A sleep vital sign monitoring method based on millimeter wave radar

Abstract

The application discloses a kind of sleep sign monitoring methods based on millimeter wave radar, first presetting the acquisition parameter of millimeter wave radar, millimeter wave radar transmits signal, receiving antenna captures reflected signal, and the intermediate frequency digital signal is obtained after sampling, and the sampling result is sent to back-end computing device;From the intermediate frequency digital signal collected, the features related to sleep posture and respiratory rate are extracted;The point cloud features and respiratory features obtained after feature extraction are used to identify the sleep posture of user.After that, the effective reflection points are extracted, and the respiratory rate is accurately estimated by using these reflection points through weighted average method.The present application aims to use low-cost, easy-to-deploy commercial millimeter wave radar equipment to realize an accurate and reliable vital sign monitoring system, to stably identify six sleep postures and accurately measure respiratory rate;Rich and stable point cloud information is obtained;Enhance the representation of point cloud to sleep posture;More accurate respiratory rate estimation results are obtained.

Description

Technical Field

[0001] This invention relates to the field of wireless sensing technology, and in particular to a method for monitoring sleep vital signs based on millimeter-wave radar. Background Technology

[0002] Sleep, as an indispensable and vital component of human life, has a profound impact on health and well-being. In recent years, research on monitoring sleep vital signs (including sleep posture, breathing, and heart rate) has received considerable attention. Developments in this field have significant background and implications for healthcare, biomedical research, and personal health management. By monitoring sleep vital signs, individuals can understand their sleep quality and habits, enabling more precise health management. Furthermore, sleep vital sign monitoring provides valuable opportunities for early disease screening and monitoring, helping doctors and patients identify and intervene in potential health problems early. For example, sleep posture is an important indicator for diagnosing postural obstructive sleep apnea syndrome (P-OSA); pressure ulcers are chronic wounds prone to recurrence; in hospitals and communities, caregivers need to ensure patients change positions regularly to prevent pressure sores caused by prolonged pressure on one side of the skin; for patients with epilepsy, incorrect sleeping posture increases the risk of sudden death. Applying sleep vital sign monitoring data to clinical research helps advance the understanding of sleep physiology, reveal the intricate links between sleep and health, and promote development in the medical and biomedical fields.

[0003] Millimeter-wave radar refers to radar systems that use the millimeter-wave frequency band (typically between 30 GHz and 300 GHz). Currently, linear frequency modulated continuous wave (FMCW) radar is widely used in the civilian sector. Internally, the radar system generates a local oscillator signal using oscillators and other devices. This signal is then frequency-modulated, phase-modulated, and amplified before being emitted as electromagnetic waves from the transmitting antenna. These electromagnetic waves are characterized by a linear frequency-time variation. After reflection from an object, they are captured by the receiving antenna and processed by low-noise amplifiers, mixers, and samples to obtain an intermediate frequency (IF) digital signal. Through a series of processing steps on the IF digital signal, information such as the target's distance, velocity, and angle can be estimated, and point clouds can be further calculated and generated. By processing and analyzing the phase of the reflected signal, the target's micro-motion information, such as respiratory rate and heart rate, can be estimated.

[0004] Convolutional Neural Networks (CNNs) are a type of deep learning algorithm that uses deep learning methods, particularly CNNs and Transformer encoders, for signal processing and classification. CNNs are responsible for local feature extraction, such as edge and texture recognition, while Transformers process sequential data, capturing global and local relationships. Together, they improve the accuracy of pose recognition.

[0005] Current research on sleep vital sign monitoring based on millimeter-wave radar can either only monitor breathing and heart rate, or only identify a limited number of sleep positions. However, comprehensive monitoring of vital signs and sleep positions is necessary to reveal the correlations between various physiological indicators and provide more accurate health diagnoses and predictions. Therefore, researching a non-contact technology that can simultaneously monitor vital signs and identify multiple sleep positions is of great significance. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a sleep vital signs monitoring method based on millimeter-wave radar. This invention extracts point cloud data from the radar signal, removes noise points through multi-frame azimuth probability clustering analysis, and obtains stable point cloud features. Respiratory features are extracted by analyzing phase changes generated by respiratory behavior, and combined with the point cloud data to describe sleep posture.

[0007] This invention is achieved through the following technical solution:

[0008] A method for monitoring sleep vital signs based on millimeter-wave radar specifically includes the following steps:

[0009] Step 1-1, Data Acquisition: First, the acquisition parameters of the millimeter-wave radar are preset. The millimeter-wave radar transmits signals according to the preset parameters, and the receiving antenna captures the reflected signals. The RF front-end of the millimeter-wave radar amplifies, mixes, and filters the received signals. After sampling, the intermediate frequency digital signal is obtained, and the sampling results are sent to the back-end computing device.

[0010] Steps 1-2, Feature extraction; extracting features related to sleep posture and respiratory rate from the acquired intermediate frequency digital signal, namely point cloud feature generation and respiratory feature extraction;

[0011] Steps 1-3: Sleep posture recognition; using point cloud features and breathing features obtained after feature extraction to identify the user's sleep posture.

[0012] Steps 1-4, respiratory rate estimation: Based on different sleep postures, specific parameters are used to extract effective reflection points from point cloud data, and these reflection points are used to accurately estimate the respiratory rate through a weighted average method based on an exponential function.

[0013] The preset millimeter-wave radar acquisition parameters mentioned in step 1-1 specifically include the selectable transmit and receive antenna parameters, the bandwidth of the transmitted frequency-modulated wave, the frequency modulation time, the number of chirps, the frame length, the ADC sampling frequency, and the number of sampling points.

[0014] Steps 1-2, which involve extracting features related to sleep posture and respiratory rate from the acquired intermediate frequency digital signal (i.e., point cloud feature generation and respiratory feature extraction), are detailed below:

[0015] In the point cloud feature generation, multi-frame azimuth probability clustering analysis is applied. For each distance cell, multi-frame cumulative clustering of the azimuth is performed with a certain error range as the radius to obtain the azimuth estimate. Finally, the corresponding elevation angle is estimated under the selected azimuth, and the point cloud in the Cartesian coordinate system is generated by combining the reflection distance.

[0016] Step 1-2-1: After receiving the data returned by the radar, the back-end computing device converts it into the format {ADC, Chirp, Horizontal, Vertical, Frame}, where ADC is the sampling result of the intermediate frequency digital signal, Chirp represents the snapshot dimension, Horizontal represents the horizontal dimension of the two-dimensional virtual antenna array, Vertical represents the height dimension, and Frame represents the frame dimension.

[0017] Step 1-2-2: Use Range-FFT to estimate the frequency components of the intermediate frequency signal, further calculate the distance information, and convert the data format to {Range,Chirp,Horizontal,Vertical,Frame}, where Range represents the distance of the radar to the target; then, extract the signal data within the effective sensing range through a preset scene.

[0018] Steps 1-2-3 involve using Doppler-FFT to estimate velocity information and converting the data format to {Range, Doppler, Horizontal, Vertical, Frame}, where Doppler represents the target's velocity. When a person sleeps, their body is relatively still, with almost only the upper body making periodic up-and-down movements with breathing. This slow movement is very low, less than the velocity range represented by a Doppler chamber, i.e., contained within the zero Doppler chamber. Therefore, by extracting data from the zero Doppler chamber, moving targets within the perception range can be filtered out, while the signals of human reflexes during sleep are not lost.

[0019] Steps 1-2-4 involve estimating the reflection angle using Angle-FFT. After performing Angle-FFT along the azimuth antenna dimension, the data format becomes {Range, Azimuth, Vertical, Frame}, where Azimuth represents the azimuth angle and Vertical represents the elevation angle. The range-azimuth estimation result can be expressed as:

[0020]

[0021] In the formula, y r,θ This represents the reflected signal of a target at a distance r and an angle θ, where A is the amplitude, j is the imaginary unit, and f is the amplitude. cHere, c is the initial frequency of chirp, and d is the speed of light. r Let Δd be the distance from the reflector to the reference element. θ This is due to the path difference caused by θ, where R is the sensing distance. Then, the actual incident angle of the reflected signal at point r is estimated, and the angular spectrum of r is ρ = |y|. r Perform a peak search to obtain n peaks. Here, θ1 and θ2 represent the angle values ​​corresponding to the first and second peaks, respectively. This paper employs a signal-to-noise ratio-based thresholding method to extract reliable peaks and sets an effective azimuth extraction range to filter out some background static target interference.

[0022] Steps 1-2-5 involve refining the azimuth angle using an algorithm based on multi-frame azimuth probability clustering. Angles estimated from a single snapshot are susceptible to random noise and multipath reflection interference, leading to inaccuracies and unreliable point cloud features. To suppress this interference, this invention designs an algorithm based on multi-frame azimuth probability clustering. The basic idea of ​​this algorithm is that the angle of arrival (ADR) of the true target appears more frequently in the estimation results of different frames, and the angle fluctuates only within a small range. In contrast, the angle estimation results of false targets caused by noise interference and multipath reflection show significant numerical variations between different frames. Furthermore, estimating human respiratory features requires data from multiple frames; therefore, this angle refinement algorithm improves the utilization rate of frame data. The process of the refinement algorithm based on multi-frame azimuth probability is as follows:

[0023] Let the total number of frames be N. frame Angle-FFT is performed on the snapshot of each range cell in the zero-Doppler cell, and the same operation is performed on each frame to obtain the azimuth angle of all frames. First, calculate θ r The probability of each angle occurring is calculated and sorted in descending order. Then, the angle with the highest probability is selected as the cluster center, assigned a new category C1, and a preset noise tolerance is used as the cluster radius γ. The remaining angles are then iterated through. If the difference Δθ between the currently iterated angle and the cluster center is less than the cluster radius γ, the angle is considered to belong to that cluster, and its category is changed to C1. After iterating through all remaining angles, the first round of clustering is complete, and the angles in category C1 are changed from θ... rRemove from the pool. Then, take the angle with the highest probability in the current angle set as the new cluster center, set the category cluster as C2, and perform a new round of clustering. After traversing all the remaining angles, obtain all angles that conform to C2. Repeat this process until the angle pool is empty. At this point, all estimated azimuth angles have been purified into several categories. The more angles a category contains, the more reliable the angles contained in the category are. Sort the categories according to the number of angles they contain, and finally, take the center angles of the categories containing the most angles as the extraction results. To ensure that the number of points in the point cloud input into the deep neural network is the same, if the number of categories estimated by the currently sampled signal is insufficient, copy the category containing the most angles to fill in the gaps. Then, perform Angle-FFT in the vertical antenna dimension to estimate the elevation angle. Specifically, firstly, the Angle-FFT result corresponding to each effective azimuth angle is extracted. Then, a peak search algorithm based on a signal-to-noise ratio threshold is used to estimate the effective elevation angle. Similar to the azimuth angle, the effective elevation angle is extracted by limiting the perceived range to filter out irrelevant object reflections. This yields the data for {Range, Azimuth, Elevation, Frame}. Finally, the distance R, azimuth θ, and elevation angle are combined... The Cartesian coordinates of the reflection point can be represented as: The entire point cloud can be generated by performing the same process on each distance bin.

[0024] In the respiratory feature extraction, the phase information of each valid reflection point in the 2D-Angle image is extracted with frames as the sampling interval, so as to obtain the phase waveform of each point cloud. The respiratory signal is obtained by dewinding and bandpass filtering the phase waveform. Then, the spectrum of the respiratory signal is analyzed by FFT and its peak value is taken as the respiratory feature of the point cloud. The respiratory features of the point cloud generated by a single measurement are normalized to obtain the relative respiratory features.

[0025] The respiratory features here are used in conjunction with point clouds to estimate the respiratory characteristics of sleep posture, and are a positively correlated representation of respiratory amplitude. Areas significantly affected by respiratory behavior, such as the torso, have larger respiratory amplitudes, and the corresponding respiratory feature values ​​are also larger. However, reflection points such as the legs and bed frame are not affected by breathing, and the phase of the reflected signal does not fluctuate, resulting in a respiratory feature value close to zero.

[0026] When generating point cloud features, 2D-Angle-FFT is used to extract the angle of arrival of the reflection points. After determining the reliable incident angle of the reflection point, the snapshot complex signal sample value of the reflection point can be obtained from the 2D-Angle-Map generated by FFT. Then, by sampling over multiple frames at a sampling rate of one snapshot per frame, the complex signal sequence of the reflection points is obtained. Let the i-th reflection point P... i The signal sequence is Breathing behavior affects the phase change of the signal, while individual differences, such as body shape and clothing, can affect the intensity of the reflected signal. Therefore, signal intensity value is not a reliable feature. This invention extracts the phase of the signal to identify fluctuations.

[0027] First, phase unwinding technology is used to address the issue of periodic phase jumps occurring near 0 or 2π, ensuring that the overall phase presents a smooth and continuous curve φ. i .

[0028] Next, a bandpass filter was used to filter out irrelevant high- and low-frequency noise, highlighting the frequency range relevant to respiratory activity. Using the typical human respiratory rate (0.1-0.6 Hz) as the passband, a bandpass filter was applied at φ... i Bandpass filtering is performed to obtain the respiratory phase curve φ i ′.

[0029] φ i The absolute value of the difference between the peak and trough values ​​of Δφ i ′ can reflect the minute displacement ΔR caused by human respiration, where Compared to static target reflections, the phase of reflections affected by breathing exhibits significant fluctuations. Therefore, this obvious difference can be used to distinguish a portion of the human body's reflection point cloud from static reflections in the environment. This helps subsequent recognition models quickly locate the human body and assign more weight to its judgment.

[0030] Then, the spectrum of the respiratory signal is analyzed using FFT, and the maximum value of the spectral peak is taken to represent the respiratory intensity, which is then used as a respiratory feature. The spectral peak value is a comprehensive representation of respiratory intensity, as it filters out the influence of weak frequency components on the waveform amplitude, making it more reliable than time-domain features.

[0031] To overcome individual differences, this invention proposes the concept of relative breathing features. Using the index of reflection points in the point cloud, the breathing feature intensities of these points are combined into a vector. Within this vector, the breathing feature of a single measurement is normalized and divided by the maximum breathing intensity of all points in the point cloud to obtain the relative breathing feature.

[0032] Steps 1-3 describe the use of point cloud features and respiratory features obtained after feature extraction to identify the user's sleep posture, specifically as follows: First, a lightweight deep neural network is used to further extract the spatial distribution information of the point cloud containing respiratory features. Two layers of spatial convolution are used to initially extract features for sleep posture recognition. Then, a Transformer encoder is used to perform deep feature fusion using a weighted average algorithm to obtain respiratory frequency information of multiple reflection points. Finally, two layers of Flattened features are used to obtain the sleep posture recognition result, and the final sleep posture classification result and the corresponding respiratory frequency estimate are output.

[0033] The method for estimating respiratory rate described in steps 1-4 relies on finding reflection points strongly correlated with respiration to accurately estimate the human respiratory rate. This invention uses the continuity of phase values ​​as an important indicator of the reliability of reflection points. Furthermore, it proposes a method based on point density on the IQ circle and phase value domain to extract effective reflection point signals. The specific method is as follows:

[0034] Step 1-4-1, in this invention, a complex signal sampling point is Ae jα Its amplitude A is the magnitude of the vector, and its phase α is the angle between the vector and the real part I-axis. Dividing each sampling point by its magnitude yields the unit IQ circle information of the sampled signal; let the frame sampling signal of the reflection point be... Where N frame Total number of frames Let be the signal sample value of the reflection point in the i-th frame, then the phase signal per unit amplitude is: Right now At this point, the coordinates of the p signal in the IQ domain all fall on the unit circle with an amplitude of 1.

[0035] Step 1-4-2: Obtain a reliable threshold T under different poses. The Rayleigh metric is used to determine whether the observation points are uniformly distributed on a circle or ring. The formula for calculating the Rayleigh metric is:

[0036]

[0037] In the formula, m is the total number of points. For each point, the unit vector is determined by the polar coordinate angle α of that point. i From formula (2), we can see that the Rayleigh index ranges from 0 to 1. A Rayleigh index closer to 0 indicates that the sampling points are more evenly distributed on the circumference, while a Rayleigh index closer to 1 indicates that the distribution is uneven.

[0038] Step 1-4-3: Calculate the Rayleigh index of the sampled signal y at each reflection point. When the index value is less than the threshold T, it is considered to be a valid reflection point affected by respiratory behavior, and the interference of noise on the signal of the reflection point can be ignored.

[0039] Step 1-4-4 involves dewinding the phase sequence of each reflection point and performing bandpass filtering to obtain the respiratory frequency signal. Finally, an FFT is used to perform spectral analysis on the signal, and the frequency index corresponding to the maximum peak value of the FFT result is taken as the respiratory frequency estimate for that reflection point. Under a given sleep posture, there are often multiple reflection points affected by breathing that satisfy the threshold T. Therefore, this invention uses an exponential function for weighted average calculation. The calculation formula is as follows:

[0040]

[0041] Where f is the final estimated respiratory rate, and n p Rayleigh represents the total number of effective reflection points. i Let f be the Rayleigh index value of the i-th reflection point. i This represents the estimated respiratory rate at the i-th reflection point. Finally, f is used as the estimated respiratory rate.

[0042] The advantages of the present invention are: (1) The present invention aims to use low-cost, easy-to-deploy commercial millimeter-wave radar equipment to realize an accurate and reliable vital signs monitoring system, stably identify 6 sleep postures and accurately measure respiratory rate;

[0043] (2) This invention proposes a point cloud purification method based on multi-frame azimuth probability clustering analysis, which obtains rich and stable point cloud information;

[0044] (3) The present invention obtains the breathing features of point cloud by extracting the phase change of each reflection point, thereby enhancing the representation of sleep posture by point cloud;

[0045] (4) This invention proposes a determination method based on the density of circumferential points of unit IQ to extract reliable reflection points, and uses these reflection points to estimate the respiratory rate through a weighted average method based on an exponential function, thereby obtaining a more accurate respiratory rate estimation result. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the deployment scenario;

[0047] Figure 2 This is a system module diagram;

[0048] Figure 3 This is a flowchart illustrating the workflow of sleep posture recognition.

[0049] Figure 4 This is a diagram illustrating six sleeping positions. Detailed Implementation

[0050] This invention is based on a scenario equipped with millimeter-wave radar. Figure 1 This is a schematic diagram of the deployment scenario of the present invention.

[0051] The system deployment process is simple. The system includes a millimeter-wave radar module, a data preprocessing module, and a sleep symptom recognition module, such as... Figure 2 As shown. The specific steps are as follows:

[0052] (1) Sleeping posture identification method

[0053] Figure 3 This is a schematic diagram of the workflow of the sleeping posture recognition method.

[0054] Step 1-1, Data Acquisition. First, radar acquisition parameters need to be preset. This includes selecting enabled transmit and receive antenna parameters, bandwidth of the transmitted FM wave, FM time, number of chirps, frame length, ADC sampling frequency, number of sampling points, etc. The millimeter-wave radar transmits signals according to the preset parameters, and the receiving antenna captures the reflected signals. The radar RF front-end amplifies, mixes, and filters the received signals, obtains an intermediate frequency digital signal after sampling, and sends the sampling results to the back-end computing device via gigabit Ethernet.

[0055] Steps 1-2, Feature Extraction. This part aims to extract features related to sleep posture and breathing rate from the collected data. It mainly consists of two parts: point cloud feature generation and breathing feature extraction.

[0056] This invention proposes an angle refinement algorithm based on multi-frame azimuth probability clustering. For each distance bin, multi-frame cumulative clustering of the azimuth is performed with a certain error range as the radius, which can obtain a relatively reliable azimuth estimate. Finally, the corresponding elevation angle is estimated at the selected azimuth, and further combined with the reflection distance to calculate and generate a point cloud in the Cartesian coordinate system.

[0057] The system extracts the phase information of each valid reflection point in the 2D-Angle image at frame sampling intervals, thus obtaining the phase waveform of each point cloud. The breathing signal is obtained by dewinding and bandpass filtering the phase waveform. Then, the system uses FFT to analyze the spectrum of the breathing signal and takes its peak value as the breathing feature of the point cloud. To eliminate individual differences, the system normalizes the breathing features of the point cloud generated from a single measurement to obtain relative breathing features.

[0058] In the point cloud feature generation, multi-frame azimuth probability clustering analysis is applied. For each distance cell, multi-frame cumulative clustering of the azimuth is performed with a certain error range as the radius to obtain the azimuth estimate. Finally, the corresponding elevation angle is estimated under the selected azimuth, and the point cloud in the Cartesian coordinate system is generated by combining the reflection distance.

[0059] Step 1-2-1: After receiving the data returned by the radar, the back-end computing device converts it into the format {ADC, Chirp, Horizontal, Vertical, Frame}, where ADC is the sampling result of the intermediate frequency digital signal, Chirp represents the snapshot dimension, Horizontal represents the horizontal dimension of the two-dimensional virtual antenna array, Vertical represents the height dimension, and Frame represents the frame dimension.

[0060] Step 1-2-2: Use Range-FFT to estimate the frequency components of the intermediate frequency signal, further calculate the distance information, and convert the data format to {Range,Chirp,Horizontal,Vertical,Frame}, where Range represents the distance of the radar to the target; then, extract the signal data within the effective sensing range through a preset scene.

[0061] Steps 1-2-3 involve using Doppler-FFT to estimate velocity information and converting the data format to {Range, Doppler, Horizontal, Vertical, Frame}, where Doppler represents the target's velocity. When a person sleeps, their body is relatively still, with almost only the upper body making periodic up-and-down movements with breathing. This slow movement is very low, less than the velocity range represented by a Doppler chamber, i.e., contained within the zero Doppler chamber. Therefore, by extracting data from the zero Doppler chamber, moving targets within the perception range can be filtered out, while the signals of human reflexes during sleep are not lost.

[0062] Steps 1-2-4 involve estimating the reflection angle using Angle-FFT. After performing Angle-FFT along the azimuth antenna dimension, the data format becomes {Range, Azimuth, Vertical, Frame}, where Azimuth represents the azimuth angle and Vertical represents the elevation angle. The range-azimuth estimation result can be expressed as:

[0063]

[0064] In the formula, y r,θ This represents the reflected signal of a target at a distance r and an angle θ, where A is the amplitude, j is the imaginary unit, and f is the amplitude. c Here, c is the initial frequency of chirp, and d is the speed of light.r Let Δd be the distance from the reflector to the reference element. θ This is due to the path difference caused by θ, where R is the sensing distance. Then, the actual incident angle of the reflected signal at point r is estimated, and the angular spectrum of r is ρ = |y|. r Perform a peak search to obtain n peaks. Here, θ1 and θ2 represent the angle values ​​corresponding to the first and second peaks, respectively. This paper employs a signal-to-noise ratio-based thresholding method to extract reliable peaks and sets an effective azimuth extraction range to filter out some background static target interference.

[0065] Steps 1-2-5 involve refining the azimuth angle using an algorithm based on multi-frame azimuth probability clustering. Angles estimated from a single snapshot are susceptible to random noise and multipath reflection interference, leading to inaccuracies and unreliable point cloud features. To suppress this interference, this invention designs an algorithm based on multi-frame azimuth probability clustering. The basic idea of ​​this algorithm is that the angle of arrival (ADR) of the true target appears more frequently in the estimation results of different frames, and the angle fluctuates only within a small range. In contrast, the angle estimation results of false targets caused by noise interference and multipath reflection show significant numerical variations between different frames. Furthermore, estimating human respiratory features requires data from multiple frames; therefore, this angle refinement algorithm improves the utilization rate of frame data. The process of the refinement algorithm based on multi-frame azimuth probability is as follows:

[0066] Let the total number of frames be N. frame Angle-FFT is performed on the snapshot of each range cell in the zero-Doppler cell, and the same operation is performed on each frame to obtain the azimuth angle of all frames. First, calculate θ r The probability of each angle occurring is calculated and sorted in descending order. Then, the angle with the highest probability is selected as the cluster center, assigned a new category C1, and a preset noise tolerance is used as the cluster radius γ. The remaining angles are then iterated through. If the difference Δθ between the currently iterated angle and the cluster center is less than the cluster radius γ, the angle is considered to belong to that cluster, and its category is changed to C1. After iterating through all remaining angles, the first round of clustering is complete, and the angles in category C1 are changed from θ... rRemove from the pool. Then, take the angle with the highest probability in the current angle set as the new cluster center, set the category cluster as C2, and perform a new round of clustering. After traversing all the remaining angles, obtain all angles that conform to C2. Repeat this process until the angle pool is empty. At this point, all estimated azimuth angles have been purified into several categories. The more angles a category contains, the more reliable the angles contained in the category are. Sort the categories according to the number of angles they contain, and finally, take the center angles of the categories containing the most angles as the extraction results. To ensure that the number of points in the point cloud input into the deep neural network is the same, if the number of categories estimated by the currently sampled signal is insufficient, copy the category containing the most angles to fill in the gaps. Then, perform Angle-FFT in the vertical antenna dimension to estimate the elevation angle. Specifically, firstly, the Angle-FFT result corresponding to each effective azimuth angle is extracted. Then, a peak search algorithm based on a signal-to-noise ratio threshold is used to estimate the effective elevation angle. Similar to the azimuth angle, the effective elevation angle is extracted by limiting the perceived range to filter out irrelevant object reflections. This yields the data for {Range, Azimuth, Elevation, Frame}. Finally, the distance R, azimuth θ, and elevation angle are combined... The Cartesian coordinates of the reflection point can be represented as: The entire point cloud can be generated by performing the same process on each distance bin.

[0067] In the respiratory feature extraction, the phase information of each valid reflection point in the 2D-Angle image is extracted with frames as the sampling interval, so as to obtain the phase waveform of each point cloud. The respiratory signal is obtained by dewinding and bandpass filtering the phase waveform. Then, the spectrum of the respiratory signal is analyzed by FFT and its peak value is taken as the respiratory feature of the point cloud. The respiratory features of the point cloud generated by a single measurement are normalized to obtain the relative respiratory features.

[0068] The respiratory features here are used in conjunction with point clouds to estimate the respiratory characteristics of sleep posture, and are a positively correlated representation of respiratory amplitude. Areas significantly affected by respiratory behavior, such as the torso, have larger respiratory amplitudes, and the corresponding respiratory feature values ​​are also larger. However, reflection points such as the legs and bed frame are not affected by breathing, and the phase of the reflected signal does not fluctuate, resulting in a respiratory feature value close to zero.

[0069] When generating point cloud features, 2D-Angle-FFT is used to extract the angle of arrival of the reflection points. After determining the reliable incident angle of the reflection point, the snapshot complex signal sample value of the reflection point can be obtained from the 2D-Angle-Map generated by FFT. Then, by sampling over multiple frames at a sampling rate of one snapshot per frame, the complex signal sequence of the reflection points is obtained. Let the i-th reflection point P... i The signal sequence is Breathing behavior affects the phase change of the signal, while individual differences, such as body shape and clothing, can affect the intensity of the reflected signal. Therefore, signal intensity value is not a reliable feature. This invention extracts the phase of the signal to identify fluctuations.

[0070] First, phase unwinding technology is used to address the issue of periodic phase jumps occurring near 0 or 2π, ensuring that the overall phase presents a smooth and continuous curve φ. i .

[0071] Next, a bandpass filter was used to filter out irrelevant high- and low-frequency noise, highlighting the frequency range relevant to respiratory activity. Using the typical human respiratory rate (0.1-0.6 Hz) as the passband, a bandpass filter was applied at φ... i Bandpass filtering is performed to obtain the respiratory phase curve φ i ′.

[0072] φ i The absolute value of the difference between the peak and trough values ​​of Δφ i ′ can reflect the minute displacement ΔR caused by human respiration, where Compared to static target reflections, the phase of reflections affected by breathing exhibits significant fluctuations. Therefore, this obvious difference can be used to distinguish a portion of the human body's reflection point cloud from static reflections in the environment. This helps subsequent recognition models quickly locate the human body and assign more weight to its judgment.

[0073] Then, the spectrum of the respiratory signal is analyzed using FFT, and the maximum value of the spectral peak is taken to represent the respiratory intensity, which is then used as a respiratory feature. The spectral peak value is a comprehensive representation of respiratory intensity, as it filters out the influence of weak frequency components on the waveform amplitude, making it more reliable than time-domain features.

[0074] To overcome individual differences, this invention proposes the concept of relative breathing features. Using the index of reflection points in the point cloud, the breathing feature intensities of these points are combined into a vector. Within this vector, the breathing feature of a single measurement is normalized and divided by the maximum breathing intensity of all points in the point cloud to obtain the relative breathing feature.

[0075] Steps 1-3: Sleep Posture Recognition. The purpose of this module is to identify the user's sleep posture using point clouds containing respiratory features obtained after signal processing. The system uses a lightweight deep neural network to further extract the spatial distribution information of the point cloud containing respiratory features. The network first performs two layers of spatial convolution to initially extract features, then uses a Transformer encoder for deep feature fusion, and finally uses two layers of Flatten to flatten the features, obtaining the sleep posture recognition result. There are a total of 6 results, as shown below. Figure 4 As shown.

[0076] (2) Methods for estimating respiratory rate.

[0077] Steps 1-4: Respiratory rate estimation. This invention proposes a method based on unit IQ circumferential point density. Valid reflection points are extracted from point cloud data using specific threshold parameters according to different sleep postures. These reflection points are then used to accurately estimate the respiratory rate through a weighted average method based on an exponential function.

[0078] This invention uses the continuity of phase values ​​as an important indicator of the reliability of reflection points. It proposes a method based on the point density and phase range on the IQ circle to extract effective reflection point signals. The method is as follows:

[0079] Step 1-4-1, in this invention, a complex signal sampling point is Ae jα Its amplitude A is the magnitude of the vector, and its phase α is the angle between the vector and the real part I-axis. Dividing each sampling point by its magnitude yields the unit IQ circle information of the sampled signal; let the frame sampling signal of the reflection point be... Where N frame Total number of frames Let be the signal sample value of the reflection point in the i-th frame, then the phase signal per unit amplitude is: Right now At this point, the coordinates of the p signal in the IQ domain all fall on the unit circle with an amplitude of 1.

[0080] This invention compares the performance of three unit amplitude phase signals in the IQ domain: signals from reflection points affected by breathing, signals from reflection points affected by breathing but also by significant noise, and signals from static reflection points. When the signal from a reflection point is affected by breathing, its phase value is theoretically continuously distributed, i.e., relatively uniformly distributed on the circumference of the unit circle, and the corresponding time-domain waveform is also uniform. Due to noise interference causing phase perturbation, the distribution of sampling points on the IQ circle becomes scattered, and the time-domain waveform of the phase value exhibits some jumps, resulting in dense peaks and valleys. If the signal is entirely from a static reflection point, its phase will not change uniformly; the phase value fluctuates only within a small range, and the range of phase time-domain waveform variation is also very small. Based on the above facts, this invention transforms the problem of determining phase continuity into the problem of determining whether the sampling points on the circumference of the unit circle in the IQ domain are uniformly distributed.

[0081] Step 1-4-2: Obtain a reliable threshold T under different poses. The Rayleigh metric is used to determine whether the observation points are uniformly distributed on a circle or ring. The formula for calculating the Rayleigh metric is:

[0082]

[0083] In the formula, m is the total number of points. For each point, the unit vector is determined by the polar coordinate angle α of that point. i From formula (2), we can see that the Rayleigh index ranges from 0 to 1. A Rayleigh index closer to 0 indicates that the sampling points are more evenly distributed on the circumference, while a Rayleigh index closer to 1 indicates that the distribution is uneven.

[0084] However, due to the varying reflective surface of the human body relative to radar in different sleep positions, the uniformity of the phase points differs. Therefore, this invention calculates a separate Rayleigh index threshold T for each sleep position to ensure the reliability of subsequent respiratory rate calculations.

[0085] This invention collected 10 sets of radar sampling data from multiple volunteers in each posture, and simultaneously recorded the actual respiratory rate corresponding to each set of data. For each posture sample, each sampling point was manually distinguished as a valid reflection point affected by breathing, and its corresponding Rayleigh index was recorded. Whether each sampling point is a valid reflection point affected by breathing was determined by combining the time domain and frequency domain characteristics of the phase signal.

[0086] In terms of time-domain characteristics: First, the phase time-domain waveform of the sampling point should be relatively smooth, with obvious periodicity between adjacent peaks. Second, the minimum absolute value of the peak-to-valley difference of the waveform should exceed a certain threshold (e.g., exceeding 15 in the supine position and exceeding 8 in the lateral position). This is to filter out weak signals falling within the respiratory frequency band due to bandpass filtering (usually the absolute value of the peak-to-valley difference is less than 3). In terms of frequency-domain characteristics: For the respiratory waveform of each sampling point, its spectrum is calculated using FFT, and the frequency corresponding to its maximum peak is taken as the calculated respiratory frequency. When the difference between the calculated respiratory frequency and the true respiratory frequency is less than 0.5 Hz, it is considered to meet the frequency-domain characteristic requirements. Only reflection points that simultaneously meet the time-domain and frequency-domain characteristic requirements are considered valid; otherwise, they are considered invalid.

[0087] When calculating the threshold T for each pose, the maximum value of the Rayleigh index for valid points and the minimum value of the Rayleigh index for invalid points in all volunteer data are selected, and the average of the two is calculated as the threshold T for that pose.

[0088] Step 1-4-3: Calculate the Rayleigh index of the sampled signal y at each reflection point. When the index value is less than the threshold T, it is considered to be a valid reflection point affected by respiratory behavior, and the interference of noise on the signal of the reflection point can be ignored.

[0089] Step 1-4-4 involves dewinding the phase sequence of each reflection point and performing bandpass filtering to obtain the respiratory frequency signal. Finally, an FFT is used to perform spectral analysis on the signal, and the frequency index corresponding to the maximum peak value of the FFT result is taken as the respiratory frequency estimate for that reflection point. Under a given sleep posture, there are often multiple reflection points affected by breathing that satisfy the threshold T. Therefore, this invention uses an exponential function for weighted average calculation. The calculation formula is as follows:

[0090]

[0091] Where f is the final estimated respiratory rate, and n p Rayleigh represents the total number of effective reflection points. i Let f be the Rayleigh index value of the i-th reflection point. i This represents the estimated respiratory rate at the i-th reflection point. Finally, f is used as the estimated respiratory rate.

[0092] This invention demonstrates a sleep vital signs monitoring system based on millimeter-wave radar. The system consists of a millimeter-wave radar, data preprocessing, and a sleep vital signs recognition module. The core innovation lies in proposing a point cloud purification method based on multi-frame azimuth probability clustering analysis, which obtains rich and stable point cloud information, and proposing a method for determining the density of circular points per unit IQ, resulting in more accurate respiratory rate estimation. This invention can utilize millimeter-wave radar to identify users' sleep vital signs. This invention has many applications in identifying sleep posture and respiratory rate, and methods based on and improved upon this invention also fall within the scope of protection of this invention.

Claims

1. A millimeter wave radar based sleep vital sign monitoring method, characterized in that: Specifically comprising the following steps: Step 1-1, data acquisition; first, preset the acquisition parameters of the millimeter wave radar, the millimeter wave radar transmits signals according to the pre-set parameters, the receiving antenna captures the reflected signals, the millimeter wave radar radio frequency front end amplifies, mixes and filters the received signals, and obtains intermediate frequency digital signals through sampling, and sends the sampling results to the back-end computing device; Step 1-2, feature extraction; the features related to the sleep posture and the breathing rate are extracted from the collected intermediate frequency digital signals, that is, the point cloud feature generation and the breathing feature extraction; Point cloud In the feature generation, multi-frame azimuth angle occurrence probability clustering analysis is applied, for each distance bin, multi-frame cumulative clustering is performed on the azimuth angle with a certain error range as the radius, the azimuth angle estimation is obtained, finally, the corresponding elevation angle is estimated under the selected azimuth angle, and the point cloud in the Cartesian coordinate system is generated by combining the reflection distance; Step 1-3, sleep posture recognition; the point cloud features and the breathing features obtained after feature extraction are used to identify the sleep posture of the user; Step 1-4, breathing frequency estimation; according to different sleep postures, specific parameters are used to extract effective reflection points from the point cloud data, and these reflection points are used to accurately estimate the breathing frequency by using the weighted average method based on the exponential function; specifically as follows: Step 1-4-1, one complex signal sampling point is , the amplitude is the modulus of the vector, the phase is the included angle between the vector and the real part axis, and the modulus of each sampling point is divided to obtain the unit IQ circle information of the sampling signal; let the frame sampling signal of the reflection point be , wherein is the total number of frames, is the signal sampling value of the reflection point of the frame, and the phase signal of the unit amplitude is , that is , at this time, the coordinate points of the signal on the IQ domain all fall on the unit circle with an amplitude of 1. Step 1-4-2, obtain reliable threshold T under different postures, Rayleigh index is used to determine whether the observation points are uniformly distributed on a circumference or a circular ring, the calculation formula of Rayleigh index is: ； wherein is the total number of points, is the unit vector of each point, which is obtained by the polar coordinate angle of the point; it is known from equation (2) that the value range of the Rayleigh index is between 0 and 1, and the Rayleigh index closer to 0 indicates that the sampling points are more uniformly distributed on the circumference, and the closer to 1 indicates that the distribution is uneven; Because the reflection surface of human body relative to radar is different in different sleeping postures, the uniformity degree of phase points has certain difference, so a Rayleigh index threshold is obtained for each sleeping posture , to ensure the reliability of subsequent breathing frequency calculation. Step 1 - 4-3, calculate the sampling signal of each reflection point Rayleigh index, when the index value is less than a threshold value, it is considered as an effective reflection point affected by respiratory behavior, and the noise interference to the reflection point signal can be ignored; Step 1-4-4, phase unwrapping and band-pass filtering are performed on the phase sequence of each reflection point to obtain the signal in the breathing frequency band; finally, FFT is used for spectrum analysis of the signal, and the frequency index corresponding to the maximum peak value of the FFT result is taken as the breathing frequency estimation of the reflection point; under a sleep posture, there are multiple reflection points affected by breathing that meet the threshold T, therefore, the exponential function is used for weighted average calculation, and the calculation formula is as follows: ； wherein is the final respiration rate estimate, is the total number of valid reflection points, is the Rayleigh index value of the th reflection point, is the respiration rate estimate result of the th reflection point, and finally is taken as the respiration rate estimate.

2. The millimeter wave radar-based sleep vital monitoring method of claim 1, wherein: The preset acquisition parameters of the millimeter wave radar in step 1-1 specifically include selectively enabled transceiver antenna parameters, bandwidth of the transmitted frequency-modulated wave, frequency-modulated time, number of chirps, frame length, ADC sampling frequency and sampling point number.

3. The millimeter wave radar-based sleep vital monitoring method of claim 1, wherein: In the point cloud feature generation, multi-frame azimuth angle occurrence probability clustering analysis is applied, for each distance bin, multi-frame cumulative clustering is performed on the azimuth angle with a certain error range as the radius, the azimuth angle estimation is obtained, finally, the corresponding elevation angle is estimated under the selected azimuth angle, and the point cloud in the Cartesian coordinate system is generated by combining the reflection distance, and the specific steps are as follows: Step 1-2-1, after the back-end computing device receives the data returned by the millimeter wave radar, the format of the data is converted to {ADC, Chirp, Horizontal, Vertical, Frame}, wherein ADC is the sampling result of the intermediate frequency digital signal, Chirp represents the snapshot dimension, Horizontal represents the horizontal dimension of the two-dimensional virtual antenna array, Vertical represents the height dimension, and Frame represents the frame dimension; Step 1-2-2, the frequency components of the intermediate frequency signal are estimated using Range-FFT, and the distance information is further calculated. The data format is converted to {Range, Chirp, Horizontal, Vertical, Frame}, where Range represents the distance from the radar to the target. Then, through the preset scene, the signal data within the effective sensing range is extracted; Step 1-2-3, the velocity information is estimated using Doppler-FFT, and the data format is converted to {Range, Doppler, Horizontal, Vertical, Frame}, where Doppler represents the velocity of the target; Step 1-2-4, the reflection angle is estimated using Angle-FFT. After performing Angle-FFT in the azimuth antenna dimension, the data format becomes {Range, Azimuth, Vertical, Frame}, where Azimuth represents the azimuth angle, and Vertical represents the elevation angle. The distance-azimuth estimation result is represented as: ； wherein, represents the reflection signal of the target at a distance of and an angle of , is the amplitude, is the imaginary unit, is the initial frequency of the chirp, is the speed of light, is the distance from the reflector to the reference element, is the wave path difference caused by , is the perceived distance, then, the actual incidence angle of the reflection signal located at is estimated, the angle spectrum of is subjected to a peak search, and peaks are obtained, wherein and respectively represent the angle values corresponding to the first peak and the second peak, a threshold method based on the signal-to-noise ratio is used to extract reliable peaks, and an effective azimuth angle extraction range is set to filter out part of the background static target interference; Step 1-2-5, the azimuth angle is purified using an algorithm based on multi-frame azimuth angle occurrence probability clustering. The flow of the purification algorithm based on multi-frame azimuth angle occurrence probability is as follows: Let the total number of frames be Angle-FFT is performed on the snapshot of each range cell in the zero-Doppler cell, and the same operation is performed on each frame to obtain the azimuth angle of all frames. First, statistics Calculate the probability of each angle appearing in the array and sort them in descending order. Then, take the angle with the highest probability as the cluster center and assign it a new category. And use a preset noise interference tolerance as the cluster radius. Then iterate through the remaining angles, and if the difference between the currently traversed angle and the cluster center is... Smaller than cluster radius The angle is considered to belong to this cluster, and its category is changed. After traversing all remaining angles, the first round of clustering is completed, and the categories are classified as follows: From the perspective Remove the angle from the set of angles, and then use the angle with the highest probability in the current set of angles as the new cluster center, setting the category cluster to... A new round of clustering is performed, and after traversing all other angles, the results are obtained. This process is repeated for all angles until the angle pool is empty. At this point, all estimated azimuth angles have been refined into several categories. The more angles a category contains, the more reliable the angles within that category are. The categories are sorted according to the number of angles they contain. Finally, the center angles of the categories containing the most angles are used as the extraction results. To ensure that the number of points in the point cloud input to the deep neural network is the same, if the number of categories estimated from the currently sampled signal is insufficient, the category containing the most angles is copied to fill the gaps. Then, Angle-FFT is performed in the vertical antenna dimension to estimate the elevation angle. Specifically, the process involves: first, extracting the Angle-FFT result for the height corresponding to each effective azimuth angle; then, using peak search based on a signal-to-noise ratio threshold to estimate the effective height angle; and finally, extracting the effective height angle by limiting the perceived range to filter out irrelevant object reflections. This yields the data for {Range, Azimuth, Elevation, Frame}. Finally, the distance is then considered... Azimuth and elevation angle The Cartesian coordinates of the reflection point are expressed as: , , The same process is performed on each distance bin to generate the entire point cloud; In the breathing feature extraction, the phase information of each valid reflection point in the 2D-Angle graph is extracted at a frame sampling interval, thereby obtaining the phase waveform of each point cloud. By unwinding and band-pass filtering the phase waveform, the breathing signal is obtained. Then, the frequency spectrum of the breathing signal is analyzed using FFT, and the peak value is taken as the breathing feature of the point cloud. The breathing features of the point clouds generated by a single measurement are normalized to obtain the relative breathing features.

4. The millimeter wave radar-based sleep vital monitoring method of claim 3, wherein: When generating point cloud features, 2D-Angle-FFT is used to extract the angle of arrival of the reflection point. After determining the reliable incident angle of the reflection point, the snapshot complex signal sample value of the reflection point is obtained from the 2D-Angle-Map generated by FFT. Then, sampling is performed on multiple frames at a sampling rate of one snapshot per frame to obtain the complex signal sequence of the reflection point. Let the nth frame be the nth frame. One reflection point The signal sequence is ; First, the phase unwrapping technique is used to solve the problem that the phase jumps periodically when approaching or to ensure that the overall phase presents a smooth and continuous curve ; Afterwards, a band-pass filter is used to filter out the irrelevant high and low frequency noise, to highlight the frequency range associated with respiratory activity, with a pass-band at the typical human respiratory frequency (0.1-0.6 Hz), to obtain the respiratory phase curve ; and ​ the absolute value of the difference between the peak and the trough reflecting the small displacement caused by the breathing of a human body wherein ; Then, the frequency spectrum of the breathing signal is analyzed using FFT, and the maximum value of the spectral peak is taken to represent the breathing intensity, which is used as the breathing feature; Using the reflection point sequence number in the point cloud as the index, the breathing feature intensities of these points are combined into a vector. In the vector, the breathing feature of a single measurement is normalized by dividing it by the maximum breathing intensity of all points in the point cloud to obtain the relative breathing feature.

5. The millimeter wave radar-based sleep vital monitoring method of claim 1, wherein: Step 1-3 uses the point cloud features and breathing features obtained after feature extraction to identify the user's sleep posture, as follows: First, a lightweight deep neural network is used to further extract the spatial distribution information of the point cloud containing the breathing features. Two layers of spatial convolution are used to preliminarily extract features for sleep posture recognition. Then, a weighted average algorithm is used through a Transformer encoder to perform deep feature fusion to obtain the breathing frequency information of multiple reflection points. Finally, two layers of Flatten are used to flatten the features to obtain the recognition result of the sleep posture, and the final sleep posture classification result and corresponding breathing frequency estimation value are output.

Images