Method and system for contactless vital sign monitoring

Abstract

The present invention relates to a method for vital sign monitoring using a radar sensor system (1), the radar sensor system (1) comprising a transmitter (2), a receiver (3) and a processing device (5), wherein: the transmitter (2) irradiates (100) at least one body region (21) of a person (20) with radar radiation (30); the receiver (3) generates (110) a receiver signal based on reflected radiation (11) from the at least one body region; the processing device (5) generates a displacement signal for each of a plurality of processing windows based on the receiver signal, and calculates an estimated fundamental frequency for at least one oscillatory motion, the displacement signal representing body motion, the body motion including respiratory motion and heartbeat motion as oscillatory motion. In order to provide a reliable and efficient means for vital sign monitoring, the present invention provides that the processing device (5) applies an adaptive nonlinear least squares method based on the displacement signal to calculate (250) a plurality of frequency estimates (f r1 、f r2 、f r3 ), each of the plurality of frequency estimates corresponds to one of the plurality of harmonics of the first oscillatory motion, wherein the processing device (5) separates the search area (A h1 、A h2 、A h3 ) for each frequency estimate (f r1 、f r2 、f r3 ), adapting at least one search area (A) for at least one processing window h1 、A h2 、A h3 ), and frequency estimation based on (f r1 、f r2 、f r3 ) calculates (300) a first estimated fundamental frequency (f h ).

Description

Technical Field

[0001] The present invention relates to a method and system for contactless vital sign monitoring. Background Art

[0002] Continuous monitoring of vital signs plays a crucial role in the early detection of conditions that could impact a patient's health. The most important vital signs are respiratory rate and heart rate. These parameters are key physiological parameters, and by continuously monitoring them, it's possible to detect various abnormalities, such as drowsiness, sleep apnea, and even depression. However, conventional monitoring devices, often connected via cables, restrict patient mobility, can cause discomfort, and even lead to skin damage. Therefore, they are insufficient for long-term monitoring.

[0003] On the other hand, it's possible to monitor vital signs using radar-based devices. This contactless monitoring offers several advantages over standard devices. Unlike cameras, radar signals can penetrate various materials and are unaffected by skin pigmentation or ambient light levels. Unlike wearable sensors, radar systems don't require the user to wear or carry any additional equipment. Furthermore, radar devices maintain privacy and can operate at low power and cost.

[0004] However, despite recent progress, accurate vital sign monitoring remains challenging in real-world scenarios, especially with regard to heart rate estimation. Radar-based vital sign processing relies on phase analysis of the backscattered signal, which corresponds to the chest wall displacement caused by the respiratory and cardiac mechanisms. The recovered displacement signal typically contains not only the respiratory and cardiac fundamental frequencies, but also their associated upper harmonics. With respect to their respective amplitudes, the cardiac motion is typically several times smaller than the respiratory motion. In addition, the cardiac fundamental frequency can be similar to the upper harmonics of the respiratory frequency. For these reasons, the heartbeat signal can be easily buried in the background noise or masked by the higher-order harmonics of the respiratory signal. In addition, additional frequency components and intermodulation products may also be present in the recovered signal, originating from different sources, including radar nonlinearities, phase demodulation problems (see B.K. Park, O. Boric-Lubecke, and V.M. Lubecke, “Arctangent demodulation with DC offset compensation in quadrature Doppler radar receiver systems,” IEEE Transactions on Microwave Theory and Techniques, vol. 55, no. 5, pp. 1073-1078, 2007), and random body movements of the monitored subject (see Q. Lv, L. Chen, K. An, J. Wang, H. Li, D. Ye, J. Huangfu, C. Li, and L. Ran, “Doppler Vital Signs Detection in the Presence of Large-Scale Random Body Movements,” IEEE Transactions on Microwave Theory and Techniques, vol. 66, no. 9, pp. 4261-4270, 2018). These interfering elements may dominate in the neighborhood of the heartbeat fundamental frequency, often preventing detection and hampering heart rate estimation.

[0005] Harmonic interference from respiration in heartbeat signals has been widely reported as one of the main problems in radar-based vital sign monitoring. While respiration estimates are typically limited only by noise, the spectral region containing the fundamental frequency of the heartbeat is dominated by higher harmonics of the respiratory motion. In this way, spectral overlap can occur, and heartbeat peaks can end up being masked by respiration harmonics. Depending on the signal-to-interference ratio (SIR) and the specific combination of fundamental respiration / heartbeat frequencies, most techniques are unable to provide robust and accurate heart rate estimates.

[0006] To avoid such interference, a method based on the concept of harmonic-based heart rate estimation has been proposed (see Y. Rong and DW Bliss, “Remote Sensing for Vital Information Based on Spectral-Domain Harmonic Signatures,” IEEE Transactions on Aerospace and Electronic Systems, vol. 55, no. 6, pp. 3454-3465, 2019). Although the fundamental heart rate frequency is limited by respiratory interference, its higher-order harmonics are noise-limited. In this method, a conventional discrete Fourier transform (DFT) is used to generate the spectrum, and the second harmonic of the heart rate signal is mainly used to perform the estimation. The first harmonic is only used when the primary estimate fluctuates violently. However, this method has some limitations. In order to detect the second-order harmonic of the heart rate signal, a relatively high signal-to-noise ratio (SNR) is required, and the tracking performance is very sensitive to background noise and random body motion. Due to these limitations, it only works when the monitored object is sitting completely still and is no more than 80 cm away from the radar.

[0007] To improve the SNR used for estimation, et al. have proposed a nonlinear least squares (NLS) method for respiratory rate monitoring (see G. R.Stutz, F.Hornberger, W.A.Martins, D.Tatarinov, M.Alaee-Kerahroodi, U.Lindner, L.Stock, E.Kaiser, S.Goedicke-Fritz, U.S.Schroeder, B.S.M.R., and M.Zemlin, “Contactless radar-based breathing monitoring of premature infants in the neonatal intensive care unit,” Scientific Reports, vol.12, no.1, pp.1-15, 2022). It uses an initial (rough) estimate as a reference for the NLS search area, which is restricted around this estimate. This initial estimate is obtained using an additional algorithm applied directly on the time domain signal. However, the NLS estimation performance may be affected by a nonlinear objective function with multiple peaks and very sharp global maxima. Therefore, finding the fundamental frequency by the search algorithm requires accurate initialization. In addition, in certain cases, especially considering heart rate estimation, the NLS search area is limited by the initial estimate. Some of the assumptions made by et al. may not hold. In particular, it is not always guaranteed that the frequency components in the NLS spectrum are distinct and well separated.

[0008] When operating at higher frequencies with larger bandwidths, radar range resolution can be less than a few centimeters. Under these conditions, the human body is an extended target and its energy can be distributed across adjacent range bins. Sun et al. reported this effect. (See L.Sun, S.Huang, Y.Li, C.Gu, H.Pan, H.Hong, and X.Zhu, “Remote Measurement of Human Vital Signs Based on Joint-Range Adaptive EEMD,” IEEE Access, vol.8, pp.68 514-68 524, 2020), who noted that the power of physiological motion from different parts of the human body will not be concentrated in one range bin, but in several consecutive bins. In this case, if only data from one range bin is used, the performance will be degraded due to the loss of signal power. To overcome this problem, Sun et al. proposed to process each range bin separately and independently, and finally combine several separated heartbeat components for subsequent frequency estimation. Despite promising results, this method increases the computational complexity of the system due to the need to process several range bins independently. Furthermore, the performance is still limited by the residual energy in each range bin.

[0009] Technical issues

[0010] The object of the present invention is to provide reliable and efficient means for vital sign monitoring. The object is achieved by a method according to claim 1 and a system according to claim 15 . Summary of the Invention

[0011] The present invention provides a method for monitoring vital signs using a radar sensor system. The radar sensor system includes a transmitter, a receiver, and a processing device.

[0012] Although the transmitter and receiver are at least partially distinct, separate components, the antenna of the radar sensor system can be used by both the transmitter (as a transmitting antenna) and the receiver (as a receiving antenna). The processing device includes hardware components, but some of its functions can be implemented in software. The processing device can also include any type of volatile or non-volatile memory device. Typically, the processing device is connected to the receiver via a wired connection. It can also be connected to the transmitter to monitor and / or control the operation of the transmitter. The processing device can include multiple components or modules that can be arranged adjacent to each other or at a certain distance from each other and connected for information transmission and / or power transmission.

[0013] According to the method, a transmitter irradiates at least one body region of a person with radar radiation. As the term "radar" implies, the radiation is an electromagnetic wave. The electromagnetic wave may comprise one or more frequencies, with at least one frequency preferably being between 1 GHz and 300 GHz. The radiation may be a continuous wave (CW), a phase modulated continuous wave (PMCM), a frequency modulated continuous wave (FMCW), an ultra-wideband wave (UWB), or other suitable waveform. The radiation from the transmitter irradiates at least one body region of the person. This may specifically be the chest region or a region including the chest region (e.g., the entire torso). Typically, the person's skin should be irradiated so that the body region should be exposed or any clothing should be largely transparent to the radiation. The person may be an adult or an infant. Since the method of the present invention is a non-contact method, it can be used on any type of person, including children, without causing any interference, irritation, or epidermal damage. Although some transmission through the person's body or absorption by the body may occur, the majority of the radiation is reflected by the irradiated body region.

[0014] The receiver generates a receiver signal based on reflected radiation from at least one body region. The reflected radiation may be a superposition of radiation reflected from different body regions and / or different portions of a single body region. Preferably, the transmitter uses the same antenna to transmit the radiation, and the receiver uses the same antenna to receive the reflected radiation. However, different antennas may be used for the transmitter and receiver. Typically, the receiver signal is an electrical signal generated by the receiver due to the reflected radiation, or an electrical signal generated by the receiver due to the portion of the reflected radiation that reaches the receiver's antenna. Optionally, the receiver may perform operations such as demodulation or amplification to generate the receiver signal.

[0015] Based on the receiver signal, a processing device generates a displacement signal for each of a plurality of processing windows, and calculates an estimated fundamental frequency for at least one oscillatory motion. The displacement signal represents body motion, including respiratory motion and cardiac motion as oscillatory motions. The displacement signal is typically a time-varying signal. It represents or indicates body motion, such as chest wall motion. For example, it can be proportional to the body motion, excluding noise and / or contributions from other sources. However, the relationship between body motion and the displacement signal can vary and be more complex. Body motion includes at least two oscillatory motions: respiratory motion and cardiac motion. Each oscillatory motion can be quasi-periodic or nearly periodic. Typically, body motion is a superposition of respiratory motion during cardiac motion and possible additional motion (such as random body motion). Other oscillatory motions may also exist, for example, if a person has tremors. The processing device can use various techniques to generate the displacement signal, some of which are discussed below. A processing window is a time window or time interval. The processing device performs the operations described herein not only for a single processing window, but also for a series of processing windows. Two consecutive processing windows can immediately follow each other, but there can also be a period of time between them, such as a "dead time." It is also conceivable that there is an overlap between the two processing windows, such that window "A" begins and window "B" begins thereafter, but before window "A" ends. In addition, for each processing window, the processing device calculates an estimated fundamental frequency for at least one oscillatory motion. Preferably, the processing device calculates an estimated fundamental frequency for the cardiac motion and an estimated fundamental frequency for the respiratory motion. Since these motions are oscillatory and generally quasi-periodic, they can be characterized by a fundamental frequency, at least within the limits of the processing window. It should be understood that none of these motions are true sinusoids, so they can be characterized by a fundamental frequency plus higher harmonics (i.e., second, third, fourth harmonics, etc.). Here and hereinafter, the terms "fundamental frequency" and "first harmonic" are synonymous. The estimated fundamental frequency is calculated by the processing device and can deviate from the true fundamental frequency, where the term "estimated" is used to indicate the possibility of deviation. However, it should be noted that the method of the present invention has been proven to be reliable and generally results in high accuracy.

[0016] According to the invention, based on the displacement signal, the processing device applies an adaptive nonlinear least squares method to calculate a plurality of frequency estimates, each frequency estimate corresponding to one of a plurality of harmonics of the first oscillatory motion, wherein the processing device uses a separate search region for each frequency estimate, adapts at least one search region for at least one processing window, and calculates a first estimated fundamental frequency for the first oscillatory motion based on the frequency estimates.

[0017] The first oscillatory motion is one of the above-mentioned oscillatory motions, typically a heartbeat motion or a breathing motion, although this can also be an oscillatory motion caused by, for example, a tremor. The estimated fundamental frequency associated with this first oscillatory motion is called the "first estimated fundamental frequency". In order to find the first estimated fundamental frequency, the processing unit calculates a plurality of frequency estimates. These estimates correspond to different harmonics of the first oscillatory motion. It can be said that the processing unit identifies the frequencies of the plurality of harmonics, which are referred to herein as "frequency estimates". In other words, the processing unit not only attempts to identify a single harmonic (e.g. the first harmonic or fundamental frequency), but also attempts to identify a plurality of harmonics, which always include at least one higher harmonic. However, preferably, one of the harmonics is the fundamental frequency, i.e. one frequency estimate corresponds to the fundamental frequency as identified by the processing unit. The processing unit applies an adaptive nonlinear least squares (NLS) method to determine the frequency estimate.

[0018] More specifically, an adaptive NLS method is used to independently calculate each frequency estimate. The processing device uses a separate search region for each frequency estimate. The search for the corresponding frequency estimate is limited to the search region, which helps speed up the identification of the frequency estimate. A search region is a frequency region or frequency interval. There is a search region for each frequency estimate, that is, a search region for each different harmonic. Typically, but not exclusively, the search regions do not overlap. The independent definition of the search regions is reasonable because lower harmonics are found in regions other than higher harmonics. Furthermore, the processing device adapts at least one search region for at least one processing window. In other words, for each processing window, at least one search region does not remain constant but is adapted. It should be understood that information from previous processing windows can be used for adaptation. Adapting the search region can mean changing its lower boundary, its upper boundary, or both. It can also mean changing the size of the search region—making it smaller or larger, or moving the search region to a higher or lower frequency without changing its size. Typically, each search region is adapted for each processing window. Because the processing unit adapts at least one search region, the NLS method is referred to as an "adaptive" NLS method. The frequency estimates can be considered as "candidates" for the first estimated fundamental frequency or its higher harmonics, and the fundamental frequency can be calculated from the higher harmonics in a conventional manner. The first estimated fundamental frequency is then calculated by the processing unit based on these frequency estimates. "Calculation" should be understood in a broad sense and can, for example, refer to identifying a frequency estimate as the first estimated fundamental frequency.

[0019] Since the method of the present invention uses frequency estimates corresponding to different harmonics, the chance of finding a reasonably accurate fundamental frequency increases. Specifically, if one of the harmonics cannot be safely identified, it is possible that this can be compensated by one of the other harmonics. This is especially true if one of the harmonics (such as the fundamental frequency) is close to at least one harmonic of another oscillatory motion. In particular, if the other harmonic has a similar or even larger amplitude, this may make correct identification impossible. However, other harmonics, in particular higher harmonics of the first oscillatory motion, may be located in an area where there is no interference from another oscillatory motion or in an area where even higher harmonics of the oscillatory motion are located, which area typically has a much lower amplitude. In addition, the definition and dynamic adaptation of independent search areas help to reduce the computational workload.

[0020] Preferably, the transmitter emits radar radiation with a frequency between 50 GHz and 100 GHz. While this method is suitable for constant-frequency CW, in other embodiments the frequency can be time-varying, for example when using PMCM or FMCW. In the case of FMCW, the center frequency can be, for example, between 70 GHz and 90 GHz, and the bandwidth can be between 3 GHz and 10 GHz. However, other options are possible. In the frequency range defined above, the wavelength is in the millimeter range. Within the range defined for the bandwidth, the distance resolution is in the range of several centimeters. Under these conditions, even the human body of an infant must be considered an extended object. This also applies to the chest area, for example. In other words, signals from different distinguishable distances may be correlated.

[0021] Preferably, at least three frequency estimates are calculated. In particular, these frequency estimates may correspond to a first harmonic (fundamental frequency), a second harmonic, and a third harmonic. Thus, preferably, the processing device uses at least one search area for the fundamental frequency and at least one different search area for higher harmonics. More specifically, it may use a first search area for the fundamental frequency and a second search area and a third search area for two higher harmonics (e.g., for the second harmonic and the third harmonic).

[0022] Specifically, the first oscillatory motion may be a heartbeat motion. The fundamental frequency of a typical heartbeat motion may be in the same frequency range as the second, third, fourth, and / or fifth harmonics of the respiratory motion. Therefore, it may be difficult to correctly identify the fundamental frequency under all conditions. On the other hand, the second and third harmonics of the heartbeat motion are typically in the frequency range of the sixth harmonic and even higher harmonics, all of which are quite weak and generally do not interfere with the identification of the heartbeat harmonics.

[0023] There are various possibilities of how the first estimated fundamental frequency can be based on the frequency estimates. For example, it can be based on several frequency estimates, e.g. a linear combination. A weighting factor can be assigned to each frequency estimate, wherein the weighting factor represents the reliability or quality of the respective frequency estimate. Another embodiment provides that the processing device calculates the first estimated fundamental frequency such that it corresponds to one of the frequency estimates. If the frequency estimate is considered to be the fundamental frequency, the first estimated fundamental frequency is set equal to this frequency estimate. On the other hand, if the frequency estimate is considered to be an nth harmonic, the first estimated frequency is set equal to 1 / n of the frequency estimate. This approach relies on the concept that only one frequency estimate can be accurate and that any inclusion of other frequency estimates will not improve the accuracy.

[0024] Preferably, the processing device generates an in-phase signal and a quadrature signal based on the receiver signal. As the term implies, the quadrature signal has a 90° (or ) phase shift. The term "based on" includes the possibility that one of the in-phase signal and the quadrature signal is identical to the receiver signal. However, typically, the processing device performs some signal processing, such as mixing and / or filtering, to obtain the in-phase signal from the receiver signal. The quadrature signal is obtained by mixing with a 90° shifted transmit signal.

[0025] Preferably, the processing device performs at least a one-dimensional discrete Fourier transform based on the in-phase signal and the orthogonal signal to obtain a slow-time signal for each of a plurality of position bins, each of the position bins corresponding to at least a one-dimensional position relative to the transmitter. Typically, a discrete Fourier transform (DFT) is performed at least with respect to the distance from the transmitter (i.e., the spacing), thereby generating a plurality of range bins. Alternatively, a DFT is typically additionally performed with respect to the angle, thereby creating a plurality of angle bins. As described above, the range resolution of the radar sensor system can be several centimeters, which then corresponds to the distance interval represented by the range bin. Thus, the illuminated body area can extend over several range bins and / or angle bins.

[0026] Additionally, the processing device can perform complex phase demodulation on the slow-time signal for each position bin to retrieve a single-bin displacement signal for the corresponding position bin. The single-bin displacement signal then represents the body motion in the corresponding position bin. However, it may also include a considerable amount of noise, i.e., the signal-to-noise ratio may be relatively low.

[0027] This problem can be alleviated in a preferred embodiment, according to which the processing device combines the single-bin displacement signals from a plurality of position bins to obtain an aggregate displacement signal. Typically, these are adjacent position bins, i.e. they form a coherent group of position bins. For example, the single-bin displacement signal of one distance bin, which is considered to correspond to the main part of the chest area, can be combined with the signal of an adjacent distance bin, which is located on the far side of the first bin relative to the transmitter. The same can be applied to different angular bins. The combination can take various forms. For example, the aggregate displacement signal can be a certain linear combination of the single-bin displacement signals. In particular, the processing device can calculate the aggregate displacement signal by adding the single-bin displacement signals. If a specific oscillatory motion is present in all these single-bin displacement signals, a suitably chosen combination (e.g. addition) enhances the SNR.

[0028] For example, if relevant parts of a body region can always be assumed to be at a given location, it would be conceivable to always combine the same single-bin displacement signals. However, in many applications, it is not possible to foresee which position bins should be combined. Therefore, it is advantageous to dynamically or adaptively find the corresponding position bins. According to one such embodiment, the processing device combines the single-bin displacement signals based on their correlation. In other words, the processing device calculates the correlation between two single-bin displacement signals, and the decision whether to combine these single-bin displacement signals depends on the correlation. Since any noise from different position bins is generally uncorrelated, it is reasonable to assume that if a body part is located in two position bins, the correlation is mainly enhanced by body motion. In this context, "correlation" should be understood in a broad sense, and it will be appreciated that various types of correlation can be used as a criterion. Specifically, the Pearson correlation coefficient can be calculated. As is known in the art, the Pearson correlation coefficient can assume values ​​between 0 (no correlation) and 1 (maximum correlation). A threshold value can be defined for the correlation, and if the correlation is above the threshold value, the single-bin signals are combined. However, other possibilities exist. For example, the weights of the single-bin displacement signals can be selected based on the correlation. In the case of the Pearson correlation coefficient, the threshold is of course between 0 and 1, and is usually above 0.5. Typically, the threshold is between 0.6 and 0.9.

[0029] Regardless of whether the single bin displacement signals are combined or not, the processing unit may apply at least one bandpass filter to the (single bin or aggregate) displacement signal. Although the signal quality may be improved by a common bandpass filter, it is advantageous to apply multiple bandpass filters, i.e. one bandpass filter for each oscillation motion. This is because the relevant frequencies of the respiratory motion are in a frequency range that is much lower than the frequency range of the cardiac motion. As an example, a suitable bandpass filter for the respiratory motion may be permeable between 0.1 Hz and 3 Hz, whereas a suitable bandpass filter for the cardiac motion may be permeable from 0.5 Hz to 5 Hz. This corresponds to the physiological range of the fundamental frequency, including some higher harmonics. The bandpass filtered signal may then be used to calculate the fundamental frequency in the following steps.

[0030] Preferably, the adaptive nonlinear least squares method is based on minimizing the difference between the displacement signal and a model function having a plurality of harmonics for each of the at least one oscillatory motion. The model function can also be viewed as a (finite) Fourier series for the at least one oscillatory motion.

[0031] Strictly speaking, the model function should have a Fourier series with a fundamental frequency and at least one higher harmonic for each oscillatory motion, i.e., one for respiratory motion and one for cardiac motion. In other words, the displacement signal d(t) can be modeled as a superposition of K sources, with the harmonic correlation L for the kth source being k The normalized angular frequency (in radians / sample) is related to the physiological frequency f k (in Hertz) related to, e.g. where fs is the slow-time sampling frequency determined by the time between transmitted frames. After sampling, the model for the chest wall displacement signal can be written as (see MG Christensen and A. Jakobsson, “Multi-pitchestimation,” Synthesis Lectures on Speech & Audio Processing, vol. 5, no. 1, pp. 1-160, 2009):

[0032]

[0033] in is the complex amplitude of the lth harmonic, and n is the sample index. Now, let us consider a single source k, and define d k , by d k [n] is a vector of N consecutive samples, which can be expressed as

[0034] d k =Z k a k , (Equation 2)

[0035] where a k is a vector containing the complex amplitudes of the harmonics, and the matrix Z k is defined as

[0036] Z k =[z(ω k ) z(2ω k )…z(L k ω k )], (Equation 3)

[0037] and z(ω) is the vector at frequency ω. Using these definitions, the signal model can be rewritten as

[0038]

[0039] In order to obtain a frequency estimate, we must find a set of fundamental frequencies that minimizes the difference between the recovered displacement signal and the signal model. In this way, the NLS optimization problem can be expressed as

[0040]

[0041] This problem is quite challenging and requires considerable computational effort. However, if the frequency of an oscillatory motion (i.e., Z k If all frequencies in are different and well separated, it is possible to solve the minimization problem for each oscillatory motion independently, as a good approximation to the general solution. The solution can be approximated by finding the fundamental frequency for each source, i.e.

[0042]

[0043] The goal of the NLS method is then to find the fundamental frequency (together with its corresponding higher harmonics) that minimizes the difference between the model function and the displacement signal. It will be appreciated that this may be an aggregated and / or bandpass filtered displacement signal. However, although the NLS method is based on minimization of the difference, there are certain embodiments in which the difference does not have to be calculated explicitly, as will be explained below.

[0044] In a preferred embodiment, the processing device calculates at least one frequency of an oscillatory motion using a cost function based on a fast Fourier transform of the displacement signal. In particular, the cost function to be minimized may be the spectral density of the displacement signal. Under certain conditions, this is at least approximately equivalent to minimizing the difference between the above-mentioned model function and the displacement signal. As shown by Christensen et al., Equation 6 is equivalent to

[0045]

[0046] The resulting cost function can be written as

[0047] It corresponds to the FFT at the harmonic frequency ω k l evaluated and summed, The power spectral density estimate of the periodogram of the frequency domain is obtained by performing a periodogram calculation. The frequency calculated in this embodiment may specifically be an estimated fundamental frequency or a frequency estimate. It will be appreciated that the use of a Fast Fourier Transform (FFT) greatly helps to reduce the numerical burden, thereby quickly providing results for the corresponding frequencies without requiring excessive computing power.

[0048] In one embodiment, the processing device uses the estimated fundamental frequency calculated in one processing window to define a search area for frequencies in a later processing window. In most cases, the fundamental frequency of the respiratory motion and the fundamental frequency of the cardiac motion do not change abruptly from one processing window to the next. Therefore, the fundamental frequency in the consecutive processing windows is mostly found near the estimated fundamental frequency calculated in the previous processing window. Similarly, the position of the higher harmonics usually changes only to a limited extent from one processing window to the next. Therefore, the search area for the fundamental frequency and higher harmonics can be defined based on the estimated fundamental frequency found previously. This greatly helps to reduce the computational burden because the search area can be kept relatively narrow.

[0049] One embodiment provides that the processing device calculates, for at least one processing window, a predicted state comprising a predicted fundamental frequency based on at least one previous processing window, and selects a frequency estimate as the estimated fundamental frequency based on a comparison between the predicted fundamental frequency and the frequency estimate. In addition to the predicted fundamental frequency, the predicted state may also include a predicted first time derivative and a predicted second time derivative of the fundamental frequency. The latter two correspond to the "speed" and "acceleration" of the possible changes of the fundamental frequency. In this case, the general state vector may have the following form:

[0050] x=[f h f′ h f″ h ] T , (Equation 9)

[0051] where f h , f′ h and f″ h Represent the heartbeat fundamental frequency, its first time derivative and its second time derivative respectively. If the corresponding state of the previous processing window is known, then if the time difference between the processing windows is also known, the predicted state can be calculated. According to the current state At the mth processing window, the state transition equation can be used to calculate the predicted state Defined as

[0052]

[0053] Where F is the state transition matrix, where Δt is the time from one processing window to the next. The above equation corresponds to a constant acceleration model. The predicted state is usually different from the actual state in the corresponding processing window. However, for short time periods, it can be expected that the actual state is at least close to the predicted state. Therefore, the frequency estimate closest to the predicted fundamental frequency will usually be the most accurate frequency estimate. Therefore, this frequency estimate can be selected as the estimated fundamental frequency. Other frequency estimates can be discarded. However, another threshold (hereinafter also referred to as the gating threshold) can also be defined for the distance between the predicted fundamental frequency and the frequency estimate. In this context, various types of distances can be used, in particular the Mahalanobis distance. If the distance is above the threshold, the corresponding frequency estimate is considered an outlier. In this case, the frequency estimate is discarded in any case. This may result in all frequency estimates being discarded for the corresponding processing window. There may be various reasons for this, such as excessive noise, making frequency identification impossible for the corresponding processing window. In addition, the threshold may be too small. Therefore, the threshold can be increased for the next processing window. For this processing window, the estimated fundamental frequency can be set to the predicted fundamental frequency.

[0054] Various techniques known in the art can be used to determine the first estimated fundamental frequency based on the frequency estimates. According to one embodiment, the processing device uses a Kalman filter to calculate the first estimated fundamental frequency. The Kalman filter is a recursive Bayesian algorithm that produces accurate estimates based on noisy or uncertain measurements. When new data arrives, it updates its estimates and parameters sequentially. Therefore, it is suitable for real-time processing using overlapping sliding window methods commonly used in vital sign processing. In this context, it can also be said that the Kalman filter is used to determine which frequency estimate of the frequency estimates is (or is) used as the basis for the first estimated fundamental frequency. Typically, one frequency estimate is used, but it can also be two or more frequency estimates that are then combined (for example, in a weighted manner). In some cases, if all estimates are discarded based on the Kalman filter, the (current) frequency estimate is not used. In three search areas and three frequency estimates for the first, second and third harmonics and In the case of h Can be defined as

[0055]

[0056] The state vector representation has a mean and covariance P m,m The extrapolated covariance can be calculated as:

[0057] P m+1,m =FP m，m FT +Q, (Equation 12)

[0058] where Q is related to the process noise uncertainty and can be modeled as in:

[0059] g=[0.5Δt 2 Δt 1] T , (Equation 13)

[0060] as well as Represents process noise. Using the observation matrix

[0061]

[0062] Measurement innovation (error)e m It can be calculated as

[0063]

[0064] Its associated covariance is given by:

[0065] S m =HP m+1，m H T +R, (Equation 16)

[0066] where R is a diagonal matrix containing the uncertainty in the frequency estimate for each search region, i.e.

[0067]

[0068] We can use predefined Q and R matrices, and have covariance P 0,0 Initial state The algorithm is initialized by . If at least one of the frequency estimates is within the above-mentioned gating threshold, the frequency estimate that minimizes the distance between the new measurement and the predicted state is selected for the first estimated fundamental frequency. Subsequently, the state vector is updated as

[0069]

[0070] where j is the index corresponding to the selected NLS estimate, and the Kalman gain k m is calculated as

[0071]

[0072] in is the jth row of H and S m (j,j) is S m Finally, the associated covariance P m+1,m+1 is calculated as

[0073]

[0074] for use in the next filter iteration.

[0075] If no frequency estimate falls within the gate, all are discarded and the Kalman gain is then set to zero. If this condition holds true during several adjacent processing windows, it may indicate that the state estimate has deviated. In this case, the updated covariance may be updated to its initial (large) value to allow the filter to reacquire. The initial predefined Q and R matrices may remain constant for each iteration. However, it is also possible to adjust Q and / or R in each iteration.

[0076] The present invention also relates to a radar sensor system for vital sign monitoring, comprising:

[0077] a transmitter for irradiating at least one body region of a person with radar radiation;

[0078] a receiver for generating a receiver signal based on reflected radiation from at least one body region; and

[0079] a processing device adapted to generate a displacement signal for each of a plurality of processing windows based on the receiver signal, and to calculate an estimated fundamental frequency for at least one oscillatory motion, the displacement signal being representative of body motion, the body motion including respiratory motion and heartbeat motion as oscillatory motions,

[0080] wherein the processing device is adapted to apply an adaptive nonlinear least squares method based on the displacement signal to calculate a plurality of frequency estimates, each frequency estimate of the plurality of frequency estimates corresponding to one of a plurality of harmonics of the first oscillatory motion, wherein the processing device is adapted to use a separate search region for each frequency estimate, to adapt at least one search region for at least one processing window, and to calculate a first estimated fundamental frequency for the first oscillatory motion based on the frequency estimates.

[0081] These terms have been described above in relation to the method of the present invention and will not be explained again.Preferred embodiments of the radar sensor system correspond to the embodiments of the vital sign monitoring method of the present invention described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Preferred embodiments of the present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0083] Figure 1 A schematic diagram showing a first embodiment of a radar sensor system and a person;

[0084] Figure 3is a block diagram illustrating various functions of a radar sensor system;

[0085] Figure 2 is a flow chart illustrating an embodiment of a method for vital signs monitoring;

[0086] Figure 4 is a diagram showing the time evolution of the displacement signal; and

[0087] Figure 5-7 is a diagram showing a frequency spectrum.

[0088] DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0089] Figure 1 The schematic diagram shows the setup of a first embodiment of a method for vital sign monitoring according to the present invention. The radar sensor system 1 for respiratory monitoring used in the method comprises a transmitter 2 and a receiver 3 connected to a single antenna 4. In other words, the transmitter 2 uses the antenna 4 to transmit radar radiation 30, and the receiver 3 uses the same antenna 4 to receive reflected radar radiation 31. Both the transmitter 2 and the receiver 3 are connected to a processing device 5, which comprises hardware components, although some of its functions may be implemented in software.

[0090] like Figure 1 As shown, person 20 is positioned opposite radar sensor system 1. Transmitter 2 illuminates chest area 21 with transmit signal 30 of radar radiation. In this embodiment, the transmit signal is an FMCW signal, but other signal types, such as PMCM or CW, may also be used. Transmitter 2 may operate at 77 GHz with a bandwidth of 4 GHz. Transmit signal 30 is reflected by chest area 21, thereby generating reflected signal 31, which is then received by receiver 3 via antenna 4.

[0091] The receiver 3 generates a receiver signal based on the reflected signal 11 from the chest area 21. The chest area 21 experiences body motions including at least two oscillatory motions, namely respiratory motion and heartbeat motion. Other oscillatory motions (e.g. due to tremors) and / or non-oscillatory motions may also be present.

[0092] Figure 2 is a block diagram of the radar detection system 1, and Figure 3 After transmitting and receiving radiation in step 100, the receiver 3 generates a receiver signal (step 110), which is mixed with a replica of the transmitted signal and, after quadrature demodulation, can be approximated as

[0093]

[0094] Among them A RIndicates the received signal amplitude, f b is the beat frequency, Δφ(t) is the residual phase noise (usually neglected for short-range applications), and Δθ(t) is the time-varying phase due to body motion. The receiver signal is then provided to an analog-to-digital converter (ADC) 6 of a processing device 5, which also generates an in-phase signal b I [n] and orthogonal signal b Q [n] (step 120) and sends these signals to the preprocessor 7. To identify and extract the slow-time signal from the monitored person 20 at a specific range and angular position, a two-dimensional discrete Fourier transform (DFT) is applied to the radar data cube across the range and angle dimensions (step 130). This results in a signal for each of a plurality of range bins and angle bins. It should be noted that other preprocessing architectures can be used. For example, such an architecture can use a two-dimensional DFT across range and Doppler, and then estimate the angle.

[0095] The phase demodulator 8 performs complex phase demodulation (step 140), in which it combines the complex samples of the slow-time signal to recover the displacement signal, which contains phase information related to chest wall movement over time. In this embodiment, arc tangent demodulation (AD) is used. Therefore, the recovered single-bin displacement signal can be obtained as follows:

[0096]

[0097] where s[n] represents the sample of the complex slow-time signal at a specific distance bin, and λ c is the chirp start wavelength of the FMCW signal. Using an expansion operation, possible phase discontinuities caused by the bounded image of the inverse tangent function can be removed. Prior to AD, possible DC offsets can be compensated using, for example, the Levenberg-Marquardt algorithm.

[0098] Due to the aforementioned frequency range of transmitter 2, a range resolution of approximately 4 cm can be achieved. This not only enables resolution of closely spaced objects but also filters out nearby interference. However, under these conditions, the body of person 20 (and even the chest portion 21 ) is an extended target, and the associated single-bin displacement signal may be distributed across several range bins. Given that the abdomen and back may also be involved in respiratory / cardiac motion, vital sign information may ultimately be detected in these additional range bins.

[0099] In order to exploit this and improve the SNR before the estimation is made, the correlator 9 of the processing unit 2 performs a slow time phase correlation process. If the signals are well correlated, the single bin displacement signals of several range bins are combined. With reference to a "detection bin" (which may correspond, for example, to the center of the chest wall), the correlation with the i-th adjacent range bin and / or angle bin may be checked. Definition and As a vector with the recovered displacement signals at the detected bin and the ith adjacent range bin, the correlator 9 calculates (at 150) the Pearson correlation coefficient, given by

[0100]

[0101] in represents the covariance between these vectors, where and is the corresponding standard deviation. At 160, a check is made to see if the correlation coefficient exceeds a predetermined threshold (e.g., 0.8). If so, the single-bin displacement signals are added. This process is repeated through steps 180 and 190 for a given number of range bins. The aggregate displacement signal is calculated by summing the sufficiently correlated single-bin displacement signals. Figure 4 By way of example, a comparison of a single bin displacement signal and the resulting aggregate displacement signal is shown. It can be seen how the displacement amplitude increases with increasing SNR, while the periodicity remains constant. Figure 5 The corresponding spectrum is shown, which also confirms that the main frequency content is preserved.

[0102] To remove any residual DC values ​​and possible high frequency noise components, the aggregate displacement signal is filtered (step 200) using two bandpass filters 10 and 11, one with a bandpass Kaiser window (β = 6.5) from 0.1 Hz to 3 Hz (6-180 breaths / minute) for respiratory motion and from 0.5 Hz to 5 Hz (30-300 heartbeats / minute) for cardiac motion. This corresponds to the physiological range of the fundamental frequency (first harmonic), including several higher harmonics. Bandpass filtered signal and This may be a good approximation of the true chest wall motion and may ultimately be used for frequency estimation, which is performed by an ANLS (Adaptive Nonlinear Least Squares) module 12 of the processing device 5. This ANLS module 12 comprises a respiratory rate estimator 13, a heart rate estimator 14 and a frequency selector 15. It will be appreciated that the elements 13 to 15 of the ANLS module 12 are typically implemented in software and are therefore not physically distinct.

[0103] Using the model function of Equation 1, the respiratory rate estimator 13 and the heart rate estimator 14 solve the NLS problem by maximizing the following cost function

[0104]

[0105] It is at the harmonic frequency ω k l evaluated and summed, Therefore, the two estimators 13, 14 are effectively implemented using the FFT algorithm.

[0106] In order to determine the fundamental breathing frequency, the search area A is defined b In order to reduce the search time. For the first processing window, the search area is initialized so that it includes the entire physiological breathing range. When the estimated respiratory fundamental frequency has been calculated in step 220, the search area can be adjusted in step 230. For any subsequent processing window, the estimated respiratory fundamental frequency of the previous processing window is used to define the search area A. b , search area A b The search area is thus limited to the reference value and will change adaptively based on the change in respiratory rate over time. Steps 220 and 230 constitute the ANLS respiratory rate estimation 210, which is performed by the respiratory rate estimator 13.

[0107] The estimated fundamental heart rate is calculated in ANLS heart rate estimation 240 performed by heart rate estimator 14 and frequency selector 15. Although ANLS heart rate estimation 240 is shown as being performed after ANLS respiratory rate estimation 210, the order may be reversed or both estimations 210, 240 may be performed simultaneously.

[0108] For the fundamental heartbeat frequency, the ANLS heartbeat frequency estimator 14 calculates three different frequency estimates f r1 、f r2 、f r3 (in step 250), which corresponds to the first harmonic f h (1) (i.e., fundamental frequency f h ), second harmonic f h (2) and the third harmonic f h (3) .like Figure 6 and 7 As shown, the basic heart rate f h (1) With relatively strong respiratory exercise b (3) -f b (5)On the other hand, the second and third heartbeat harmonics f h (2) 、f h (3) With the strongly attenuated seventh, eighth and ninth order respiratory harmonics f b (7) -f b (9) share their spectral positions. Therefore, including the second and third heartbeat harmonics can significantly improve the calculation of the estimated heartbeat fundamental frequency. However, since the amplitude of these harmonics depends on the specific characteristics of the respiratory and cardiac motions, as well as many other processing parameters, there is no guarantee that these higher-order heartbeat harmonics will always be detectable. Therefore, conventional estimates based on the heartbeat fundamental frequency are still helpful.

[0109] like Figure 7 As shown, a search area A is defined for each heartbeat harmonic h1 -A h3 Separate search areas may be selected because the physiological range of heart rate for a resting healthy person 20 is typically from 50 bpm to 90 bpm. Thus, the first search area A h1 The second search area A can be defined by these values. h2 It can be defined as from 100bpm to 180bpm, and the third search area A h3 It can also be defined as from 150bpm to 270bpm.

[0110] When the frequency estimate f has been calculated r1 、f r2 、f r3 , they are provided to the frequency selector 15 which applies a Kalman filter as described above with reference to equations 9-20. In step 260, the frequency selector 15 checks whether the distance (e.g., the Mahalanobis distance, but a different definition of distance may be used) between the corresponding estimate and the predicted frequency is above a certain gating threshold. If so, the frequency estimate is discarded, and if it is not found to be the last (i.e., "third") estimate in step 270, the next estimate is selected in step 280. If the distance is not above the threshold, it is checked in step 290 whether the distance is the smallest distance so far, i.e., whether the frequency estimate is closest to the predicted frequency. If so, it is selected in step 300 as the estimated fundamental frequency of the heartbeat fh (until perhaps another frequency estimate is closer to the predicted frequency). Fundamental frequency f b 、f h It may, for example, be provided to a display (not shown).

[0111] In step 310, the state vector is updated The Kalman gain k is calculated according to equations 18-20 (although other variations of the Kalman equation can also be used) m and the associated covariance P m+1,m+1 . In addition, a new predicted state can be calculated, making it possible to adapt the search area for the next processing window entered at step 320 before the method returns to step 100. If no frequency estimate falls within the gate, that is, all frequency estimates are discarded, then the Kalman gain is set to zero and the final estimate is based only on the predicted state. If this condition holds during several adjacent processing windows, it may indicate that the state estimate has drifted. In this case, the updated covariance is reset to its initial (large) value to allow the filter to reacquire.

Claims

1. A method for vital sign monitoring using a radar sensor system (1), the radar sensor system (1) comprising a transmitter (2), a receiver (3) and a processing device (5), wherein: The transmitter (2) irradiates (100) at least one body region (21) of a person (20) with radar radiation (30); The receiver (3) generates (110) a receiver signal based on the reflected radiation (31) from the at least one body region; The processing device (5) generates a displacement signal for each of a plurality of processing windows based on the receiver signal, and calculates an estimated fundamental frequency for at least one oscillatory motion, the displacement signal representing body motion including respiratory motion and heartbeat motion as oscillatory motions, wherein the processing device (5) applies an adaptive nonlinear least squares method to calculate (250) a plurality of frequency estimates (f r1 、f r2 、f r3 ), each of the plurality of frequency estimates corresponds to one of the plurality of harmonics of the first oscillatory motion, wherein the processing device (5) separates the search area (A h1 、A h2 、A h3 ) for each frequency estimate (f r1 、f r2 、f r3 ), adapting at least one search area (A) for at least one processing window h1 、A h2 、A h3 ), and based on the frequency estimate (f r1 、f r2 、f r3 ) calculate (300) a first estimated fundamental frequency (f h ).

2. The method according to claim 1, characterized in that The processing device (5) sets at least one search area (A h1 、A h2 、A h3 ) for the base frequency, and at least one different search area (A h1 、A h2 、A h3 ) for higher harmonics.

3. The method according to claim 1 or 2, characterized in that The first oscillatory motion is the heartbeat motion.

4. The method according to claim 1 or 2, characterized in that The processing device (5) calculates (300) the first estimated fundamental frequency (f h ), so that the first estimated fundamental frequency (f h ) corresponds to the frequency estimate (f r1 、f r2 、f r3 ) is a frequency estimate in .

5. The method according to claim 1 or 2, characterized in that The processing device (5) generates (120) an in-phase signal and a quadrature signal based on the receiver signal.

6. The method according to claim 5, characterized in that The processing device (5) performs (130) at least one-dimensional discrete Fourier transform based on the in-phase signal and the quadrature signal to obtain a slow-time signal for each of a plurality of position bins, each of the position bins corresponding to at least one-dimensional position relative to the transmitter (2).

7. The method according to claim 6, characterized in that The processing device (5) performs (140) complex phase demodulation on the slow-time signal of each position bin to retrieve the single-bin displacement signal of the corresponding position bin.

8. The method according to claim 7, characterized in that The processing device (5) combines (170) the single-bin displacement signals from a plurality of position bins to obtain an aggregate displacement signal.

9. The method according to claim 8, characterized in that The processing device (5) combines the single-bin displacement signals according to the correlation between the single-bin displacement signals (170).

10. The method according to claim 1 or 2, characterized in that The adaptive nonlinear least squares method is based on minimizing the difference between the displacement signal and a model function having a plurality of harmonics for each of at least one oscillatory motion.

11. The method according to claim 1 or 2, characterized in that The processing device (5) calculates (220, 250) at least one frequency of an oscillatory motion using a cost function based on a fast Fourier transform of the displacement signal.

12. The method according to claim 1 or 2, characterized in that The processing device (5) uses a Kalman filter to determine the first estimated fundamental frequency.

13. The method according to claim 1 or 2, characterized in that The processing device (5) uses the estimated fundamental frequency calculated in one processing window to define a search area for frequencies in a later processing window.

14. The method according to claim 1 or 2, characterized in that The processing device (5) calculates (310) for at least one processing window a predicted state including a predicted fundamental frequency based on at least one previous processing window, and selects (300) a frequency estimate as the estimated fundamental frequency based on a comparison between the predicted fundamental frequency and the frequency estimate.

15. A radar sensor system (1) for vital sign monitoring, comprising: a transmitter (2) for irradiating (100) at least one body region (21) of a person (20) with radar radiation (30); a receiver (3) for generating (110) a receiver signal based on reflected radiation (31) from the at least one body region; as well as a processing device (5) adapted to generate a displacement signal for each of a plurality of processing windows based on the receiver signal, and to calculate (250) an estimated fundamental frequency for at least one oscillatory motion, the displacement signal being representative of body motion including respiratory motion and heartbeat motion as oscillatory motions, wherein the processing device is adapted to apply an adaptive nonlinear least squares method to calculate a plurality of frequency estimates (f r1 、f r2 、f r3 ), each frequency estimate of the plurality of frequency estimates corresponds to one of the plurality of harmonics of the first oscillatory motion, wherein the processing device (5) is adapted to separate the independent search area (A h1 、A h2 、A h3 ) for each frequency estimate (f r1 、f r2 、f r3 ), adapting at least one search area (A) for at least one processing window h1 、A h2 、A h3 ), and based on the frequency estimate (f r1 、f r2 、f r3 ) calculate (300) a first estimated fundamental frequency (f h ).

Images