Heartbeat timing state detection method based on radar signal

Abstract

The invention relates to a heartbeat timing state detection method based on radar signals, comprising: obtaining a baseband I / Q orthogonal signal including a heartbeat signal and a breathing signal of a detection object, and synchronously collecting an electrocardiogram signal of the detection object; processing the baseband I / Q orthogonal signal based on a heartbeat mathematical model to obtain three heartbeat signal characteristic waves, wherein the heartbeat mathematical model is used to characterize a radar phase signal including two main harmonics of contraction and diastole and only affected by the heartbeat, and the three heartbeat signal characteristic waves can respectively represent the two main harmonics and the superposition waves of the two main harmonics; taking the three heartbeat signal characteristic waves as inputs, taking the wave crest position of a "Q-R-S-T" wave of the synchronously collected electrocardiogram signal as a label, training a recurrent neural network model, and using the trained network model to perform time series division on the unlabeled radar heartbeat signal characteristic waves, thereby realizing heartbeat timing state detection, and further realizing accurate evaluation of heart health conditions.

Description

Technical Field

[0001] The invention relates to the technical field of non-contact radar detection of vital signs, in particular to a heartbeat timing state detection method based on radar signals. Background Art

[0002] Non-contact radar detection technology is suitable for detecting physical signs of people with sleep disorders, skin contact problems or infants because it does not bring a sense of restraint to the body. The human body sign signals detected by radar mainly include heartbeat signals and breathing signals. Since the chest fluctuation caused by the heart is weaker than the respiratory movement, the separation of the heartbeat signal component is easily interfered by the respiratory signal component.

[0003] In the prior art, patent (CN113440120A) discloses a method for detecting human breathing and heartbeat based on millimeter wave radar, including providing a specific radar signal preprocessing method and a method for reconstructing the waveform of breathing and heartbeat. The calculation process of waveform reconstruction is complicated, and the detection of the heartbeat timing state cannot be realized. Patent (CN113951856A) discloses a non-contact accurate heartbeat detection method based on Doppler radar. The extracted radar signal is separated and extracted from the heart signal by using the Mallett algorithm and the matching algorithm filtering, and then the heartbeat spectrum result is obtained by the FTPR-TWV algorithm. However, no solution is provided for reducing the separation and extraction of the heart signal by respiratory motion, and no specific and effective method for estimating the heart timing state and the change of the heartbeat interval is provided. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a heartbeat timing state detection method based on radar signals, the purpose of which is to achieve accurate separation of radar heartbeat signals and detect the heart timing change state according to the obtained heartbeat signals.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A heartbeat timing state detection method based on radar signals, comprising:

[0007] Acquire the baseband I / Q orthogonal signal including the heartbeat signal and the breathing signal of the detection object, and synchronously collect the electrocardiogram signal of the detection object;

[0008] The baseband I / Q orthogonal signal is processed based on a heartbeat mathematical model to obtain three heartbeat signal characteristic waves, wherein the heartbeat mathematical model is used to characterize a radar phase signal only affected by the heartbeat and including two main harmonics of contraction and diastole, and the three heartbeat signal characteristic waves can respectively represent the two main harmonics and the superposition wave of the two main harmonics;

[0009] Taking the three heartbeat signal characteristic waves as input and the peak position of the "QRST" wave of the synchronously collected electrocardiogram signal as the label, a recurrent neural network learning model is trained. The trained recurrent neural network learning model is used to perform time series division on the unlabeled radar heartbeat signal characteristic waves to realize heartbeat timing state detection.

[0010] Further technical solutions are:

[0011] The heartbeat mathematical model is constructed based on the mutual modulation of the heartbeat harmonics of the Gaussian pulse signal emitted by the radar and the radar phase signal received by the radar that is only affected by the heartbeat. The mathematical expression is:

[0012]

[0013] In the formula, x h (t) is the radar phase signal affected only by heartbeat, A1 and A2 are the amplitudes of the main peaks of the two main harmonics of systole and diastole, respectively. and The frequencies of the two main peaks respectively; and are the phase shifts of the two main peaks respectively; t is time, b and c are constants.

[0014] The baseband I / Q orthogonal signal is processed based on the heartbeat mathematical model to obtain three heartbeat signal characteristic waves, including:

[0015] S1. Preprocessing the baseband I / Q quadrature signal to obtain a phase difference signal with enhanced heartbeat signal component;

[0016] S2. Select the enhanced phase difference signal of the heartbeat signal component within a specific frequency range, and use a bimodal Gaussian model to calculate the Gaussian weighted average value in the sliding window for spectrum fitting, and fit two center frequencies in each sliding window to achieve the fitting of the main peak frequencies of the two main harmonics;

[0017] S3, taking the frequency of a series of harmonics in the fitted spectrum as the center, filtering the enhanced phase difference signal of the heartbeat signal component obtained in S1 to obtain a series of radar phase component signals, and then normalizing the radar phase component signals and calculating the Shannon entropy to obtain a series of Shannon entropy signals;

[0018] S4, filtering the series of Shannon entropy signals to obtain a series of pulse signals with enhanced heartbeat harmonics;

[0019] S5. Clustering the series of heartbeat harmonic enhanced pulse signals using a k-means clustering algorithm, and outputting three cluster centers, which are the three heartbeat signal characteristic waves.

[0020] Taking the frequencies of a series of harmonics in the fitted spectrum as the center, filtering the enhanced phase difference signal of the heartbeat signal component obtained by S1, including:

[0021] The zero-phase bandpass filtering method is used for filtering, and the filtering range is [w c -Δw,w c +Δw],w c is the frequency of the series of harmonics, and Δw is the radius of the filtering range.

[0022] The series of Shannon entropy signals are filtered, including:

[0023] The Shannon entropy signal is smoothed using zero-phase filtering with rectangular pulse window lengths of T1 and T2 and then subtracted. T1 and T2 are empirical values ​​obtained based on one heartbeat cycle and one respiratory cycle, respectively.

[0024] The baseband I / Q quadrature signal is preprocessed to obtain a phase differential signal with enhanced heartbeat signal components, including:

[0025] The baseband I / Q quadrature signals are calculated using an extended differential cross-multiplication algorithm.

[0026] The specific frequency range is far away from the basic frequency of heartbeat.

[0027] The heartbeat timing state detection method based on radar signals also includes: detecting the peak positions of the three heartbeat signal characteristic waves, calculating the peak-to-peak interval, obtaining the heartbeat interval time information, and obtaining an estimated value of the heart rate.

[0028] The beneficial effects of the present invention are as follows:

[0029] The present invention can use the extracted heartbeat signal to predict the changes in heart rate and heart state based on machine learning methods, so as to achieve accurate assessment of heart health status.

[0030] The present invention takes the heartbeat mathematical model as the basis for signal processing and can accurately separate the harmonic components of the radar heartbeat signal through double-peak Gaussian fitting, thereby accurately extracting the heart cycle motion characteristics.

[0031] The heartbeat mathematical model proposed in the present invention accurately describes the regular contraction and relaxation movement information of the heart in the radar signal affected only by the heartbeat, and provides a reference for processing radar signals affected by breathing and heartbeat and separating heartbeat signals. It is also widely applicable to non-contact radar sensors and radar signal types that detect human heartbeat signals through echo reflection and phase modulation.

[0032] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The figure is a flow chart of a method for detecting heartbeat timing state based on radar signals according to an embodiment of the present invention.

[0034] Figure 2 The present invention is a flowchart of processing a baseband I / Q orthogonal signal based on a heartbeat mathematical model to obtain three heartbeat signal characteristic waves.

[0035] Figure 3 Schematic diagram of the principle of detecting human body signs by radar according to an embodiment of the present invention.

[0036] Figure 4 Schematic diagram of three heartbeat signal characteristic waves obtained by an embodiment of the present invention.

[0037] Figure 5 This is a schematic diagram of LSTM model prediction using three heartbeat signal characteristic waves as input according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention is described below with reference to the accompanying drawings.

[0039] See also Figure 1 , a heartbeat timing state detection method based on radar signals in this embodiment includes:

[0040] Acquire the baseband I / Q orthogonal signal including the heartbeat signal and the breathing signal of the detection object, and synchronously collect the electrocardiogram signal of the detection object;

[0041] The baseband I / Q orthogonal signal is processed based on a heartbeat mathematical model to obtain three heartbeat signal characteristic waves, wherein the heartbeat mathematical model is used to characterize a radar phase signal only affected by the heartbeat and including two main harmonics of contraction and diastole, and the three heartbeat signal characteristic waves can respectively represent the two main harmonics and the superposition wave of the two main harmonics;

[0042] Taking the three heartbeat signal characteristic waves as input and the peak position of the "QRST" wave of the synchronously collected electrocardiogram signal as a label, a recurrent neural network learning model is trained, and the unlabeled radar heartbeat signal characteristic waves are divided into time series using the trained recurrent neural network learning model to realize heartbeat time series state detection;

[0043] See also Figure 2 The baseband I / Q orthogonal signal is processed based on the heartbeat mathematical model to obtain three heartbeat signal characteristic waves, including:

[0044] S1. Preprocessing the baseband I / Q quadrature signal to obtain a phase difference signal with enhanced heartbeat signal component;

[0045] S2. Select the enhanced phase difference signal of the heartbeat signal component within a specific frequency range, and use a bimodal Gaussian model to calculate the Gaussian weighted average value in the sliding window for spectrum fitting, and fit two center frequencies in each sliding window to achieve the fitting of the main peak frequencies of the two main harmonics;

[0046] S3, taking the frequency of a series of harmonics in the fitted spectrum as the center, filtering the phase difference signal obtained in S1 with the heartbeat signal component enhanced, to obtain a series of radar phase component signals (which will contain possible heartbeat harmonic signals), and then normalizing the radar phase component signals and calculating the Shannon entropy to obtain a series of Shannon entropy signals;

[0047] S4, filtering the series of Shannon entropy signals to obtain a series of pulse signals with enhanced heartbeat harmonics;

[0048] S5. Clustering the series of heartbeat harmonic enhanced pulse signals using a k-means clustering algorithm, and outputting three clustering centers, which are the three heartbeat signal characteristic waves.

[0049] In the above embodiment, the heartbeat mathematical model is constructed based on the mutual modulation of the heartbeat harmonics of the Gaussian pulse signal emitted by the radar and the radar phase signal received by the radar which is only affected by the heartbeat. The mathematical expression is:

[0050]

[0051] In the above formula, t is time, x is h (t) is the radar phase signal affected only by the heartbeat, f0 is the heart rate; k is the heartbeat harmonic number of the radar phase signal affected only by the heartbeat; K min , K max are the minimum and maximum harmonic levels in the frequency range of 3-40 Hz, a k and θ k are the amplitude and phase shift of the heartbeat harmonic components, respectively, and b and c are constants.

[0052] According to Shannon entropy calculation and analysis, the heartbeat harmonic components contain two main peaks, contraction and diastole. The above formula can be further simplified as follows:

[0053]

[0054] In the above formula, A1 and A2 are the amplitudes of the main peaks of the two main harmonics of contraction and relaxation, respectively. and The frequencies of the two main peaks are and are the phase shifts of the two main peaks, respectively.

[0055] The heartbeat mathematical model of this embodiment is obtained by analyzing the radar phase signal that is only affected by the heartbeat motion under ideal conditions. By describing the frequency, phase shift and amplitude of the two main peaks, it can accurately reflect the regular contraction and relaxation motion of the atrium and ventricle in one heart cycle. The heartbeat mathematical model can be widely used to describe the signal obtained by the non-contact radar sensor that detects the human heartbeat signal through echo reflection and phase modulation.

[0056] In the case of Figure 3 In the actual situation shown, when the radar phase signal affected by both heartbeat and respiratory movements is processed, after separating the respiratory signal, according to the description of the mathematical model of the heartbeat, the remaining signal is purposefully spectrally fitted through step S2, and it is fitted toward the form of two harmonic main peaks per cycle, so that the two main peak frequencies corresponding to the mathematical model can be accurately screened out later. The phase difference signal enhanced by the heartbeat signal component obtained in step S1 is filtered with the center frequency obtained by fitting, and the interference signals such as respiratory movement are basically removed through normalization and Shannon entropy calculation in step S3, and the heartbeat signal is separated more accurately. In order to obtain a complete heartbeat rhythm feature that can be used as input data for the machine learning model, the signal is processed into a pulse form through step S4, and these pulse signals are clustered through step S5. Those skilled in the art can understand that the pulse signals with the same frequency characteristics are divided into one category. After the processing of step S2, the signals of the pulse signals with frequencies close to the main peaks of the two main harmonics will be clustered around the main peak frequencies, and the remaining signals will be clustered around the main frequencies of the two main harmonic superposition signals, that is, the clustering algorithm can classify all pulse signals into three categories according to the frequency and peak position of the waveform, and then obtain three characteristic waves of the heartbeat signal, such as Figure 4 shown.

[0057] Since the human body's ECG signal exhibits the characteristics of a "QRST" wave, by comparing the three heartbeat signal characteristic waves obtained with the synchronously collected ECG signal, it can be found that the peak positions of the three heartbeat signal characteristic waves basically correspond to the positions of the "R peak", "T peak", and "T wave end" in the "QRST" wave, indicating that the heartbeat signal obtained by the processing of this embodiment is relatively accurate.

[0058] As described in the above embodiment, three heartbeat signal characteristic waves are used as input, and the peak positions of the "QRST" waves of the synchronously collected electrocardiogram signals (i.e., "R peak", "T peak", and "T wave end") are used as labels to train a recurrent neural network learning model. The unlabeled radar heartbeat signal characteristic waves are input into the trained recurrent neural network learning model, and labeled (with peak positions marked and diastole and contraction cycles recorded) heartbeat characteristic waves are output, thereby obtaining the time series division of the unlabeled radar heartbeat signal characteristic waves, thereby realizing heartbeat timing state detection.

[0059] The above embodiment is a heartbeat timing state detection method based on radar signals, which can extract heartbeat interval changes by unsupervised learning, including: detecting the peak positions of the three heartbeat signal characteristic waves, calculating the peak-to-peak interval, obtaining heartbeat interval time information, and obtaining an estimated value of the heart rate.

[0060] Specifically, the peak position of each heartbeat is calculated and the interval between two heartbeat peaks is obtained by difference, and the heart rate is estimated based on the heartbeat interval.

[0061] In the above embodiment, step S1 preferably uses an extended differential cross-multiplication algorithm to pre-process the baseband I / Q orthogonal signal to obtain a phase difference signal with enhanced heartbeat signal components and containing heartbeat and respiration information.

[0062] Generally speaking, chest displacement caused by heartbeat and respiratory motion is proportional to radar phase change. Although the phase signal can be obtained by calculating the I / Q quadrature signal through the inverse tangent, calculating the phase difference signal through the extended differential cross-multiplication algorithm can further highlight the details of the heartbeat motion change and reduce the interference of the respiratory signal.

[0063] In the above embodiment, the specific frequency range described in step S2 is far away from the basic frequency of the heartbeat.

[0064] Since the heartbeat fundamental frequency (1Hz-2.5Hz) is easily interfered by the fundamental frequency of the respiratory signal and its harmonics, it cannot be separated accurately. Therefore, the phase difference signal in the specific frequency range of 3Hz-40Hz is selected for unilateral spectrum analysis. The bimodal Gaussian model is used to fit the spectrum signal in the sliding window. The window length of the sliding window is preferably 5Hz, and the step length is 2Hz.

[0065] In the above embodiment, in step S3, the frequency of a series of harmonics in the fitted spectrum is taken as the center, and a zero-phase bandpass filtering method is specifically used to filter the phase difference signal of the heartbeat signal component enhanced obtained in step S1, and the filtering range is [w c -Δw,w c +Δw],w c is the frequency of the series of harmonics, and Δw is the radius of the filtering range.

[0066] Because the harmonic frequency is generally an integer multiple of the fundamental frequency, the filter range radius Δw of this embodiment has a value range of approximately one heartbeat fundamental frequency, and according to the empirical method, the value can be 1.2 Hz and then fine-tuned.

[0067] In the above embodiment, in step S4, the series of Shannon entropy signals are filtered, including: using zero-phase filtering with rectangular pulse window lengths of T1 and T2 to smooth the Shannon entropy signals and then subtracting them, where T1 and T2 are empirical values ​​obtained based on a heartbeat cycle and a respiratory cycle, respectively.

[0068] In this embodiment, T1 and T2 can be specifically set to 0.4s and 1.5s respectively.

[0069] In the above embodiment, the cluster distance in step S5 is the Pearson correlation coefficient.

[0070] In the above embodiment, the recurrent neural network may specifically adopt an LSTM model.

[0071] like Figure 5 As shown in the figure, it is a schematic diagram of using LSTM model for training. The three heartbeat signal waves obtained by clustering are used as input features, and the peak position of the synchronously collected ECG signal is used as the label. The label can represent the diastole and contraction of the heart cycle. 1，tn 、x 2，tn 、x 3，tn They respectively represent the signal values ​​of the three heartbeat signal waves corresponding to time node n.

[0072] When training the recurrent neural network learning model, the learning rate was initially set to 0.01 and updated every two epochs. All models had 30 training iterations. The tolerance window length can be set according to the signal sampling frequency. For the same subject, the ratio of training set to test set is 7:3. If the amount of training data collected for a subject is small, fine-tuning training can be performed based on the model of other subjects, and then used for subsequent cardiac cycle state detection.

[0073] The present invention can use the extracted heartbeat signal to predict the changes in heart rate and heart state based on a machine learning method, and can achieve an accurate assessment of heart health status.

[0074] Those skilled in the art can understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A heartbeat timing state detection method based on radar signals, characterized in that: include: Acquire the baseband I / Q orthogonal signal including the heartbeat signal and the breathing signal of the detection object, and synchronously collect the electrocardiogram signal of the detection object; The baseband I / Q orthogonal signal is processed based on a heartbeat mathematical model to obtain three heartbeat signal characteristic waves, wherein the heartbeat mathematical model is used to characterize a radar phase signal only affected by the heartbeat and including two main harmonics of contraction and diastole, and the three heartbeat signal characteristic waves can respectively represent the two main harmonics and the superposition wave of the two main harmonics; Taking the three heartbeat signal characteristic waves as input and the peak position of the "QRST" wave of the synchronously collected electrocardiogram signal as a label, a recurrent neural network learning model is trained, and the unlabeled radar heartbeat signal characteristic waves are divided into time series using the trained recurrent neural network learning model to realize heartbeat time series state detection; The heartbeat mathematical model is constructed based on the mutual modulation of the heartbeat harmonics of the Gaussian pulse signal emitted by the radar and the radar phase signal received by the radar that is only affected by the heartbeat. The mathematical expression is: In the formula, x h (t) is the radar phase signal affected only by heartbeat, A1 and A2 are the amplitudes of the main peaks of the two main harmonics of systole and diastole, respectively. and The frequencies of the two main peaks respectively; and are the phase shifts of the two main peaks, respectively; t is time, b and c are constants; The baseband I / Q orthogonal signal is processed based on the heartbeat mathematical model to obtain three heartbeat signal characteristic waves, including: S1. Preprocessing the baseband I / Q quadrature signal to obtain a phase difference signal with enhanced heartbeat signal component; S2. Select the enhanced phase difference signal of the heartbeat signal component within a specific frequency range, and use a bimodal Gaussian model to calculate the Gaussian weighted average value in the sliding window for spectrum fitting, and fit two center frequencies in each sliding window to achieve the fitting of the main peak frequencies of the two main harmonics; S3, taking the frequency of a series of harmonics in the fitted spectrum as the center, filtering the enhanced phase difference signal of the heartbeat signal component obtained in S1 to obtain a series of radar phase component signals, and then normalizing the radar phase component signals and calculating the Shannon entropy to obtain a series of Shannon entropy signals; S4, filtering the series of Shannon entropy signals to obtain a series of pulse signals with enhanced heartbeat harmonics; S5. Clustering the series of heartbeat harmonic enhanced pulse signals using a k-means clustering algorithm, and outputting three clustering centers, which are the three heartbeat signal characteristic waves.

2. The heartbeat timing state detection method based on radar signal according to claim 1, characterized in that: Taking the frequencies of a series of harmonics in the fitted spectrum as the center, filtering the enhanced phase difference signal of the heartbeat signal component obtained by S1, including: The zero-phase bandpass filtering method is used for filtering, and the filtering range is [w c -Δw,w c +Δw],w c is the frequency of the series of harmonics, and Δw is the radius of the filtering range.

3. The heartbeat timing state detection method based on radar signal according to claim 1, characterized in that: The series of Shannon entropy signals are filtered, including: The Shannon entropy signal is smoothed using zero-phase filtering with rectangular pulse window lengths of T1 and T2 and then subtracted. T1 and T2 are empirical values ​​obtained based on one heartbeat cycle and one respiratory cycle, respectively.

4. The heartbeat timing state detection method based on radar signal according to claim 1, characterized in that: The baseband I / Q quadrature signal is preprocessed to obtain a phase differential signal with enhanced heartbeat signal components, including: The baseband I / Q quadrature signals are calculated using an extended differential cross-multiplication algorithm.

5. The heartbeat timing state detection method based on radar signal according to claim 1, characterized in that: The specific frequency range is far away from the basic frequency of heartbeat.

6. The heartbeat timing state detection method based on radar signal according to any one of claims 1 to 5, characterized in that: Also includes: The peak positions of the three heartbeat signal characteristic waves are detected, the peak-to-peak intervals are calculated, the heartbeat interval time information is obtained, and the estimated value of the heart rate is obtained.

Images