Heart rate and breath prediction method and system, program product, equipment and storage medium

Abstract

The invention provides a heart rate and breath prediction method, a program product, electronic equipment and a storage medium. The method comprises the following steps: acquiring an original signal sequence acquired from a human body through an ultra-wideband radar; determining a thoracic cavity displacement signal according to a mapping relation between the phase change and thoracic cavity displacement; performing multi-domain feature extraction on the thoracic cavity displacement signal to obtain multi-domain features; the multi-domain features comprise time domain features, frequency domain features and time-frequency domain features; on the basis of the multi-domain features, prediction values are obtained through a pre-trained prediction model, and the prediction values comprise the heart rate and the respiratory rate in the future time period. The robustness of the algorithm is enhanced through multi-dimensional feature fusion, high precision is still kept in an electromagnetic interference or multi-target scene, and the reliability of heart rate and respiratory rate prediction in a complex environment is improved. The physical signal advantage of the ultra-wideband radar is complementary with the environmental adaptability of the neural network, and the prediction accuracy of the heart rate and respiratory rate is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a heart rate and respiration prediction method, system, program product, electronic device and storage medium. BACKGROUND

[0002] Monitoring vital signs such as heart rate (HR) and respiration rate (RR) plays a crucial role in home healthcare, sports health, and smart home, etc. Traditional health monitoring methods usually use a contact method to collect respiration and heart rate data, such as wearing a heart rate monitor or manual auscultation. This method is usually not comfortable and has low convenience. SUMMARY

[0003] The purpose of the embodiments of the present application is a heart rate and respiration prediction method, program product, electronic device and storage medium, which are used to improve the above problems.

[0004] In a first aspect, the embodiments of the present application provide a heart rate and respiration prediction method, comprising: acquiring an original signal sequence collected by an ultra-wideband radar on a human body; the original signal sequence includes phase information and amplitude information; determining a chest displacement signal according to a mapping relationship between phase change and chest displacement; the chest displacement signal reflects the displacement of the chest with the heartbeat or respiration; performing multi-domain feature extraction on the chest displacement signal to obtain multi-domain features; the multi-domain features include time domain features, frequency domain features and time-frequency domain features; based on the multi-domain features, using a pre-trained prediction model to obtain a prediction value, the prediction value including heart rate and respiration rate in a future period.

[0005] In the above implementation process, the original signal sequence is collected by the ultra-wideband radar on the human body, providing high-resolution raw data for subsequent processing, avoiding privacy leakage of visual solutions, and improving the inconvenience and interference of traditional contact monitoring devices, and improving the convenience of home medical care. The multi-dimensional feature fusion enhancement algorithm enhances the robustness, and still maintains high accuracy in electromagnetic interference or multi-target scenarios, improving the reliability of heart rate and respiration frequency prediction in complex environments. The physical signal advantage of the ultra-wideband radar and the environmental adaptability of the neural network are complementary, improving the prediction accuracy of heart rate and respiration frequency.

[0006] Optionally, in the embodiments of the present application, the threshold method and / or the derivative method are used to detect extreme points of the chest displacement signal to obtain a sequence of peak time points of the respiratory cycle and / or the heartbeat cycle; time domain features are calculated according to the sequence of peak time points, the time domain features including peak interval features and waveform complexity features of the respiration or heartbeat; the peak interval features are used to reflect rhythm stability; the waveform complexity features are used to reflect signal morphology; the chest displacement signal is subjected to Fourier transform to obtain a frequency domain power spectrum; frequency domain features are calculated based on the frequency domain power spectrum; the frequency domain features include frequency energy distribution and main frequency energy proportion; multi-resolution analysis is performed on the chest displacement information to obtain a plurality of sub-bands; energy entropy of each sub-band is calculated to obtain time-frequency domain features.

[0007] In the implementation process described above, the multi-domain features cooperatively depict the essence of physiological state, the time domain features reveal physiological rhythm stability, the frequency domain features ensure signal purity, and the time-frequency features capture transient abnormal events (coughing / turning over), which are complementary to each other, thereby improving the robustness of the respiratory rhythm prediction under complex activity state and improving the accuracy of the respiratory rhythm prediction.

[0008] Optionally, in the embodiments of the present application, the prediction model includes an input layer, a bidirectional long short-term memory network, a convolutional neural network, a feature fusion layer and an output layer; the bidirectional long short-term memory network and the convolutional neural network are arranged in parallel; the input layer is used to receive multi-domain features of a continuous preset time window; the bidirectional long short-term memory network is used to capture the dependency between features by forward or reverse scanning of the multi-domain features to obtain long-time sequence dependency features; the convolutional neural network is used to extract features within a single time window in the multi-domain features in parallel to obtain spatial local features; the feature fusion layer is used to splice the long-time sequence dependency features and the spatial local features to generate global features; and the output layer is used to receive the global features and map the global features through a fully connected layer to generate a prediction value.

[0009] In the implementation process described above, the bidirectional long short-term memory network scans bidirectionally to capture gradual changes of the respiratory rhythm, such as slow deterioration trend of sleep apnea, thereby improving the long-term dependency decay problem of the traditional network. The convolutional neural network detects transient coupling of heartbeat and respiration to enhance short-time physiological interaction sensitivity. The bidirectional long short-term memory network and the convolutional neural network are arranged in parallel to realize cross-dimension complementation and form a holographic state representation. The spatio-temporal model reduces the prediction error of the respiratory frequency under motion interference and improves the accuracy of the prediction.

[0010] Optionally, in the embodiments of the present application, based on the multi-domain features, a pre-trained prediction model is used to obtain a prediction value, including: segmenting continuous multi-domain features based on a preset time window to obtain a plurality of feature sequences corresponding to the windows; and inputting the plurality of feature sequences into the pre-trained prediction model to obtain a prediction value.

[0011] In the implementation process described above, the continuous multi-domain feature ensures the statistical significance of feature extraction and reduces short-term fluctuation misjudgment; the sliding window is used for segmentation, the dynamic segmentation mechanism is used to solve the contradiction between continuity and real-time of long-term monitoring, the sliding window retains the continuity of historical data, and the truncation effect of the fixed window is improved.

[0012] Optionally, in the embodiment of the present application, the original signal sequence collected by the ultra-wideband radar on the human body is obtained, including: sending a pulse signal to the human body through the ultra-wideband radar, receiving a reflected echo, and generating an original signal sequence of a preset time window; calculating a channel impulse response based on the original signal sequence.

[0013] In the implementation process described above, the channel impulse response is calculated based on the original signal sequence, the channel impulse response separates the direct signal and the multipath noise, greatly improves the signal-to-noise ratio, and eliminates the false detection in the smart home multi-object environment.

[0014] Optionally, in the embodiment of the present application, the chest displacement signal is determined according to the mapping relationship between the phase change and the chest displacement, including: using a phase unwrapping algorithm to obtain the chest displacement signal based on the original signal sequence.

[0015] In the implementation process described above, the displacement signal directly quantifies the chest micro-motion, overcomes the measurement distortion caused by skin slip of the contact type device, and provides a physical true value for monitoring. The chest displacement signal is obtained by the phase unwrapping algorithm, and the displacement signal can accurately distinguish the physiological signals of deep breathing and vigorous exercise, and provide reliable data support for model prediction.

[0016] Optionally, in the embodiment of the present application, after obtaining the original signal sequence collected by the ultra-wideband radar on the human body, the method further includes: pre-processing the original signal sequence to obtain a pre-processed phase signal; the pre-processing includes at least one of offset correction, band-pass filtering, threshold denoising, variance screening and Chebyshev windowing; the offset correction is used to eliminate the signal baseline drift caused by the hardware coupling of the ultra-wideband radar; the band-pass filtering is used to filter out environmental static clutter and supraphysiological frequency band noise; the threshold denoising is used to suppress background electromagnetic noise; the variance screening is used to retain the high-variance signal segment of the target human body; and the Chebyshev windowing is used to eliminate the edge mutation in the original signal sequence.

[0017] In the implementation process, the offset correction eliminates hardware drift, so that the weak breathing signal of the baby can also be stably detected. The band-pass filter suppresses environmental mechanical vibration noise, ensuring the accuracy of the breathing frequency under the interference of the fitness equipment. The wavelet denoising can still recover the effective physiological signal in the negative signal-to-noise ratio environment, solving the pain point of electromagnetic interference of multiple devices in smart home. The variance screening reduces redundant interference and locks the main monitoring target chest signal. The Chebyshev windowing reduces the energy of spectral leakage, and reduces the frequency multiplication misjudgment caused by the spectral aliasing of heart rate. Through preprocessing, the signal-to-noise ratio of the chest displacement signal is greatly improved, providing a physical basis for prediction.

[0018] Optionally, in the embodiment of the present application, after obtaining the prediction value, the method further comprises: matching the heart rate and the breathing frequency in the future period in the prediction value with the activity feature type to obtain an activity feature; and acquiring a threshold corresponding to the activity feature, and triggering an alarm in a case where the prediction value exceeds the threshold.

[0019] In the implementation process, the phase change rate detection accurately locates the source of motion artifacts, reducing the misjudgment of physiological signals as jumps; the Kalman filter adaptive smoothing breaks through the radar half-wavelength displacement limit, reducing the heart rate monitoring error in high dynamic scenes such as running and standing up, but at the same time retaining the physiological response of real vigorous motion, and having reliability and motion adaptability. In addition, the Kalman filter smoothing has low computational complexity, is suitable for real-time processing scenarios, does not affect the timeliness of processing, and improves the accuracy of prediction.

[0020] Optionally, in the embodiment of the present application, after the chest displacement signal, the method further comprises: performing phase change rate detection on the chest displacement signal, and obtaining a jump position in the chest displacement signal if the detection result represents that the instantaneous change amplitude exceeds a preset threshold; performing Kalman filter smoothing processing on the chest displacement signal to adjust the signal value of the jump position, and obtaining a smoothed chest displacement signal; and performing multi-domain feature extraction on the chest displacement signal to obtain multi-domain features, including: performing multi-domain feature extraction on the smoothed chest displacement signal to obtain multi-domain features.

[0021] In the implementation process, the phase change rate detection accurately locates the source of motion artifacts, reducing the misjudgment of physiological signals as jumps; the Kalman filter adaptive smoothing breaks through the radar half-wavelength displacement limit, reducing the heart rate monitoring error in high dynamic scenes such as running and standing up, but at the same time retaining the physiological response of real vigorous motion, and having reliability and motion adaptability. In addition, the Kalman filter smoothing has low computational complexity, is suitable for real-time processing scenarios, does not affect the timeliness of processing, and improves the accuracy of prediction.

[0022] Secondly, embodiments of this application also provide a heart rate and respiratory prediction system, comprising: an ultra-wideband radar module for acquiring raw signal sequences collected from the human body via ultra-wideband radar; the raw signal sequences include phase information and amplitude information; a signal processing module for determining a chest cavity displacement signal based on the mapping relationship between phase change and chest cavity displacement; the chest cavity displacement signal reflects the displacement of the chest cavity with heartbeat or respiration; a feature extraction module for performing multi-domain feature extraction on the chest cavity displacement signal to obtain multi-domain features; the multi-domain features include time domain features, frequency domain features, and time-frequency domain features; and a prediction module for obtaining predicted values ​​based on the multi-domain features and using a pre-trained prediction model, the predicted values ​​including heart rate and respiratory rate in the future time period.

[0023] Thirdly, embodiments of this application also provide a computer program product, including computer program instructions, which are executed by a processor to perform the method provided in the first aspect or any implementation thereof.

[0024] Fourthly, embodiments of this application also provide an electronic device, including: a processor and a memory, the memory storing computer program instructions, which are executed by the processor to perform the method provided in the first aspect or any implementation thereof.

[0025] Fifthly, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, perform the method provided in the first aspect or any implementation thereof.

[0026] This application provides a heart rate and respiration prediction method, program product, electronic device, and storage medium. It utilizes ultra-wideband radar to collect raw signal sequences from the human body, providing high-resolution raw data for subsequent processing. This avoids privacy leaks associated with visual solutions and enhances the convenience of home-based medical care. Multi-dimensional feature fusion enhances the algorithm's robustness, maintaining high accuracy even under electromagnetic interference or multi-target scenarios, thus improving the reliability of heart rate and respiration frequency prediction in complex environments. The physical signal advantages of ultra-wideband radar complement the environmental adaptability of neural networks, improving the accuracy of heart rate and respiration frequency prediction. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1A flowchart of a heart rate and respiration prediction method provided by an embodiment of the present application is shown in FIG. 1. Figure 2 A structural diagram of a heart rate and respiration prediction system provided by an embodiment of the present application is shown in FIG. 2. Figure 3 A structural diagram of an electronic device provided by an embodiment of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0029] The embodiments of the technical solutions of the present application will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0031] In the description of the embodiments of the present application, the technical terms "first", "second", etc. are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified.

[0032] Traditional vital sign monitoring faces two major pain points: contact devices such as ECG patches may cause cutaneous hypersensitivity, long-term wearing discomfort, and cannot cover scenes such as bathing; non-contact solutions include camera visual analysis, but this way has the risk of privacy leakage, and fails in dark light and shielding environment. Although existing radar monitoring technology solves the privacy problem, it is limited by environmental noise interference and insufficient algorithm generalization, and it is difficult to meet the medical level precision requirement.

[0033] The embodiments of the present application provide a heart rate and respiration prediction method, which collects original signal sequences of human body through ultra-wideband radar, provides high-resolution original data for subsequent processing, avoids privacy leakage of visual solutions, and improves the convenience of home medical care. Multi-dimensional feature fusion enhances the robustness of the algorithm, and still maintains high precision in electromagnetic interference or multi-target scene, improves the reliability of heart rate and respiration frequency prediction in complex environment. The physical signal advantage of ultra-wideband radar and the environmental adaptability of neural network are complementary, which improves the prediction accuracy of heart rate and respiration frequency.

[0034] Before introducing the embodiments of the present application, the ultra-wideband radar in the embodiments of the present application is introduced first: the ultra-wideband radar of the present application refers to a UWB radar, which is a radar system using an ultra-wideband (UWB) signal. It is significantly different from a traditional narrowband radar (such as a car radar, a weather radar) in working principle and application scene.

[0035] The following is a detailed introduction to the UWB radar: 1. Definition: Ultra-wideband UWB signal refers to a radio technology that transmits very low power radio signals in a very wide frequency range (usually > 500 MHz). Its relative bandwidth (ratio of bandwidth to center frequency) is usually greater than 20%, or the absolute bandwidth is at least 500 MHz.

[0036] 2. Key features of ultra-wideband: including extremely wide bandwidth, low power spectral density, strong penetration ability, high range resolution, good multipath resolution. Among them, the extremely wide bandwidth refers to the extremely high time resolution. The strong penetration ability refers to the certain penetration ability of the UWB signal to non-metal materials (such as walls, clothes, human tissues, plastics, wood, soil, etc.). This makes the UWB radar have application potential in wall penetration imaging, living body detection, buried target detection, etc.

[0037] Please refer to Figure 1 A flowchart of a heart rate and respiration prediction method provided by the embodiments of the present application is shown. The heart rate and respiration prediction method provided by the embodiments of the present application can be applied to an electronic device, which can include a server, a PC, a tablet computer, a smart phone, or the like physical device, or can also be a virtual machine or a container, or the like virtual device. The electronic device can be a single device, or a combination of multiple devices, or a cluster of a large number of devices. The heart rate and respiration prediction method can include: Step S110: acquiring an original signal sequence collected by an ultra-wideband radar on a human body; the original signal sequence includes phase information and amplitude information.

[0038] Step S120: determining a chest displacement signal according to a mapping relationship between phase change and chest displacement; the chest displacement signal reflects the displacement of the chest with heartbeat or respiration.

[0039] Step S130: performing multi-domain feature extraction on the chest displacement signal to obtain multi-domain features; the multi-domain features include time domain features, frequency domain features, and time-frequency domain features.

[0040] Step S140: based on the multi-domain features, using a pre-trained prediction model to obtain a prediction value, the prediction value including heart rate and respiration frequency in a future period.

[0041] In step S110, the ultra-wideband radar is a radar that transmits nanosecond-level narrow pulses. The ultra-wideband radar can be a UWB radar. UWB is a short-range wireless communication method. The pulse of the ultra-wideband radar can penetrate obstacles such as clothes. After being reflected by the human chest cavity, the micro-motion displacement can be calculated by the phase difference and amplitude attenuation of the received echo.

[0042] The ultra-wideband radar can be integrated into a bedside terminal or a wall-mounted device and placed within a preset range from the human chest cavity to emit a pulse beam with a preset center frequency (for example, 7.98 GHz). The chest cavity region can be focused by directional beam orientation to suppress lateral interference. A distance gate can be set to filter reflected signals outside the range, such as wall or furniture signals. The human chest cavity micro-motion reflected echo can be received to separate the phase difference (caused by the change in optical path) and the amplitude attenuation (tissue reflection characteristics) to generate an original signal sequence.

[0043] The original signal sequence refers to the original echo data received by the radar. The original echo data includes phase information and amplitude information. The amplitude information is used to represent the intensity of the reflected signal. The phase information is used to represent the phase shift between the standard reflected wave and the transmitted wave.

[0044] In step S120, the chest cavity micro-motion causes the change in the optical path of the echo, and the phase change and the chest cavity displacement satisfy the mapping condition. The chest cavity displacement signal can be determined according to the mapping relationship between the phase change and the chest cavity displacement. In an optional embodiment, the chest cavity displacement signal can also separate the respiratory component and the heartbeat component. The respiratory component can be captured by band-pass filtering to capture slow fluctuations. The heartbeat component can be captured by band-pass filtering to capture high-frequency micro-motion.

[0045] In step S130, the time-domain feature reflects the waveform characteristics of the chest cavity displacement signal in time; the frequency-domain feature reflects the Fourier transform result of the chest cavity displacement signal; and the time-frequency domain feature reflects the time-frequency joint distribution of the chest cavity displacement signal in the wavelet domain.

[0046] For the time-domain feature, the extreme points of the chest cavity displacement signal can be located, and the time-domain feature can be calculated according to the sequence of the peak time points. The time-domain feature includes the peak interval feature of the respiration or heartbeat and the waveform complexity feature. The peak interval feature is used to reflect the rhythm stability, and the waveform complexity feature is used to reflect the signal form.

[0047] For the frequency-domain feature, the chest cavity displacement signal can be subjected to Fourier transform to obtain a frequency-domain power spectrum. The frequency-domain feature can be calculated based on the frequency-domain power spectrum. The frequency-domain feature includes the frequency-domain energy distribution and the main frequency energy proportion. The main peak frequency directly corresponds to the heart rate or the respiratory frequency, and the main frequency energy proportion is used to distinguish the main and secondary physiological signal components.

[0048] For time-frequency domain features: time-frequency analysis (such as short-time Fourier transform STFT, continuous wavelet transform CWT or Wigner-Ville distribution) is performed on the chest displacement signal to obtain a time-frequency graph; on this basis, time-frequency domain features are extracted, such as: wavelet energy entropy curve changing with time (reflecting the change of energy distribution in the frequency band with time); instantaneous frequency trajectory (such as the change of respiratory / heartbeat frequency with time); time-frequency ridge energy ratio (the proportion of the main frequency component in the time-frequency graph); time-frequency entropy: based on the time-frequency graph, the information entropy is calculated to reflect the complexity of the signal in the time-frequency plane.

[0049] In step S140, the model input is constructed based on the multi-domain features, for example, the continuous multi-domain features can be segmented based on a preset time window to obtain a plurality of feature sequences corresponding to the windows, and the plurality of feature sequences are input into a pre-trained prediction model to obtain a prediction value.

[0050] For example, 30 seconds of continuous multi-domain features can be obtained, segmented into 60 segments with a time window of 0.5 seconds each, each segment including the multi-domain features within 0.5 seconds to form a feature sequence, which can be in the form of a matrix. The plurality of feature sequences are input into a pre-trained prediction model to obtain a prediction value.

[0051] In the implementation process of the above embodiments: the original signal sequence of the human body is collected by the ultra-wideband radar to provide high-resolution original data for subsequent processing, avoid privacy leakage of visual solutions, and improve the inconvenience and interference of traditional contact monitoring devices, and improve the convenience of home medical care. The multi-dimensional feature fusion enhancement algorithm enhances the robustness, and still maintains high accuracy in electromagnetic interference or multi-target scenarios, and improves the reliability of heart rate and respiratory frequency prediction in complex environments. The physical signal advantage of the ultra-wideband radar and the environmental adaptability of the neural network are complementary, and the prediction accuracy of the heart rate and respiratory frequency is improved.

[0052] Optionally, in the embodiments of the present application, the multi-domain features are extracted from the chest displacement signal to obtain the multi-domain features, including: (1) Extraction of time domain features: threshold method and / or derivative method are used to detect extreme points of the chest displacement signal to obtain a sequence of peak time points of the respiratory period and / or heartbeat period; the extreme points include peak values and trough values. Taking the derivative method as an example, the derivative operation is performed on the chest displacement signal, and the zero crossing point of the first derivative and the negative second derivative represent the peak value; the zero crossing point of the first derivative and the positive second derivative represent the trough value. Through the derivative operation, a sequence of peak time points of the respiratory period and / or heartbeat period is obtained. As an implementation manner, the threshold method can be an adaptive threshold method.

[0053] The time domain features are calculated according to the sequence of peak time points, and the time domain features include peak interval features of respiration or heartbeat and waveform complexity features; the peak interval features are used for reflecting rhythm stability; and the waveform complexity features are used for reflecting signal morphology.

[0054] The peak interval features include four types: average peak interval, peak interval standard deviation, minimum peak interval and maximum peak interval. The average peak interval refers to the average of differences between adjacent peak times, and is used for representing average respiration or heartbeat period. The peak interval standard deviation refers to statistical dispersion of period fluctuation, and is used for quantifying rhythm variability. The minimum peak interval is used for identifying an extreme short period. The maximum peak interval is used for identifying a long period event.

[0055] The waveform complexity features also include four types: zero-crossing rate, kurtosis, skewness and root mean square value. The zero-crossing rate refers to the number of times of waveform crossing zero axis per unit time, and is used for reflecting signal oscillation frequency. The kurtosis refers to sharpness of waveform distribution, and high value in the kurtosis indicates a sudden and violent displacement (such as cough). The skewness reflects asymmetry of waveform distribution, and is used for identifying difference between inhalation and exhalation phases. The root mean square value reflects the ratio of effective value to absolute average value, and is used for comprehensively evaluating waveform distortion degree.

[0056] Exemplarily, the extracted time domain features can be 8-dimensional time domain features, including four types of peak interval features and four types of waveform complexity features.

[0057] (2) Extraction of frequency domain features: Fourier transform is performed on the chest displacement signal to obtain a frequency domain power spectrum; frequency domain features are calculated based on the frequency domain power spectrum; the frequency domain features include frequency energy distribution and main frequency energy proportion.

[0058] The chest displacement signal is added with a Hanning window to reduce spectrum leakage, and fast Fourier transform (FFT) is performed to obtain a complex spectrum. A square amplitude spectrum is calculated based on the complex spectrum to obtain a frequency domain power spectrum, and the frequency domain power spectrum represents energy distribution of the signal in the frequency domain.

[0059] The frequency energy distribution is used for reflecting power spectrum density (PSD) characteristics, and includes five types: PSD mean, PSD standard deviation, PSD maximum, PSD minimum and main peak frequency. The PSD mean represents average energy of the whole frequency band, and is used for evaluating signal overall strength. The PSD standard deviation represents dispersion degree of energy distribution, and is used for identifying spectrum stability. The PSD maximum and the PSD minimum represent positions of extreme energy values, and are used for capturing abnormal frequency points. The main peak frequency represents a frequency corresponding to the highest peak in the power spectrum, and is used for directly corresponding to a respiration rate.

[0060] The first main frequency energy proportion corresponds to the proportion of the energy corresponding to the main peak frequency in the total frequency energy; the second main frequency energy proportion corresponds to the proportion of the energy corresponding to the secondary peak frequency; and the third main frequency energy proportion corresponds to the proportion of the energy corresponding to the third peak frequency.

[0061] Exemplarily, the extracted frequency domain features can be 8-dimensional frequency domain features, including 5 frequency energy proportions and 3 main frequency energy proportions.

[0062] (3) Extraction of time-frequency domain features: multi-resolution analysis is performed on the chest displacement information to obtain multiple layers of sub-frequency bands; and energy entropy calculation is performed on each sub-frequency band to obtain time-frequency domain features.

[0063] The multi-resolution analysis can be a Daubechies wavelet basis analysis algorithm, which divides the chest displacement information in the time-frequency plane into multiple layers of sub-frequency bands, including a first layer of sub-frequency bands, a second layer of sub-frequency bands, a third layer of sub-frequency bands, and a fourth layer of sub-frequency bands. The first layer of sub-frequency bands is high-frequency details, mainly containing heartbeat components. The second layer of sub-frequency bands is medium-high frequency, which is a mixture of heartbeat and respiration, the third layer of sub-frequency bands is medium-low frequency, mainly containing respiratory components, and the fourth layer is low frequency, reflecting the basic displacement trend. Each layer of sub-bands retains time axis information and frequency information at the same time, realizing time-frequency localization.

[0064] Energy entropy calculation is performed on each sub-frequency band layer by layer: the energy entropy of the kth layer is sum(kth layer wavelet coefficient)^2. The entropy value reflects the degree of energy distribution disorder in different frequency bands. For example, an entropy value close to 0 indicates that the energy is concentrated in a single frequency band (ideal periodic signal), and an entropy value > 1 indicates the presence of non-stationary interference (such as cough). The energy entropies of the four layers are taken as four items in the time-frequency domain features.

[0065] The energy entropy can be bound to a time window: each energy entropy value is bound to the entropy value of the window in which the kth layer of sub-frequency bands is located, giving the energy entropy a time dimension. The entropy value sequence of the four layers of sub-frequency bands is aligned according to the time window, facilitating subsequent bidirectional long short-term memory network learning of the change law of the intelligent device over time.

[0066] Then, the proportion of each layer of energy in the total energy is calculated, and the proportion of the kth layer is kth layer energy entropy / total energy entropy, to obtain four layers of energy entropy proportions.

[0067] Exemplarily, the extracted time-frequency domain features can be 8-dimensional time-frequency domain features, including four layers of energy entropies and four layers of energy entropy proportions.

[0068] The multi-domain features can be 24-dimensional multi-domain features formed by combining 8-dimensional time domain features, 8-dimensional frequency domain features, and 8-dimensional time-frequency domain features.

[0069] In the implementation of the above embodiments: the multi-domain feature cooperatively depicts the essence of the physiological state, the time-domain feature reveals the stability of the physiological rhythm, the frequency-domain feature ensures the purity of the signal, and the time-frequency feature captures transient abnormal events (coughing / turning over), which are complementary to each other, thereby improving the robustness of the respiratory rhythm prediction under complex activity states and improving the accuracy of the respiratory rhythm prediction.

[0070] Optionally, in the embodiments of the present application, the prediction model comprises an input layer, a bidirectional long short-term memory network, a convolutional neural network, a feature fusion layer, and an output layer; wherein the bidirectional long short-term memory network and the convolutional neural network are arranged in parallel.

[0071] The input layer is configured to receive multi-domain features of a continuous preset time window. For example, the input layer receives a multi-domain feature sequence derived from a chest displacement signal for 30 seconds, which is divided into 2 time windows per second, and a total of 60 time windows. Each time window corresponds to a 24-dimensional feature vector (including time-domain peak interval, frequency-domain main frequency energy proportion, time-frequency wavelet entropy, etc.), forming a 60x24-dimensional feature sequence. This layer serves as a data inlet, ensuring that the features are input into the subsequent network in a time-sequential structure, and providing a standardized data basis for spatio-temporal analysis.

[0072] The bidirectional long short-term memory network is configured to scan the multi-domain features in a forward or reverse direction, capture the dependency between the features, and obtain long-time sequence dependent features.

[0073] The bidirectional long short-term memory network (Bi-LSTM) adopts a double-layer 128-neuron structure, scans the multi-domain features in a forward direction to learn the historical feature dependency, such as the continuity of the respiratory cycle, and scans in a reverse direction to capture potential future correlations, such as compensatory respiratory changes after heart rate acceleration. The long-time sequence regularity is selectively memorized through a gating mechanism (forgetting gate, input gate, and output gate), and 60x256 long-time sequence dependent features are output, which include long-period evolution patterns of the signal, such as progressive respiratory lengthening of suffocation risk.

[0074] The convolutional neural network is configured to extract features within a single time window in the multi-domain features in parallel, and obtain spatial local features.

[0075] The convolutional neural network (CNN) scans the input matrix through 64 one-dimensional convolution kernels (kernel size 3, step 1), each convolution kernel covers 3 adjacent time windows, i.e., the multi-domain features corresponding to 1.5 seconds of physiological signals. The convolution operation extracts spatial coupling features within the local window, such as the phase locking relationship between the heart rate peak and the respiratory trough, and the ReLU activation function enhances the non-linear expression ability. The output is 60x64-dimensional spatial features, which obtain spatial local features focusing on short-time correlations of transient physiological events, such as the instantaneous disturbance of coughing on heart rate.

[0076] The feature fusion layer is used for splicing long time sequence dependent features and spatial local features to generate global features. The feature fusion layer splices (Concatenation) the 60x256 dimensional long time sequence dependent features output by the Bi-LSTM and the 60x64 dimensional spatial local features output by the CNN along the feature dimension to generate a 60x320 dimensional global fusion feature matrix. The splicing operation preserves the integrity of the space-time features: the time sequence features carry rhythm evolution laws, such as the slow failure trend of respiration; and the spatial features encode transient interaction patterns, such as the enhanced coupling of heart and respiration after exercise, which complement each other to form a holographic physiological state representation.

[0077] The output layer is used for receiving the global features and mapping the global features through a fully connected layer to generate prediction values. The output layer maps the 320 dimensional global features to the target space through the fully connected layer: for example, first, weight matrix calculation: 320 dimensional global feature input x 12 dimensional output weight matrix, bias item optimization fitting deviation; then linear activation to ensure continuous physiological parameter values; and finally prediction value analysis: the first 2 dimensions of the 12 dimensional output represent the current heart rate (times / minute) and respiratory rate (times / minute); and the last 10 dimensions represent the heart rate and respiratory rate prediction values (2 dimensions per time point) at 2, 4, 6, 8 or 10 minutes in the future.

[0078] The essence of the output layer is a regressor that decodes the global features into interpretable current observation values and prediction values of heart rate and respiratory rate, realizing integrated output from real-time monitoring to trend prediction.

[0079] As an implementation: the model adopts a loss function combining mean square error (MSE) and cross entropy (CE) to adjust the relative importance of MSE and CE in the total loss. By optimizing this combined loss function, the model can improve the prediction accuracy of heart rate and respiratory rate while enhancing the early warning ability for abnormal situations. Mean square error (MSE) is used to measure the difference between predicted values and true values, which is used here to calculate the prediction error of heart rate and respiratory rate, so that the predicted values are as close to the true values as possible, thereby improving the accuracy of heart rate and respiratory rate prediction; cross entropy (CE) is commonly used in classification problems to measure the prediction accuracy of abnormal labels, which can divide the physiological state of the human body into normal and abnormal situations, and use labels to represent them.

[0080] In the implementation process of the above embodiment: the bidirectional long short-term memory network scans bidirectionally to capture the gradual changes of respiratory rhythm, such as the slow deterioration trend of sleep apnea, improving the long-term dependence decay problem of traditional networks. The convolutional neural network detects the transient coupling of heartbeats and respiration, enhancing the sensitivity of short-term physiological interaction. The bidirectional long short-term memory network and the convolutional neural network are arranged in parallel to realize cross-dimensional complementarity and form a holographic state representation. The space-time model reduces the prediction error of respiratory rate under motion interference and improves the accuracy of prediction.

[0081] Optionally, in the embodiments of the present application, based on the multi-domain features, a prediction model is trained in advance to obtain a prediction value, comprising: based on a preset time window, the continuous multi-domain features are segmented to obtain a plurality of window corresponding feature sequences; the plurality of feature sequences are input into the pre-trained prediction model to obtain the prediction value.

[0082] For example, the multi-domain feature flow extracted from the continuous 300-second chest displacement signal (2 24-dimensional feature vectors per second are generated) is obtained; based on a preset sliding window, the 300-second chest displacement signal is divided into 30-second units, and the window interval is 1 second; single window data: 30 seconds x 2 windows / second = 60 windows can generate a 60 x 24-dimensional feature matrix; the plurality of feature sequences are input into the pre-trained prediction model to obtain the prediction value. Of course, the continuous multi-domain feature duration and the preset sliding window described above can be set according to actual requirements.

[0083] In the implementation process of the above embodiments: the continuous multi-domain features ensure the statistical significance of feature extraction and reduce short-term fluctuation misjudgment; the sliding window is used for segmentation, the dynamic segmentation mechanism is used to solve the contradiction between continuity and real-time of long-term monitoring, the sliding window retains the continuity of historical data, and the truncation effect of the fixed window is improved.

[0084] Optionally, in the embodiments of the present application, the original signal sequence collected by the ultra-wideband radar on the human body is obtained, comprising: The ultra-wideband radar (center frequency 7.98 GHz) transmits a narrow pulse to the human chest, and the penetration depth of the clothes can reach 30 cm; the receiving end captures the reflected echo through an 11-channel antenna array, focuses on the chest region by using beamforming technology, and suppresses the reflection of static objects such as furniture; 12 seconds are continuously collected at a sampling rate of 20 Hz to generate an original signal sequence containing phase information and amplitude information.

[0085] The channel impulse response is calculated based on the original signal sequence. The channel impulse response (CIR) is a mathematical model that describes the channel characteristics experienced by the radar signal from the transmitting end to the receiving end. The core idea is that the complex transmission channel is regarded as a linear time-varying system, and the response of the linear time-varying system to the input signal (transmitted pulse) is the received signal.

[0086] For example, the radar echo is a two-dimensional matrix composed of fast time and slow time, and the processing of the original signal sequence is generally determined according to actual application, for example, 150 CIR data are generally required for respiratory and heart rate extraction to see the rule. The channel impulse response is solved from the original signal sequence by using pulse compression or deconvolution algorithm.

[0087] In the implementation process of the above embodiment: the channel impulse response is calculated based on the original signal sequence, the channel impulse response separates the direct signal from the multipath noise, greatly improves the signal-to-noise ratio, and eliminates the false detection in the smart home multi-object environment.

[0088] Optionally, in the embodiment of the application, the thoracic displacement signal is determined according to the mapping relationship between the phase change and the thoracic displacement, comprising: using a phase unwrapping algorithm, based on the original signal sequence, to obtain the thoracic displacement signal.

[0089] The phase unwrapping algorithm formula is:

[0090] Wherein, represents the thoracic displacement signal, and the speed of light is =3 e 8m / s, represents the center frequency (for example, 7.98GHz), represents the phase difference relative to the reference time.

[0091] In the implementation process of the above embodiment: the displacement signal directly quantifies the thoracic micro-motion, overcomes the measurement distortion caused by skin slip of the contact type device, and provides a physical true value for monitoring. The thoracic displacement signal is obtained by the phase unwrapping algorithm, and the displacement signal can accurately distinguish the physiological signals of deep breathing and vigorous exercise, and provide reliable data support for model prediction.

[0092] Optionally, in the embodiment of the application, after obtaining the original signal sequence collected by the ultra-wideband radar on the human body, the method further comprises: Pretreating the original signal sequence to obtain a pretreated phase signal; the pretreatment includes at least one of offset correction, band pass filtering, threshold denoising, variance screening and Chebyshev windowing; the offset correction is used to eliminate the signal baseline drift caused by the hardware coupling of the ultra-wideband radar; the band pass filtering is used to filter out the environmental static clutter and the super-physiological frequency band noise; the threshold denoising is used to suppress the background electromagnetic noise; the variance screening is used to retain the high-variance signal segment of the target human body; and the Chebyshev windowing is used to eliminate the edge mutation in the original signal sequence.

[0093] Offset correction: separate the direct current component from the original signal sequence, calculate the signal mean value, and obtain the optimized signal Xdc(n) by subtracting the signal mean value from the original signal sequence, thereby reducing the baseline drift (such as antenna standing wave effect) caused by radar hardware coupling, and ensuring that the phase zero is aligned with the thoracic static position.

[0094] Band-pass filtering: Design a finite impulse response filter with impulse response length of 100 taps, perform convolution: y(n) = Xdc(n) * h(n); where h(n) is the impulse response, y(n) is the filtered signal, Xdc(n) is the offset-corrected optimized signal. Through band-pass filtering, suppress ultra-low frequency vibration (<0.1 Hz, such as building sway) and high-frequency noise (>5 Hz, such as muscle tremor), and retain the effective frequency band of breathing and heartbeat.

[0095] Variance screening: Calculate the variance of each time row (radar channel): where, is the variance, M is the total number of time points of the signal, n is the time point index, is the filtered signal value of the i-th radar fast time row (channel) at time n, is the signal mean value of the i-th channel.

[0096] The rows with variance in the top pre-set quantile (such as Top 30%) can be screened out by variance screening, which can better reflect significant physiological activity and eliminate low-energy interference targets such as clothing fluttering.

[0097] Threshold denoising: Remove background electromagnetic noise through wavelet threshold denoising algorithm, threshold function process detail coefficients, and reconstruct signal to remove WiFi / Bluetooth impulse noise.

[0098] Chebyshev windowing: Remove discontinuity and outliers at the edges of the signal through Chebyshev window function, R is the sidelobe attenuation parameter, TN(x) is the Nth order first-class Chebyshev polynomial.

[0099] In an optional embodiment, the process of pre-processing the original signal sequence, the pre-processing sequence is: offset correction → band-pass filtering → variance screening → threshold denoising → Chebyshev windowing. The optimization principle is realized from global to local and from amplitude to phase. First, correct the baseband offset, the DC offset will distort the signal amplitude distribution, the first step eliminates and avoids introducing nonlinear errors in subsequent processing. Band-pass filtering is placed in front, limiting the signal bandwidth before denoising to prevent high-frequency noise from polluting low-frequency physiological components in wavelet decomposition. Variance screening precedes denoising, and effective channels are selected based on energy distribution to avoid redundant denoising calculations on ineffective channels. Wavelet denoising is performed after filtering, and band-pass filtering concentrates noise energy in the stop band, improving the adaptability of the wavelet threshold. Chebyshev windowing is finally implemented, and the windowing operation weakens the signal edge energy, which is placed at the end to prevent it from affecting the statistical effectiveness of variance screening and denoising. This process reduces the heart rate prediction error under motion interference and improves data accuracy.

[0100] ​​In the implementation process of the above embodiments: offset correction eliminates hardware drift, so that the weak breathing signal of the baby can also be stably detected. The band-pass filter suppresses environmental mechanical vibration noise, ensuring the accuracy of the breathing frequency under the interference of the fitness equipment. The wavelet denoising can still recover the effective physiological signal in the negative signal-to-noise ratio environment, solving the pain point of electromagnetic interference of multiple devices in smart home. The variance screening reduces redundant interference and locks the main monitoring target chest signal. The Chebyshev windowing reduces the energy of spectral leakage, and reduces the frequency multiplication misjudgment caused by the spectral aliasing of heart rate. Through preprocessing, the signal-to-noise ratio of the chest displacement signal is greatly improved, providing a physical basis for prediction.

[0101] Optionally, in the embodiments of the present application, after obtaining the prediction value, the method further comprises: Matching the heart rate and the breathing frequency in the future period in the prediction value with the activity feature type to obtain the activity feature.

[0102] The activity feature type can be a dynamic rule base, which records the heart rate and breathing frequency in different activity scenarios, such as sleep scenarios, exercise scenarios, or stress scenarios. By matching the predicted heart rate and breathing frequency in the future period with the activity feature type, the activity feature of the user can be discovered in advance.

[0103] For example, in the sleep scenario, the amplitude of limb displacement suddenly increases in a short time, and there is no regularity. Here it is necessary to distinguish from active movements such as turning over. At the same time, single or multiple, single convulsion duration ≤5s with frequency ≥1 / 5min, or multiple continuous convulsion duration ≤5s with frequency ≥1 / 2min, the activity feature type can be sleep limb convulsion type. In the exercise scenario, the heart rate slope > 1.2 times / second and the breathing depth increases by 50%, which can match the activity feature type as high-intensity intermittent training. In the stress scenario, the respiratory rhythm index > 0.4 and the heart and respiratory coupling degree < 0.1 can match the activity feature type as anxiety attack Suppose the prediction value is the heart rate [78, 82, 86] and the breathing [18, 19, 21] in the next 2 minutes; then the matching heart rate slope = 4 times / min / min (> threshold 1.2), breathing depth increases by 16.7%, and the matching activity feature type is moderate-intensity exercise.

[0104] Obtain the threshold value corresponding to the activity feature, and trigger an alarm if the prediction value exceeds the threshold value.

[0105] The threshold values corresponding to different types of activity features are different. For example, in the sleep scenario, the threshold value can be: turning over frequency > 3 times / minute + apnea > 15 seconds. If the threshold value is exceeded, sleep apnea syndrome warning can be performed.

[0106] The threshold of the motion overload activity feature can be: heart rate > x 0.9 + breath > 30 x 0.9 + breath > 30 times / min. When the threshold is exceeded, it indicates that the running heart rate exceeds the safe upper limit, and a motion warning needs to be given.

[0107] The threshold of the acute stress activity feature can be: heart-breath coupling < 0.08 for > 5 minutes. When the threshold is exceeded, it can be a test anxiety that causes palpitations, and a corresponding warning can be given.

[0108] In the implementation process of the above embodiments: by matching with the activity feature type and combining the corresponding threshold for warning, the coverage of multiple activity scenarios is realized in a single device, and the user does not need to wear multiple sensors. By using the predicted heart rate and breathing frequency in the future period, a progressive reminder is triggered when apnea occurs, improving user experience; it can also warn before the motion exceeds the safe load, realize the warning type health management, and improve the safety of the motion.

[0109] In an optional embodiment, a dynamic threshold model can be constructed based on the individual differences of exercise physiology. First, the basic threshold is calculated, using the age-corrected maximum heart rate data of the International Society of Sports Medicine, combined with the gender adjustment coefficient: women have a higher resting heart rate, and the threshold is adjusted down by 5-7%. For example, the maximum heart rate of a 30-year-old male is 187, and the maximum heart rate of a female is 175.

[0110] Then, based on the maximum heart rate data, static factors and dynamic factors, a personalized threshold model is generated using linear regression; the static factors are, for example, basal metabolic rate (BMR), BMI, medical history, etc.; the dynamic factors are, for example, the average resting heart rate in the past 7 days, and the sleep quality index.

[0111] The threshold model includes age-gender-health three-dimensional thresholds, for example, a 60-year-old male with diabetes (maximum heart rate of 166), historical resting heart rate of 72bpm, BMI=28, safe threshold set to 140bpm (static activity) / 120bpm (anaerobic exercise); if the predicted heart rate reaches 125bpm (>120bpm) when lifting iron, a warning is triggered.

[0112] The threshold model is calibrated by age-gender-health, giving a threshold that is more consistent with the user's own situation, reducing false alarms caused by using fixed thresholds for different gender users, different age users, or different health level users, greatly reducing the false alarm rate and misjudgment rate. Some sudden and severe movements (such as jumping, waving hands, and suddenly getting up from lying down) can cause the displacement of the chest to change instantaneously by an amplitude exceeding a threshold (for example, λ / 4, λ being the wavelength of the radar). The phase unwrapping algorithm can misjudge the waving hands and the like as a jump, resulting in an error in the displacement calculation. In the embodiments of the present application, Kalman filtering is used to smooth the chest displacement signal in this case, reducing the misjudgment of the jump.

[0113] Optionally, in the embodiments of the present application, after the chest displacement signal, the method further comprises: The phase change rate of the chest displacement signal is detected, and if the detection result represents that the instantaneous change amplitude exceeds a preset threshold, the jump position in the chest displacement signal is obtained. For example, the instantaneous change rate of the chest displacement signal can be calculated by first-order difference, and the instantaneous change rate can also be calculated by the formula: instantaneous change rate = Δφ / Δt, wherein Δφ is the phase difference between adjacent sampling points, and Δt is the sampling point time. By setting a threshold, if the instantaneous change rate exceeds π / 2 (corresponding to λ / 4 displacement), it is determined that a jump occurs, and the jump position in the chest displacement signal is determined.

[0114] The Kalman filtering is used to smooth the chest displacement signal, and the signal value of the jump position is adjusted to obtain a smoothed chest displacement signal.

[0115] First, a state model is established, and the chest displacement is regarded as uniform motion (state variables: displacement x, velocity v), and the sudden motion is regarded as a sudden disturbance. Then, an observation model is established: the phase change measured by the radar → displacement (including jump noise). The Kalman filtering steps are as follows: the current displacement is predicted according to the state at the last time (assuming uniform speed); if the observation value (caused by phase jump) is too different from the predicted value, the observation weight is reduced, and the influence of the jump is suppressed. In an optional embodiment, the process noise covariance can also be increased to allow the filter to track the real motion more quickly.

[0116] The multi-domain features of the chest displacement signal are extracted to obtain the multi-domain features, including: the multi-domain features of the smoothed chest displacement signal are extracted to obtain the multi-domain features.

[0117] In the implementation process of the above embodiments: the phase change rate detection accurately locates the source of the motion artifact, reducing the misjudgment of the physiological signal as a jump; the Kalman filtering adaptively smooths the sudden motion, breaking through the half-wavelength displacement limit of the radar, reducing the heart rate monitoring error in high dynamic scenes such as running and getting up, but at the same time retaining the physiological response of real sudden motion, and having both reliability and motion adaptability. Moreover, the Kalman filtering smoothing has low computational complexity, is suitable for real-time processing scenarios, does not affect the timeliness of processing, and improves the prediction accuracy.

[0118] See Figure 2A structure diagram of a heart rate and respiration prediction system is shown in the embodiments of the present application; the embodiments of the present application provide a heart rate and respiration prediction system 200, comprising: An ultra-wideband radar module 210 is configured to acquire an original signal sequence collected by an ultra-wideband radar on a human body; the original signal sequence comprises phase information and amplitude information; A signal processing module 220 is configured to determine a chest displacement signal according to a mapping relationship between phase variation and chest displacement; the chest displacement signal reflects displacement of a chest with heartbeats or respiration; A feature extraction module 230 is configured to perform multi-domain feature extraction on the chest displacement signal to obtain multi-domain features; the multi-domain features comprise time-domain features, frequency-domain features and time-frequency domain features; A prediction module 240 is configured to obtain a prediction value by using a pre-trained prediction model based on the multi-domain features; the prediction value comprises heart rate and respiration frequency in a future period.

[0119] Optionally, in the embodiments of the present application, the heart rate and respiration prediction system 200, the feature extraction module 230 is configured to detect extreme points of the chest displacement signal by using a threshold method and / or a derivative method to obtain a sequence of peak time points of a respiration period and / or a heartbeat period; time-domain features are calculated according to the sequence of peak time points; the time-domain features comprise peak interval features and waveform complexity features of respiration or heartbeat; the peak interval features are used to reflect rhythm stability; the waveform complexity features are used to reflect signal morphology; the chest displacement signal is subjected to Fourier transform to obtain a frequency-domain power spectrum; frequency-domain features are calculated based on the frequency-domain power spectrum; the frequency-domain features comprise frequency energy distribution and main frequency energy proportion; multi-resolution analysis is performed on the chest displacement information to obtain a plurality of sub-frequency bands; energy entropy of each sub-frequency band is calculated to obtain time-frequency domain features.

[0120] Optionally, in the embodiments of the present application, the heart rate and respiration prediction system 200, the prediction model comprises an input layer, a bidirectional long short-term memory network, a convolutional neural network, a feature fusion layer and an output layer; the bidirectional long short-term memory network and the convolutional neural network are arranged in parallel; the input layer is configured to receive multi-domain features of a continuous preset time window; the bidirectional long short-term memory network is configured to capture a dependency relationship between features by scanning the multi-domain features in a forward direction or a reverse direction to obtain long-time sequence dependency features; the convolutional neural network is configured to extract features in a single time window in the multi-domain features in parallel to obtain spatial local features; the feature fusion layer is configured to splice the long-time sequence dependency features and the spatial local features to generate global features; and the output layer is configured to receive the global features and map the global features through a fully connected layer to generate a prediction value.

[0121] Optionally, in the embodiment of the present application, the heart rate and respiration prediction system 200, the prediction module 240 is configured to segment the continuous multi-domain features based on a preset time window to obtain a plurality of feature sequences corresponding to the plurality of windows; and input the plurality of feature sequences into a pre-trained prediction model to obtain a prediction value.

[0122] Optionally, in the embodiment of the present application, the heart rate and respiration prediction system 200, the ultra-wideband radar module is configured to send a pulse signal to a human body through an ultra-wideband radar, receive a reflected echo, and generate an original signal sequence of a preset time window; and calculate a channel impulse response based on the original signal sequence.

[0123] Optionally, in the embodiment of the present application, the heart rate and respiration prediction system 200, the signal processing module is configured to obtain a thoracic displacement signal based on the original signal sequence by using a phase unwrapping algorithm.

[0124] Optionally, in the embodiment of the present application, the heart rate and respiration prediction system 200 further comprises a preprocessing module configured to preprocess the original signal sequence to obtain a preprocessed phase signal; the preprocessing comprises at least one of offset correction, band-pass filtering, threshold denoising, variance screening, and Chebyshev windowing; the offset correction is configured to eliminate signal baseline drift caused by ultra-wideband radar hardware coupling; the band-pass filtering is configured to filter out environmental static clutter and supraphysiological frequency band noise; the threshold denoising is configured to suppress background electromagnetic noise; the variance screening is configured to retain a high-variance signal segment of the target human body; and the Chebyshev windowing is configured to eliminate edge mutations in the original signal sequence.

[0125] Optionally, in the embodiment of the present application, the heart rate and respiration prediction system 200 further comprises an activity matching module configured to match an activity feature type according to a heart rate and a respiration rate in a future period in the prediction value to obtain the activity feature; obtain a threshold value corresponding to the activity feature, and trigger an alarm in a case where the prediction value exceeds the threshold value.

[0126] Optionally, in the embodiment of the present application, the heart rate and respiration prediction system 200 further comprises a smoothing module configured to perform phase change rate detection on the thoracic displacement signal, obtain a jump position in the thoracic displacement signal if a detection result represents that an instantaneous change amplitude exceeds a preset threshold value, perform Kalman filter smoothing processing on the thoracic displacement signal to adjust a signal value of the jump position, and obtain a smoothed thoracic displacement signal; and perform multi-domain feature extraction on the thoracic displacement signal to obtain the multi-domain features, comprising: performing multi-domain feature extraction on the smoothed thoracic displacement signal to obtain the multi-domain features.

[0127] It should be understood that the system corresponds to the heart rate respiration prediction method embodiments described above, and can perform each step involved in the above method embodiments. The specific functions of the system can be referred to the description above, and the detailed description is appropriately omitted here to avoid repetition. The system includes at least one software function module that can be stored in the form of software or firmware in the memory or solidified in the operating system (OS) of the system.

[0128] Please refer to Figure 3 The electronic device provided by the embodiments of the present application is shown in the structural schematic diagram. The electronic device 300 provided by the embodiments of the present application includes a processor 310 and a memory 320. The memory 320 stores machine readable instructions executable by the processor 310. When the machine readable instructions are executed by the processor 310, the method described above is performed.

[0129] Figure 3 The components shown in the above embodiments can be realized by hardware, software or a combination thereof. The electronic device 300 can be a physical device such as a server, a PC, etc., or a virtual device such as a virtual machine, a virtual container, etc. In addition, the electronic device 300 is not limited to a single device, but can also be a combination of multiple devices or a cluster of a large number of devices.

[0130] The embodiments of the present application also provide a storage medium, which stores a computer program. When the computer program is run by a processor, the method described above is performed.

[0131] The storage medium can be realized by any type of volatile or non-volatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0132] The embodiments of the present application also provide a computer program product, which includes computer program instructions. When the computer program instructions are run by a processor, the method described above is performed.

[0133] It should be understood that the disclosed system and method can also be implemented in other ways. The system embodiments described above are only illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible implementation architectures, functions and operations of the system, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in a different order from that shown in the accompanying drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0134] In addition, each functional module in each of the embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0135] The above description is only an optional implementation of the embodiments of the present application, but the protection scope of the embodiments of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the embodiments of the present application, which should be covered within the protection scope of the embodiments of the present application.

Claims

1. A method for predicting heart rate and respiration, characterized in that, include: Acquire the raw signal sequence collected from the human body by ultra-wideband radar; the raw signal sequence includes phase information and amplitude information; The chest cavity displacement signal is determined based on the mapping relationship between phase change and chest cavity displacement; the chest cavity displacement signal reflects the displacement of the chest cavity with heartbeat or respiration. Multi-domain features are extracted from the thoracic displacement signal to obtain multi-domain features; the multi-domain features include time-domain features, frequency-domain features, and time-frequency-domain features. Based on the multi-domain features, a pre-trained prediction model is used to obtain predicted values, which include heart rate and respiratory rate in the future time period.

2. The method according to claim 1, characterized in that, Multi-domain feature extraction is performed on the thoracic displacement signal to obtain multi-domain features, including: The extreme points of the thoracic displacement signal are detected using a threshold method and / or a derivative method to obtain the peak time point sequence of the respiratory cycle and / or heartbeat cycle; the time-domain features are calculated based on the peak time point sequence, and the time-domain features include the peak interval features of respiration or heartbeat and the waveform complexity features; the peak interval features are used to reflect rhythm stability; the waveform complexity features are used to reflect the signal morphology. The thoracic displacement signal is subjected to Fourier transform to obtain the frequency domain power spectrum; the frequency domain features are calculated based on the frequency domain power spectrum; the frequency domain features include frequency domain energy distribution and the proportion of energy at the dominant frequency; Multi-resolution analysis is performed on the thoracic displacement information to obtain multiple sub-bands; energy entropy is calculated for each sub-band to obtain the time-frequency domain features.

3. The method according to claim 1, characterized in that, The prediction model includes an input layer, a bidirectional long short-term memory network, a convolutional neural network, a feature fusion layer, and an output layer; wherein the bidirectional long short-term memory network and the convolutional neural network are configured in parallel. The input layer is used to receive the multi-domain features of consecutive preset time windows; The bidirectional long short-term memory network is used to capture the dependencies between features by scanning the multi-domain features in a forward or backward manner, and to obtain long-term time-dependent features. The convolutional neural network is used to extract features within a single time window from the multi-domain features in parallel to obtain spatial local features. The feature fusion layer is used to concatenate the long temporal dependency features and the spatial local features to generate global features; The output layer is used to receive the global features and map the global features through a fully connected layer to generate the predicted value.

4. The method according to claim 1, characterized in that, After acquiring the raw signal sequence collected from the human body by ultra-wideband radar, the method further includes: The original signal sequence is preprocessed to obtain a preprocessed phase signal. The preprocessing includes at least one of offset correction, bandpass filtering, threshold denoising, variance screening, and Chebyshev windowing. The offset correction is used to eliminate signal baseline drift caused by ultra-wideband radar hardware coupling. The bandpass filtering is used to filter out environmental static clutter and supraphysiological frequency noise. The threshold denoising is used to suppress background electromagnetic noise. The variance screening is used to retain high-variance signal segments of the target human body. The Chebyshev windowing is used to eliminate edge abrupt changes in the original signal sequence.

5. The method according to claim 1, characterized in that, After obtaining the predicted value, the method further includes: Based on the predicted heart rate and respiratory rate for future time periods, the activity characteristics are matched with the activity characteristic types to obtain the activity characteristics. Obtain the threshold corresponding to the activity feature, and trigger an alarm if the predicted value exceeds the threshold.

6. The method according to claim 1, characterized in that, Following the thoracic displacement signal, the method further includes: The phase change rate of the chest cavity displacement signal is detected. If the detection result indicates that the instantaneous change amplitude exceeds a preset threshold, the jump position in the chest cavity displacement signal is obtained. The chest cavity displacement signal is smoothed using a Kalman filter, and the signal value at the jump position is adjusted to obtain a smoothed chest cavity displacement signal. Multi-domain feature extraction is performed on the thoracic displacement signal to obtain multi-domain features, including: Multi-domain feature extraction is performed on the smoothed thoracic displacement signal to obtain the multi-domain features.

7. A heart rate and respiration prediction system, characterized in that, include: An ultra-wideband radar module is used to acquire the raw signal sequence collected from the human body by an ultra-wideband radar; the raw signal sequence includes phase information and amplitude information. The signal processing module is used to determine the chest cavity displacement signal based on the mapping relationship between phase change and chest cavity displacement; the chest cavity displacement signal reflects the displacement of the chest cavity with heartbeat or respiration. The feature extraction module is used to extract multi-domain features from the thoracic displacement signal to obtain multi-domain features; the multi-domain features include time-domain features, frequency-domain features, and time-frequency-domain features; The prediction module is used to obtain predicted values ​​based on the multi-domain features and using a pre-trained prediction model. The predicted values ​​include heart rate and respiratory rate in the future time period.

8. A computer program product, characterized in that, It includes computer program instructions that are executed by a processor to perform the method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, perform the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, perform the method as described in any one of claims 1 to 6.

Images