Multi-person respiratory signal detection method and system based on LFMCW radar and underdetermined blind source separation

Abstract

The application discloses a kind of based on LFMCW radar and underdetermined blind source separation Multi-person respiratory signal detection method and system.Method includes: first, the radar echo signal of two-way receiving antenna is rearranged, the phase of rearranged echo signal is extracted, then the noise and interference in phase signal are filtered out by band-pass filter, and the mixed signal of two-way multi-person breathing is obtained;Afterwards, the two-way mixed signal is transformed into time-frequency domain using short-time Fourier transform, two-way time-frequency points constitute clustering plane, and improved near neighbor propagation algorithm based on angle information and energy information fusion is used for clustering analysis, to obtain the estimate of number and mixed matrix;Finally, according to the mixed matrix, the respiratory signal in time-frequency domain is separated out by L1 norm minimization algorithm, and is recovered to time domain by inverse short-time Fourier transform, to obtain the respiratory signal of multiple persons.The method of the application is effective and feasible, and the performance is reliable, and the same distance gate, very close multi-person respiratory signal can be obtained using double receiving antenna LFMCW radar system.

Description

Technical Field

[0001] This invention belongs to the field of vital sign monitoring, and in particular, it is a non-contact method and system for detecting respiratory signals in multiple people based on LFMCW radar and underdetermined blind source separation technology. Background Technology

[0002] In recent years, microwave-based radar systems have been widely used for detecting vital signs. Compared to traditional contact-based respiratory detection devices, non-contact methods offer advantages such as no skin irritation, no discomfort, and freedom from the constraints of electrodes and cables. With the development of digital signal processing technology and the improvement of hardware system performance, multi-antenna systems can monitor multiple objects simultaneously and can be widely used in various scenarios such as hospital wards, bedrooms, and post-disaster search and rescue.

[0003] Currently, radar-based multi-target respiratory monitoring technology has been studied in considerable depth, and its implementation methods can be mainly divided into the following three types:

[0004] One approach involves separating multiple targets by angle, using antenna beams pointing towards each individual human object. This method primarily utilizes direction-of-arrival (DOA) estimation and digital beamforming techniques to locate multiple human targets and detect their breathing signals. However, this method is limited by the angular resolution of the radar system. In scenarios where multiple human targets are very close together, DOA estimation cannot accurately estimate angular information, and digital beamforming is no longer suitable for such situations. Especially in radar systems with a small number of receiving antennas, wide beams cannot distinguish between multiple closely spaced human targets.

[0005] Secondly, multiple targets can be separated based on the different range gates of the human target. This method mainly utilizes the characteristics of FMCW radar to distinguish multiple human targets by distance, but this method is difficult to separate multiple targets located at the same range gate.

[0006] Thirdly, respiratory signals from multiple human subjects are extracted using separation algorithms. This method has received relatively little research both domestically and internationally. The main separation algorithms include Variational Mode Decomposition (VMD) and blind source separation algorithms. VMD can separate respiratory signals from multiple targets using a single receiving antenna, but this algorithm relies on various initial parameters, yielding drastically different results under different parameters. Without prior knowledge of the parameters, the algorithm exhibits poor robustness. Research on blind source separation algorithms has focused entirely on positive definite conditions, requiring the number of receiving antennas to equal the number of human targets. As the number of subjects increases, this method significantly increases the hardware cost of the radar system. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in the prior art by providing a non-contact method and system for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation technology. This invention is the first to apply the underdetermined blind source separation algorithm to the detection of multiple respiratory signals and proposes an improved nearest neighbor propagation algorithm based on the fusion of angle and energy information, aiming to more accurately estimate the mixing matrix and thus accurately recover the respiratory signals of multiple individuals.

[0008] The technical solution to achieve the purpose of this invention is: a method for detecting respiratory signals from multiple people based on LFMCW radar and underdetermined blind source separation, comprising the following steps:

[0009] Step 1: Use the dual-receiving antenna LFMCW radar to detect the chest movement of multiple human targets within the gate at the same distance, and obtain the echo signals from the two receiving antennas.

[0010] Step 2: Preprocess the radar echo signal corresponding to each receiving antenna to obtain a mixed signal of two multiple breathing signals. The mixed signal of each multiple breathing signal is a linear superposition of the multiple breathing signals of that signal. The mixed signal is represented in matrix form as follows:

[0011]

[0012] In the formula, The mixed signal vector representing the two mixed signals based on the blind source separation model. Source signal vector s1,s2,…,s P Let A be the respiratory signals of P individuals, and let A be a 2-row, P-column mixed matrix.

[0013] Step 3, Estimate the mixing matrix: Use short-time Fourier transform to transform the mixed signal vector Transform to the time-frequency domain to obtain the mixed signal vector in the time-frequency domain. Time-frequency points X1 and X2 form a clustering plane. An improved nearest neighbor propagation algorithm based on the fusion of angle and energy information is used for cluster analysis to obtain the number of clusters, cluster centers, and cluster count. This is an estimate of the number of people to be tested, and the cluster centers are an estimate of the mixture matrix, expressed as:

[0014]

[0015] In the formula, (X1(k i ),X2(k i )) represents the cluster center of the i-th class on the clustering plane, k i This is the index value of the cluster center of the i-th cluster;

[0016] Step 4, based on the estimated mixing matrix Respiratory signal recovery is performed. Under underdetermined conditions, when the number of targets exceeds the number of receiving antennas, respiratory signal recovery transforms into an optimization problem. By introducing constraints and minimizing the L1 norm, the optimal solution is obtained, leading to an estimate of the respiratory signal in the time-frequency domain. Finally, Performing an inverse short-time Fourier transform yields a time-domain estimate of the respiratory signal.

[0017] Furthermore, step 1 specifically includes:

[0018] Acquire echo signal s r (t) and the local oscillator signal s t (t), and the difference frequency signal s obtained by downconversion. b (t).

[0019] Furthermore, step 2, which involves preprocessing the radar echo signal corresponding to each receiving antenna, specifically includes:

[0020] Step 2-1: Rearrange the difference frequency signal according to the radar repetition period;

[0021] Step 2-2: Calculate the phase of the rearranged echo signal using the differential cross-multiplication algorithm;

[0022] Steps 2-3 involve filtering the phase signal to obtain a mixed signal.

[0023] Further, step 2-1, which involves rearranging the difference frequency signal according to the radar repetition period, specifically includes:

[0024] ① For frequency modulation period T r The frequency modulation signal has an ADC sampling period of T. s If M samples are collected in each frequency modulation cycle, and echo data is collected for a total of N frequency modulation cycles, then the difference frequency signal can be rearranged into an N-row, M-column matrix R, represented as:

[0025] R[n,m]=s b (t=m·T s +n·T r )

[0026] In the formula, n is the row index of the matrix, indicating that the index is T. r The number of samples in the sampling period is denoted by 'm', ranging from 1 to N; 'm' is the column index of the matrix, indicating that the column index is T. s The number of sample points in the sampling period, ranging from 1 to M;

[0027] ② Perform N operations on each row of matrix R[n,m]. fft The fast Fourier transform yields matrix R. f[n,v], where v is the FFT index value, ranging from 1 to N. fft ;

[0028] ③ Calculate matrix R f The energy of each column [n, v] is calculated using the following formula:

[0029]

[0030] The FFT index value V corresponding to the maximum energy value represents the distance gate number of the target location. Therefore, the target distance r can be represented by the FFT index value V:

[0031]

[0032] In the formula, K is the frequency modulation coefficient of the linear frequency modulated continuous wave radar, and f s The sampling rate is used; the above formula enables the localization of the target.

[0033] The slow time difference frequency signal s(n) at the target location is represented as:

[0034] s(n)=R f (n,V).

[0035] Further, in step 2-2, the phase of the rearranged echo signal is calculated using the differential cross-multiplication algorithm, and the calculation formula is as follows:

[0036]

[0037] In the formula, For the phase of the slow time difference frequency signal, s I (n) and s Q (n) represents the quadrature branch IQ signal, and n is the frequency modulation period T. r The number of sample points in the sampling period.

[0038] Furthermore, the filtering of the phase signal described in steps 2-3 specifically involves:

[0039] A bandpass filter h(t) of order L⁻¹ with a passband frequency of 0.15–0.55 Hz is used to filter the phase signal. This frequency range corresponds to the basic breathing band. The resulting mixed signal x(t) is calculated using the following formula:

[0040]

[0041] Further, step 3, estimating the mixture matrix, specifically includes:

[0042] Step 3-1 involves transforming the mixed signal of multiple respirations to the time-frequency domain to sparsify the mixed signal for cluster analysis; the specific process includes:

[0043] ① Perform a short-time Fourier transform on the mixed signal x(t) of multiple people's breathing corresponding to the two receiving antennas:

[0044] X i (t,f)=∫x i (τ)·g(t-τ)·e -j2πft dτt∈(0,L t -1), f∈(0,L f -1)

[0045] In the formula, g(t) is a Gaussian window function, x i (t) represents the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, X i (t,f) is the time-frequency matrix of the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, L f L represents the number of rows in the matrix and the number of FFT points. t The number of columns in the matrix represents the number of time frames.

[0046] ② Take the positive frequency band of the time-frequency matrix, that is, the frequency f takes a value from 0 to L. f / 2-1, let matrix X i (t,f) are rearranged into a one-dimensional vector X. i (q), the rearrangement formula is:

[0047] X i (q)=X i (t,f)q=t+f·L t

[0048] In the formula, q is the index number of the time-frequency vector. The time-frequency vectors corresponding to the two receiving antennas are X1(q) and X2(q). X1(q) and X2(q) constitute a two-dimensional clustering plane. A point X on the plane is defined as... * (q) = (X1(q),X2(q));

[0049] Step 3-2: Based on the improved nearest neighbor propagation algorithm that integrates angle and energy information, estimate the number of people to be tested and the mixing matrix;

[0050] The improved nearest neighbor propagation algorithm process includes:

[0051] ① Calculate the similarity matrix s(i,k). The similarity matrix measures the similarity of points X in the clustering plane. * (i) and point X * The similarity of (k) is defined in the similarity matrix as follows:

[0052] s(i,k)=abs[X * (k)]·exp(-ρ·(1-cos(∠[X * (i),X *(k)])))

[0053] In the formula, abs[X * [k] represents the cluster plane point X. * The modulus of (k); ∠[X * (i),X * [k] represents point X * (i) and point X * The angle between (k); ρ is the attenuation coefficient of the nonlinear function. The larger the value of ρ, the faster the attenuation and the greater the effect of insufficiently sparse points on clustering.

[0054] ② Initialize the attraction matrix r(i,k) and the membership matrix a(i,k) to 0; where the attraction matrix represents the suitability of data point k as the cluster center of data point i; and the membership matrix represents whether data point i chooses data point k as its cluster center.

[0055] ③ Update the attraction matrix r(i,k)

[0056] r t+1 (i,k)=s(i,k)-max j≠k [a t (i,j)+s(i,j)]

[0057] In the formula, r t+1 (i,k) represents the attractiveness in the next iteration, a t (i,k) represents the degree of belonging in the current iteration;

[0058] ④ Update the membership matrix a(i,k)

[0059]

[0060] ⑤ Use the attenuation coefficient λ to weight the current result of steps ③ and ④ with the result of the previous iteration. Usually, the attenuation coefficient λ = 0.5. The weighting formula is as follows:

[0061] r t+1 (i,k)=λ·r t (i,k)+(1-λ)·r t+1 (i,k)

[0062] a t+1 (i,k)=λ·a t (i,k)+(1-λ)·a t+1 (i,k)

[0063] ⑥ Repeat steps ③, ④, and ⑤ until the matrix converges or the maximum number of iterations is reached, at which point the algorithm terminates;

[0064] ⑦ Determine data point X * (i) Cluster center X* (k)=(X1(k),X2(k)), k should satisfy the formula:

[0065] max{r(i,k)+a(i,k)}

[0066] Number of cluster centers This is an estimate of the number of people to be tested. The cluster centers, or mixture matrix, are represented as follows:

[0067]

[0068] In the formula, X * (k i )=(X1(k i ),X2(k i )) represents the cluster center of the i-th class, k i This is the index value of the cluster center of the i-th cluster; This is an estimate of the number of people to be tested.

[0069] Further, step 4 involves using the estimated mixing matrix... Restoring respiratory signals specifically includes:

[0070] Step 4-1: Obtain the optimal solution to the underdetermined problem based on the mixing matrix. When the number of receiving antennas is less than the number of people to be tested, the blind source separation problem is underdetermined, the number of unknown sources is greater than the number of equations, and the equations do not have a unique solution. Introducing constraints to minimize the L1 norm of the solution transforms source recovery into an optimization problem. The objective function can be expressed as:

[0071]

[0072] In the formula, s(t,f) is the time-frequency domain estimate of each respiratory signal source, and X(t,f) is the mixed respiratory signal in the time-frequency domain. The mixture matrix is ​​calculated in step 3-2;

[0073] For a two-antenna receiving system, the steps of the L1 norm minimization algorithm are as follows:

[0074] ① Calculate the mixture matrix A full-rank submatrix, which is 2 rows and 2 columns long and has a total length of 260 ... 1, denoted as a k ,

[0075] ② Find all possible solutions for a point X(t,f) in the sparse domain.

[0076] ③ Calculate the L1 norm of each solution and take the one with the smallest norm. As the optimal estimate of the source signal, that is:

[0077]

[0078] ④ Repeat steps ①②③ to find the optimal solution for all points in the sparse domain;

[0079] Step 4-2: Restore the time-frequency domain respiratory signal separated in Step 4-1 to the time domain, using the following formula:

[0080]

[0081] In the formula, h(τ-t) is a rectangular window. The recovered time-domain signal.

[0082] A multi-person respiratory signal detection system based on LFMCW radar and underdetermined blind source separation, the system comprising:

[0083] The radar echo signal acquisition module is used to detect the chest movement of multiple human targets within a gate at the same distance using a dual-receiving antenna LFMCW radar, and obtain the echo signals from the two receiving antennas.

[0084] The preprocessing module is used to preprocess the radar echo signal corresponding to each receiving antenna to obtain a mixed signal of two multiple breathing signals. The mixed signal of each multiple breathing signal is a linear superposition of the multiple breathing signals of that signal. The mixed signal is represented in matrix form as follows:

[0085]

[0086] In the formula, The mixed signal vector representing the two mixed signals based on the blind source separation model. Source signal vector s1,s2,…,s P Let A be the respiratory signals of P individuals, and let A be a 2-row, P-column mixed matrix.

[0087] The mixing matrix estimation module is used to estimate the mixing matrix by using a short-time Fourier transform to transform the mixed signal vector. Transform to the time-frequency domain to obtain the mixed signal vector in the time-frequency domain. Time-frequency points X1 and X2 form a clustering plane. An improved nearest neighbor propagation algorithm based on the fusion of angle and energy information is used for cluster analysis to obtain the number of clusters, cluster centers, and cluster count. This is an estimate of the number of people to be tested, and the cluster centers are an estimate of the mixture matrix, expressed as:

[0088]

[0089] In the formula, (X1(k i ),X2(ki )) represents the cluster center of the i-th class on the clustering plane, k i This is the index value of the cluster center of the i-th cluster;

[0090] Source recovery module, used to recover the estimated mixture matrix Respiratory signal recovery is performed. Under underdetermined conditions, when the number of targets exceeds the number of receiving antennas, respiratory signal recovery transforms into an optimization problem. By introducing constraints and minimizing the L1 norm, the optimal solution is obtained, leading to an estimate of the respiratory signal in the time-frequency domain. Finally, Performing an inverse short-time Fourier transform yields a time-domain estimate of the respiratory signal.

[0091] Compared with the prior art, the significant advantages of this invention are:

[0092] 1) Non-contact respiratory monitoring is achieved using LFMCW radar. It can penetrate obstacles such as clothing and bedding. Compared with traditional contact monitoring, it is more convenient to operate, reduces human discomfort, and has no limitations on electrode and cable length.

[0093] 2) The multi-person respiratory signal detection method based on the underdetermined blind source separation algorithm is not limited by the angular resolution of the radar system and can be adapted to scenarios where multiple people are in close proximity, such as overnight sleep monitoring of couples, mothers and infants;

[0094] 3) In the implementation of the underdetermined blind source separation algorithm, an improved nearest neighbor propagation algorithm based on the fusion of angle information and energy information is proposed to estimate the mixing matrix, which reduces the influence of insufficiently sparse points with small energy and large deviations between the angle and the cluster center on the clustering results, and makes the estimation of the mixing matrix more accurate.

[0095] 4) The multi-person respiratory signal detection method based on the underdetermined blind source separation algorithm can realize the detection and separation of respiratory signals of more than 2 people using only dual receiving antennas. The system has a simple structure, low cost and reliable performance, and is easy to implement.

[0096] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0097] Figure 1 This is a flowchart of the non-contact multi-person respiratory signal detection method based on underdetermined blind source separation according to the present invention.

[0098] Figure 2Figure 1 shows the time-domain and frequency-domain waveforms of the mixed signal after preprocessing of the two receiving antennas in one embodiment. Figure 2a) is the time-domain signal obtained by preprocessing the first receiving antenna, Figure 3b) is the frequency-domain signal obtained by preprocessing the first receiving antenna, Figure 4c) is the time-domain signal obtained by preprocessing the second receiving antenna, and Figure 5d) is the frequency-domain signal obtained by preprocessing the second receiving antenna.

[0099] Figure 3 This is a clustering result diagram obtained using an improved nearest neighbor propagation algorithm based on nonlinear projection in one embodiment.

[0100] Figure 4 The following is a comparison of the time-domain waveforms of the respiratory signals of three subjects recovered using the algorithm proposed in this invention, and the respiratory band reference signal, in one embodiment. Figure a) shows the time-domain waveforms of the respiratory signal and respiratory band signal recovered by the first subject, and Figure b) shows the time-domain waveforms of the respiratory signal and respiratory band signal recovered by the second subject. Figure 4 =c) is the time-domain waveform of the respiratory signal and respiratory band signal recovered by the third subject.

[0101] Figure 5 Figure a) shows the frequency domain waveforms of the respiratory signals of three subjects recovered using the algorithm proposed in this invention, compared with the respiratory band reference signal. Figure b) shows the frequency domain waveforms of the respiratory signals and respiratory band signals recovered by the first subject, and Figure c) shows the frequency domain waveforms of the respiratory signals and respiratory band signals recovered by the second subject. Detailed Implementation

[0102] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0103] In one embodiment, a dual-receiver antenna LFMCW radar system is used to monitor the respiratory signals of three people inside a door at the same distance. Combined with... Figure 1 This paper presents a non-contact method for detecting respiratory signals in multiple individuals based on LFMCW radar and underdetermined blind source separation technology. The steps are as follows:

[0104] Step 1: Use the dual-receiving antenna LFMCW radar to detect the chest movement of three people inside the door at the same distance, and obtain the echo signals from the two receiving antennas.

[0105] Step 2 involves preprocessing the radar echo signal corresponding to each receiving antenna, specifically as follows:

[0106] Step 2-1: Rearrange the difference frequency signal according to the radar repetition period. Extract the slow time difference frequency signal at the target range threshold. The specific process includes:

[0107] The specific process includes:

[0108] ① For frequency modulation period T r The frequency modulation signal has an ADC sampling period of T. s If M samples are collected in each frequency modulation cycle, and echo data are collected for a total of N frequency modulation cycles, then the difference frequency signal can be rearranged into an N-row, M-column matrix R, which can be represented as:

[0109] R[n,m]=s b (t=m·T s +n·T r )

[0110] In the formula, n is the row index of the matrix, indicating that the index is T. r The number of samples in the sampling period is denoted by 'm', ranging from 1 to N; 'm' is the column index of the matrix, indicating that the column index is T. s The number of sample points in the sampling period ranges from 1 to M.

[0111] ② Perform N operations on each row of matrix R[n,m]. fft The fast Fourier transform yields matrix R. f [n,v], where v is the FFT index value, ranging from 1 to N. fft .

[0112] ③ Calculate matrix R f The energy of each column [n, v] is calculated using the following formula:

[0113]

[0114] The FFT index value V corresponding to the maximum energy value represents the range gate number of the target location. This is because the difference frequency f of the linear frequency modulated continuous wave radar... b Proportional to the echo delay τ, the target distance r can be represented by the FFT index value V:

[0115]

[0116] In the formula, K is the frequency modulation coefficient of the linear frequency modulated continuous wave radar, and f s Let be the sampling rate. The target location is achieved using the above formula. The slow time difference frequency signal s(n) at the target's location can be expressed as:

[0117] s(n)=R f (n,V)

[0118] Step 2-2: Calculate the phase of the slow-time difference frequency signal. The chest cavity motion signal is modulated into the phase of the radar echo slow-time signal. The Differential Cross Multiplication (DACM) algorithm is based on the orthogonal branch IQ signal s. I (n) and s Q (n) Calculate the phase of the slow time difference frequency signal The formula for the differential cross-multiplication algorithm is as follows:

[0119]

[0120] In the formula, n is the frequency modulation period T r The number of sample points in the sampling period.

[0121] Steps 2-3: Filter out the DC component and noise interference in the phase of the echo signal. A bandpass filter h(t) of order L⁻¹ with a passband frequency of 0.15-0.55 Hz is used; this frequency range corresponds to the basic breathing band. The resulting mixed signal x(t) is calculated using the following formula:

[0122]

[0123] After preprocessing, the mixed breathing signals of three people received by the two antennas are obtained, such as... Figure 2 As shown.

[0124] Step 3, estimate the mixing matrix of the respiratory mixed signal, specifically:

[0125] Step 3-1 transforms the mixed signal of multiple respirations to the time-frequency domain to sparsify the mixed signal for cluster analysis. The specific process includes:

[0126] ① Perform a short-time Fourier transform on the mixed signal x(t) of multiple people's breathing corresponding to the two receiving antennas:

[0127] X i (t,f)=∫x i (τ)·g(t-τ)·e -j2πft dτt∈(0,L t -1), f∈(0,L f -1)

[0128] In the formula, g(t) is a Gaussian window function, x i (t) represents the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, X i (t,f) is the time-frequency matrix of the mixed breathing signal of multiple people corresponding to the i-th receiving antenna. f L represents the number of rows in the matrix and the number of FFT points. t The column number represents the number of time frames.

[0129] ② Take the positive frequency band of the time-frequency matrix, that is, the frequency f takes a value from 0 to L.f / 2-1, let matrix X i (t,f) are rearranged into a one-dimensional vector X. i (q), the rearrangement formula is:

[0130] X i (q)=X i (t,f)q=t+f·L t

[0131] In the formula, q is the index number of the time-frequency vector, and the time-frequency vectors corresponding to the two receiving antennas are X1(q) and X2(q). X1(q) and X2(q) constitute a two-dimensional clustering plane. A point X on the plane is defined as... * (q)=(X1(q),X2(q))

[0132] Step 3-2: Based on the improved nearest neighbor propagation algorithm that integrates angle and energy information, estimate the number of people to be tested and the mixing matrix.

[0133] The improved nearest neighbor propagation algorithm process includes:

[0134] ① Calculate the similarity matrix s(i,k). The similarity matrix measures the similarity of points X in the clustering plane. * (i) and point X * The similarity is calculated for (k). The similarity matrix is ​​defined as follows:

[0135] s(i,k)=abs[X * (k)]·exp(-ρ·(1-cos(∠[X * (i),X * (k)])))

[0136] In the formula, abs[X * [k] represents the cluster plane point X. * The modulus of (k); ∠[X * (i),X * [k] represents point X * (i) and point X * The angle between points (k) is used to measure similarity. ρ is the attenuation coefficient of the nonlinear function. The larger the value of ρ, the faster the attenuation, and the better it can reduce the influence of insufficiently sparse points on clustering. Unlike the traditional nearest neighbor propagation algorithm, which measures similarity by calculating the Euclidean distance between points on the clustering plane, this invention considers that respiratory signals are relatively sparse in the time-frequency domain. That is, each respiratory signal in the time-frequency domain determines a straight line on the clustering plane. Therefore, the angle between points is used to measure similarity, and a nonlinear function is used to reduce the influence of insufficiently sparse points on clustering. At the same time, the energy information of time-frequency points is considered to reduce the influence of low-energy points on clustering.

[0137] ② Initialize the attraction matrix r(i,k) and the membership matrix a(i,k) to 0. The attraction matrix represents the suitability of data point k as the cluster center of data point i; the membership matrix represents whether data point i chooses data point k as its cluster center.

[0138] ③ Update the attraction matrix r(i,k)

[0139] r t+1 (i,k)=s(i,k)-max j≠k [a t (i,j)+s(i,j)]

[0140] In the formula, r t+1 (i,k) represents the attractiveness in the next iteration, a t (i,k) represents the degree of belonging in the current iteration.

[0141] ④ Update the membership matrix a(i,k)

[0142]

[0143] ⑤ To avoid numerical oscillations, a decay coefficient λ is used to weight the current result of steps ③ and ④ with the result of the previous iteration. Typically, the decay coefficient λ = 0.5 is chosen, and the weighting formula is as follows:

[0144] r t+1 (i,k)=λ·r t (i,k)+(1-λ)·r t+1 (i,k)

[0145] a t+1 (i,k)=λ·a t (i,k)+(1-λ)·a t+1 (i,k)

[0146] ⑥ Repeat steps ③, ④, and ⑤ until the matrix converges or the maximum number of iterations is reached, at which point the algorithm terminates.

[0147] ⑦ Determine data point X * (i) Cluster center X * (k)=(X1(k),X2(k)), k should satisfy the formula:

[0148] max{r(i,k)+a(i,k)}

[0149] Number of cluster centers This is an estimate of the number of people to be tested. The cluster centers, or mixture matrix, can be represented as:

[0150]

[0151] In the formula, X* (k i )=(X1(k i ),X2(k i )) represents the cluster center of the i-th class, k i This is the index value of the cluster center of the i-th cluster; This is an estimate of the number of people to be tested.

[0152] Combination Figure 3 Three cluster centers were obtained, which were used as estimates of the mixture matrix.

[0153] Step 4: Recover the respiratory signals of multiple individuals based on the hybrid matrix, specifically as follows:

[0154] Step 4-1: Obtain the optimal solution to the underdetermined problem based on the mixing matrix. When the number of receiving antennas is less than the number of people to be measured, the blind source separation problem is underdetermined, the number of unknown sources is greater than the number of equations, and the equations do not have a unique solution. Introducing constraints to minimize the L1 norm of the solution transforms source recovery into an optimization problem. The objective function can be expressed as:

[0155]

[0156] In the formula, s(t,f) is the time-frequency domain estimate of each respiratory signal source. X(t,f) is the mixed respiratory signal in the time-frequency domain. This is the mixture matrix calculated in step 3-2.

[0157] For a two-antenna receiving system, the steps of the L1 norm minimization algorithm are as follows:

[0158] ① Calculate the mixture matrix A full-rank submatrix. This submatrix is ​​2 rows and 2 columns long, with a total of... 1, denoted as a k ,

[0159] ② Find all possible solutions for a point X(t,f) in the sparse domain.

[0160] ③ Calculate the L1 norm of each solution and take the one with the smallest norm. As the optimal estimate of the source signal, i.e.

[0161]

[0162] ④ Repeat steps ①②③ to find the optimal solution for all points in the sparse domain;

[0163] Step 4-2: Restore the time-frequency domain respiratory signal separated in Step 4-1 to the time domain using the following formula:

[0164]

[0165] In the formula, h(τ-t) is a rectangular window. The recovered time-domain signal.

[0166] Combination Figure 4 , Figure 5 The recovered respiratory signals of the three individuals were consistent with the respiratory belt signals, demonstrating reliable performance.

[0167] In summary, the non-contact multi-person respiratory signal detection method based on LFMCW radar and underdetermined blind source separation technology proposed in this invention is used to detect vital signs of multiple objects at close range, overcoming the limitation of radar system angular resolution. This invention's system is simple, effective, feasible, and reliable, and can be widely applied in various scenarios such as hospital wards, bedrooms, and post-disaster search and rescue.

[0168] The foregoing has shown and described the basic principles, main steps, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting respiratory signals from multiple individuals based on LFMCW radar and underdetermined blind source separation, characterized in that, Includes the following steps: Step 1: Use the dual-receiving antenna LFMCW radar to detect the chest movement of multiple human targets within the gate at the same distance, and obtain the echo signals from the two receiving antennas. Step 2: Preprocess the radar echo signal corresponding to each receiving antenna to obtain a mixed signal of two multiple breathing signals. The mixed signal of each multiple breathing signal is a linear superposition of the multiple breathing signals of that signal. The mixed signal is represented in matrix form as follows: In the formula, The mixed signal vector representing the two mixed signals based on the blind source separation model. Source signal vector s1,s2,…,s P Let A be the respiratory signals of P individuals, and let A be a 2-row, P-column mixed matrix. Step 3, Estimate the mixing matrix: Use short-time Fourier transform to transform the mixed signal vector Transform to the time-frequency domain to obtain the mixed signal vector in the time-frequency domain. Time-frequency points X1 and X2 form a clustering plane. An improved nearest neighbor propagation algorithm based on the fusion of angle and energy information is used for cluster analysis to obtain the number of clusters, cluster centers, and cluster count. This is an estimate of the number of people to be tested, and the cluster centers are an estimate of the mixture matrix, expressed as: In the formula, (X1(k i ),X2(k i )) represents the cluster center of the i-th class on the clustering plane, k i This is the index value of the cluster center of the i-th cluster; Step 4, based on the estimated mixing matrix Respiratory signal recovery is performed. Under underdetermined conditions, when the number of targets exceeds the number of receiving antennas, respiratory signal recovery transforms into an optimization problem. By introducing constraints and minimizing the L1 norm, the optimal solution is obtained, leading to an estimate of the respiratory signal in the time-frequency domain. Finally, Performing an inverse short-time Fourier transform yields a time-domain estimate of the respiratory signal. Step 3, estimating the mixture matrix, specifically includes: Step 3-1 transforms the mixed signal of multiple people's breathing into the time-frequency domain to sparsify the mixed signal for cluster analysis; the specific process includes: ① Perform a short-time Fourier transform on the mixed signal x(t) of multiple people's breathing corresponding to the two receiving antennas: X i (t,f)=∫x i (τ)·g(t-τ)·e -j2πft dπt∈(0,L t -1),f∈0,L f -1) In the formula, g(t) is a Gaussian window function, x i (t) represents the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, X i (t,f) is the time-frequency matrix of the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, L f L represents the number of rows in the matrix and the number of FFT points. t The number of columns in the matrix represents the number of time frames. ② Take the positive frequency band of the time-frequency matrix, that is, the frequency f takes a value from 0 to L. f / 2-1, let matrix X i (t,f) are rearranged into a one-dimensional vector X. i (q0, the rearrangement formula is: X i (q)=X i (t,f) q=t+f·L t In the formula, q is the index number of the time-frequency vector. The time-frequency vectors corresponding to the two receiving antennas are X1(q) and X2(q). X1(q) and X2(q) constitute a two-dimensional clustering plane. A point X on the plane is defined as... * (q) = (X1(q),X2(q)); Step 3-2: Based on the improved nearest neighbor propagation algorithm that integrates angle and energy information, estimate the number of people to be tested and the mixing matrix; The improved nearest neighbor propagation algorithm process includes: ① Calculate the similarity matrix s(i,k). The similarity matrix measures the similarity of points X in the clustering plane. * (i) and point X * The similarity of (k) is defined in the similarity matrix as follows: s(i,k)=abs[X * (k)]·exp(-ρ·(1-cos(∠[X * (i),X * (k)]))) In the formula, abs[X * [k] represents the cluster plane point X. * The modulus of (k); ∠[X * (i),X * [k] represents point X * (i) and point X * The angle between (k); ρ is the attenuation coefficient of the nonlinear function. The larger the value of ρ, the faster the attenuation and the greater the effect of insufficiently sparse points on clustering. ② Initialize the attraction matrix r(i,k) and the membership matrix a(i,k) to 0; where the attraction matrix represents the suitability of data point k as the cluster center of data point i; and the membership matrix represents whether data point i chooses data point k as its cluster center. ③ Update the attraction matrix r(i,k) r t+1 (i,k)=s(i,k)-max j≠k [a t (i,j)+s(i,j)] In the formula, r t+1 (i,k) represents the attractiveness in the next iteration, a t (i,k) represents the degree of belonging in the current iteration; ④ Update the membership matrix a(i,k) ⑤ Use the attenuation coefficient λ to weight the current result of steps ③ and ④ with the result of the previous iteration. Usually, the attenuation coefficient λ = 0.

5. The weighting formula is as follows: r t+1 (i,k)=λ·r t (i,k)+(1-λ)·r t+1 (i,k) a t+1 (i,k)=λa t (i,k)+(1-λ)·a t+1 (i,k) ⑥ Repeat steps ③, ④, and ⑤ until the matrix converges or the maximum number of iterations is reached, at which point the algorithm terminates; ⑦ Determine data point X * (i) Cluster center X * (k)=(X1(k),X2(k)), k should satisfy the formula: max{r(i,k)+a(i,k)} Number of cluster centers This is an estimate of the number of people to be tested. The cluster centers, or mixture matrix, are represented as follows: In the formula, X * (k i )=(X1(k i ),X2(k i )) represents the cluster center of the i-th class, k i This is the index value of the cluster center of the i-th cluster; This is an estimate of the number of people to be tested.

2. The method for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation according to claim 1, characterized in that, Step 1 specifically includes: Acquire echo signal s r (t) and the local oscillator signal s t (t), and the difference frequency signal s obtained by downconversion. b (t).

3. The method for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation according to claim 1 or 2, characterized in that, Step 2 involves preprocessing the radar echo signal corresponding to each receiving antenna, specifically including: Step 2-1: Rearrange the difference frequency signal according to the radar repetition period; Step 2-2: Calculate the phase of the rearranged echo signal using the differential cross-multiplication algorithm; Steps 2-3 involve filtering the phase signal to obtain a mixed signal.

4. The method for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation according to claim 3, characterized in that, Step 2-1, which involves rearranging the difference frequency signal according to the radar repetition period, specifically includes: ① For frequency modulation period T r The frequency modulation signal has an ADC sampling period of T. s If M samples are collected in each frequency modulation cycle, and echo data is collected for a total of N frequency modulation cycles, then the difference frequency signal can be rearranged into an N-row, M-column matrix R, represented as: R[n,m]=s b (t=m·T s +n·T r ) In the formula, n is the row index of the matrix, indicating that the index is T. r The number of samples in the sampling period is denoted by 'm', ranging from 1 to N; 'm' is the column index of the matrix, indicating that the column index is T. s The number of sample points in the sampling period, ranging from 1 to M; ② Perform N operations on each row of matrix R[n,m]. fft The fast Fourier transform yields matrix R. f [n,v], where v is the FFT index value, ranging from 1 to N. fft ; ③ Calculate matrix R f The energy of each column [n, v] is calculated using the following formula: The FFT index value V corresponding to the maximum energy value represents the distance gate number of the target location. Therefore, the target distance r can be represented by the FFT index value V: In the formula, K is the frequency modulation coefficient of the linear frequency modulated continuous wave radar, and f s The sampling rate is used; the above formula enables the localization of the target. The slow time difference frequency signal s(n) at the target location is represented as: s(n)=R f (n,V)。 5. The method for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation according to claim 4, characterized in that, Step 2-2 describes the calculation of the phase of the rearranged echo signal using the differential cross-multiplication algorithm. The calculation formula is as follows: In the formula, For the phase of the slow time difference frequency signal, s I (n) and s Q (n) represents the IQ signal of the quadrature branch, and n is the frequency modulation period T. r The number of sample points in the sampling period.

6. The method for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation according to claim 5, characterized in that, Step 2-3, which involves filtering the phase signal, specifically includes: A bandpass filter h(t) of order L⁻¹ with a passband frequency of 0.15–0.55 Hz is used to filter the phase signal. This frequency range corresponds to the basic breathing band. The resulting mixed signal x(t) is calculated using the following formula:

7. The method for detecting multiple respiratory signals based on LFMCW radar and underdetermined blind source separation according to claim 6, characterized in that, Step 4 describes the process based on the estimated mixing matrix. Restoring respiratory signals specifically includes: Step 4-1: Obtain the optimal solution to the underdetermined problem based on the mixing matrix. When the number of receiving antennas is less than the number of people to be tested, the blind source separation problem is underdetermined, the number of unknown sources is greater than the number of equations, and the equations do not have a unique solution. Introducing constraints to minimize the L1 norm of the solution transforms source recovery into an optimization problem. The objective function can be expressed as: In the formula, s(t,f) is the time-frequency domain estimate of each respiratory signal source, and X(t,f) is the mixed respiratory signal in the time-frequency domain. The mixture matrix is ​​calculated in step 3-2; For a two-antenna receiving system, the steps of the L1 norm minimization algorithm are as follows: ① Calculate the mixture matrix A full-rank submatrix, which is 2 rows and 2 columns long and has a total length of 260 ... 1, denoted as a k , ② Find all possible solutions for a point X(t,f) in the sparse domain. ③ Calculate the L1 norm of each solution and take the one with the smallest norm. As the optimal estimate of the source signal, that is: ④ Repeat steps ①②③ to find the optimal solution for all points in the sparse domain; Step 4-2: Restore the time-frequency domain respiratory signal separated in Step 4-1 to the time domain, using the following formula: In the formula, h(τ-t) is a rectangular window. The recovered time-domain signal.

8. A multi-person respiratory signal detection system based on LFMCW radar and underdetermined blind source separation, characterized in that, The system includes: The radar echo signal acquisition module is used to detect the chest movement of multiple human targets within a gate at the same distance using a dual-receiving antenna LFMCW radar, and obtain the echo signals from the two receiving antennas. The preprocessing module is used to preprocess the radar echo signal corresponding to each receiving antenna to obtain a mixed signal of two multiple breathing signals. The mixed signal of each multiple breathing signal is a linear superposition of the multiple breathing signals of that signal. The mixed signal is represented in matrix form as follows: In the formula, The mixed signal vector representing the two mixed signals based on the blind source separation model. Source signal vector s1,s2,…,s P Let A be the respiratory signals of P individuals, and let A be a 2-row, P-column mixed matrix. The mixing matrix estimation module is used to estimate the mixing matrix by using a short-time Fourier transform to transform the mixed signal vector. Transform to the time-frequency domain to obtain the mixed signal vector in the time-frequency domain. Time-frequency points X1 and X2 form a clustering plane. An improved nearest neighbor propagation algorithm based on the fusion of angle and energy information is used for cluster analysis to obtain the number of clusters, cluster centers, and cluster count. This is an estimate of the number of people to be tested, and the cluster centers are an estimate of the mixture matrix, expressed as: In the formula, (X1(k i ),X2(k i )) represents the cluster center of the i-th class on the clustering plane, k i This is the index value of the cluster center of the i-th cluster; Source recovery module, used to recover the estimated mixture matrix Respiratory signal recovery is performed. Under underdetermined conditions, when the number of targets exceeds the number of receiving antennas, respiratory signal recovery transforms into an optimization problem. By introducing constraints and minimizing the L1 norm, the optimal solution is obtained, leading to an estimate of the respiratory signal in the time-frequency domain. Finally, Performing an inverse short-time Fourier transform yields a time-domain estimate of the respiratory signal. The estimation of the mixture matrix specifically includes: Step 3-1: Transform the mixed signal of multiple people's breathing into the time-frequency domain to sparsify the mixed signal for cluster analysis; the specific process includes: ① Perform a short-time Fourier transform on the mixed signal x(t) of multiple people's breathing corresponding to the two receiving antennas: X i (t,f)=∫x i (τ)·g(t-τ)·e -j2πft dτt∈(0,L t -1),f∈(0,L f -1) In the formula, g(t0) is a Gaussian window function, x i (t0 is the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, X) i (t,f) is the time-frequency matrix of the mixed breathing signal of multiple people corresponding to the i-th receiving antenna, L f L represents the number of rows in the matrix and the number of FFT points. t The number of columns in the matrix represents the number of time frames. ② Take the positive frequency band of the time-frequency matrix, that is, the frequency f takes a value from 0 to L. f / 2-1, let matrix X i (t,f) are rearranged into a one-dimensional vector X. i (q), the rearrangement formula is: X i (q)=X i (t,f) q=t+f·L t In the formula, q is the index number of the time-frequency vector. The time-frequency vectors corresponding to the two receiving antennas are X1(q) and X2(q). X1(q) and X2(q) form a two-dimensional clustering plane. The point X*(q) on the plane is defined as (X1(q), X2(q)). Step 3-2: Based on the improved nearest neighbor propagation algorithm that integrates angle and energy information, estimate the number of people to be tested and the mixing matrix; The improved nearest neighbor propagation algorithm process includes: ① Calculate the similarity matrix s(i,k). The similarity matrix measures the similarity of points X in the clustering plane. * (i) and point X * The similarity of (k) is defined in the similarity matrix as follows: s(i,k)=abs[X * (k)]·exp(-ρ·(1-cos(∠[X * (i),X * (k)]))) In the formula, abs[X * [k] represents the cluster plane point X. * The modulus of (k); ∠[X * (i),X * [k] represents point X * (i) and point X * The angle between (k); ρ is the attenuation coefficient of the nonlinear function. The larger the value of ρ, the faster the attenuation and the greater the effect of insufficiently sparse points on clustering. ② Initialize the attraction matrix r(i,k) and the membership matrix a(i,k) to 0; where the attraction matrix represents the suitability of data point k as the cluster center of data point i; and the membership matrix represents whether data point i chooses data point k as its cluster center. ③ Update the attraction matrix r(i,k) r t+1 (i,k)=s(i,k)-max j≠k [a t (i,j)+s(i,j)] In the formula, r t+1 (i,k) represents the attractiveness in the next iteration, a t (i,k) represents the degree of belonging in the current iteration; ④ Update the membership matrix a(i,k) ⑤ Use the attenuation coefficient λ to weight the current result of steps ③ and ④ with the result of the previous iteration. Usually, the attenuation coefficient λ = 0.

5. The weighting formula is as follows: r t+1 (i,k)=λ·r t (i,k)+(1-λ)·r t+1 (i,k) a t+1 (i,k)=λ·a t (i,k)+(1-λ)·a t+1 (i,k) ⑥ Repeat steps ③, ④, and ⑤ until the matrix converges or the maximum number of iterations is reached, at which point the algorithm terminates; ⑦ Determine data point X * (i) Cluster center X * (k)=(X1(k),X2(k)), k should satisfy the formula: max{r(i,k)+a(i,k)} Number of cluster centers This is an estimate of the number of people to be tested. The cluster centers, or mixture matrix, are represented as follows: In the formula, X * (k i )=(X1(k i ),X2(k i )) represents the cluster center of the i-th class, k i This is the index value of the cluster center of the i-th cluster; This is an estimate of the number of people to be tested.

Images