Human respiration and heartbeat signal estimation method based on multi-channel millimeter wave radar

Abstract

A human body breathing and heartbeat signal estimation method based on a multi-channel frequency modulation continuous wave millimeter wave radar (FMCW) is characterized by comprising the following steps of (1) obtaining an intermediate frequency signal of a multi-channel frequency modulation continuous wave millimeter wave radar (FMCW) system, (2) carrying out distance dimension FFT and phase extraction and unwrapping on a Chirp intermediate frequency signal channel by channel, and (3) carrying out phase extraction and unwrapping on the Chirp intermediate frequency signal channel by channel. (3) calculating signal energy of interested breathing and heartbeat frequency bands in the phase time sequence under all distances, and weighting amplitudes of all distance dimension FFT under all distances to realize accurate positioning of a human body target, and (4) regarding the phase time sequence under the accurate distance and the adjacent distance of each channel as potential multipath signals, and performing fusion enhancement on the signals to realize accurate positioning of the human body target. And (5) processing the enhanced phase time sequences of the channels on the basis of multivariate variational mode decomposition, separating and estimating respiratory signals and heartbeat signals of the human body, and calculating an average respiratory rate and an average heart rate in an observation time period.

Description

Technical Field

[0001] This invention relates to the field of non-contact vital sign monitoring technology, and in particular to a method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW).

[0002] The high incidence of cardiovascular diseases and the limitations of traditional contact-based monitoring technologies (such as electrocardiograms and pulse oximeters) (requiring contact wearing, poor comfort, and difficulty in long-term continuous use) have spurred an urgent need for non-contact vital sign monitoring technologies. Among numerous sensing technologies, frequency-modulated continuous wave millimeter-wave radar, due to its high range resolution, certain penetration, and good protection of visual privacy, is considered one of the ideal technologies for achieving high-precision non-contact extraction of human respiratory and heartbeat signals. However, it still faces significant challenges:

[0003] 1. The chest cavity micro-movements caused by heartbeats are extremely weak (submillimeter level), and the resulting radar phase change signals have small amplitudes, making them easily drowned out by environmental noise, system hardware noise, and unconscious random body movements of the subjects.

[0004] 2. The signal amplitude generated by respiratory movements is usually 1-2 orders of magnitude stronger than that generated by heartbeats. Its rich harmonic components often fall into the fundamental frequency or even harmonic range of the heartbeat signal, resulting in severe spectral aliasing and interfering with the accurate estimation of heart rate.

[0005] 3. The performance of single-channel radar echo signal processing methods is heavily dependent on the signal quality of a single channel. Once the channel fails due to multipath fading, target deviation from the beam center, or excessively low signal-to-noise ratio, the entire monitoring process becomes unsustainable.

[0006] To improve the accuracy and anti-interference capability of signal separation, the research focus has gradually shifted from single-channel to multi-channel signal fusion, aiming to enhance system robustness and measurement accuracy through spatial diversity gain. Background Technology

[0007] The mathematical principles of linear frequency modulated continuous wave (FMCW) radar. The radar transmitter signal can be expressed as:

[0008]

[0009] Where f c Let B be the carrier frequency of the transmitted signal, B be the bandwidth of the transmitted signal, and T be the frequency sweep time of the transmitted signal. This is the frequency modulation slope. The echo signal obtained after reflection from the target is expressed as:

[0010]

[0011] Where t dThis represents the time delay of the echo signal. The transmitted and received signals are mixed in a mixer to obtain an intermediate frequency (IF) signal. Mathematically, this is equivalent to multiplying the conjugate of the received and transmitted signals and then performing a low-pass filter (LP). Its expression is:

[0012]

[0013] Where c is the speed of light and R is the target distance. The frequency f is based on the intermediate frequency signal. b and phase φ b Then the target distance R can be calculated:

[0014]

[0015] In practice, due to the winding problem in phase estimation, it is generally necessary to first estimate the frequency f. b To initially calculate the distance R, then use the phase φ b The study investigates subtle changes in distance R, such as the subtle changes in distance caused by human respiration and heartbeat.

[0016] When FMCW radar detects the human chest cavity, the slight movements of the chest cavity caused by breathing and heartbeat will cause the echo distance R(t) to change over time:

[0017] R(t)=R0+ΔR(t)≈R0+r b (t)+r h (t)

[0018] Where R0 is the static distance, r b (t) represents the distance change caused by respiration, r h Δφ(t) represents the distance change caused by the heartbeat, and ΔR(t) represents the total distance change. At this point, the relationship between the echo phase change Δφ(t) and ΔR(t) is:

[0019]

[0020] By estimating the phase change Δφ(t), a micro-motion signal containing information about respiration and heartbeat can be obtained:

[0021]

[0022] Multi-channel radar systems typically employ Multiple-Input Multiple-Output (MIMO) technology, which expands the virtual aperture of the antenna array to obtain rich multi-channel signals. For example, a typical MIMO architecture uses three transmit antennas (TX) and four receive antennas (RX), forming twelve virtual channels through virtual element expansion and TDMA (Time Division Multiple Access) strategy. The formation of the MIMO virtual array is based on the coordinated operation of the transmit and receive arrays: assuming the receive antenna array consists of four antennas spaced by d, with adjacent receive antennas generating a phase shift of ω; when multiple transmit antennas (such as TX1, TX2, TX3) are arranged with a spacing of 4d, the signals generated from different transmit antennas introduce additional phase shifts at the receiver. By concatenating the received phase shift sequences corresponding to each transmit antenna, an expanded array consisting of 12 virtual elements can be formed, equivalent to obtaining 12 channels of radar signal.

[0023] To separate signals from different transmit antennas at the receiver, MIMO radar needs to ensure that the signals from each transmit antenna are orthogonal and do not interfere with each other. The TDMA (Time Division Multiplexing) mechanism achieves channel separation through time separation: each transmit antenna transmits signals in turn, with only one transmit antenna active at any given moment, while all receive antennas simultaneously receive echo signals. The TDMA framework divides time into multiple equal-length time slots, allocating a dedicated time slot to each transmit antenna. Furthermore, the chirp signal parameters (such as start frequency, bandwidth, and duration) within each time slot must be strictly consistent to ensure that signals from different transmit antennas can be processed uniformly at the receiver.

[0024] In existing technologies, to improve signal separation accuracy and anti-interference capabilities, many studies have attempted to utilize multi-channel information fusion and enhance system robustness through spatial diversity. For example, Wu et al. proposed a multi-channel Kalman smoother (MCKS) that achieved respiratory rate and heart rate estimation errors of less than 2 BPM under random body motion environments; Chen et al. used the expectation-maximization method with adaptive parameter selection (APSEM) to achieve a mean absolute error of less than 2.5 BPM in heart rate detection; Zhang et al., based on a multi-channel cross-correlation algorithm, were able to accurately estimate heart rate even under low signal-to-noise ratio conditions; Yu et al. used MVDR beamforming technology to achieve a mean respiratory rate error of less than 1.5 BPM within a 2.5-meter range; and Qu et al. combined maximum ratio combining and multivariate variational mode decomposition (MVMD) to achieve better respiratory and heart rate frequency estimation accuracy than single-channel data on multi-channel data. Although the above methods have achieved certain results in the utilization of multi-channel information, they still face challenges in terms of signal separation accuracy, environmental adaptability, and universality of individual differences in complex environments. In particular, there is still room for improvement in performance when dealing with scenarios such as rapid heart rate changes and complex body movement interference. Summary of the Invention

[0025] Purpose of the invention

[0026] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for accurate separation and estimation of heartbeat and respiratory signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), which can effectively resist noise interference and has good robustness and estimation accuracy.

[0027] Technical solution

[0028] A method for estimating human respiratory and heartbeat signals based on multi-channel frequency-modulated continuous wave millimeter-wave radar (FMCW) is characterized by the following steps: (1) real-time acquisition or reading of the intermediate frequency (IF) signals of the acquired multi-channel FMCW system; (2) preprocessing the acquired FMCW IF signals, performing range-dimensional FFT on the IF signals channel by channel and chirp by chirp, and performing phase extraction and unwinding; (3) calculating the signal energy of the respiratory and heartbeat frequency bands of interest in all range-based phase time series, and estimating the respiratory and heartbeat frequency bands of interest in a range-based phase time series. The proportion of the energy of the heartbeat frequency band in the total energy of the frequency band of interest in the phase time series at all distances is determined as the weight at that distance. Based on this, the amplitudes at each distance of the FFT in all distance dimensions are weighted to achieve accurate positioning of the human target. (4) The phase time series at each channel's precise distance and its neighboring distances are regarded as potential multipath signals and are fused and enhanced. (5) The enhanced phase time series of each channel are processed based on multivariate variational mode decomposition to separate and estimate the human respiratory signal and heartbeat signal, and the average respiratory rate and average heart rate during the observation period are calculated. See Figure 1 As shown.

[0029] According to the above-described method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), step (1) involves real-time acquisition or reading of the acquired intermediate frequency (IF) signal from the multi-channel FMCW system. The characteristic of this method is that the multi-channel IF signal originates from the configuration of the FMCW radar hardware system. Taking the TIIWR1843 millimeter-wave radar evaluation board hardware platform as an example, it is configured with 3 transmit antennas (TX) and 4 receive antennas (RX). Each transmit antenna is activated sequentially through the TDMA (Time Division Multiple Access) working mode. During each activation period, all four receiving antennas synchronously receive the echoes, thus forming an equivalent linear array of 12 virtual array elements in space, corresponding to 12 channels of intermediate frequency (IF) signals. Each channel's IF signal consists of frames of data, with two parameters—frame count and frame rate—that need to be configured in advance. For physiological monitoring, a frame rate of tens of hertz is generally sufficient. Each frame contains the IF signal of one or more chirp signals configured accordingly. It also has two parameters that need to be configured in advance: the number of samples and the sampling rate. The sampling rate is generally configured to be above megahertz. In the subsequent experiment, 512 points were sampled for each chirp IF signal, and 1500 frames were continuously acquired at a frame rate of 25Hz. The constructed original data cube had dimensions of 512 ADC sampling points × 1 chirp / frame × 1500 frames × 12 channels, completely recording the temporal echo of the detection scene.

[0030] According to the above-described method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), step (2) preprocesses the obtained multi-channel FMCW intermediate frequency signal, performs range-dimensional FFT on the intermediate frequency signal channel by channel and chirp by chirp, and performs phase extraction and unwinding. The method is characterized by: (a) Range-dimensional FFT: For each frame of ADC data, along the fast time dimension, i.e., for the number of sampling points of each chirp intermediate frequency signal, such as 512, a fast Fourier transform (FFT) is performed to convert the time domain signal to the frequency domain. Each frequency point corresponds to a distance, and the horizontal axis is called a range gate (RangeBin). After each channel is processed by chirp, a "range-time topographic map" (Range-time topographic map) is generated. (a) TimeMap), where one axis is the distance gate index and the other axis is the frame number (time series). The value of each pixel in the topographic map represents the signal strength of the distance unit at the corresponding time. (b) Phase extraction and unwrapping: For each distance unit in each channel, the phase sequence φ(t) is extracted from its complex form distance-time map. Since the original phase value is wrapped in the interval [-π,π], there is a 2π jump. Phase unwrapping is required to obtain continuous phase changes. An unwrapping algorithm based on first-order differential accumulation is adopted. By detecting the difference between adjacent phase points and correcting the jumps exceeding π by integer multiples of 2π, a smooth and continuous phase change sequence Δφ(t) is finally obtained. This sequence directly corresponds to the thoracic micro-displacement caused by breathing and heartbeat.

[0031] According to the above-described method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), step (3) calculates the signal energy of the respiratory and heartbeat frequency band of interest in the phase time series at all distances. The proportion of the energy of the respiratory and heartbeat frequency band of interest in the phase time series at a certain distance to the total energy of the frequency band of interest in the phase time series at all distances is determined as the weight at that distance. Based on this, the amplitude of each distance in all distance-dimensional FFT is weighted to achieve accurate positioning of human targets. The method is characterized by: (a) constructing a radar energy matrix: calculating the square of the signal amplitude of each pixel in the distance-time graph |R(r,t)| 2(a) Obtain the original radar energy matrix Ra, which reflects the original energy distribution of each scatterer in the environment; (b) Calculate the physiological signal energy weights: For the phase sequence Δφ(t) of each range unit, the Welch method is used to estimate its power spectral density (PSD); Set the respiratory signal frequency band to [0.1, 0.5] Hz and the heartbeat signal frequency band to [0.8, 2.0] Hz, calculate the power integrals Eb(r) and Eh(r) of each range unit in these two characteristic frequency bands respectively, and add them together to obtain the total physiological energy Etotal(r) of the unit; Subsequently, normalize the Etotal(r) of the unit by dividing it by the sum of Etotal(r) of all range units to obtain the physiological signal energy weights Rs. This matrix is ​​used in areas rich in vital signs (such as the human body). (c) Weighted fusion: The original radar energy matrix Ra is weighted according to Rs under each range cell to generate a weighted energy matrix W. This operation effectively suppresses the interference of static clutter and highlights the echo characteristics of living organisms. (d) Target localization: The matrix W is averaged along the time dimension to obtain a one-dimensional average weighted energy distribution curve. The global maximum value of the curve is searched within the preset effective detection range (e.g., 0.5m to 2.0m). The corresponding range cell is used as the candidate human target position r for that channel. (e) Final target position determination: All position candidates {r_1, r_2, ..., r_12} for 12 channels are collected, sorted, and the median is taken as the final human target position r_target.

[0032] According to the above-described method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), in step (4), the phase time series of each channel's precise distance and its neighboring distances are regarded as potential multipath signals, and they are fused and enhanced. The method is characterized by: (a) multipath signal extraction: for each channel signal, taking the target distance unit located in step (3) as the center, extracting its phase time series and its adjacent ±2 distance units (a total of 5 units), forming an initial multipath micro-motion signal set; (b) dual-band hierarchical clustering: processing the filtered signals of the respiratory frequency band (0.2-0.5Hz) and the heart rate frequency band (0.5-2.5Hz) respectively; for each micro-motion signal, calculating its normalized power spectral density (PSD) vector in the target frequency band as its feature, and then using a cohesive hierarchical clustering algorithm, with the cosine distance between PSD vectors as the similarity measure, clustering signals from different distances. The signals of the units are clustered; during the clustering process, spatial adjacency constraints are introduced, that is, spatially adjacent units are merged first to ensure that the signals in the merged clusters originate from the same scattering center, thereby improving the physical rationality of the clustering; (c) Optimal cluster selection and signal fusion: in each frequency band, the average signal-to-noise ratio (SNR) of each cluster is calculated, and the cluster with the highest SNR is selected as the target cluster (Cbreath and Cheart) of that frequency band. The signal-to-noise ratio is defined as the ratio of the peak power in the target frequency band to the average power of the noise outside the frequency band. The final set of high-quality signal units is obtained by taking the intersection of the two target clusters, Coptimal = Cbreath ∩ Cheart. The signals in this set perform well in both the respiratory and heart rate frequency bands; (d) The phases of all signals in the set are weighted and summed, and the weights are determined by their respective SNRs, so that each channel outputs a fused and enhanced high signal-to-noise ratio phase signal Sfused(t).

[0033] According to the above-described method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), in step (5), the enhanced phase time series of each channel are processed based on multivariate variational mode decomposition to separate and estimate human respiratory and heartbeat signals, and the average respiratory rate and average heart rate during the observation period are calculated. The method is characterized by: (a) multivariate variational mode decomposition: for each of the 12 channels, the above steps are used to process 12 enhanced signals Sfused(t) as multivariate inputs, which are then fed into the multivariate variational mode decomposition algorithm; MVMD solves a constrained variational problem to adaptively decompose the input signal into K (e.g., K=5) quasi-orthogonal modal components Uk(t). Its core idea is to force the same mode to share a center frequency ωk in different channels, thereby ensuring the consistency of the decomposed modes in a physical sense; (b) respiratory harmonic identification and elimination: in the... Among the decomposed modes, the mode with the strongest energy and center frequency located in the respiratory fundamental frequency range (0.2-0.5Hz) is identified and marked as the respiratory fundamental frequency mode Uresp. Subsequently, all other modes are traversed. If the center frequency fi of a certain mode is approximately equal to an integer multiple of the respiratory fundamental frequency fresp (fi≈n·fresp, n=2,3,...), then the mode is determined to be a respiratory harmonic and is removed from the candidate set. (c) Heartbeat signal reconstruction and estimation: From the remaining modes after removing respiratory harmonics, modes with consistent energy distribution on all 12 channels (small energy variance between channels) and center frequencies falling within the normal heart rate range (0.8-2.0Hz) are selected. These qualified modes are weighted and averaged on each channel to reconstruct the final pure heartbeat signal sheart(t). The heartbeat signal is subjected to spectral analysis. By detecting the main peak value in its power spectrum, the instantaneous heart rate value can be accurately estimated.

[0034] Beneficial effects:

[0035] Experiment. This invention uses the TI IWR1843 Boost 77GHz millimeter-wave radar as the core sensor. Four healthy subjects (four males, aged 22-25 years) were recruited for the experiment, and data acquisition was conducted in a standard laboratory environment. Each subject underwent static measurements for one minute at distances of 0.7m and 1.5m, while ECG signals were simultaneously acquired as a reference. The radar hardware configuration consisted of three transmit antennas (TX) and four receive antennas (RX), forming 12 virtual channels through a TDMA (Time Division Multiple Access) mechanism. Table 1 shows the millimeter-wave radar system configuration.

[0036] Table 1. Millimeter-wave radar system configuration

[0037]

[0038]

[0039] The experiment was conducted in a laboratory measuring 5m × 4m × 3m. The IWR1843 radar was mounted on a tripod at a height of 1.2 meters, level with the subject's chest. Subjects remained seated, breathing naturally, and avoiding significant body movements. Measurements were repeated 6 times under each configuration condition, resulting in 96 sets of valid data (4 subjects × 2 distances × 2 bandwidths × 6 replicates).

[0040] The experimental results are shown in Table 2, which compares the average heart rate detection error under different distances and resolutions. The results show that the average heart rate detection error of the method described in this invention remains at a low level under different conditions, with a comprehensive average of approximately 1.19 BPM, significantly better than the traditional single-channel method. This method performs excellently at two typical detection distances of 0.7 meters and 1.5 meters, and under different resolutions. It also surpasses the estimation accuracy reported in the literature.

[0041] Table 2: Comparison of average heart rate detection error at different distances and resolutions (unit: BPM)

[0042]

[0043] The beneficial effects shown in the experiment can be summarized as follows: (1) Accurate positioning: By weighted positioning by fusing radar energy and physiological signal frequency domain energy, static clutter interference is effectively suppressed, and the accuracy of human target identification in complex environments is improved; (2) High signal-to-noise ratio: By using multi-channel spatial diversity gain and dual-band joint clustering fusion, the signal-to-noise ratio of heart rate and respiratory signals is significantly improved; (3) Good separation effect: By adopting the MVMD algorithm and combining respiratory harmonic identification and elimination mechanism, the problem of spectral aliasing between respiratory harmonics and heart rate signals can be effectively solved, and accurate extraction of heart rate signals can be achieved; (4) Strong robustness: The entire method framework does not rely on a single channel. Through deep fusion and decision-making of multi-channel information, the stability and adaptability of the system under complex environments and individual differences are improved. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the principle of MIMO virtual array formation.

[0046] Figure 3 This is a timing diagram for the TDMA time division multiplexing mechanism.

[0047] Figure 4 Generating a diagram for TDMA time division multiplexing mechanism

[0048] Figure 5This is a comparison chart showing the performance of human body localization algorithms based on physiological signal energy weighting.

[0049] Figure 6 This is a flowchart illustrating the multipath signal fusion enhancement method.

[0050] Figure 7 This is a comparison chart showing the effects of multipath signal fusion enhancement methods.

[0051] Figure 8 Comparison of heart rate detection error distribution at different distances and resolutions in experimental results Detailed Implementation

[0052] Example. The specific implementation steps of the method are as follows:

[0053] (1) Signal acquisition

[0054] This embodiment uses the TI IWR1843 Boost 77GHz millimeter-wave radar evaluation board as the core sensor. The radar system transmits a linear frequency modulated continuous wave (Chirp) signal according to preset parameters and receives the echo reflected from the target. After mixing and analog-to-digital conversion (ADC), the raw intermediate frequency signal data is obtained.

[0055] To achieve multi-channel measurement, the hardware configuration consists of 3 transmit antennas (TX) and 4 receive antennas (RX). Using TDMA (Time Division Multiple Access) mode, the 3 transmit antennas are activated sequentially in turn, and during the activation period of each transmit antenna, all 4 receive antennas synchronously receive the echo signal. This mechanism spatially forms an equivalent linear array composed of 12 virtual array elements, thus corresponding to 12 channels of intermediate frequency (IF) signals.

[0056] The acquired raw data forms a data cube with dimensions of 512 ADC sampling points × 1 Chirp / frame × 1500 frames × 12 channels. Specific parameter configurations are as follows:

[0057] Frame rate: Configured to 25Hz, which means acquiring 25 frames of data per second, meeting the Nyquist sampling requirements for physiological signal monitoring.

[0058] Total frames: 1500 frames were continuously captured, with a recording duration of 60 seconds.

[0059] Chirp count: Each frame contains 1 chirp signal.

[0060] Number of ADC sampling points: 512 sampling points for the intermediate frequency signal of each Chirp.

[0061] Sampling rate: 10,000 ksps or 3,850 ksps depending on the bandwidth configuration to ensure accurate capture of distance-dimensional signals.

[0062] This data cube fully records the temporal echo information of each target in the detection scene, which is the basis for all subsequent processing.

[0063] (2) Signal preprocessing

[0064] Preprocessing operations were performed on each of the 12 channels, including distance-dimensional FFT, phase extraction, and phase unwinding.

[0065] (a) Range-Dimensional FFT Processing: For each chirp's ADC data, a Fast Fourier Transform (FFT) is performed along the fast time dimension to convert the time-domain signal into a range-frequency domain signal. The formula for the range-dimensional FFT is:

[0066]

[0067] Where x(n) is the ADC sampling sequence, N = 512 is the number of FFT points, and R(r) is the range frequency domain output. Each of the 1500 frames is processed in this way to form a range-time map: R, which has a size of 1500 (time dimension) * 512 (range dimension), where each value in the range dimension corresponds to a range unit.

[0068] Phase extraction: Extract the phase of the complex signal for each distance unit in each channel. The phase signal φ(t) of the complex signal z(t) is calculated using arctangent operation:

[0069]

[0070] Where Re(z(t)) and Im(z(t)) are the real and imaginary parts of the signal, respectively.

[0071] (b) Phase unwrapping: Since the original phase is limited to the interval [-π, π], there are abrupt transition points. An unwrapping algorithm is needed to recover the continuous phase change sequence Δφ(t). This embodiment uses an unwrapping algorithm based on first-order difference accumulation:

[0072]

[0073] Where cumsum represents cumulative summation and round is the rounding function. The unwrapped phase sequence Δφ(t) directly reflects the actual displacement changes caused by chest wall micromotion, serving as the basic input for subsequent physiological signal analysis.

[0074] (3) Human target localization based on physiological signal energy weighting

[0075] This step aims to accurately locate human chest targets by combining radar energy information and physiological signal frequency domain energy for weighted fusion.

[0076] (a) Construction of the original radar energy matrix: Calculate the amplitude values ​​of the range-dimensional FFT result obtained in step S1 to obtain the original radar energy matrix R. a For each range cell r and time frame t, the energy value is:

[0077] R a (r, t) = |R(r, t)|

[0078] Where R(r, t) is the complex value at position (r, t) in the distance-time plot.

[0079] (b) Calculation of the physiological signal energy weight matrix: For the phase signal Δφ(t) of each distance unit, the power spectral density (PSD) is calculated using the sliding window Welch method. The Welch method segments the signal (in this embodiment, the window length is 8 seconds and the overlap rate is 50%). Specifically, the sliding window length is set to 8 seconds and the overlap rate is 50%. After applying a Hanning window to each signal segment, the periodogram is calculated, and then the PSD is obtained by averaging. The expression for the Hanning window function is as follows:

[0080]

[0081] Where L is the window length. The respiratory frequency band is defined as [0.1, 0.5] Hz, and the heart rate frequency band as [0.8, 2.0] Hz. The integrated energy of each distance cell within these two frequency bands is calculated:

[0082]

[0083] Where P r (f) represents the PSD of the distance unit r. The total physiological energy is:

[0084] E total =E breath +E heart

[0085] The final physiological signal energy weight Rs is obtained after normalizing the total energy.

[0086]

[0087] (c) Weighted fusion: Multiply the original radar energy matrix Ra by the weights Rs to obtain the weighted energy matrix W:

[0088] W(r,t)=Ra(r,t)·Rs(r)

[0089] (d) Target localization: Average matrix W along the time dimension to obtain the average energy distribution.

[0090]

[0091] Where T is the total number of frames. The global maximum value is searched within the effective detection range (0.5m to 2m), and the corresponding distance cell is the location of the human target. This corresponding distance cell is used as the candidate human target location r for that channel.

[0092]

[0093] (e) Final target location determination: Apply the above process to all 12 channels to obtain candidate locations {r1, r2, ..., r...} 12 After sorting them, the median is taken as the final target human body position r. target :

[0094] r target =median{r1, r2, ..., r 12}

[0095] This method significantly enhances the response of regions containing vital signs through physiological energy weighting, and can effectively distinguish between people and other stationary objects in the environment.

[0096] (4) Multipath signal fusion enhancement

[0097] To improve the signal-to-noise ratio and robustness, this step performs fusion processing on the multi-channel signals based on power spectral density clustering.

[0098] (a) Multipath signal extraction: After steps S1 and S2, the target position is determined, and the phase sequence of each range cell is obtained. The phase signals Δφ of multiple range cells near the target (e.g., the main range cell and its adjacent ±2 cells) are extracted. i (t), where i = 1, 2, ..., 5 represents the distance cell index.

[0099] (b) Dual-band processing and feature extraction: Filtered signals in the respiratory band (0.2-0.5 Hz) and heart rate band (0.5-2.5 Hz) were processed separately. For each signal in each band, its normalized power spectral density (PSD) was calculated as a feature vector. The PSD was estimated using the Welch method and normalized to the unity norm.

[0100]

[0101] Where p i Let be the PSD vector of the i-th signal in the target frequency band.

[0102] Hierarchical clustering and similarity measurement: Cosine distance is used as the similarity measure to calculate the cosine distance between PSD features of different channels.

[0103]

[0104] Based on the cosine distance matrix, an agglomerative hierarchical clustering algorithm (AGNES) is implemented. Spatial adjacency constraints are introduced during the clustering process: only signal clusters that are spatially adjacent distance cells and / or channels are allowed to be merged, thereby avoiding the erroneous fusion of unrelated signals.

[0105] (c) Optimal cluster selection and signal fusion: Within each frequency band, the cluster with the highest signal-to-noise ratio and the most stable signal is selected as the target cluster C. breath (Respiratory frequency) and C heart (Heart rate frequency band). The signal-to-noise ratio is calculated as follows:

[0106]

[0107] Among them, P signal P represents the peak power within the target frequency band. noise This represents the out-of-band noise power. The final set of high-quality signal units is obtained through a set intersection operation:

[0108] C flnal =C breath ∩C heart

[0109] Among them, C breath For the respiratory frequency band target cluster, C heart For the target cluster in the heart rate frequency band,

[0110] (d) Sum the phases of all signals in the set using a weighted average, with the weights assigned based on the SNR of each signal:

[0111]

[0112] The enhanced phase signal after fusion is:

[0113]

[0114] This fusion strategy makes full use of multi-channel spatial diversity, significantly improving signal quality.

[0115] (5) Signal separation and estimation based on multivariate variational mode decomposition

[0116] This step employs an improved multivariate variational mode decomposition (MVMD) algorithm to separate heart rate and respiratory components from the fused signal.

[0117] (a) Multivariable Variational Mode Decomposition: Steps S1-3 above process the signal of each receiving channel, resulting in s2 channels. fused (t) Signal (considered as multi-channel input) is input to the MVMD algorithm.

[0118] The core of the MVMD algorithm is to construct and solve a constrained variational problem.

[0119] Extended Autovariable Mode Decomposition (MVMD) aims to minimize the sum of estimated bandwidths for all modes while constraining the same mode to have a consistent center frequency across different channels. The objective function is expressed as:

[0120]

[0121] Constraints:

[0122]

[0123] Among them, u k，c (t) represents the component of the k-th mode in the c-th channel, ω k Let t be the center frequency of the k-th mode (consistent across channels), K = 5 is the number of modes, C = 12 is the number of channels, δ(t) is the Dirac function, and * denotes convolution.

[0124] MVMD solves the above variational problem iteratively using the alternating direction multiplier method (ADMM). The main steps include:

[0125] 1) Modal update:

[0126]

[0127] 2) Center frequency update:

[0128]

[0129] 3) Lagrange multiplier update:

[0130]

[0131] (b) Respiratory Harmonic Identification and Removal: The mode with the highest energy and center frequency in the range of 0.2-0.5Hz is identified as the respiratory fundamental frequency mode μ. resp Record its center frequency f resp Traversing other modes, if its center frequency f i satisfy:

[0132] Where n = 2, 3, 4, 5

[0133] If δ = 0.1 is the tolerance threshold, then the mode is determined to be a breathing harmonic and marked for removal.

[0134] (c) Heartbeat signal reconstruction and estimation: From the remaining modes, modes with consistent energy distribution across all 12 channels (inter-channel energy variance less than a preset threshold) and center frequencies within the range of 0.8-2.0 Hz are selected, and weighted fusion is performed to reconstruct the final heartbeat signal s. heart(t) The weighting is based on the energy consistency allocation of the modes across channels:

[0135]

[0136] Among them, Kh eart For a qualified mode set, α k Let be the weight. For s heart (t) Perform spectral analysis to estimate heart rate through peak detection:

[0137] f heart =argmax f |F{s heart (t))|

[0138] Where F represents the Fourier transform.

Claims

1. A method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW), characterized in that, The steps include: (1) real-time acquisition or reading of the intermediate frequency signal of the acquired multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) system; (2) preprocessing of the acquired multi-channel FMCW intermediate frequency signal, performing range-dimensional FFT on the intermediate frequency signal channel by channel and chirp by channel, and performing phase extraction and unwinding; (3) calculating the signal energy of the respiratory and heartbeat frequency bands of interest in the phase time series at all distances, determining the proportion of the energy of the respiratory and heartbeat frequency bands of interest in the phase time series at a certain distance to the total energy of the frequency bands of interest in the phase time series at all distances as the weight at that distance, and weighting the amplitude of each distance under all distance-dimensional FFTs accordingly to achieve accurate positioning of human targets; (4) treating the phase time series at each channel's precise distance and its neighboring distances as potential multipath signals, and fusing and enhancing them; (5) processing the enhanced phase time series of each channel based on multivariate variational mode decomposition, separating and estimating the human respiratory signal and heartbeat signal, and calculating the average respiratory rate and average heart rate during the observation period.

2. The method for estimating human respiratory and heartbeat signals based on a multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) according to claim 1, wherein step (1) involves real-time acquisition or reading of the intermediate frequency signal of the acquired multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) system, characterized in that, The multi-channel intermediate frequency (IF) signal originates from the configuration of the FMCW radar hardware system. Taking the TIIWR1843 millimeter-wave radar evaluation board hardware platform as an example, it is configured with 3 transmit antennas (TX) and 4 receive antennas (RX). Each transmit antenna is activated sequentially through the TDMA (Time Division Multiple Access) working mode. During each activation period, all 4 receive antennas synchronously receive echoes, thereby forming an equivalent linear array composed of 12 virtual array elements in space, corresponding to 12 channels of IF signal. The IF signal of each channel is composed of frames of data, and there are two parameters that need to be configured in advance: the number of frames and the frame rate. For physiological monitoring, the frame rate is generally configured to be tens of hertz. Each frame of data contains the IF signal of one or more chirp signals that are configured. It also has two parameters that need to be configured in advance: the number of samples and the sampling rate. The sampling rate is generally configured to be above megahertz. In the subsequent experiments, 512 points of the intermediate frequency signal were sampled for each Chirp, and 1500 frames were continuously acquired at a frame rate of 25Hz. The original data cube was constructed with dimensions of 512 ADC sampling points × 1 Chirp / frame × 1500 frames × 12 channels, which completely recorded the temporal echo of the detection scene.

3. The method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) according to claim 1, wherein step (2) preprocesses the obtained multi-channel FMCW intermediate frequency signal, performs range-dimensional FFT on the intermediate frequency signal channel by channel and chirp by chirp, and performs phase extraction and unwinding, characterized in that, (a) Range-Dimensional FFT: For each frame of ADC data, a Fast Fourier Transform (FFT) is performed along the fast time dimension, i.e., the number of sampling points of the intermediate frequency signal in each chirp, such as 512, to transform the time-domain signal into the frequency domain. Each frequency point corresponds to a distance, and the horizontal axis is called a range gate. After performing FFT on each channel chirp, a "range-time topography" is generated. (a) Map), where one axis is the distance gate index and the other axis is the frame number (time series). The value of each pixel in the topographic map represents the signal strength of the distance unit at the corresponding time. (b) Phase extraction and unwrapping: For each distance unit in each channel, the phase sequence φ(t) is extracted from its complex form distance-time map. Since the original phase value is wrapped in the interval [-π, π], there is a 2π jump. Phase unwrapping is required to obtain continuous phase changes. An unwrapping algorithm based on first-order differential accumulation is adopted. By detecting the difference between adjacent phase points and correcting the jumps exceeding π by integer multiples of 2π, a smooth and continuous phase change sequence Δφ(t) is finally obtained. This sequence directly corresponds to the thoracic micro-displacement caused by breathing and heartbeat.

4. A method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) according to claim 1, wherein step (3) calculates the signal energy of the respiratory and heartbeat frequency band of interest in the phase time series at all distances, and determines the proportion of the energy of the respiratory and heartbeat frequency band of interest in the phase time series at one distance to the total energy of the frequency band of interest in the phase time series at all distances as the weight at that distance, and accordingly weights the amplitude at each distance of the FFT in all distance dimensions to achieve accurate positioning of the human target, characterized in that, (a) Constructing the radar energy matrix: Calculate the squared signal amplitude |R(r,t)| at each pixel in the range-time graph. 2 (a) Obtain the original radar energy matrix Ra, which reflects the original energy distribution of each scatterer in the environment; (b) Calculate the physiological signal energy weights: For the phase sequence Δφ(t) of each range cell, estimate its power spectral density (PSD) using the Welch method; Set the respiratory signal frequency band to [0.1, 0.5] Hz and the heartbeat signal frequency band to [0.8, 2.0] Hz, calculate the power integrals Eb(r) and Eh(r) of each range cell in these two characteristic frequency bands respectively, and add them together to obtain the total physiological energy of the cell. Etotal(r); Subsequently, the Etotal(r) of this unit is divided by the sum of Etotal(r) of all range units and normalized to obtain the physiological signal energy weight Rs. This matrix has a high weight in areas rich in vital signs (such as the human chest cavity) and a very low weight in static background or inanimate areas; (c) Weighted fusion: The original radar energy matrix Ra is weighted according to Rs of each range unit to generate a weighted energy matrix W. This operation effectively suppresses the interference of static clutter and highlights the echo characteristics of living organisms; (d) Target localization: Average matrix W along the time dimension to obtain a one-dimensional average weighted energy distribution curve, and search for the global maximum value of the curve within a preset effective detection range (e.g., 0.5m to 2.0m). The corresponding distance unit is used as the candidate human target position r for that channel; (e) Final target position determination: Collect all position candidates {r_1, r_2, ..., r_12} for 12 channels, sort them, and take the median as the final human target position r_target.

5. The method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) according to claim 1, wherein in step (4), the phase time series of each channel at its precise distance and its neighboring distances are regarded as potential multipath signals, and they are fused and enhanced, characterized in that, (a) Multipath signal extraction: For each channel signal, taking the target distance cell located in step (3) as the center, extract its phase time series and its adjacent ±2 distance cells (a total of 5 cells) to form an initial multipath micro-motion signal set; (b) Dual-band hierarchical clustering: Process the filtered signals of the respiratory band (0.2-0.5Hz) and heart rate band (0.5-2.5Hz) respectively; For each micro-motion signal, calculate its normalized power spectral density (PSD) vector in the target band as its feature, and then use the agglomerative hierarchical clustering algorithm to cluster the signals from different distance cells with the cosine distance between PSD vectors as the similarity measure; In the clustering process, spatial adjacency constraint is introduced, that is, spatially adjacent distance cells are merged first to ensure that the merged clusters are (c) Optimal cluster selection and signal fusion: In each frequency band, the average signal-to-noise ratio (SNR) of each cluster is calculated, and the cluster with the highest SNR is selected as the target cluster (Cbreath and Cheart) of that frequency band. The SNR is defined as the ratio of the peak power in the target frequency band to the average power of the noise outside the frequency band. The final set of high-quality signal units is obtained by taking the intersection of the two target clusters, Coptimal = Cbreath ∩ Cheart. The signals in this set perform well in both the respiratory and heart rate frequency bands. (d) The phases of all signals in the set are weighted and summed. The weights are determined by their respective SNRs, so that each channel outputs a fused and enhanced high SNR phase signal Sfused(t).

6. The method for estimating human respiratory and heartbeat signals based on multi-channel frequency modulated continuous wave millimeter-wave radar (FMCW) according to claim 1, in step (5), the enhanced phase time series of each channel are processed based on multivariate variational mode decomposition to separate and estimate the human respiratory and heartbeat signals, and the average respiratory rate and average heart rate during the observation period are calculated, characterized in that, (a) Multivariate Variational Mode Decomposition: For a 12-channel signal, the above steps are applied to each channel to obtain 12 enhanced signals Sfused(t), which are then fed into the multivariate variational mode decomposition algorithm. MVMD adaptively decomposes the input signal into K (e.g., K=5) quasi-orthogonal modal components Uk(t) by solving a constrained variational problem. Its core idea is to force the same mode to share a center frequency ωk in different channels, thereby ensuring the physical consistency of the decomposed modes. (b) Breathing Harmonic Identification and Removal: Among all the decomposed modes, the mode with the strongest energy and a center frequency located in the breathing fundamental frequency range (0.2-0.5Hz) is identified. The mode is labeled as the respiratory fundamental frequency mode Uresp; then, all other modes are traversed. If the center frequency fi of a certain mode is approximately equal to an integer multiple of the respiratory fundamental frequency fresp (fi≈n·fresp, n=2,3,...), then the mode is determined to be a respiratory harmonic and is removed from the candidate set; (c) Heartbeat signal reconstruction and estimation: From the remaining modes after removing respiratory harmonics, modes with consistent energy distribution on all 12 channels (small energy variance between channels) and center frequencies falling within the normal heart rate range (0.8-2.0Hz) are selected. These qualified modes are weighted and averaged on each channel to reconstruct the final pure heartbeat signal sheart(t); By performing spectral analysis on the heartbeat signal and detecting the main peak value in its power spectrum, the instantaneous heart rate can be accurately estimated.

Images