each subclass to 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 adaptation model.

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

[0132] S610, pre-building a model pool containing multiple regression models, wherein one regression model is optimized for one pulse wave category, and selecting a regression model matching the pulse wave category from the model pool according to the classification result of the pulse wave;

[0133] For the i-th pulse wave PWS 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, namely NN act =NN i .

[0134] S620 , downsampling and normalizing the pulse wave, and inputting the processed pulse wave into the selected regression model for calculation to obtain a blood pressure detection result of the pulse wave.

[0135] The original pulse wave is downsampled by 1 in 5, and the pulse wave PWS i The result after downsampling is recorded as PWD i, the number of points of a single pulse wave is reduced from 1280 to 256 points, and then a normalization operation is performed, that is, the pulse wave PWD i The value of is normalized to [0,1] to obtain the normalized pulse wave PWN i , the method is as follows

[0136]

[0137] The normalized pulse waveform PWN i Input to the currently activated neural network model NN act In the test, you can get the single pulse wave blood pressure test result SBP i and DBP i , the test results SBP i and DBP i Buffered into SBPL and DBPL queues.

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

[0139] S107, 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.

[0140] Since non-contact measurement is difficult to avoid the influence of noise, clutter and body movement, the single pulse wave detection result may have a large deviation, and this scheme will output N each time it is detected. nndet Blood pressure test results, so multiple groups of results can be combined to improve the detection accuracy. For this, the sliding window histogram method is used to calculate N nndet The statistical filtering is performed on the results, and the statistical filtering results are used as the final blood pressure measurement results, so as to achieve the purpose of improving the final detection accuracy. The steps are as follows:

[0141] S710, setting the size and step length of the sliding window, sliding the sliding window with the set step length, and collecting statistics on the blood pressure monitoring results of each pulse wave in the window to generate a histogram;

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

[0143] SBPL and DBPL contain N nndet Systolic blood pressure results and N nndetThe diastolic blood pressure results are counted by sliding window histogram for the two, that is, the number of detection results falling within the window is counted by sliding with the window length of Wlen, and the systolic blood pressure detection histogram HIST is obtained respectively. sbp and diastolic pressure detection histogram HIST dbp ,like Figure 6 As shown, in one embodiment, Wlen is twice the sample standard deviation and is an integer value. In addition, the minimum value of Wlen is 1, that is:

[0144] For systolic blood pressure:

[0145] Wlen=max(round(2×std(SBPL)),1)

[0146] HIST sbp The height of each bin unit represents the SBPL at [x sbp -wlen,x sbp ] The number of detection points in the interval, where x sbp is the horizontal axis of the bin unit, representing systolic blood pressure.

[0147] For diastolic blood pressure:

[0148] Wlen=max(round(2×std(DBPL)),1)

[0149] HIST dbp The height of each bin unit represents the position of [x dbp -wlen,x dbp ] The number of detection points in the interval, where x dbp It is the horizontal axis of the bin unit, representing the diastolic pressure.

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

[0151] Select the window corresponding to the maximum number as the target window, remove the values ​​outside the window as abnormal points, and use the average of the test results within the window as the final blood pressure test result. The method is as follows:

[0152] 1) In the systolic blood pressure detection histogram HIST sbp The peak value is searched, and the detection window corresponding to the peak value is determined, and the average value of all systolic blood pressure detection results within the detection window is taken as the final systolic blood pressure detection result SBP_OUT.

[0153] 2) In the diastolic pressure detection histogram HIST dbpThe peak value is searched, and the detection window corresponding to the peak value is determined, and the average value of all diastolic pressure detection results in the detection window is taken as the final diastolic pressure detection result DBP_OUT.

[0154] The present invention can be generally 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 pulse wave explicit features; (4) multi-model layer processing: blood pressure regression detection based on target subclasses and model pools; (5) statistical layer processing: statistical filtering of the detection results and output of the final results; The present invention designs a hierarchical multi-model detection architecture. In the "classification layer", with the help of primary or secondary classification, it makes full use of prior information to realize the detection scene division, so that the problem complexity is decomposed and downgraded, and the system computing power overhead is reduced; in the "multi-model layer", the best matching model is established for each subclass to improve the detection accuracy and achieve the optimization of the overall system performance; finally, in the "statistical layer", the results are filtered based on statistical distribution to eliminate abnormal samples, further improve the measurement accuracy, and also improve the detection stability.

[0155] The non-contact blood pressure measurement method based on millimeter wave radar and layered multi-model architecture of the present invention can be applied to a smart card device based on an integrated millimeter wave radar module, wherein the radar module includes 3 transmitting channels and 4 receiving channels, and a virtual array of 1 transmit and 12 receive is formed through MIMO.

[0156] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture, characterized in that: include: 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, build 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 in the AOI area and generate a radar point cloud; The pulse wave propagation path of the target area is estimated based on the radar point cloud data, 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 a pulse wave sequence of a key node. The pulse wave transmission time between key nodes is calculated by cross-correlation to generate a transmission time matrix, and the relative position matrix between key nodes is calculated. Based on the transmission time matrix and the relative position matrix between key nodes, the pulse wave transmission speed is calculated. Building several classification models based on the explicit features of the pulse wave, and training the several classification models based on the corresponding feature subsets and sample sets; The pulse wave sequence of key nodes is segmented, each pulse wave is extracted separately and the pulse waves with obvious features are screened to form a pulse wave set. Based on the features in the feature subset, the features of each pulse wave in the pulse wave set are extracted to obtain a feature matrix, which is then input into the classification model to obtain the classification result of each pulse wave. According to the classification result of each pulse wave, a regression model that best matches each pulse wave is selected from the pre-trained model pool, and the pulse wave is processed and input into the regression model to obtain the blood pressure detection result; Perform sliding window histogram statistics on the blood pressure detection results of each pulse wave, and output the final blood pressure detection results based on the statistical results.

2. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1 is characterized in that: 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, build 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 in the AOI area, and generate a radar point cloud, including: The millimeter wave radar device includes several virtual receiving channels configured to transmit and receive pulse signals; The millimeter wave radar device transmits a pulse signal and receives the signal reflected from the target area, generating raw ADC data containing channel data of one pulse per frame. The received raw ADC data is saved in a matrix according to several virtual receiving channels. Each row of the matrix represents all distance sampling points of a virtual receiving channel, and each column represents the data of multiple virtual channels at a sampling time. The number of columns represents the number of distance sampling points. Perform windowing on the saved original ADC data, perform one-dimensional FFT on the windowed data, convert the time domain signal into a frequency domain signal, and generate a one-dimensional FFT result; Combining the one-dimensional FFT results of multiple virtual receiving channels and the pre-calibrated azimuth steering vector table, azimuth DBF processing is performed on all range cells to generate a two-dimensional range-azimuth map; Based on the generated range-azimuth map, the CFAR algorithm is used to detect the target, obtain the target distance and azimuth information, and determine the AOI area based on the CFAR detection results; For the target in the AOI area, pitch DBF processing is performed based on the pre-calibrated pitch steering vector table to obtain the target pitch angle information; The radar point cloud of the AOI area is generated by combining the distance, azimuth and elevation information.

3. The non-contact blood pressure measurement method based on millimeter wave radar and hierarchical multi-model architecture according to claim 1 is characterized in that: The pulse wave propagation path of the target area is estimated based on the radar point cloud data, 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 including: Using the radar point cloud data in 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; Extract all points in each closed area and calculate its geometric center point, form a center point sequence with the center points of all areas, perform linear interpolation on the center point sequence, and generate a denser path point sequence, i.e., the estimated pulse wave propagation path; Selecting several detection points from the point cloud data in sequence along the estimated pulse wave propagation path as pulse wave key nodes, and the Euclidean distance between each key node and the previous key node except the first key node is greater than the set threshold; 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 from all key nodes constitutes a phase matrix, where each row of the phase matrix corresponds to a pulse wave sequence of a key node, and each column corresponds 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 key nodes is calculated by cross-correlation, a transmission time matrix is ​​generated, and a relative position matrix between key nodes is calculated. Based on the transmission time matrix and the relative position matrix between key nodes, the pulse wave transmission speed 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 Among them, PW i represents the pulse wave sequence of the i-th key node, PW j represents the pulse wave sequence of the jth key node, k is an integer, representing the pulse wave sampling point offset number, N frm represents the total number of frames, that is, the total number of sampling points of the pulse wave sequence, and n represents the pulse wave sampling point index; The pulse wave transmission time between each key node is calculated using the cross-correlation function. The calculation formula is as follows: Among them, PWTT ij represents the pulse wave transmission time between the i-th key node and the j-th key node, T frm represents the frame period, that is, the pulse wave sampling period, represents the cross-correlation function R ij (k) k value corresponding to the peak position; The pulse wave transmission time between all key nodes constitutes the pulse wave transmission time matrix Where N keyp Indicates the number of key nodes; The radar point cloud data is used to determine the specific coordinates of each key node in three-dimensional space, and the Euclidean distance between each key node is calculated. The calculation formula is as follows: Among them, LOC ij represents the Euclidean distance from the i-th key node to the j-th key node, x i ,y i and z i Represent the three-dimensional coordinates of the i-th key node, x j ,y j and z j Respectively represent the three-dimensional coordinates of the j-th key node; The Euclidean distances between all key nodes form a relative position matrix The transmission speed matrix PWVM of the pulse wave between each key node is calculated based on the transmission time matrix and the relative position matrix, where the calculation formula of the transmission speed PWVM(i,j) between each key node is: Take out the upper or lower triangular part of the transmission velocity matrix PWVM and exclude the diagonal elements, and calculate its mean to get the pulse wave transmission velocity PWV. 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: Several classification models are constructed based on the explicit features of the pulse wave, and several classification models are trained based on the corresponding feature subsets and sample sets, including: Extracting explicit features from the pulse wave signal, performing correlation analysis based on the explicit features, and selecting a feature subset with high correlation with the target classification task, wherein the feature subset includes multiple explicit features; Several classification models are constructed based on feature subsets, the feature subsets in the sample set are input into the classification model for training, the model parameters are adjusted to minimize the classification error, and a trained classification model is obtained. The sample set includes sample data corresponding to the feature subset and a target classification label.

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 key nodes is segmented, each pulse wave is extracted separately and the pulse waves with obvious features are screened to form a pulse wave set. Based on the features in the feature subset, each pulse wave in the pulse wave set is extracted to obtain a feature matrix, which is then input into the classification model to obtain the classification results of each pulse wave, including: The pulse wave sequence of the key nodes is segmented, and the pulse wave sequence is segmented into individual pulse waves by setting a fixed length, and the segmented individual pulse waves are extracted to form multiple independent pulse waves, and pulse waves with obvious characteristics are screened out to form a pulse wave set; Extracting corresponding features from each pulse wave in the pulse wave set based on a pre-selected feature subset, arranging all extracted features into a feature matrix, and then inputting the feature matrix into a 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 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 results including: A model pool including multiple regression models is pre-built, one regression model is optimized for one pulse wave category, and a regression model matching the pulse wave category is selected from the model pool according to the classification result of the pulse wave; 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.

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: Perform sliding window histogram statistics on the blood pressure test results of each pulse wave, and output the final blood pressure test results based on the statistical results, 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 results of each pulse wave in the window are counted to generate a histogram; For all window units in the histogram, the blood pressure interval with the highest frequency is identified, and the mean of the blood pressure detection results in the sliding window is calculated within the blood pressure interval with the highest frequency, and the mean is output as the final blood pressure detection result.

Images