Millimeter wave radar health monitoring method and system based on wearable device

Abstract

The invention discloses a millimeter wave radar health monitoring method and system based on wearable equipment, and relates to the technical field of non-contact physiological signal monitoring, the specific steps are as follows: low-frequency life signals of a monitored object are collected based on frequency modulation continuous waves emitted by a radar system, and the low-frequency life signals are time sequence signals of multiple channels; performing wavelet packet decomposition and adaptive filtering on the low-frequency life signal to obtain a filtered signal; and inputting the filtered signal into a trained feature extraction and separation model, carrying out feature extraction and evaluation, and outputting a heartbeat signal and a respiration signal. According to the method, a strong anti-interference preprocessing process is constructed by fusing wavelet packet dynamic decomposition and a sub-band energy entropy threshold method and combining an NLMS adaptive filtering technology, motion artifacts and environmental noise can be effectively stripped, the signal-to-noise ratio of original millimeter wave radar signals is increased, and the anti-interference performance of the millimeter wave radar signals is improved. And a solid foundation is laid for subsequent accurate extraction of weak vital sign signals (respiration and heart rate).

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of non-contact physiological signal monitoring, and more particularly to a millimeter wave radar health monitoring method and system based on a wearable device. BACKGROUND

[0002] At present, the non-contact physiological signal monitoring technology based on millimeter wave radar has great potential in the field of wearable health monitoring (such as smart elderly care) due to its penetration, high sensitivity and high precision, and breaks through the limitations of space and time in traditional contact monitoring, providing the possibility of realizing remote real-time monitoring of users. However, this technology still faces significant challenges in the application process: first, the feature extraction capability is insufficient, and the traditional scheme relies on fixed window length FFT spectrum analysis, resulting in a loss of about 90% of the original information, making it difficult to restore key details such as respiratory rhythm fluctuations and heartbeat waveform variations; and when the respiratory and heartbeat spectra overlap, the FIR / IIR filter with a fixed cutoff frequency is easy to cause signal confusion, which seriously reduces the waveform fidelity. Second, the noise suppression capability is insufficient, and the multipath interference signals caused by metal objects often overlap with vital sign spectra, making time domain denoising methods ineffective; at the same time, human micro-movements have strong time-varying characteristics, and the existing adaptive filter has convergence delay due to fixed step size, poor motion artifact suppression effect, and lack of robust adaptive noise reduction capability. Third, the signal separation precision is not high, especially in the motion state, the measurement error increases significantly; the existing method often presets a respiratory harmonic template, which is easy to fail when the respiratory mode is abnormal, and the normal measurement range is limited; and the overall lack of effective multi-modal signal decoupling capability greatly limits the practicality of the device. Therefore, how to realize high-precision and robust separation and enhancement of respiratory and heartbeat signals and improve the accuracy and practicality of millimeter wave radar physiological monitoring is a problem that needs to be solved by those skilled in the art. SUMMARY

[0003] Therefore, the present application provides a millimeter wave radar health monitoring method and system based on a wearable device, which overcomes the above-mentioned defects.

[0004] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0005] A millimeter wave radar health monitoring method based on a wearable device, the specific steps are:

[0006] Collecting low-frequency life signals of a monitoring object based on a frequency-modulated continuous wave emitted by a radar system, the low-frequency life signals being time sequence signals of multiple channels;

[0007] Wavelet packet decomposition and adaptive filtering are performed on the low-frequency life signals to obtain a filtered signal;

[0008] The filtered signal is input into a trained feature extraction and separation model to perform feature extraction and evaluation, and output heartbeat signals and respiratory signals; the feature extraction and separation model includes a preprocessing layer, a feature extraction module, a quasi-periodicity construction module, and a feature state evaluation module.

[0009] Optionally, the radar system adopts a three-transmitting and four-receiving configuration and adopts a beamforming technology.

[0010] Optionally, the step of obtaining the filtered signal is:

[0011] The low-frequency life signal of any channel is subjected to wavelet packet decomposition to generate a plurality of sub-bands;

[0012] The energy entropy of the plurality of sub-bands is calculated, the plurality of sub-bands are dynamically classified into a first entropy zone, a second entropy zone, and a third entropy zone according to the energy entropy, and the sub-bands of each entropy zone are processed according to a preset rule;

[0013] Inverse wavelet packet transform is performed on the processed sub-bands to reconstruct time domain signals of any channel;

[0014] The step size of the NLMS filter is dynamically adjusted based on the energy of the time domain signals, and then the NLMS weight is calculated;

[0015] The time domain signals of each channel are fused to generate the filtered signal according to the NLMS weight.

[0016] Optionally, the energy entropy calculation formula is:

[0017]

[0018] In the formula, m is the number of coefficients of a sub-band; |c k,i | 2 is the energy of the i-th coefficient in the k-th sub-band; p k is the energy proportion of the k-th sub-band; E k is the energy of the k-th sub-band; E j is the energy of the j-th sub-band; and H is the Shannon entropy value.

[0019] Optionally, the preset rule of the second entropy zone is that a soft threshold method is used to perform amplitude attenuation and signal smoothing processing on the sub-bands of the second entropy zone based on an adaptive threshold.

[0020] Optionally, the calculation formula of the adaptive threshold is:

[0021] T k = T global *(1+αH k );

[0022] In the formula, Tglobal is the Donoho basic threshold; a is an adjustment factor; H k represents the current sub-band local energy entropy.

[0023] Optionally, the expression of the soft threshold method is:

[0024]

[0025] wherein, c k,i is the i-th coefficient in the k-th sub-band; T k is an adaptive threshold; and d is an entropy compensation term.

[0026] Optionally, the expression of the NLMS weight is:

[0027] w(n+1)=w(n)+μ(n)x(n)e(n);

[0028] wherein, w(n) is the weight of the input wavelet packet sub-band signal at the n-th moment; μ(n) is a time-varying step factor at the n-th moment; x(n) is the vector of the input wavelet packet sub-band signal at the n-th moment; and e(n) is the difference between the expected signal and the filtered output at the n-th moment.

[0029] Optionally, the expression of the filtered signal is:

[0030]

[0031] wherein, w i is the weight value of the i-th channel; x i (t) is the time-domain signal of the i-th channel at the t-th moment; μ i is the mean value of the i-th channel signal, and s i is the standard deviation of the i-th channel signal.

[0032] A wearable device-based millimeter wave radar health monitoring system, comprising:

[0033] A signal acquisition module configured to acquire a low-frequency life signal of a monitoring object based on a frequency-modulated continuous wave emitted by a radar system, the low-frequency life signal being a time-series signal of multiple channels;

[0034] A signal processing module configured to perform wavelet packet decomposition and adaptive filtering on the low-frequency life signal to obtain a filtered signal;

[0035] A feature extraction module configured to input the filtered signal into a trained feature extraction and separation model to perform feature extraction and evaluation, and output a heartbeat signal and a breathing signal; the feature extraction and separation model comprises a preprocessing layer, a feature extraction module, a quasi-periodicity construction module, and a feature state evaluation module.

[0036] Through the technical solution, the application provides a wearable device-based millimeter wave radar health monitoring method and system, which has the following beneficial effects compared with the prior art:

[0037] 1. By fusing wavelet packet dynamic decomposition and sub-band energy entropy threshold method, and combining NLMS adaptive filtering technology, a powerful anti-interference preprocessing procedure is constructed, which can effectively strip motion artifacts and environmental noise, and improve the signal-to-noise ratio of the original millimeter wave radar signal, thereby laying a solid foundation for subsequent accurate extraction of weak vital sign signals (respiration, heart rate).

[0038] 2. The dual-flow 1D-CNN+BiLSTM lightweight network architecture can not only separate basic signals such as heart rate and respiration with high precision, but also effectively learn and identify potential abnormal patterns, thereby improving the health risk warning capability.

[0039] 3. The diversified multi-scene sample library constructed based on the three-transmitting four-receiving antenna configuration and the beamforming FMCW radar provides a rich and practical data basis for the deep learning model, and effectively enhances the generalization capability and universality of the model in different application environments. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0041] Figure 1 The method flowchart provided by the present application is shown in the figure.

[0042] Figure 2 The acquisition flowchart of the filtered signal provided by the present application is shown in the figure.

[0043] Figure 3 The data processing flowchart of the feature extraction and separation model provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] The embodiment of the present application discloses a wearable device-based millimeter wave radar health monitoring method, as shown inFigure 1 As shown, the specific steps are:

[0046] Step 1, collecting the low-frequency vital signal of the monitoring object based on the frequency-modulated continuous wave emitted by the radar system, the low-frequency vital signal being a time sequence signal of multiple channels;

[0047] Step 2, wavelet packet decomposition and adaptive filtering are performed on the low-frequency vital signal to obtain a filtered signal;

[0048] Step 3, inputting the filtered signal into a trained feature extraction and separation model to perform feature extraction and evaluation, and outputting a heartbeat signal and a breathing signal.

[0049] In an embodiment, the step of obtaining the filtered signal is:

[0050] Wavelet packet decomposition is performed on the low-frequency vital signal of any channel to generate multiple sub-bands;

[0051] The energy entropy of the multiple sub-bands is calculated, the multiple sub-bands are dynamically classified into a first entropy region, a second entropy region and a third entropy region according to the energy entropy, and the sub-bands of each entropy region are processed according to a preset rule;

[0052] Inverse wavelet packet transform is performed on the processed sub-bands to reconstruct a time domain signal of any channel;

[0053] The step size of the NLMS filter is adjusted based on the energy of the time domain signal, and then the NLMS weight is calculated;

[0054] The time domain signals of the multiple channels are fused to generate the filtered signal according to the NLMS weight.

[0055] Further, as shown in the following formula (2), the wavelet packet decomposition recursively decomposes the high-frequency and low-frequency components to form finer frequency band division. j In contrast to the wavelet decomposition that decomposes the scale space V j+1 of the jth layer into the scale space V j+1 of the next layer and the wavelet space W j of the next layer; the wavelet packet decomposition further decomposes the wavelet space W s of the jth layer into finer subspaces. The low-frequency vital signal has strong time sequence, and sym8 is selected as the wavelet base function, which has roll-off characteristics of the passband edge to prevent sub-band leakage and 8th order vanishing moment to realize transient response. For the number of decomposition layers, the frequency of the breathing signal is about 0.1-0.5Hz, and the frequency of the heartbeat signal is about 0.7-3Hz, with a frequency difference of about 0.20Hz. According to the sub-band division frequency formula (1), assuming the sampling frequency is 50Hz, j is taken as 7, the signal is divided into 256 sub-bands, and the frequency band of the low-frequency vital signal is separated with high resolution.

[0056] Δf = f s / 2 j+1 (1);

[0057] where Δf is the width of each sub-band, f s is the sampling frequency.

[0058] Further, the energy entropy dynamic threshold replaces the fixed threshold, and the threshold is dynamically adjusted using the sub-band energy entropy. The higher the entropy value, that is, the more random the energy distribution, the larger the threshold, and the noise needs to be suppressed. And the lower the entropy value, that is, the more concentrated the energy distribution, the smaller the threshold, and the key features need to be preserved. According to the mathematical formula (2) energy entropy, the correction formula (3) is derived.

[0059]

[0060] In the formula, m is the number of coefficients of each sub-band, which is determined by the length of the signal segment; |c k,i | 2 is the energy of the i-th coefficient in the k-th sub-band; p k is the energy probability proportion of the k-th sub-band; E k is the energy of the k-th sub-band, so as to obtain the specific value H of Shannon entropy and perform dynamic classification; E j is the energy of the j-th sub-band.

[0061] The dynamic classification of the sub-band can be divided into a low-entropy area dominated by life signals, a high-entropy area dominated by noise, and a mixed area of both. The coefficients in the high-entropy area are forced to be zero, the threshold in the medium-entropy area is linearly increased, and the alpha is optimized by genetic algorithm, to realize accurate processing and classification.

[0062] T k = T global *(1+αH k )(3);

[0063] In the formula, T global is the Donoho basic threshold; alpha is the adjustment factor; H k represents the local energy entropy of the current sub-band.

[0064] where,

[0065] σ = MAD / 0.6745 (5);

[0066] In the formula, T is the Donoho global threshold, N represents the total number of points of the signal sampling segment, and sigma is the standard deviation of the estimated noise. MAD (median absolute deviation) is the median of the absolute deviation.

[0067] Further, the soft threshold method is used to attenuate the amplitude of the strong noise signal of the non-stationary signal and to smooth the signal. In the radar life signal, the life signal in the chest region needs a continuous waveform. The soft threshold method is used to avoid additional vibration interference frequency detection.

[0068]

[0069] wherein δ = β (1-H k ) is an entropy compensation term, when H k tends to 0, it indicates that the signal is relatively pure, which may cause overcompensation and inhibit features, and when it tends to 1, it indicates that the signal is noise-dominated, which needs to avoid amplifying noise.

[0070] Further, the weight update step is dynamically adjusted by the signal energy entropy joint adaptive NLMS algorithm, and the convergence speed and steady accuracy are balanced by real-time feedback of the error and the input signal. First, the sub-band signals after classification processing are integrated into a continuous waveform, starting from the 256 sub-bands at the bottom layer, and according to the parent-child sub-band relationship during decomposition, the sub-band signals are merged layer by layer upwards through the inverse transform operator corresponding to sym8 used in the decomposition process, and finally the complete time domain signal is reconstructed. For the complete signal of each channel after integration, p global is calculated according to the definition of energy entropy. H out is obtained, and the calculation formula is shown in formula (2). The NLMS update formula is formula (7), and the step is adjusted by introducing energy entropy feedback, which effectively avoids the contradiction between convergence speed and steady error.

[0071] w(n+1) = w(n) + μ(n)x(n)e(n) (7);

[0072] wherein x(n) represents the vector of the input wavelet packet sub-band signal at the n th moment; e(n) represents the difference between the expected signal and the filtered output at the n th moment, and μ(n) represents the time-varying step factor at the n th moment;

[0073]

[0074] wherein σ v is the standard deviation of noise, i.e. the estimation of the lifeless signal period; σ e (n) is the standard deviation of the error signal sliding; is the introduced entropy feedback term. When H out is large and tends to 1, i.e. in a high noise state, μ(n) is reduced to suppress the divergence risk, and vice versa, i.e. in a signal pure state, μ(n) is increased to accelerate convergence. After normalization processing of each channel, the signal segment is obtained by weight addition fusion as shown in formula (9). By adaptively and dynamically adjusting the parameters, the contradiction between convergence speed and stability is effectively avoided, and the waveform fidelity and detection accuracy are improved.

[0075]

[0076] where w i is the weight value obtained by the NLMS algorithm; μ i is the mean of the i-th channel signal, σ i is the standard deviation of the i-th channel signal.

[0077] In an embodiment, in step 3, before training the feature extraction and separation model, a standard data set is made based on the time series samples of the low-frequency vital signals collected in the chest region, and a training data set is constructed according to the standard data set; specifically:

[0078] A configuration of 3 transmitters and 4 receivers is adopted, and the technology of beamforming is used to reduce environmental interference. Frequency-modulated continuous wave (FMCW) is transmitted, and 5000 groups of sample data random length fragments are collected, including signal samples of different material clothes, different postures including sitting, standing and lying, and different environments, to ensure the integrity of the sample data. The collected data is stored as a csv file type, with rows representing the time axis and columns representing the position change and the number of radar transmission channels (in this embodiment, the number of radar transmission channels is 3*4=12). The specific phase and amplitude can be read from each row and column, and 80% of the standard data set is randomly taken as the training set, 10% as the validation set, and 10% as the test set, which facilitates subsequent data training.

[0079] Further, the feature extraction and separation model adopts a 1D-CNN+BiLSTM lightweight network, which puts the signal data after noise reduction and adaptive filtering into the network for training, and trains the most suitable model parameters to realize accurate separation of breathing and heartbeat.

[0080] In an embodiment, the feature extraction and separation model mainly includes a preprocessing layer (Preprocessing Layer), a feature extraction module (Feature-extraction Module), a quasi-periodic construction module (Quasi-periodic construction Module), and a signal quality assessor (Signal Quality Assessor). The data processing steps are as shown in Figure 3 The feature learning is guided by physiological characteristics, the low-frequency vital signal is divided into a heartbeat branch and a breathing branch, then the quasi-periodic characteristics are used to construct a state memory bank by BiLSTM, the future state is captured by bidirectional fusion through forward LSTM and reverse LSTM, and the signal data is effectively fitted.

[0081] Further, in the preprocessing layer, the filtered single-channel mixed time series signal data is converted into a Numpy array. First, the signal length is standardized using RobustScaler, which is scaled by the median and interquartile range, and has strong robustness to outliers. The time steps of all samples are unified to a fixed length (max_len=500). If the sample length exceeds max_len, it is compressed to max_len by step sampling (e.g., ::int(X_norm.shape[1] / max_len)). If the sample length is less than max_len, it is padded with zeros at the end. When the sample length is less than half of max_len, the sample is discarded directly. Second, the two-dimensional array (samples, timesteps) is adjusted to a three-dimensional array (samples, timesteps, 2) to meet the requirements of the convolutional neural network (CNN) for time series data.

[0082] Further, in the feature extraction module, a deeper convolutional layer is used to extract and identify high-frequency weak features in the heartbeat signal; a wider convolution kernel is used to directly capture the overall envelope of the respiratory waveform.

[0083] Further, in the heartbeat branch, the first layer of convolution: the convolution kernel size is 5 (kernel_size=5), the stride is 1 (stride=1), the padding is 1 (padding=1), the number of channels is 16, and the bias (bias) is set to True. This layer is used to capture the local detail features of the heartbeat signal, and Batch Normalization is used to batch normalize the input x to speed up the model training and maintain its stability, apply the ReLU activation function to introduce non-linear features and enhance the model learning ability; the second layer of convolution: the convolution kernel size is 3 (kernel_size=3), the stride is 1 (stride=1), the padding is 1 (padding=1), the number of channels is 32, and a smaller convolution kernel is used for weak feature extraction. After Batch Normalization and ReLU activation function, the pooling process (MaxPooling1D) is performed, the pooling window size is 2 (pool_size=2), the step is 2 (strides=2), the feature dimension is reduced, and the main features of the heartbeat signal are extracted; since there are position information, emission radar frame serial number and measurement target ID and other multi-class data, a multi-scale feature fusion module is used to capture different scale feature information by using convolution layers with different expansion rates in parallel. The standard convolution with a convolution kernel size of 3 (kernel_size=3), a stride of 1 (stride=1), a padding of 1 (padding=1), and a receptive field size of 3 (dilation_rate=1) is used. The expansion convolution with a receptive field size of 9 (3+3*(3-1)=9) (dilation_rate=3) is used, and the features of the two branches are spliced by Concatenate, and Batch Normalization and ReLU activation function are used.

[0084] Further, in the breathing branch, the first layer of convolution: the convolution kernel size is 25 (kernel_size=25), the stride is 1 (stride=1), the padding is 1 (padding=1), the number of channels is 24, and the bias (bias) is set to True. A wide convolution is used to cover 1-2 breathing periods, and then AveragePooling1D is used for pooling to adapt to low-frequency smooth signals to directly capture the overall envelope of the breathing waveform. The second layer of convolution: to capture different scale feature information, a convolution kernel size of 15 (kernel_size=15) with an output of 32 channels is used, the stride is 1 (stride=1), the padding is 1 (padding=1), and then Batch Normalization and ReLU activation function are used.

[0085] Further, in the quasi-periodic construction module, the two physiological signal features of heartbeats and breaths are processed respectively using bidirectional LSTM. A quasi-periodic fusion model is constructed, which includes heartbeat feature dimension (heart_feat_dim), respiratory feature dimension (resp_feat_dim), LSTM hidden layer size (hidden_size) and output class number (num_classes). The input feature dimension of the heartbeat signal is 96, the hidden layer size is set to 128, the number of layers is 2, and the Dropout regularization is set to 0.2 to improve the model expression ability and generalization ability. The final output is the heartbeat feature sequence (batch_size, seq_len, heart_feat_dim). Similarly, the same operation is performed on the respiratory feature dimension, and the input feature dimension is 32. Using the bidirectional structure, the forward LSTM processes the sequence in time order (t = 1→t = seq_len), and the backward LSTM processes the sequence in reverse time order (t = seq_len→t = 1). The forward and backward hidden states are concatenated to form the final output.

[0086] Further, in the signal quality assessment module, the heartbeat signal feature and the respiratory signal feature are independently input, and two independent quality assessment networks (heart_quality_net) and (resp_quality_net) are defined, corresponding to the heartbeat and respiratory signals respectively. For the quality assessment of the heartbeat signal, the peak value of the heartbeat signal is detected by scipy.signal.find_peaks to calculate the RR interval, i.e. the time interval between adjacent heartbeat peaks, including the average RR interval, the RR interval standard deviation, i.e. the heart rate variability, to judge the stability of the heartbeat rhythm. For the respiratory signal, after band-pass filtering preprocessing of the signal, the inspiration / expiration peak value is detected, the average respiratory rate and the respiratory interval standard deviation are calculated, and the stability of the respiratory rhythm is judged. Secondly, both networks calculate the average amplitude, signal standard deviation and signal change rate to judge the integrity of the signal and noise interference. The calculation method of the signal quality score (Quality Score, QS) is shown in formula (9).

[0087] QS = 0.6 * rhythm stability + 0.4 * signal integrity (9);

[0088] The rhythm stability of the heartbeat signal is (1-RR interval standard deviation normalized value) * 0.5 + RR interval efficiency * 0.5, and the rhythm stability of the respiratory signal is (1-breathing interval standard deviation normalized value) * 0.5 + breathing cycle efficiency * 0.5, and the signal integrity is (average amplitude normalized value * 0.4 + signal noise ratio normalized value * 0.4 + signal change rate normalized value * 0.2). When QS is greater than 0.6, it is proved that the processing signal quality is qualified, the phase and amplitude are converted into complex signals by using the phase_amplitude_to_complex function, and then the inverse fast Fourier transform (IFFT) is performed on the complex signals by using np.fft.ifft, so as to convert the complex representation in the frequency domain into real signals in the time domain. Through the network, the input sequence is processed step by step, and the forward and reverse dependencies of the sequence are captured at the same time, and the periodic ability of the heartbeat and respiratory signals is enhanced.

[0089] In another aspect, the embodiment also discloses a wearable device-based millimeter wave radar health monitoring system, comprising:

[0090] A signal acquisition module is configured to acquire low-frequency life signals of a monitoring object based on a frequency-modulated continuous wave emitted by a radar system, wherein the low-frequency life signals are time sequence signals of multiple channels.

[0091] A signal processing module is configured to perform wavelet packet decomposition and adaptive filtering on the low-frequency life signals to obtain filtered signals.

[0092] A feature extraction module is configured to input the filtered signals into a trained feature extraction and separation model to perform feature extraction and evaluation, and output heartbeat signals and respiratory signals. The feature extraction and separation model comprises a preprocessing layer, a feature extraction module, a quasi-period construction module, and a feature state evaluation module.

[0093] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0094] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A health monitoring method based on millimeter-wave radar using wearable devices, characterized in that, The specific steps are as follows: The low-frequency life signals of the monitored objects are collected based on the frequency-modulated continuous wave emitted by the radar system, and the low-frequency life signals are time-series signals from multiple channels. The low-frequency life signal is subjected to wavelet packet decomposition and adaptive filtering to obtain the filtered signal; The filtered signal is input into the trained feature extraction and separation model for feature extraction and evaluation, and outputs heartbeat and respiratory signals. The feature extraction and separation model includes a preprocessing layer, a feature extraction module, a quasi-periodic construction module, and a feature state evaluation module.

2. The millimeter-wave radar health monitoring method based on wearable devices according to claim 1, characterized in that, The radar system adopts a three-transmitter, four-receiver configuration and employs beamforming technology.

3. The millimeter-wave radar health monitoring method based on wearable devices according to claim 1, characterized in that, The steps for obtaining the filtered signal are as follows: Wavelet packet decomposition is performed on the low-frequency life signal of any channel to generate multiple sub-bands; Calculate the energy entropy of multiple sub-bands, dynamically classify the multiple sub-bands into a first entropy region, a second entropy region, and a third entropy region based on the energy entropy, and process the sub-bands in each entropy region according to preset rules; Perform inverse wavelet packet transform on the processed sub-bands to reconstruct the time-domain signal of any channel; The NLMS filter step size is dynamically adjusted based on the energy of the time-domain signal, and then the NLMS weights are calculated. The time-domain signals from each channel are fused according to the NLMS weights to generate the filtered signal.

4. The millimeter-wave radar health monitoring method based on a wearable device according to claim 3, characterized in that, The formula for calculating energy entropy is: In the formula, m represents the number of coefficients in the sub-band; |c k,i | 2 p represents the energy of the i-th coefficient in the k-th sub-band. k E represents the energy percentage of the k-th sub-band. k E represents the energy of the k-th sub-band. j Let H be the energy of the j-th sub-band; H is the Shannon entropy value.

5. A millimeter-wave radar health monitoring method based on a wearable device according to claim 3, characterized in that, The preset rule for the second entropy region is: based on an adaptive threshold, a soft thresholding method is used to perform amplitude attenuation and signal smoothing processing on the sub-frequency bands of the second entropy region.

6. A millimeter-wave radar health monitoring method based on a wearable device according to claim 5, characterized in that, The formula for calculating the adaptive threshold is: T k =T global *(1+αH k ); In the formula, T global The Donoho baseline threshold; α is the adjustment factor; H k This represents the local energy entropy of the current sub-band.

7. A millimeter-wave radar health monitoring method based on a wearable device according to claim 5, characterized in that, The expression for the soft thresholding method is: In the formula, c k,i T is the i-th coefficient in the k-th sub-band; k δ is the adaptive threshold; δ is the entropy compensation term.

8. A millimeter-wave radar health monitoring method based on a wearable device according to claim 3, characterized in that, The expression for NLMS weights is: w(n+1)=w(n)+μ(n)x(n)e(n); In the formula, w(n) is the weight of the input wavelet bun frequency band signal at time n; μ(n) is the time-varying step size factor at time n; x(n) is the vector of the input wavelet bun frequency band signal at time n; and e(n) is the difference between the desired signal and the filtered output at time n.

9. A millimeter-wave radar health monitoring method based on a wearable device according to claim 1, characterized in that, The expression for the filtered signal is: In the formula, w i x is the weight value of the i-th channel; i (t) represents the time-domain signal of the i-th channel at time t; μ i Let σ be the mean of the i-th channel signal. i Let be the standard deviation of the i-th channel signal.

10. A millimeter-wave radar health monitoring system based on wearable devices, characterized in that, include: The signal acquisition module is used to acquire low-frequency life signals of the monitored object based on the frequency-modulated continuous wave emitted by the radar system. The low-frequency life signals are time-series signals from multiple channels. The signal processing module is used to perform wavelet packet decomposition and adaptive filtering on the low-frequency life signal to obtain a filtered signal; The feature extraction module is used to input the filtered signal into the trained feature extraction and separation model, perform feature extraction and evaluation, and output heartbeat and respiratory signals. The feature extraction and separation model includes a preprocessing layer, a feature extraction module, a quasi-periodic construction module, and a feature state evaluation module.

Images