A millimeter wave radar vital sign modeling method based on time-frequency characteristic decoupling

Abstract

The present application relates to a kind of millimeter wave radar vital sign modeling method based on time-frequency feature decoupling.The method is by deploying millimeter wave radar array, acquires reflected echo signal, obtains pure initial signal data;Based on the time-frequency transformation of initial signal data by multi-scale sliding window, generate time-frequency energy distribution atlas;According to the energy concentration degree and stability difference of different frequency components in time-frequency atlas, extract the frequency component with high energy concentration degree and continuous stable periodic variation on time axis as the effective signal component corresponding to human respiration and heartbeat feature;According to effective signal component, extract corresponding frequency variation track, establish the vital sign signal track model based on time-frequency feature variation trend, determine the optimal vital sign track path according to the stability and continuity of track on time axis;Using optimal vital sign track path, establish vital sign monitoring model, realize high-precision monitoring to human respiration and heart rate.

Description

Technical Field

[0001] This invention relates to the field of vital sign modeling technology, and in particular to a method for modeling vital signs using millimeter-wave radar based on time-frequency feature decoupling. Background Technology

[0002] With the development of millimeter-wave radar technology, non-contact vital sign monitoring methods based on radar signals have been extensively studied. Current technologies typically use single-channel or multi-channel millimeter-wave radar equipment to acquire reflected echo signals, and then combine this with techniques such as time-domain filtering, wavelet analysis, and Fourier transform to extract periodic vital sign information such as respiration and heart rate. In static or weakly interfered environments, these methods can achieve a certain level of accuracy in detecting respiratory and heart rates, and are gradually being applied in health monitoring, smart healthcare, and security fields.

[0003] However, in complex environments such as non-line-of-sight (NLOS), subtle human movements, multipath reflections, and environmental clutter, existing technologies often struggle to accurately distinguish between periodic vital signs and non-periodic interference signals. This leads to decreased detection accuracy and increased false positives, especially in dynamic scenarios where vital signs are easily overwhelmed by interference, and robust and effective signal extraction and modeling mechanisms are lacking. Here, NLOS refers to an environment where there is no direct, unobstructed straight-line propagation path between the millimeter-wave radar and the monitored object (e.g., the human body). During propagation, the radar signal is reflected, diffracted, or scattered by various objects such as walls, furniture, obstacles, or the human body itself, forming stray echoes along multiple paths, resulting in signal superposition and interference.

[0004] Therefore, a new method for modeling vital signs using millimeter-wave radar is urgently needed. Summary of the Invention

[0005] This application provides a method for modeling vital signs in millimeter-wave radar based on time-frequency feature decoupling, in order to improve the robustness of vital sign detection.

[0006] This application provides a method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling, including:

[0007] By deploying millimeter-wave radar arrays, reflected echo signals are continuously acquired synchronously. The raw signals of each receiving channel are then subjected to time-domain filtering and noise reduction to obtain clean initial signal data.

[0008] The initial signal data is transformed by time-frequency transformation based on a multi-scale sliding window to generate a time-frequency energy distribution map that can characterize the time-varying features of human vital signs signals.

[0009] For the non-periodic interference signals present in the time-frequency energy distribution spectrum, based on the differences in energy concentration and stability of different frequency components in the time-frequency spectrum, frequency components with high energy concentration and continuous and stable periodic changes on the time axis are extracted as effective signal components corresponding to human breathing and heartbeat characteristics.

[0010] Based on the extracted effective signal components, the corresponding frequency change trajectory in the time-frequency spectrum is extracted, a vital sign signal trajectory model based on the time-frequency characteristic change trend is established, and the optimal vital sign trajectory path is determined based on the stability and continuity of the trajectory on the time axis.

[0011] A vital signs monitoring model is established using the determined optimal vital signs trajectory path.

[0012] The beneficial effects of the technical solution provided in this application include:

[0013] (1) This application extracts vital sign signals through time-frequency feature decoupling, which can effectively suppress interference caused by small human movements, multipath reflections, and environmental clutter in non-line-of-sight environments, and significantly improve the robustness of vital sign detection. (2) This application generates time-frequency energy distribution maps through multi-scale sliding window time-frequency transformation, and filters them based on the energy concentration and stability differences of frequency components, which can improve the extraction accuracy of respiratory and heartbeat signals. (3) This application establishes a vital sign signal trajectory model based on the changing trend of time-frequency features, and determines the optimal trajectory path according to the trajectory stability and continuity criteria, which can realize continuous tracking and dynamic modeling of the changing trend of vital sign signals. (4) This application can realize high-precision real-time monitoring of human respiration and heart rate by millimeter-wave radar in complex non-line-of-sight environments, expanding the application scenarios of millimeter-wave radar in smart healthcare, emergency rescue, and safety monitoring. Attached Figure Description

[0014] Figure 1 This is a flowchart of a millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling provided in the first embodiment of this application. Detailed Implementation

[0015] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0016] The first embodiment of this application provides a method for modeling vital signs in millimeter-wave radar based on time-frequency feature decoupling. Please refer to... Figure 1 This figure is a schematic diagram of the first embodiment of this application. The following is in conjunction with... Figure 1The first embodiment of this application provides a detailed description of a millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling.

[0017] The millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling addresses the abnormal amplitude and phase fluctuations of millimeter-wave radar signals in non-line-of-sight environments caused by minute human movements, multipath reflections, and environmental clutter. It achieves robust extraction and real-time monitoring of vital sign signals through the following steps:

[0018] Step S101: The millimeter-wave radar array is deployed to continuously acquire reflected echo signals in a synchronous manner. The original signals of each receiving channel are subjected to time-domain filtering and noise reduction processing to obtain clean initial signal data.

[0019] In step S101, a millimeter-wave radar array needs to be constructed first. This array should include at least two transceiver channels, each containing a transmitting antenna and a receiving antenna, and all channels should be able to operate synchronously. The operating frequency band of the millimeter-wave radar is preferably set between 24 GHz and 77 GHz, with the specific frequency determined based on the actual application scenario. For example, in indoor vital sign monitoring, the 60 GHz band is preferred to balance penetration capability and resolution. The array layout should be determined based on the size and layout of the monitoring area, selecting a suitable array arrangement, such as a rectangular array or a linear array, ensuring that the array's spatial resolution covers the range of changes in the monitored target's position. Each channel in the radar array should be synchronized using a unified clock source to ensure that all received signals have a unified temporal reference, avoiding data distortion caused by sampling timing offsets.

[0020] After the radar array is deployed, synchronous and continuous acquisition of reflected echo signals is performed. The acquisition method employs periodic triggering, with the sampling time interval Δt preferably set between 0.5 and 5 milliseconds. This ensures a sufficiently high sampling frequency to capture subtle movements caused by breathing and heartbeat, while avoiding excessive data volume and unnecessary processing burden. Each sampling should record at least one complete pulse reflection waveform, including the transmission timestamp, reception timestamp, received signal amplitude, and phase information. The acquired echo signal can be the raw IQ data (i.e., in-phase and quadrature components) or the envelope signal after preliminary demodulation, depending on the radar hardware system design.

[0021] After obtaining the raw acquisition data, the signals from each receiving channel need to undergo time-domain filtering to remove high-frequency noise and low-frequency drift components. Preferably, a finite impulse response (FIR) filter or an infinite impulse response (IIR) filter can be used, with the filter's cutoff frequency range covering the characteristic frequencies of human respiration and heartbeat, i.e., the range of 0.1Hz to 3Hz. For example, the cutoff frequency of a low-pass filter can be set to 5Hz to fully preserve vital sign signal components while suppressing noise above 5Hz. To eliminate background static clutter, a moving average filtering or background modeling subtraction method can be used. This involves calculating the background mean curve over a preset time window (e.g., 10 seconds) and then subtracting the corresponding background mean from the original signal sampled each time to obtain the dynamically changing components.

[0022] In addition to filtering, noise suppression is also required. Common methods include median filtering for each signal segment, noise threshold truncation, or wavelet denoising. In this invention, a simple and effective median filtering method is preferred. The median of each consecutive N sampling points (e.g., N=5) is taken and replaced with the value of the original center sampling point, thereby effectively removing isolated abrupt noise without significantly weakening periodic signal components. Furthermore, to enhance the signal quality for subsequent time-frequency analysis, amplitude normalization can be performed on the signal after time-domain filtering, that is, scaling the signal amplitude of each sampling segment to the [-1,1] interval to reduce interference caused by channel gain fluctuations.

[0023] Through the above continuous acquisition, synchronous control, time-domain filtering and noise suppression processing, pure initial signal data of each receiving channel can be obtained. This initial signal data not only completely preserves the phase and amplitude changes caused by the minute movements caused by human breathing and heartbeat, but also effectively suppresses environmental clutter and system noise, providing a reliable data foundation for subsequent time-frequency feature extraction and trajectory modeling.

[0024] Furthermore, the time-domain filtering and denoising processing of the raw signal for each receiving channel includes:

[0025] A background modeling subtraction method based on a nonlinear adaptive background update mechanism is adopted to remove stable clutter components from the dynamically updated background template;

[0026] The signal after background removal is subjected to amplitude normalization, and a two-stage median filter is applied to the normalized signal. The first stage median filter is used to eliminate isolated pulse interference, and the second stage median filter adaptively adjusts the size of the filter window according to the local noise energy level of the sliding window, thereby enhancing the signal-to-noise ratio of the periodic vital signs signal.

[0027] In this embodiment, the raw signal of each receiving channel undergoes time-domain filtering and denoising. First, a background modeling subtraction based on a nonlinear adaptive background update mechanism is used to remove stable clutter components. Specifically, for each frame of echo signal continuously acquired by the millimeter-wave radar, a dynamic background template is constructed. This background template is not a static average but is dynamically updated based on the statistical characteristics of the local minimum energy or minimum amplitude change within each time window. When the change in a signal position within several consecutive frames is lower than a set change threshold, that position is determined to belong to a stable background region, and its amplitude is included in the background template update. The background template update process preferably uses the exponentially weighted moving average (EWMA) method, where the background update rate parameter is adaptively adjusted in real time according to environmental stability, typically ranging from 0.01 to 0.1. This ensures tracking capability for slowly changing clutter while avoiding the inclusion of short-term micro-motion signals in the background template. In practical applications, when strong micro-motion is detected in a certain time segment, the background update at that position can be paused to prevent dynamic vital signs signals from being incorrectly erased.

[0028] After completing the background modeling, the current frame signal is differentially analyzed point by point with the dynamic background template to remove static background components, resulting in a purified signal containing weak dynamic variations. To further improve the subsequent processing effect, amplitude normalization is required for the signal after background removal. Normalization methods can include local maximum normalization or root mean square (RMS) normalization. Using each time window as a unit, the signal amplitude is adjusted to a standard range (e.g., -1 to 1 or RMS = 1), thereby eliminating amplitude deviations caused by inconsistent antenna gain across different channels or equipment fluctuations, ensuring good comparability of data from different time periods and channels.

[0029] After amplitude normalization, a two-stage median filter is applied to the normalized signal to remove remaining random noise and burst interference. In the first stage, a standard median filter is performed using a fixed window size (e.g., 5 sampling points) to eliminate isolated, single-point impulse noise, maintaining the continuity and smoothness of the signal curve at a fine-grained scale. In the second stage, the median filter window size is adaptively adjusted based on the local noise energy level within the sliding window. Specifically, the local variance or root mean square value of the signal is calculated as a noise energy index within each sliding time window. When the local noise level is high, the median filter window is increased (e.g., expanded to 7 to 9 sampling points); when the local noise level is low, a smaller window is maintained (e.g., 3 to 5 sampling points) to avoid over-smoothing and loss of effective periodic components. This two-stage median filtering process not only effectively suppresses various random interferences but also dynamically adapts to signal changes under different environmental noise conditions while preserving weak periodic signals of vital signs, improving the accuracy and robustness of subsequent time-frequency feature extraction.

[0030] Through the above specific processing flow, this step can ensure that in complex non-line-of-sight environments, a millimeter-wave radar initial clean signal with high signal-to-noise ratio, prominent periodic components, and good environmental clutter suppression is obtained, providing a stable and reliable data foundation for subsequent time-frequency energy map generation, effective signal extraction, and trajectory modeling.

[0031] Furthermore, the background modeling subtraction based on a nonlinear adaptive background update mechanism, which removes stable clutter components from the dynamically updated background template, includes:

[0032] Within each continuous acquisition period, the local rate of change of the signal is detected based on the sliding time window, and an exponentially weighted moving average is performed on the signal area with a rate of change lower than a preset threshold to dynamically update the background template. During the background template update process, if a change area is detected where the signal amplitude increases instantaneously and the duration exceeds a preset threshold, the background update of the corresponding area is paused and marked as a dynamic activity protection area to prevent weak vital signs signals from being mistakenly included in the background.

[0033] After background subtraction, local energy comparison analysis is applied to the purified signal to calculate the energy difference between the local area of ​​the purified signal and the corresponding background area. Areas with significant energy differences are selected as the key compensation objects for subsequent normalization processing.

[0034] During the normalization stage, a dynamic normalization baseline adjustment mechanism is introduced for regions with significant energy differences. The normalization parameters are adaptively adjusted based on the ratio of the local background energy level to the local energy level of the purified signal, so as to enhance the discriminability of weak periodic signals relative to residual background noise in subsequent processing.

[0035] In this embodiment, a background modeling subtraction method based on a nonlinear adaptive background update mechanism is employed to remove stable clutter components from the dynamically updated background template. First, the local rate of change of the signal needs to be detected and processed based on a sliding time window in the continuously acquired millimeter-wave radar echo signal. Specifically, for each sampling time point, a fixed-length time segment is selected before and after the current time point as a sliding window, for example, the window length is preferably set to 1 to 3 seconds, and the amplitude change rate between adjacent time points is calculated within this window. For the rate of change sequence within each local time window, the mean and variance of the rate of change are further calculated. When the mean rate of change is lower than a set threshold (e.g., less than 5% relative change) and the variance is lower than a set range, the signal in that local area is determined to be stable and belongs to the background clutter region. In the determined stable region, an exponentially weighted moving average method is used to dynamically update the background template, that is, the signal amplitude in the current window is weighted and superimposed with the existing background template according to a certain update weight. The update weight is preferably set between 0.01 and 0.1 to ensure that the background template can adaptively adjust with slow environmental changes, but will not respond quickly to instantaneous micro-motion signals.

[0036] During the continuous dynamic updating of the background template, if a sudden increase in signal amplitude is detected in a local area and its duration exceeds a preset threshold (e.g., a signal amplitude growth rate greater than 50% and lasting for more than 500 milliseconds), the area is identified as a dynamic activity protection zone. For dynamic activity protection zones, background template updates for the corresponding areas are paused to prevent the erroneous inclusion of weak vital signs signals into the background average, thus protecting the independence of the micro-movement components. Areas marked as dynamic activity protection zones are only allowed to be re-included in the background update process after the signal changes stabilize (e.g., the rate of change returns to the normal background range and persists for a certain period).

[0037] After performing background modeling subtraction, local energy contrast analysis is applied to the obtained purified signal to further identify potential weak components of vital signs signals. Specifically, the purified signal is divided into multiple local time windows, and the energy of the purified signal within each local window is integrated, while the energy integral of the background template signal within the corresponding window is extracted. Regions with energy differences greater than a set threshold, such as local regions where the purified signal energy is more than 30% higher than the background template energy, are marked as regions of significant energy difference. Regions of significant energy difference are prioritized for compensation in subsequent normalization processing because these regions are more likely to contain periodic micro-motion components related to vital signs.

[0038] In the normalization stage, a dynamic normalization baseline adjustment mechanism is introduced for regions with significant energy differences. Unlike traditional uniform normalization methods, this method dynamically adjusts the normalization parameters based on the ratio of local background energy level to the local energy level of the purified signal. When the local background energy level is low and the purified signal energy is high, the normalization compression ratio is appropriately reduced to enhance signal amplitude preservation; when the background energy level is high, a stronger amplitude normalization suppression strategy is adopted to weaken the influence of residual background noise. The normalization parameter adjustment is preferably based on adaptive optimization of the local signal-to-noise ratio, making the weakly periodic vital signs signal more recognizable relative to background noise in subsequent time-frequency transformation and trajectory extraction processes, thereby significantly improving the sensitivity and robustness of the overall monitoring system.

[0039] Step S102: Perform time-frequency transformation on the initial signal data based on a multi-scale sliding window to generate a time-frequency energy distribution map that can characterize the time-varying features of human vital signs signals.

[0040] In step S102, the clean initial signal data obtained in step S101 needs to undergo time-frequency transformation based on a multi-scale sliding window approach to extract the energy distribution characteristics of the signal over time. First, a sliding window needs to be defined on the initial signal data. The length of the sliding window should be set according to the characteristic frequency range of the target vital sign signal. Preferably, for respiratory signals, the sliding window length is set between 5 and 10 seconds; for heartbeat signals, the sliding window length is set between 1 and 3 seconds to ensure a balance between frequency resolution and time resolution. The sliding step size, i.e., the time interval for window sliding, is preferably set to 10% to 20% of the window length. For example, a step size of 1 second is set in a 10-second window to ensure sufficient temporal continuity and time-frequency resolution.

[0041] For each signal segment captured within a sliding window, a local spectral analysis is performed using Short-Time Fourier Transform (STFT). Specifically, a Hamming or Hanning window is used to window each signal segment to reduce spectral leakage caused by boundary effects. The windowed signal is then processed using Fast Fourier Transform (FFT) to obtain the frequency energy distribution corresponding to each window. The frequency range obtained by the transform should cover the characteristic frequency band corresponding to vital signs signals, i.e., it should contain at least frequency components in the range of 0.1 Hz to 3 Hz, to ensure that the characteristics of respiratory and heartbeat signals can be fully observed.

[0042] During the sliding process of the entire initial signal data, the above signal truncation, windowing, and Fourier transform steps are repeated to finally obtain a continuously changing two-dimensional time-frequency energy distribution map, where the horizontal axis represents time, the vertical axis represents frequency, and the amplitude represents the energy magnitude of the frequency component corresponding to each time segment. To enhance the resolvability of the map, the amplitude can be normalized, usually compressed to the [0,1] interval, so that feature extraction can be directly performed based on energy magnitude in subsequent processing. At the same time, in order to suppress abnormally high-energy isolated points caused by radar system noise or background multipath spurious signals, mean filtering or local mean smoothing can be applied to the generated time-frequency energy distribution map. The filtering window is preferably set to the neighborhood of 3 sampling steps on the time axis and 5 frequency sampling points on the frequency axis, thereby effectively eliminating isolated high-energy pseudo-peaks and improving the smoothness and stability of the overall map.

[0043] Throughout the multi-scale sliding window time-frequency transformation process, to simultaneously consider the characteristics of different vital sign signals, two sets of sliding window parameters can be used in parallel to independently generate two sets of time-frequency energy maps for low-frequency (respiration) and high-frequency (heartbeat) features, respectively. This avoids the problem of insufficient feature extraction capability for a certain frequency band due to the selection of a single scale window, and improves the accuracy of complete modeling of vital sign signals. Finally, based on the above processing, the construction of a time-frequency energy distribution map reflecting the dynamic characteristics of vital sign changes from pure initial signal data is completed, laying the foundation for subsequent effective signal component extraction and trajectory modeling.

[0044] Furthermore, the time-frequency transformation of the initial signal data based on a multi-scale sliding window includes:

[0045] Dynamically adaptive sliding window lengths are set for respiratory signals and heartbeat signals respectively. The respiratory signal window length is adjusted in real time according to the energy distribution trend of the target respiratory frequency range, while the heartbeat signal window length is dynamically adjusted according to the energy peak of the frequency components of the heart rate range.

[0046] Independent time-frequency energy distribution maps are generated for each scale, and a local energy enhancement mechanism is introduced during the map generation process. Nonlinear gain compensation is used for low-energy frequency regions to improve the detectability of weak vital signs signals.

[0047] In this embodiment, the multi-scale sliding window time-frequency transformation processing of the initial signal data requires setting dynamically adaptive sliding window lengths based on the frequency characteristics of different vital sign signals to balance the requirements of frequency resolution and time resolution. Specifically, for respiratory signal extraction, the energy distribution trend within the target respiratory frequency range is first determined during the initial analysis of the initial signal data. Typically, the respiratory frequency is between 0.1Hz and 0.6Hz; therefore, the sliding window length selected in the low-frequency region should be sufficiently long to provide high frequency resolution. Preferably, the initial sliding window length for the respiratory signal can be set to 10 to 20 seconds, and then dynamically adjusted based on the observed changes in respiratory frequency energy distribution. When large energy fluctuations are detected within the respiratory frequency range or a weakening of respiratory periodicity, the window length is appropriately extended to enhance frequency resolution; when respiratory energy is concentrated and signal changes are stable, the window can be appropriately shortened to improve time tracking capability.

[0048] For heartbeat signal extraction, since the heartbeat frequency is relatively high, typically ranging from 0.8Hz to 2.0Hz, a relatively short sliding window is required to avoid insufficient time resolution due to frequency spread. The initial sliding window length is preferably set to 3 to 5 seconds and dynamically adjusted based on the energy peaks of the frequency components within the heart rate range. When the heartbeat signal energy is concentrated and stable, the original window length is maintained; when the heartbeat frequency fluctuates drastically or multiple energy peaks exist, the window is appropriately shortened to improve the response speed to heart rate changes. The respiratory and heartbeat signals are processed using independent scale windows to ensure optimal capture of time-frequency characteristics under multi-scale analysis.

[0049] After setting up the multi-scale sliding window and segmenting the initial signal data, a short-time Fourier transform or other time-frequency transform methods need to be performed for each scale to generate an independent time-frequency energy distribution map. The time-frequency map corresponding to each scale is calculated independently, maintaining consistency between its time axis and frequency axis, and the energy distribution of frequency components within each time segment is recorded synchronously for subsequent feature extraction and trajectory construction.

[0050] To further enhance the detectability of weak vital signs in complex environments during the energy spectrum generation process, a local energy enhancement mechanism is introduced. Specifically, for low-energy frequency regions in the time-frequency energy spectrum, a nonlinear gain compensation method is used for enhancement. Nonlinear gain compensation can be achieved by setting an energy threshold. When the energy of a frequency component is lower than the threshold, the energy is boosted according to a preset nonlinear function (such as a square root function, logarithmic function, or exponential enhancement function), thereby amplifying weak but continuously existing periodic signals while suppressing the influence of isolated burst noise on the overall shape of the energy spectrum. The gain compensation parameters are dynamically set according to different scales. For example, a milder gain curve is used for the respiratory signal scale, while a stronger gain boost is used for the heartbeat signal scale to fully demonstrate the weak variation characteristics of different vital signs.

[0051] Through the comprehensive processing of dynamic adaptive sliding window length control, independent scale map generation, and local energy enhancement compensation, this step can effectively extract weak and stable breathing and heartbeat features from the initial signal of millimeter-wave radar, laying a solid and reliable time-frequency analysis foundation for subsequent signal component extraction and vital sign trajectory modeling.

[0052] Furthermore, the step of setting dynamically adaptive sliding window lengths for respiratory and heartbeat signals respectively includes:

[0053] In the initial stage, by analyzing the local energy peak density in the time-frequency energy distribution spectrum, the approximate range of respiratory and heart rate under the current environment is estimated;

[0054] Based on the preliminary prediction results, initial sliding window lengths were set for respiratory and heart rate signals, and adjustable window length ranges and dynamic adjustment step sizes were defined.

[0055] During continuous acquisition, the changes in local energy concentration are monitored in real time. When the local energy density fluctuates beyond the preset threshold within the sliding time window, the window length is adaptively shortened to improve the time domain response capability.

[0056] When the local energy density fluctuates below a preset threshold within the sliding time window and the frequency components remain stable, the window length is adaptively extended to improve the frequency resolution, and the window adjustment trajectory is recorded simultaneously for reference in subsequent feature extraction and anomaly detection.

[0057] In this embodiment, to set dynamically adaptive sliding window lengths for respiratory and heartbeat signals respectively, the initial stage requires analyzing the local energy peak density in the time-frequency energy distribution map to estimate the approximate range of respiratory and heartbeat frequencies under the current environment. Specifically, local energy peak detection is performed on the generated time-frequency energy distribution map using a peak extraction method based on energy thresholds. An appropriate energy detection threshold is set to identify significant local maxima on the frequency axis. For the detected local peaks, a preliminary classification is performed based on the frequency range and peak amplitude. High energy peaks with frequencies below 0.6Hz can be preliminarily classified into the respiratory signal range, and high energy peaks with frequencies between 0.8Hz and 2.0Hz can be classified into the heartbeat signal range. By statistically analyzing the number and distribution density of peaks in each range, the main frequency ranges and stability characteristics of respiratory and heartbeat signals in the current environment are further determined, providing a basis for the initial setting of the sliding window.

[0058] Based on the preliminary estimates above, initial sliding window lengths are set for both respiratory and heart rate signals. Due to the longer period of respiratory signals, a longer initial sliding window length is preferred, for example, within the range of 10 to 20 seconds, with the specific length further fine-tuned based on the concentration of respiratory rate distribution. Heart rate signals, with their shorter period, are preferably set with an initial sliding window length of 3 to 5 seconds. Simultaneously, an adjustable window length range and a dynamic adjustment step size need to be defined for each signal type, allowing the window length to dynamically shorten or lengthen based on subsequent energy changes, starting from the initial setting. The adjustable range is, for example, set to fluctuate within 30% of the initial value, and the dynamic step size is set to 5% to 10% of the initial window length, ensuring that the adjustment is flexible without causing drastic fluctuations that could affect the stability of subsequent analyses.

[0059] During continuous data acquisition, changes in local energy concentration are monitored in real time, serving as the basis for dynamically adjusting the sliding window length. For each sliding time window, the local energy integral within the respiratory or heart rate interval is calculated, and the degree of energy density fluctuation is also determined. If the fluctuation of local energy density within the current sliding window exceeds a preset threshold, such as exceeding 20% ​​of the local average energy level, it is considered that the current signal change rate has accelerated or environmental interference has increased. The system automatically shortens the sliding window length for the corresponding signal to improve time-domain response capabilities and better capture the characteristics of rapid changes in vital signs.

[0060] When the local energy density fluctuates below a preset threshold within the sliding time window, and the frequency components remain stable within continuous time segments, the current signal environment is determined to be stable. The system then adaptively extends the sliding window length for the corresponding signal to improve frequency resolution, thereby more accurately characterizing the periodic changes in vital sign signals. Simultaneously, during each sliding window length adjustment, the system needs to synchronously record the window adjustment trajectory, including the time point, direction, and magnitude of each change. These window adjustment trajectories serve as crucial references for subsequent feature extraction and anomaly detection, helping to identify the continuous characteristics of vital sign signal changes or potential environmental disturbances, thus supporting subsequent high-precision vital sign modeling.

[0061] Step S103: For the non-periodic interference signals present in the time-frequency energy distribution spectrum, based on the differences in energy concentration and stability of different frequency components in the time-frequency spectrum, extract the frequency components with high energy concentration and continuous and stable periodic changes on the time axis as effective signal components corresponding to human breathing and heartbeat characteristics.

[0062] In step S103, it is necessary to extract effective signal components that reflect the changes in human vital signs from the time-frequency energy distribution map generated in step S102. First, energy assessment is performed on each time-frequency point in the time-frequency energy distribution map, and the energy magnitude of different frequency components within each time segment is statistically analyzed. To ensure the representativeness of the extracted feature signals, frequency components with higher energy need to be selected based on the degree of energy concentration. Preferably, a dynamic energy threshold can be set for the energy values ​​of all frequency components within each time segment. The dynamic energy threshold can be determined based on the median or upper quartile of the frequency energy distribution of the current time segment. For example, the energy threshold can be set to more than 1.5 times the median energy value, retaining only frequency components with energy values ​​greater than this threshold, thereby eliminating frequency components with weak energy that may be generated by noise.

[0063] Based on energy-based frequency component screening, further screening is needed based on the stability differences of the frequency components over time. Specifically, for each frequency component retained through energy screening, its existence and variation trend are tracked over continuous time segments. If a frequency component can be detected in multiple consecutive time windows, and its center frequency varies little over time (preferably within ±0.05Hz), then the frequency component is determined to have continuous and stable periodic variation characteristics. Conversely, if a frequency component appears only in a few time segments, or its frequency fluctuates significantly over time, then the component is considered to likely originate from non-periodic interference and should be eliminated.

[0064] To facilitate implementation, a sliding time window method can be used to calculate the probability of existence and frequency fluctuation range of each frequency component. A sliding time window is defined as 10 seconds, with each step being 1 second. Within each window, the proportion of time each frequency component appears is counted, and the maximum frequency change amplitude of that component is calculated. When the probability of existence of a frequency component is greater than 80%, and the maximum frequency change amplitude is less than 0.05 Hz, the frequency component is considered to meet the requirements of continuous stable periodic variation.

[0065] The frequency components, after being screened for energy concentration and stability, are considered effective signal components corresponding to human respiration and heartbeat characteristics. Respiratory signals typically correspond to components with a frequency range of 0.1Hz to 0.6Hz, while heartbeat signals typically correspond to components with a frequency range of 0.8Hz to 2.0Hz. Therefore, when screening for effective signal components, frequency range can be used as an auxiliary constraint, retaining only effective frequency components within the typical frequency ranges of the aforementioned vital signs, further improving the accuracy and robustness of the screening.

[0066] Finally, through the dual screening process based on the differences in energy concentration and stability, effective signal components that can stably characterize human respiration and heartbeat were extracted from the time-frequency energy distribution map, laying a solid foundation for subsequent trajectory extraction and vital sign modeling.

[0067] Furthermore, based on the differences in energy concentration and stability of different frequency components within the time-frequency spectrum, frequency components with high energy concentration and continuous, stable periodic changes on the time axis are extracted as effective signal components corresponding to human respiration and heartbeat characteristics, including:

[0068] The ratio of the energy integral result within the local time window to the mean of the global energy distribution for each frequency component is standardized.

[0069] Time-domain smoothing filtering is applied to the standardized energy ratio to suppress energy anomalies caused by sudden noise.

[0070] Based on the smoothed energy trajectory, a multidimensional weighted score is performed using the combined frequency fluctuation standard deviation and existence probability, and the frequency component with the highest score is selected as the effective signal component for respiration and heartbeat.

[0071] In this embodiment, to accurately extract effective signal components corresponding to human respiration and heartbeat characteristics from the time-frequency energy distribution map, it is necessary to perform screening based on the differences in energy concentration and stability of different frequency components within the time-frequency map. Specifically, firstly, energy integration is calculated for each frequency component within a local time window. The length of the local time window is determined based on the changing characteristics of the target vital sign signal, preferably set between 10 and 20 seconds to cover the complete respiratory or heartbeat cycle. For each local time window, the energy values ​​of that frequency component in each time segment are accumulated to obtain the local energy integration result. Simultaneously, based on all global time segments, the average energy distribution of the same frequency component is calculated as a reference benchmark.

[0072] After obtaining the local energy integral and the global energy distribution mean, the ratio between the two is calculated and standardized. Standardization methods can include linear normalization, mapping the ratio to the [0,1] interval, or Z-score standardization, to eliminate incomparability caused by differences in baseline energy levels between different frequency components. This standardization step allows for direct cross-sectional comparison of the energy concentration changes of each frequency component over different time periods, facilitating subsequent stability analysis.

[0073] After standardization, to further suppress isolated energy spikes caused by radar system noise or environmental multipath interference, time-domain smoothing filtering needs to be applied to the standardized energy ratio. Preferably, moving average filtering or median filtering is used, with the sliding window length set to 2 to 5 seconds, adjusted according to the rate of change of the target signal. Moving average filtering effectively smooths small fluctuations within continuous time segments, while median filtering suppresses sudden energy spikes. The two can be used in combination for better filtering results. Time-domain smoothing filtering ensures that the subsequently extracted energy trajectory accurately reflects the periodic changes caused by human respiration or heartbeat, rather than accidental noise interference.

[0074] After obtaining the smoothed energy trajectory, further screening based on the stability characteristics of the frequency components is necessary. First, for each frequency component, the standard deviation of its frequency fluctuation over the entire monitoring period is calculated. This standard deviation can be obtained by tracking the frequency change trajectory across continuous time segments and calculating the standard deviation, thus assessing the temporal stability of the frequency component. A smaller standard deviation indicates greater stability of the frequency component over time, better reflecting the periodicity of vital signs. Second, the probability of each frequency component's existence within the monitoring period is statistically analyzed; that is, the proportion of times the energy of that frequency component exceeds a set threshold within the sliding time window out of the total number of time segments is calculated, reflecting its continuous existence.

[0075] After obtaining the standard deviation of frequency fluctuation and the probability of existence, a multi-dimensional weighted scoring mechanism is used to comprehensively score each frequency component. The weighted scoring can be set as follows: the standard deviation of frequency fluctuation is normalized to the [0,1] interval, and its reciprocal is taken as the stability score. At the same time, the probability of existence is directly used as the continuity score. The two are added together according to a preset weight (e.g., each accounting for 50%) to obtain the comprehensive score. The frequency components with the highest scores are selected as the effective signal components corresponding to human respiration and heartbeat characteristics, and are used for subsequent vital sign trajectory extraction and monitoring model construction.

[0076] Through the above processing flow, this step can accurately and reliably extract effective signal components reflecting the periodic changes of human vital signs from millimeter-wave radar time-frequency data in complex environments, greatly improving the robustness and accuracy of overall modeling and detection.

[0077] Furthermore, the standardization process for the ratio of the energy integral result of each frequency component within a local time window to the mean of the global energy distribution includes:

[0078] For each frequency component within the sliding time window, the local energy integral, local mean, and local energy variance are calculated respectively, and the ratio of the local integral to the local mean is used as the preliminary energy index.

[0079] Applying a normalization function based on dynamic adjustment of local energy variance to the preliminary energy index automatically enhances the ability to suppress energy anomalies in high-noise environments.

[0080] The normalized energy trajectory is smoothed using a two-scale method: short-scale smoothing is used to filter out isolated spikes, while long-scale smoothing is used to highlight periodic stable trends.

[0081] After multi-scale smoothing, frequency components with small curvature changes and high temporal continuity are selected as candidate components for vital sign signals based on local curvature change analysis of the trajectory.

[0082] In this embodiment, to standardize the energy integral result based on the ratio of the energy integral result within a local time window to the global energy distribution mean, it is necessary to calculate the local energy integral, local energy mean, and local energy variance for each frequency component within a sliding time window. Specifically, a fixed-length sliding time window is set, such as a continuous time segment of 5 to 10 seconds. For each frequency component, the energy values ​​corresponding to each time sampling point are accumulated within this window to obtain the local energy integral. Simultaneously, the energy of the frequency component within the window is averaged to obtain the local energy mean, and the variance of the energy values ​​within the window is further calculated to reflect the stability of the energy change of the frequency component. The local energy integral is mainly used to describe the overall energy intensity of the frequency component over a certain period of time, the local mean reflects the background baseline level, and the local variance reflects the energy fluctuation characteristics in the short term.

[0083] Subsequently, the ratio of the local energy integral to the local energy mean is used as a preliminary energy index. This preliminary energy index can intuitively reflect the degree of energy enhancement of a certain frequency component relative to the background baseline within a local time period. To further suppress abnormally high energy errors caused by isolated spikes or random fluctuations in high-noise environments, a normalization function based on dynamic adjustment of local energy variance is applied when standardizing the preliminary energy index. Specifically, when the local energy variance is small and the energy change is stable, conventional linear normalization is used; while when the local energy variance is large and the energy fluctuation is drastic, logarithmic compression or soft-threshold normalization functions are applied to weaken the influence of abnormal energy peaks. This adaptively adjusts the energy standardization results under different noise levels, ensuring that the normalized energy trajectory is more stable and reliable.

[0084] After standardization, the normalized energy trajectory needs to be smoothed using a dual-scale method to further eliminate isolated noise spikes and highlight periodic trends. Short-timescale smoothing preferably uses a small sliding window of 1 to 2 seconds, employing a moving average or median filter to filter out isolated instantaneous spikes and maintain signal continuity over short timescales. Long-timescale smoothing uses a larger sliding window of 5 to 10 seconds to smooth the overall trend of the energy trajectory, highlighting the stable changes in vital signs over longer periods. This combination of dual-scale smoothing simultaneously addresses fine-grained noise suppression and macroscopic trend extraction, improving the accuracy of subsequent feature selection.

[0085] After performing dual-scale smoothing, further screening was conducted based on local curvature change analysis of the trajectory. For each frequency component, the local curvature change of the trajectory was calculated within a sliding time window; that is, the smoothness of the trajectory change was measured by evaluating the second-order rate of change of the trajectory in continuous time segments. When the local curvature change of the trajectory is small and the direction of change is consistent, it indicates that the frequency component has good temporal continuity and periodic variation characteristics, which conforms to the typical characteristics of vital sign signals. Therefore, frequency components with small curvature changes and high temporal continuity were selected as candidate components of vital sign signals. Ultimately, the extracted frequency components not only have energy concentration and noise suppression, but also temporal continuity and smooth variation characteristics, providing a high-quality input foundation for subsequent trajectory modeling and vital sign monitoring.

[0086] Furthermore, the application of a normalization function based on dynamic adjustment of local energy variance to the preliminary energy index includes:

[0087] According to Formula 1 below, the local standard energy index E is calculated for each frequency component f within a sliding time window. norm (f,t):

[0088]

[0089] Among them, E local (f,t) represents the local energy integral of the frequency component f within the current sliding time window t; μ global (f) represents the average energy of frequency component f over the global time period; This represents the energy variance of the frequency component f within the current sliding time window t; λ represents the energy variance of the frequency component f over the global time period; λ is a normalization adjustment factor used to increase the global stability weight when the local noise level is high, with a preferred range of 0.1 to 0.5.

[0090] Calculate the dynamic signal stability weighting factor W(f,t) according to the following formula 2:

[0091]

[0092] in, E represents the normalized energy index norm The second time derivative of (f,t), where γ is a stability inhibition factor, preferably ranging from 0.01 to 0.1.

[0093] With E norm(f,t)×W(f,t) serves as the normalized energy feature and participates in subsequent frequency component screening. By introducing local noise adaptive normalization and second-order trend weighting, not only is the energy stability enhanced in low signal-to-noise ratio environments, but the robustness of identifying periodic vital signs signals is further improved.

[0094] In this embodiment, to improve the detection stability of weak human vital signs signals by millimeter-wave radar in complex environments, a normalization processing strategy based on dynamic adjustment of local energy variance is proposed. This strategy combines local noise adaptive normalization and dynamic stability weighting to enhance signal detectability and robustness. Specifically, the normalized local energy index is first calculated based on the signal characteristics within the local time window. This normalization process is implemented based on Equation 1:

[0095]

[0096] In Formula 1, E local (f,t) represents the local energy integral of frequency component f within the sliding time window t. Specifically, it is calculated by integrating or summing the energy values ​​corresponding to the frequency component at all times within the current sliding time window, thus characterizing the total energy performance of the frequency component within the local time scale. E local (f,t) can be achieved by extracting the energy value corresponding to the frequency component from the time-frequency energy spectrum and then using integration or summation. The sliding time window length is preferably set between 2 and 10 seconds, depending on the frequency characteristics of the target vital signs.

[0097] μ global (f) represents the average energy of frequency component f over the entire observation period. It is calculated by taking the arithmetic mean of the energy sequence of the frequency component over the global time range, thereby reflecting the average energy level of the frequency component on a long time scale, which is used as a benchmark reference for local changes.

[0098] The variance represents the energy variance of frequency component f within the current sliding time window t. It is obtained by averaging the squared deviations of the energy values ​​within the window from the local mean, reflecting the degree of energy variation of this frequency component within the local time scale. A smaller variance indicates a more stable signal, while a larger variance indicates stronger fluctuations or interference in the signal.

[0099] It represents the energy variance of frequency component f over the global time range. The calculation method is the same as that of local variance, but the integration range covers the entire observation time. It is used to describe the fluctuation level of the frequency component over the overall time range and serves as a benchmark for global noise stability.

[0100] λ is a normalization adjustment factor, a weighting coefficient used to adjust the contribution relationship between local noise and global noise to the standardization denominator. When the local noise level is high, compensation is achieved by multiplying the global energy variance by λ, avoiding instability of the standardization results due to high local noise. The preferred value range for λ is 0.1 to 0.5, with 0.2 recommended for general low-noise environments and 0.4 to 0.5 for high-noise environments.

[0101] After processing with Formula 1, the obtained E norm (f,t) reflects the degree of energy anomaly of the frequency component f within the current time window relative to the global average state and after noise stability adjustment, which can effectively suppress the spurious signal amplification phenomenon caused by local instantaneous noise.

[0102] After obtaining the standardized energy index E norm Following (f,t), a dynamic signal stability weighting factor W(f,t) is further introduced to suppress drastic changes in the energy trajectory of frequency components. The stability weighting factor is calculated based on the following formula 2:

[0103]

[0104] In Formula 2, The standardized energy index E norm The second derivative of (f,t) with respect to time is the rate of change of the trajectory curvature. A larger second derivative value indicates more drastic energy changes and greater fluctuations, while a smaller absolute value of the second derivative indicates smoother and more stable signal changes. The second derivative can be calculated using conventional numerical differentiation methods, such as the central difference method.

[0105] γ is a stability suppression factor, a positive proportionality coefficient used to control the degree of influence of the rate of change on the weighting factor. A larger γ will have a stronger suppression effect on the rate of change, thus favoring the selection of frequency components with stable energy trajectories. The preferred value range for γ is 0.01 to 0.1, with 0.02 to 0.05 recommended for applications with stable signal changes, and 0.05 to 0.08 recommended for applications with drastic changes or significant interference.

[0106] By using an exponential function, the weight of drastically changing frequency components is exponentially suppressed, while components with stable changes are given a higher weight, thus effectively screening out vital signs with stable frequencies and obvious periodicity.

[0107] Finally, the standardized energy index is multiplied by the stability weighting factor to obtain the weighted normalized energy characteristic, namely:

[0108] E final (f,t)=E norm (f,t)×W(f,t)

[0109] The weighted normalized energy characteristic E final (f,t) serves as the basic data input for subsequent frequency component screening, trajectory extraction, and vital sign modeling. Through a combination of local noise adaptive standardization and dynamic trend weighting, it not only improves the signal stability in low signal-to-noise ratio environments but also significantly enhances the detectability and robustness of periodic breathing and heartbeat signals in complex environments, thereby significantly improving the overall system's detection accuracy and practicality.

[0110] Step S104: Based on the extracted effective signal components, extract the corresponding frequency change trajectory in the time-frequency spectrum, establish a vital sign signal trajectory model based on the time-frequency characteristic change trend, and determine the optimal vital sign trajectory path based on the stability and continuity of the trajectory on the time axis.

[0111] In step S104, based on the effective signal components extracted in step S103, it is necessary to further extract the corresponding frequency change trajectories from the time-frequency energy distribution map. Specifically, firstly, time series analysis is performed on each effective signal component, tracing the trajectory of its frequency component changes over time along the time axis. To ensure the continuity of trajectory extraction, in each time segment, the frequency point that is closest to the known frequency position in the previous time segment and still meets the effective component screening criteria should be selected as the continuation point of the current trajectory. If there are multiple candidate frequency points in the current time segment, a unique trajectory continuation path can be determined by the minimum frequency offset priority matching principle.

[0112] When extracting the trajectory, the smoothness of frequency changes needs to be considered. Preferably, during trajectory extension, the frequency change within continuous time segments is constrained, limiting the frequency change per second to no more than 0.1 Hz, thereby avoiding abrupt trajectory changes due to occasional noise. For trajectory segments with breaks or jumps, linear interpolation correction can be performed based on the trajectory trends of surrounding time segments to ensure good continuity of the entire trajectory.

[0113] After trajectory extraction, a vital sign signal trajectory model based on the time-frequency characteristic variation trend needs to be established. This model uses time as the horizontal axis and frequency as the vertical axis, with the trajectory curve reflecting the variation trend of the vital sign signal over time. During the modeling process, the trajectory energy distribution, i.e., the energy intensity corresponding to the trajectory at each time segment, should be recorded simultaneously for subsequent trajectory stability and continuity analysis. The trajectory energy can be directly taken from the amplitude of the corresponding frequency point in the time-frequency energy distribution map, or local smoothing can be performed to reduce the impact of short-term energy fluctuations.

[0114] Based on the established trajectory model, it is necessary to further evaluate the stability and continuity of each trajectory over time. Trajectory stability can be assessed by calculating the standard deviation of the trajectory's frequency fluctuations over a specified time period. A standard deviation threshold of 0.05 Hz is preferred; a trajectory is considered stable if its standard deviation is less than this threshold for most time periods. Trajectory continuity is assessed based on the coverage ratio of the trajectory throughout the entire observation period. A coverage ratio threshold of 80% is preferred, meaning the trajectory exists continuously without interruption for at least 80% of the time segments.

[0115] Based on the evaluation results of the stability and continuity of the comprehensive trajectories, the optimal vital sign trajectory path is selected from all extracted trajectories. If multiple trajectories meet the requirements, the trajectory with the highest overall energy level is prioritized as the final optimal trajectory to ensure that the selected trajectory has strong vital sign signal energy support. Ultimately, the determined optimal trajectory path will be used to establish a subsequent vital sign monitoring model, serving as the core basis for characterizing the trends of changes in human respiratory rate and heart rate.

[0116] Furthermore, determining the optimal vital sign trajectory path based on the stability and continuity of the trajectory along the time axis includes:

[0117] For each frequency variation trajectory, the local energy density, time coverage, and frequency variation trend consistency index are calculated.

[0118] The above three indicators are fused and weighted according to a preset nonlinear combination function to form a comprehensive trajectory stability score.

[0119] Among all trajectories, the trajectory with the highest comprehensive score and a single-peak local energy distribution is selected as the optimal vital sign trajectory path to avoid misjudgment due to frequency drift caused by multi-peak overlap.

[0120] In this embodiment, to determine the optimal vital sign trajectory path based on the stability and continuity of the trajectory along the time axis, each extracted frequency change trajectory needs to be analyzed and processed in detail. First, for each trajectory, the local energy density, time coverage, and frequency change trend consistency index are calculated. The local energy density is calculated by statistically analyzing the time-frequency energy integral value corresponding to the trajectory within a sliding time window, and then standardizing this value with the total energy of all frequency components within the window to obtain the energy proportion of the trajectory in a local time period. Local energy density reflects the continuity and prominence of the trajectory's energy intensity during monitoring and is an important basis for judging the representativeness of the trajectory.

[0121] The time coverage rate is calculated as the ratio of the number of time segments in which the trajectory exists to the total number of time segments throughout the monitoring period. Specifically, in each time segment, if the frequency component energy of the trajectory is higher than a set detection threshold and the frequency change is continuous, the trajectory is considered to exist. The number of all existing segments is counted, and the ratio is calculated to the total number of segments. A high time coverage rate indicates that the trajectory has strong continuity and stable existence, and can reliably reflect the periodic change characteristics of vital signs signals.

[0122] The consistency index for frequency change trends is calculated by evaluating the coherence of frequency changes along the time axis of a trajectory. For each trajectory, a first-order difference sequence is first calculated based on the frequency changes in adjacent time segments. Then, a consistency score for the frequency change trend is calculated globally or locally, for example, by quantifying it through linear fitting slope consistency or directional continuity. A higher consistency score is achieved when the trajectory's frequency change trend is smooth and the direction of change is consistent; conversely, a lower score is achieved if the frequency change exhibits abrupt changes, jumps, or directional reversals.

[0123] After calculating the trajectory local energy density, time coverage, and frequency variation trend consistency index separately, these three indices need to be fused and weighted according to a preset nonlinear combination function. The preferred nonlinear combination function is a weighted normalized exponential function or a weighted logarithmic function, which can strengthen the nonlinear contribution of each index to the overall score and prevent a single index from abnormally inflating or deflating the overall score. The weight allocation can be adjusted according to specific scenario requirements; for example, the weight of the frequency consistency index can be increased in noisy environments, and the weight of local energy density can be increased when signal strength changes significantly. The standardized result of the overall score is used to unify the trajectory stability evaluation standard, ensuring the comparability of scores between different trajectories.

[0124] After calculating the comprehensive score for all trajectories, the trajectory with the highest comprehensive score is selected as the candidate trajectory. Simultaneously, to further avoid misjudgments due to trajectory frequency drift caused by frequency multi-peak overlap, the local energy distribution characteristics of the candidate trajectories need to be examined. Specifically, within each local time window, the distribution of trajectory energy along the frequency axis is analyzed. If the local energy exhibits a unimodal distribution, i.e., concentrated near a specific frequency without other significant secondary peaks, the trajectory's local energy distribution is considered to meet the unimodal characteristic requirement. If the energy exhibits a significant multimodal phenomenon, the trajectory is excluded, and the next trajectory with the second highest comprehensive score that meets the unimodal distribution condition is selected as the final optimal vital sign trajectory path.

[0125] Through the above-mentioned detailed and interconnected processing, the optimal trajectory path can be effectively screened out, which exhibits excellent performance in energy density, time coverage, and frequency variation trends, and has stable local energy distribution without aliasing. This allows for accurate tracking of human respiratory and heartbeat signal changes in complex non-line-of-sight environments and under multipath interference conditions, providing highly reliable basic data support for subsequent vital sign monitoring.

[0126] Furthermore, determining the optimal vital sign trajectory path based on the stability and continuity of the trajectory along the time axis also includes:

[0127] Among all the extracted frequency change trajectories, a preliminary screening is performed based on the time coverage of the trajectory, and only trajectories with a coverage higher than a set threshold are retained as valid candidate trajectories.

[0128] For each valid candidate trajectory, calculate the consistency score of the direction of frequency change within a local time window, and remove trajectories that frequently reverse direction.

[0129] Among the remaining trajectories, based on local energy distribution analysis, trajectories with a clear unimodal energy distribution pattern and a skewness greater than a preset positive value are selected to ensure energy concentration.

[0130] In this embodiment, to determine the optimal vital sign trajectory path based on the stability and continuity of the trajectory along the time axis, further screening and evaluation of all extracted frequency variation trajectories are required. First, preliminary screening of each trajectory is performed based on time coverage. Specifically, for each frequency variation trajectory, the number of time segments actually existing within the entire monitoring period is counted, and this number is calculated as a ratio to the total number of time segments to obtain the trajectory's time coverage. If the time coverage of a trajectory is lower than a preset threshold, such as below 70%, it is considered that the trajectory exhibits discontinuity, intermittency, or instability, possibly caused by noise or occasional interference. Therefore, it is discarded, and only trajectories with a time coverage higher than the set threshold are retained as valid candidate trajectories, ensuring that subsequent analysis is based on relatively continuous and reliable data.

[0131] For each valid candidate trajectory obtained from the initial screening, its frequency change direction consistency score needs to be further calculated within a local time window. Specifically, an appropriate sliding time window is selected, for example, 3 to 5 seconds. Within each window, the frequency change trend of the trajectory is tracked. By calculating the frequency change direction between adjacent time segments, it is determined whether the trajectory exhibits a consistent change on the local time scale. If the frequency direction reverses multiple times within a window (i.e., the frequency first rises and then falls, or changes repeatedly), it indicates that the trajectory may be affected by noise disturbance and lacks a stable change trend; therefore, the direction consistency score of this trajectory will be reduced. A threshold can be set based on the direction consistency score, for example, requiring that at least 80% of the frequency change direction remains consistent within the local window. If a trajectory fails to meet this requirement, it is discarded. In this way, discontinuous trajectories caused by environmental noise, system errors, or interference signals can be effectively removed.

[0132] To further improve the energy concentration and representativeness of the remaining trajectories, screening based on local energy distribution characteristics is necessary. Specifically, within each sliding time window, the local energy distribution near the corresponding frequency point of the trajectory is extracted, and the energy distribution pattern along the frequency axis is analyzed. By calculating the skewness index of the local energy distribution, it is determined whether the energy exhibits a clear single-peak pattern. When the skewness is positive and exceeds a preset positive threshold, for example, a skewness greater than 1.0, it indicates that the energy is concentrated on one side of the frequency axis, exhibiting a single concentrated peak distribution, consistent with the typical characteristics of vital signs such as breathing or heartbeat. If the skewness is close to zero or negative, it indicates that the energy distribution is scattered or has a multi-peak structure, possibly containing multipath interference or other stray signal components; such trajectories will be eliminated.

[0133] Through the above screening process, the trajectory that is ultimately retained has a high time coverage, good frequency change continuity and stability at the local scale, and the energy distribution shows obvious single-peak concentration characteristics. It can accurately and stably represent the human body's respiratory rate or heart rate, providing a solid foundation for finally determining the optimal vital sign trajectory path.

[0134] Step S105: Establish a vital signs monitoring model using the determined optimal vital signs trajectory path.

[0135] In step S105, a complete vital sign monitoring model needs to be established based on the optimal vital sign trajectory path determined in step S104. First, the optimal trajectory path should be used as the input basis, and real-time parameters should be extracted based on the trajectory's changes over time. Specifically, for respiratory signal monitoring, the instantaneous respiratory rate for each time period can be calculated directly based on the evolution of the center frequency of the trajectory within the respiratory frequency range over time. A sliding window method is preferred for local fitting, with the sliding window length set to 10 to 20 seconds and the step size set to 10% to 20% of the window length to smooth the respiratory rate change trend and suppress instantaneous fluctuations. For heart rate monitoring, the same sliding window local fitting method is used to extract the instantaneous heart rate based on the changes of the optimal trajectory within the heart rate range.

[0136] During the extraction of respiratory rate and heart rate, the frequency estimate at each moment should be amplitude-weighted. The amplitude weighting coefficient is taken from the energy value of the corresponding trajectory point in the time-frequency energy distribution spectrum, so that moments with high energy and high signal-to-noise ratio occupy a higher weight in the final monitoring curve, further improving the stability and accuracy of vital sign parameters. To suppress abnormal peaks or frequency jumps caused by instantaneous signal fluctuations, median filtering or local extremum suppression strategies are preferred for post-processing the extracted respiratory rate and heart rate sequences. The filtering window size is set to 3 to 5 data points, which can effectively filter out isolated abnormal points while maintaining the trend of vital sign changes.

[0137] At the output of the monitoring model, respiratory rate, heart rate, and their corresponding confidence indices should be recorded simultaneously. The confidence score can be calculated jointly based on trajectory energy level, trajectory continuity, and frequency stability. For example, by normalizing the trajectory energy and combining it with the trajectory existence rate, a monitoring confidence score can be given for each moment, with the score range set from 0 to 1. When the confidence score is lower than a set threshold (preferably 0.6), it can indicate that the reliability of the current monitoring results has decreased, avoiding misjudgment or misuse of low-quality data.

[0138] Furthermore, the vital signs monitoring model should support continuous output and anomaly detection. During continuous output, the respiratory rate and heart rate curves should be updated in real-time or near real-time, with the update cycle preferably set between 1 and 5 seconds to ensure the timeliness of the monitoring data. For anomaly detection, reasonable fluctuation ranges for vital sign parameters can be set, such as a respiratory rate between 8 and 30 beats per minute and a heart rate between 40 and 180 beats per minute. If the monitoring results exceed these ranges within a short period, an anomaly warning signal should be generated to alert the system or user to potential risks.

[0139] Through the above modeling process, a vital signs monitoring model based on millimeter-wave radar data and time-frequency characteristic trajectories was finally formed. This model can not only track the changes in human respiration and heart rate in real time, but also maintain high detection stability and robustness in complex non-line-of-sight scenarios with micro-motion interference, multipath reflection and environmental clutter, providing reliable data support for various health monitoring, smart healthcare and security applications.

[0140] The second embodiment of the application provides an electronic device, the electronic device comprising:

[0141] processor;

[0142] The memory is used to store a program, which, when read and executed by the processor, executes a millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling provided in the first embodiment of this application.

[0143] The third embodiment of this application provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it executes a millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling provided in the first embodiment of this application.

[0144] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling, characterized in that, include: By deploying millimeter-wave radar arrays, reflected echo signals are continuously acquired synchronously. The raw signals of each receiving channel are then subjected to time-domain filtering and noise reduction to obtain clean initial signal data. The initial signal data is transformed by time-frequency transformation based on a multi-scale sliding window to generate a time-frequency energy distribution map that can characterize the time-varying features of human vital signs signals. For the non-periodic interference signals present in the time-frequency energy distribution spectrum, based on the differences in energy concentration and stability of different frequency components in the time-frequency spectrum, frequency components with high energy concentration and continuous and stable periodic changes on the time axis are extracted as effective signal components corresponding to human breathing and heartbeat characteristics. Based on the extracted effective signal components, the corresponding frequency change trajectory in the time-frequency spectrum is extracted, a vital sign signal trajectory model based on the time-frequency characteristic change trend is established, and the optimal vital sign trajectory path is determined based on the stability and continuity of the trajectory on the time axis. A vital signs monitoring model is established using the determined optimal vital signs trajectory path.

2. The method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling according to claim 1, characterized in that, The step of performing time-domain filtering and noise reduction processing on the raw signal of each receiving channel includes: A background modeling subtraction method based on a nonlinear adaptive background update mechanism is adopted to remove stable clutter components from the dynamically updated background template; The signal after background removal is subjected to amplitude normalization, and a two-stage median filter is applied to the normalized signal. The first stage median filter is used to eliminate isolated pulse interference, and the second stage median filter adaptively adjusts the size of the filter window according to the local noise energy level of the sliding window, thereby enhancing the signal-to-noise ratio of the periodic vital signs signal.

3. The method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling according to claim 1, characterized in that, The time-frequency transformation of the initial signal data based on a multi-scale sliding window includes: Dynamically adaptive sliding window lengths are set for respiratory signals and heartbeat signals respectively. The respiratory signal window length is adjusted in real time according to the energy distribution trend of the target respiratory frequency range, while the heartbeat signal window length is dynamically adjusted according to the energy peak of the frequency components of the heart rate range. Independent time-frequency energy distribution maps are generated for each scale, and a local energy enhancement mechanism is introduced during the map generation process. Nonlinear gain compensation is used for low-energy frequency regions to improve the detectability of weak vital signs signals.

4. The method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling according to claim 1, characterized in that, Based on the differences in energy concentration and stability of different frequency components within the time-frequency spectrum, frequency components with high energy concentration and continuous, stable periodic changes on the time axis are extracted as effective signal components corresponding to human respiration and heartbeat characteristics, including: The ratio of the energy integral result within the local time window to the mean of the global energy distribution for each frequency component is standardized. Time-domain smoothing filtering is applied to the standardized energy ratio to suppress energy anomalies caused by sudden noise. Based on the smoothed energy trajectory, a multidimensional weighted score is performed using the combined frequency fluctuation standard deviation and existence probability, and the frequency component with the highest score is selected as the effective signal component for respiration and heartbeat.

5. The method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling according to claim 1, characterized in that, The determination of the optimal vital sign trajectory path based on the stability and continuity of the trajectory on the time axis includes: For each frequency variation trajectory, the local energy density, time coverage, and frequency variation trend consistency index are calculated. Based on a preset combination function, the trajectory local energy density, time coverage and frequency change trend consistency index are fused and weighted to form a comprehensive trajectory stability score. Among all trajectories, the trajectory with the highest comprehensive score and a single-peak local energy distribution is selected as the optimal vital sign trajectory path to avoid misjudgment due to frequency drift caused by multi-peak overlap.

6. The method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling according to claim 2, characterized in that, The background modeling subtraction method based on a nonlinear adaptive background update mechanism removes stable clutter components from the dynamically updated background template, including: Within each continuous acquisition period, the local rate of change of the signal is detected based on the sliding time window, and an exponentially weighted moving average is performed on the signal area with a rate of change lower than a preset threshold to dynamically update the background template. During the background template update process, if a region is detected where the signal amplitude increases instantaneously and the duration exceeds a preset threshold, the background update of the corresponding region is paused and marked as a dynamic activity protection zone to prevent weak vital signs signals from being mistakenly included in the background. After background subtraction, local energy comparison analysis is applied to the purified signal to calculate the energy difference between the local area of ​​the purified signal and the corresponding background area. Areas with significant energy differences are selected as the key compensation objects for subsequent normalization processing. During the normalization stage, a dynamic normalization baseline adjustment mechanism is introduced for regions with significant energy differences. The normalization parameters are adaptively adjusted based on the ratio of the local background energy level to the local energy level of the purified signal, so as to enhance the discriminability of weak periodic signals relative to residual background noise in subsequent processing.

7. The millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling according to claim 3, characterized in that, The method of setting dynamically adaptive sliding window lengths for respiratory and heartbeat signals respectively includes: In the initial stage, by analyzing the local energy peak density in the time-frequency energy distribution spectrum, the approximate range of respiratory and heart rate under the current environment is estimated; Based on the preliminary prediction results, initial sliding window lengths were set for respiratory and heart rate signals, and adjustable window length ranges and dynamic adjustment step sizes were defined. During continuous acquisition, the changes in local energy concentration are monitored in real time. When the local energy density fluctuates beyond the preset threshold within the sliding time window, the window length is adaptively shortened to improve the time domain response capability. When the local energy density fluctuates below a preset threshold within the sliding time window and the frequency components remain stable, the window length is adaptively extended to improve the frequency resolution, and the window adjustment trajectory is recorded simultaneously for reference in subsequent feature extraction and anomaly detection.

8. The millimeter-wave radar vital sign modeling method based on time-frequency feature decoupling according to claim 4, characterized in that, The standardization process for the ratio of the energy integral result of each frequency component within a local time window to the mean of the global energy distribution includes: For each frequency component within the sliding time window, the local energy integral, local mean, and local energy variance are calculated respectively, and the ratio of the local integral to the local mean is used as the preliminary energy index. Applying a normalization function based on dynamic adjustment of local energy variance to the preliminary energy index automatically enhances the ability to suppress energy anomalies in high-noise environments. The normalized energy trajectory is smoothed using a two-scale method: short-scale smoothing is used to filter out isolated spikes, while long-scale smoothing is used to highlight periodic stable trends. After multi-scale smoothing, frequency components with small curvature changes and high temporal continuity are selected as candidate components for vital sign signals based on local curvature change analysis of the trajectory.

9. The method for modeling vital signs of millimeter-wave radar based on time-frequency feature decoupling according to claim 5, characterized in that, The method of determining the optimal vital sign trajectory path based on the stability and continuity of the trajectory on the time axis also includes: Among all the extracted frequency change trajectories, a preliminary screening is performed based on the time coverage of the trajectory, and only trajectories with a coverage higher than a set threshold are retained as valid candidate trajectories. For each valid candidate trajectory, calculate the consistency score of the direction of frequency change within a local time window, and remove trajectories that frequently reverse direction. Among the remaining trajectories, based on local energy distribution analysis, trajectories with a clear unimodal energy distribution pattern and a skewness greater than a preset positive value are selected to ensure energy concentration.

Images