Fine-grained vital sign signal reconstruction method based on encoder-decoder network

Abstract

The application discloses a fine-grained vital sign signal reconstruction method based on an encoder-decoder network, extracts fine-grained vital sign signal waveforms from original wireless radio frequency signals, models the basic distribution of fine-grained vital sign signals based on original radio frequency signals, and automatically models the nonlinear relationship between the original radio frequency signals and the real signal waveforms under the driving of the original radio frequency signals and the real signal waveforms. No additional complex features need to be extracted from the signals, the daily behavior of the user is not limited, data augmentation transformation operations are performed on the extracted phase information caused by the target user, the availability of the data is improved to a greater extent, the latent space is normalized by encoding the latent vector into a probability distribution, the network can be sampled in a more meaningful way, overfitting of the network is avoided, and the input beyond the range is better processed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of millimeter wave radar sensing vital signs, and relates to a contactless vital sign monitoring method, in particular to a fine-grained vital sign signal reconstruction method based on a millimeter wave radar and an encoder-decoder network. BACKGROUND

[0002] The rapid development of the Internet of Things injects many new applications into daily life, such as behavior recognition and identity authentication of target individuals through densely deployed wireless devices in multiple environments, or fall detection and vital sign monitoring of the elderly in home environments, which greatly brings convenience to people. Therefore, contactless sensing has attracted the attention of researchers in academia and the industry. In these studies, the use of various commercial-grade radar RF sensing methods has proven its development prospects due to its characteristics of anti-noise, easy deployment, and reasonable cost. Many research works differ in the RF technology used, including Impulse Radio-UltraWide-Band (IR-UWB), Frequency Modulated Continuous Wave (FMCW) radar, and WiFi. Among these technologies, FMCW has the advantages of high resolution, low transmission power, and separation of different motion sources in the environment, and is more robust in extracting vital sign information, and can monitor multiple users at the same time, so it is more widely used by academia and industry.

[0003] Existing RF sensing systems usually assume that users are static objects and cannot provide continuous monitoring. In the actual environment, it is not expected that the user will remain static during the period, on the one hand, the user himself will have a sustained motion, such as typing, turning pages, etc. On the other hand, unintentional posture deviations and unconscious movements (such as nodding, shaking the body, etc.) may always occur. When there is a user body motion, the fine-grained breathing, heartbeat, and other vital sign information in the signal obtained by reflection may be masked, and effective information cannot be obtained.

[0004] Traditional signal separation schemes seek conditions that can separate mixed signals from the perspective of time domain or frequency domain. Common methods include band-pass filters and modal decomposition. The band-pass filter refers to a filter that can pass frequency components within a certain frequency range, but attenuate frequency components of other ranges to a very low level. However, the signal waveform generated by the band-pass filter cannot describe the characteristics of the motion cycle of vital sign signals to a great extent: a narrower range of filter results in only the fundamental frequency without details, and a wider range of filter results in the waveform being distorted by harmonic frequency components. Compared with the filtering method focusing on spectral analysis, the modal decomposition method for time domain analysis has better adaptability in signal separation, such as Ensemble Empirical Mode Decomposition (EEMD), Variational Mode Decomposition (VMD) and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). These methods decompose the signal into components called Intrinsic Mode Function (IMF). Such methods roughly recover the fundamental frequency components of vital sign signals, and the obtained waveform does not have the detailed part that needs to be analyzed. Although there are nonlinear filtering methods, they may not help to reconstruct the fine-grained waveform of recovered breathing and heartbeat due to two reasons. On the one hand, the recursive approximation method cannot achieve satisfactory performance. On the other hand, the design of nonlinear filters specific to the task (such as median filters and linear internal parameter nonlinear filters) requires rich professional experience and domain knowledge, but they are difficult to generalize.

[0005] In summary, the existing methods are difficult to extract fine-grained vital sign waveforms from signals containing multiple interference sources, because these methods all involve linear operations and are more for processing linearly synthesized signals. The reflection signals caused by target body movements, breathing and heartbeat are mixed in a nonlinear way, and the traditional signal decomposition method is not suitable for extracting fine-grained waveform data of each mode for such obtained signals. SUMMARY

[0006] In view of the above problems of the prior art, the present application provides a fine-grained vital sign signal reconstruction method based on an encoder-decoder network, to solve the technical problem that the existing signal decomposition method is difficult to obtain fine-grained vital sign signal waveform.

[0007] In order to solve the above technical problems, the present application adopts the following technical solutions to achieve it:

[0008] A fine-grained vital sign signal reconstruction method based on an encoder-decoder network, the method comprising the following steps:

[0009] Step 1: Device deployment, deploy radar devices and wearable devices;

[0010] Step 2: Data acquisition, collect wearable device wireless signals;

[0011] Step 3: Wireless signal corresponding phase information extraction and preprocessing: process the collected wireless signal data to obtain clean phase information;

[0012] Step 4: Wireless signal phase information data enhancement processing:

[0013] The clean phase information obtained in step 3 is subjected to data enhancement processing, and the number and diversity of the data set are expanded through data transformation;

[0014] Step 5: Training fine-grained vital sign signal reconstruction network:

[0015] The enhanced data obtained in step 4 is used as the original input of the fine-grained vital sign signal reconstruction network, and the corresponding wearable device synchronous collected signal is used for reconstruction network training;

[0016] Step 6: Fine-grained vital sign signal reconstruction:

[0017] The reconstruction network obtained in step 5 is used to recover the fine-grained vital sign signal; the reconstruction network uses the phase information obtained in step 4 and the vital sign waveform provided by the wearable device as labels during the training process, and the deep neural network obtained after training.

[0018] Further, the radar device in step 1 is placed around the user, and the wearable device at least includes a breathing belt and a PPG sensor, wherein the breathing belt is deployed on the user's chest, and the PPG sensor is deployed at the user's fingertip position.

[0019] The radar device adopts a one-transmit-four-receive mode.

[0020] Step 3: Wireless signal corresponding phase information extraction and preprocessing specifically includes:

[0021] Step 3.1: Perform target detection to determine the distance interval change range corresponding to the target motion;

[0022] Step 3.2: Extract the original phase information on the distance interval where the target position is located;

[0023] Step 3.3: Eliminate abnormal noise data in the original phase information to obtain clean phase information.

[0024] Step 4: Wireless signal phase information data enhancement processing, data enhancement processing uses one or several combinations of the following methods,

[0025] 1) Add white Gaussian noise to the wireless signal phase information data;

[0026] 2) Select a random interval in the phase information signal to replace with white Gaussian noise;

[0027] 3) Use time offset to offset the sample point position of 0 to 50 randomly selected samples of the input phase information signal;

[0028] 4) Perform linear time transformation on the time offset processed phase information signal, expand or shrink in the time dimension, and randomly select the expansion range between 0.8 and 1.2.

[0029] Step 5: Fine-grained vital sign signal reconstruction network training: the reconstruction network includes an encoder, a latent space, and a decoder, the input of the encoder is clean phase information, and the input clean phase information is compressed and mapped to the latent space, and the decoder is used for dimensionality reduction operation, and finally the fine-grained breathing / heartbeat signal waveform is reconstructed.

[0030] The encoder network includes multiple stack blocks, each stack sequentially includes a one-dimensional convolution layer, a batch normalization layer, and a self-gating activation function layer; the one-dimensional convolution layer extracts time features related to fine-grained breathing / heartbeat pattern data; the batch normalization layer accelerates model training by reducing internal covariate shift, and the self-gating activation function layer increases the nonlinearity of the network through the activation function.

[0031] The output of the encoder network follows a Gaussian distribution, and the output information of the encoder network is reparameterized and mapped to the latent space, and the specific reparameterization process is as follows:

[0032] z=μ z +σ z ⊙∈

[0033] Where ⊙ represents the Hadamard product of elements, ∈ is random noise used to maintain the randomness of z, a unit Gaussian distribution is randomly sampled, and then the samples are transferred through the respective mean and respective covariance to obtain the final z, σ z represents the variance of the latent variable z, μ z represents the mean of the latent variable z.

[0034] The decoder includes a one-dimensional transpose convolution, a batch normalization layer, and a self-gating activation function layer, and each one-dimensional transpose convolution layer receives m input channels and uses a kernel size of n.

[0035] The training loss function of the reconstruction network is as follows:

[0036] Loss=loss rc +γ·loss rl

[0037] Wherein, loss rc represents the reconstruction loss, loss rl represents the regularization loss, and γ is the weight of the regularization loss.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] (1) The present application extracts fine-grained vital sign signal waveform r ′ from the original wireless radio frequency signal r, which is regarded as modeling the basic distribution p(r'|r) of r based on r', and under the driving of the original radio frequency signal and the real signal waveform, the method can automatically model the nonlinear relationship between the two.

[0040] (2) The present application does not need to extract additional complex features from the signal, does not limit the user's daily behavior, and at the same time performs data enhancement transformation operation on the extracted phase information caused by the target user, which greatly improves the availability of data;

[0041] (3) The method of the present application combines the signal waveform provided by the wearable device to reconstruct and recover the fine-grained vital sign signal, which greatly suppresses the influence of environmental, noise and other interference;

[0042] (4) The method of the present application normalizes the latent space by encoding the latent vector into a probability distribution, so that the network can be sampled in a more meaningful way, avoiding overfitting of the network, and better handling of out-of-range inputs. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 It is a schematic diagram of experimental scene deployment.

[0044] Figure 2 It is a schematic diagram of the reconstruction network structure.

[0045] Figure 3 It is a schematic diagram of the reconstruction result of the present application in a real environment and the comparison with the traditional method.

[0046] Figure 4The figure of the cosine similarity results of the reconstructed respiration signal results and the real respiration waveform in different activity states in the real environment of the present application, in which the five states are: playing on the phone (PoP), typewriting (TW), shaking legs (SL), standing up / sitting down (SUSD), and speaking (SK).

[0047] Figure 5 The figure of the cosine similarity results of the reconstructed heartbeat signal results and the real heartbeat waveform in different activity states in the real environment of the present application, in which the five states are: playing on the phone (PoP), typewriting (TW), shaking legs (SL), standing up / sitting down (SUSD), and speaking (SK). DETAILED DESCRIPTION

[0048] The embodiment gives a fine-grained vital sign reconstruction method based on an encoder-decoder network. The method uses a Frequency Modulated Continuous Wave (FMCW) millimeter wave radar sensor IWR1843BOOST (operating frequency band: 77-81 GHz) of Texas Instruments and a real-time data acquisition adapter DCA1000EVM for radar sensing. In the experiment, the relevant detailed parameters of the signal are configured using mmWave Studio. The DCA1000EVM is connected to the computer through an Ethernet cable, and the captured data is sent to the computer using mmWave Studio software. For the collected original radar radio frequency signals, the target position is first located, and then the corresponding phase change information is extracted, so as to obtain the clean phase information r(t) of the wireless signal. Similarly, the synchronously collected wearable device signals are processed to obtain the fine-grained vital sign signal data information r(t) provided by the wearable device. gt In the embodiment, the method of a sliding window with an overlap window of 50% is used to split the data. Specifically, the length of each window is 20 seconds. Thus, the input sample set of the reconstruction network can be obtained as follows:

[0049]

[0050] wherein R represents the input sample set data collection, W represents the total number of window samples divided, i represents the current i-th window sample, r(t) i represents the clean phase information of the wireless signal in the i-th window, and r(t) represents the fine-grained vital sign signal data information provided by the wearable device in the i-th window.gt (t) i t represents the time sequence index, and represents the fine-grained vital sign signal data information provided by the wearable device in the ith window.

[0051] Step 1, device deployment:

[0052] The radar device and the corresponding wearable device are deployed. The radar device is placed around the user, and the respiratory belt and the photoplethysmography (PPG) pulse sensor are respectively deployed on the chest and the fingertip of the user.

[0053] In this embodiment, the Frequency Modulated Continuous Wave (FMCW) millimeter wave radar sensor IWR1843BOOST (operating frequency band: 77-81 GHz) of Texas Instruments and a real-time data acquisition adapter DCA1000EVM for radar signal acquisition are used to collect radar wireless signals; the BIOPAC MP160 type 16-channel physiological instrument is used to provide the ground truth, specifically, the PPGED-R and RSP100C modules provided by the device are used to provide real-time monitoring of respiration and heartbeat. In the experiment, the related detailed parameters of the wireless signal are configured using mmWaveStudio. Specifically, the experiment uses a one-to-four receiving mode, the frame period of the transmitted signal is T m = 10 ms, 20 Chirp signals are transmitted in each frame, the slope of the Chirp signal is S = 75.461 MHz / μs, the ADC sampling frequency is f ADCsample = 4 Msps, and the sampling period is T c = 50 μs. The specific deployment experiment diagram is shown in Figure 1 .

[0054] Step 2, data acquisition:

[0055] Under the deployment of step 1, long-time acquisition of wireless signals and wearable device signals is carried out without limiting the state of the user's daily life behavior, and synchronous data acquisition of the millimeter wave radar device wireless signal and the wearable device can be realized.

[0056] In this embodiment, 5 volunteers participated in the experimental data collection (2 males and 3 females), with an age range of 21-51 years old, a height range of 156-187 cm, and a weight range of 45-70 kg. In the experiment, the volunteers wore their daily clothes, such as T-shirts, shirts, and jackets of different materials, and carried out learning, work, or mobile phone playing, etc. In the office (37 m 2 ), the self-study room (30 m 2 ), and the teaching and research room (48 m2 The experiment was carried out in the experiment in the experiment, the millimeter wave device is always placed in the position close to the main target of the experiment, the distance is within 1.5m of its working area, and the BIOPAC MP160 wearable device module is used as the experimental baseline for synchronous collection, specifically, the PPG sensor is placed on the volunteer's fingertips, and the chest strap is worn on the volunteer's chest cavity where the movement is most obvious. During data collection, the state of the volunteer is not limited, and the behavior in daily life (such as reading, typing, listening to music, speaking, etc.) is maintained.

[0057] Step 3: Wireless signal corresponding phase information extraction and preprocessing:

[0058] The wireless signal data obtained in step 2 is processed to extract the original phase information corresponding to the target user, and the data is preprocessed to remove the interference of abnormal values and the like, to obtain clean phase information;

[0059] In this embodiment, before extracting the phase corresponding to the target motion, target detection is first performed to determine the position. The continuously collected multiple swept frequency signals can constitute a two-dimensional matrix, wherein the swept frequency sampling time dimension is the fast time dimension (Fast Time), and the multiple swept frequency signal accumulation dimension constitutes the slow time dimension (Slow Time). The fast time dimension sampling points are Fourier transformed (FFT) to obtain the frequency spectrum of the intermediate frequency signal, that is, the distance distribution information of the radar field of view, and then a range-time matrix r(d,t) (Range Time Matrix, RTM) is constructed. In In the matrix, the maximum average power of different distance intervals is calculated to obtain the distance interval change range corresponding to the human target motion. The displacement motion of the target usually causes the reflected signal to be enhanced, and has a large average power, so the power change can be used to distinguish the target motion displacement interval and the stationary object interval, so as to determine the range where the target motion is located. After determining the position area of the target, the original phase information in the distance interval where the user position is located is extracted; after the Hampel identifier (Hampel Indentifier) is used to eliminate the interference of abnormal noise data, the clean phase information caused by the current target user is extracted.

[0060] Step 4: Wireless signal phase information data enhancement processing:

[0061] The clean phase information obtained in step 3 is subjected to data enhancement processing, and the number and diversity of the data set are expanded through data transformation.

[0062] In this embodiment, in order to prevent overfitting in the training of the signal reconstruction network and increase the diversity of data samples, the training mechanism of the signal reconstruction network includes various data enhancement techniques. Specifically, the following four data enhancement operations are performed on the wireless signal data in a window, and the fine-grained vital sign waveform of the corresponding wearable device remains consistent. The data enhancement operations are as follows: (1) Additive White Gaussian Noise (AWGN) is added to the signal, and the mean value is 0, and the variance is randomly selected between 0.1 and 0.4 of the signal variance. (2) White Gaussian noise replacement: this enhancement technique makes the network robust to the erasure of perceived wireless signals, which may come from sudden, large body movements such as sudden standing. For this purpose, a random interval in the signal is selected and replaced with white noise. In addition, in order to represent large movements, the noise variance of this enhancement method is selected to be 5 to 10 times larger than the signal noise (unlike method (1), the original data in the interval is deleted).(3) Time offset: this method helps to maintain the displacement invariance of the network, i.e. robustness to time movement of the input signal sequence. For this purpose, the input signal is offset by a randomly selected sample point position of 0 to 50 samples.(4) Linear time transformation: this method helps the network to handle changes in the signal period. The input signal is resampled by a standard anti-aliasing low-pass filter, which expands or shrinks the time dimension, and the expansion range is randomly selected between 0.8 and 1.2, ensuring that the signal period after the transformation is still within the normal range.

[0063] Step 5: Fine-grained vital sign signal reconstruction network training:

[0064] The data obtained in step 4 is used as the original input of the fine-grained vital sign signal reconstruction network, and the corresponding wearable device simultaneously collected signal is used for reconstruction network training. The entire training process makes the input result of the reconstruction network approach the signal waveform provided by the wearable device, until the similarity of the two no longer increases significantly.

[0065] The reconstruction network includes an encoder network and a decoder network. The encoder network decomposes the input wireless signal, extracts features and compresses the dimensions of the target signal, encodes the information into a compact representation in the latent space, and then samples in the latent space to expand the dimension, which is expanded to a longer sequence and matched with the output dimension, and finally generates the reconstruction result, thereby realizing the recovery of fine-grained vital signs.

[0066] Specifically, in this embodiment, the reconstruction network is constructed according to the system requirements. For example, Figure 2As shown, the reconstruction network consists of an encoder, a latent space and a decoder. r is the original input signal phase data, and r' is the reconstruction data that can be observed after passing through the network, and r' is generated by the latent variable z, i.e. z→r' is the decoder p θ (r'|z), to generate the reconstructed data result; and r→z is the encoder q φ (z|r) that maps the original input data into the latent space. z is generated from the sampling of the original input data r and parameters, which contains the information of r and satisfies the Gaussian distribution.

[0067] In this embodiment, the encoder q φ (z|r) is the clean phase information related to the target extracted from the original received signal, and then this input is compressed and mapped to the latent space z, and then sampled to train the model parameters and structure of the modeling decoder, so as to reconstruct the fine-grained breathing / heartbeat waveform. The purpose of the entire convolutional filter is to extract useful features, which can be considered as a demixing function to decompose the influence between the target vital sign signal and the nonlinear motion interference. The encoder network consists of multiple stack blocks, each of which contains a one-dimensional convolutional layer (1D-CNN), a batch normalization layer (BN) and a self-gated activation function layer (Swish) in order. Specifically, the encoder applies (1) 1D-CNN to extract time features related to fine-grained breathing / heartbeat pattern data, which is constantly refined from the mixed signal; (2) the BN layer reduces the strong dependence on the initialization parameters, which is equivalent to the existence of regularization, allowing a larger learning rate, accelerating the model training process by reducing the internal covariate shift, solving the problem of gradient disappearance in the propagation process, and making the training process more effective and stable; (3) the Swish layer as an activation function to increase the nonlinearity of the network, which has the function characteristics of smoothness, non-monotonicity, no upper bound and lower bound. Finally, the encoded features are mapped to the latent space by the fully connected layer. Each one-dimensional convolutional layer Conv1D(m, n) represents receiving m input channels and using a kernel of size n; the convolution step size for each convolutional layer is set to 4, that is, each convolutional layer performs 4 times downsampling on the input, which constantly maps the input data to the latent space, so the feature dimension is gradually reduced. Finally, the encoded features are mapped to the latent space distribution by the fully connected layer.

[0068] In this embodiment, the output of the encoder obeys the Gaussian distribution with their respective mean and variance. To enable the gradient descent algorithm for network training, a reparameterization method is applied to reorganize the way of obtaining the derivative of latent space. Specifically, a sample is randomly drawn from a unit Gaussian distribution

[0069] z = μ z + σ z ⊙ ∈

[0070] Here, ∈ can be considered as random noise to maintain the randomness of z. ⊙ denotes the Hadamard product of elements. After each input phase encoding distribution is sampled as z, the result becomes the input of the first fully connected layer of the decoder for further reconstruction of fine-grained respiratory / heartbeat signal waveform.

[0071] In this embodiment, the decoder p θ (r'|z) is used as a dimension-increasing model to reconstruct the desired fine-grained respiratory / heartbeat signal waveform. The decoder network is actually the reverse composition of the encoder, using the architecture of one-dimensional transposed convolution (TransposedConvolution) layer, batch normalization layer and self-gating activation function layer. Specifically, each one-dimensional transposed convolution layer ConvTranspose1d(m, n) receives m input channels and uses a kernel of size n; contrary to the encoder, since each transposed convolution layer performs 4 times up-sampling on the input, the dimension of the feature data also increases, and the latent space representation is expanded to a longer sequence. The decoded features are converted to match the output dimension and use the Sigmoid activation function, and finally produce the reconstructed result r'.

[0072] In this embodiment, the training strategy of the reconstruction network is as follows, by maximizing the likelihood p θ (r') can achieve the best performance of respiratory / heartbeat waveform reconstruction, and its intuitive explanation is that The reconstruction process of fine-grained respiratory / heartbeat signal waveform is described, D KL (q φ (z|r)||p θ (z)) measures the KL divergence between the approximate posterior distribution and the true prior p θ (z). These two terms are represented as reconstruction loss and regularization loss:

[0073] (1) Reconstruction loss: As the actual measurement result of , the reconstruction loss loss rc ​The reconstruction loss loss is defined on L2 norm, which measures the sum of differences between corresponding sampling points of the reconstructed signal waveform r' and the real waveform r provided by the wearable device gt

[0074] loss rc =||r gt -r′||2

[0075] (2) Regularization Loss: Corresponding to D KL (q φ (z|r)||p θ (z)), there is a loss for the regularization of latent space distribution. In this paper, it is assumed that its real prior distribution is unit Gaussian distribution, so its regularization loss loss rl is defined as:

[0076]

[0077] Combining the above loss function, the overall loss function of the reconstruction network training becomes:

[0078] Loss=loss rc +γ·loss rl

[0079] Wherein, gamma is the weight of the regularization loss, used to balance the balance between the reconstruction loss and the regularization loss.

[0080] Step 6: Fine-grained vital signal reconstruction:

[0081] The reconstruction network is used to recover the fine-grained vital signal, wherein the reconstruction network is obtained by using the phase information obtained in step 4 and the vital signal waveform provided by the wearable device as labels in the reconstruction network training process, and the reconstruction network required for training is obtained after training.

[0082] According to the experimental results in the real environment: as shown in Figs. Figure 3 、 4 and 5, we can observe that the performance of the reconstruction network is greatly improved compared with the traditional scheme, and the monitoring of the fine-grained vital signal can be realized without limiting the user's behavior. Through the above experiments, the feasibility of the reconstruction network of the present application is verified. Among them Figure 3 ​The results of the reconstructed breathing pattern and heartbeat pattern under different body activities are shown, from top to bottom, the original radio frequency signal, the real waveform provided by the wearable device, the fine-grained signal waveform reconstructed by the application and the results obtained by the decomposition method. It can be found by observation that the application reconstructs and recovers the accurate shape of the breathing and heartbeat pattern, which is largely consistent with the real waveform result, while the CEEMDAN decomposition method usually cannot provide results with correct details.

[0083] Referring to Figure 4 The application studies and compares the cosine similarity between the reconstructed breathing signal and the actual waveform under different body movements. It can be observed that ST and SL have the least effect on the performance of the model, because the body is static or the part involved in the movement is far from the chest of the target user. As expected, the SUSD movement makes the performance of the model worse, because it can cause serious interference with the chest movement. On the whole, the average cosine similarity between the reconstructed breathing signal waveform and the real waveform provided by the wearable device is 0.9302, indicating that the reconstructed signal effect is satisfactory, because the similarity greater than 0.8 indicates a strong positive correlation.

[0084] Referring to Figure 5 As can be seen in the application, the reconstructed heartbeat signal waveform and the real heartbeat waveform provided by the wearable device have a median cosine similarity of more than 0.93. Among them, the SUSD type of body movement makes the performance of the model become worse, because this type of movement can seriously interfere with the chest movement, so that the heartbeat, such a weak movement, is covered, which is consistent with the result in the breathing waveform recovery.

Claims

1. A method for fine-grained vital sign signal reconstruction based on an encoder-decoder network, characterized in that: The method comprises the following steps: Step 1: device deployment, deploying millimeter wave radar devices and wearable devices; Step 2: data collection, collecting wearable device wireless signals; Step 3: wireless signal corresponding phase information extraction and preprocessing: processing the collected wireless signal data to obtain clean phase information; Step 4: data enhancement processing on the clean phase information obtained in step 3, which adopts the following combination of methods: 1) adding white Gaussian noise to the wireless signal phase information data; 2) selecting a random interval in the white Gaussian noise to replace a random interval in the phase information signal; 3) using time offset to offset the sample point positions of 0 to 50 randomly selected samples of the input phase information signal; 4) performing linear time transformation on the phase information signal after time offset processing, expanding or shrinking in the time dimension, and the expansion range is randomly selected between 0.8 and 1.2; Step 5: constructing a fine-grained vital sign signal reconstruction network, wherein the reconstruction network includes an encoder, a latent space and a decoder, the input of the encoder is clean phase information, and the clean phase information of the input is compressed and mapped to the latent space, and the dimension is increased through the decoder to finally obtain the fine-grained vital sign signal reconstruction network; the enhanced data obtained in step 4 is used as the original input of the fine-grained vital sign signal reconstruction network, and the corresponding wearable device signal is collected simultaneously to train the reconstruction network: Wherein the encoder network includes a plurality of stack blocks, each stack block sequentially includes a one-dimensional convolution layer, a batch normalization layer and a self-gating activation function layer; the one-dimensional convolution layer extracts time features related to fine-grained breathing / heartbeat pattern data; the batch normalization layer accelerates model training by reducing internal covariate shift, and the self-gating activation function layer increases the nonlinearity of the network through the activation function; The output of the encoder network is subject to Gaussian distribution, and the output information of the encoder network is reparameterized and mapped to the latent space, and the specific reparameterization process is as follows: where denotes the Hadamard product of the elements, is used to maintain randomness, a unit Gaussian distribution randomly sampled in the mean and covariance, respectively, and then the samples are shifted by their respective means and respective covariances to obtain the final z, denotes the variance of the latent variable z, denotes the mean of the latent variable z; The training loss function of the reconstruction network is as follows: Loss = lossrc + γ ∙ lossrl Wherein lossrc represents the reconstruction loss, lossrl represents the regularization loss, and γ is the weight of the regularization loss; Step 6: fine-grained vital sign signal reconstruction: Using the reconstruction network obtained in step 5 to recover the fine-grained vital sign signal.

2. The encoder-decoder network-based fine-grained vital signs signal reconstruction method of claim 1, wherein: The radar device in step 1 is placed around the user, and the wearable device at least includes a breathing belt and a PPG sensor, wherein the breathing belt is deployed on the user's chest, and the PPG sensor is deployed at the user's fingertip position.

3. The encoder-decoder network-based fine-grained vital signs signal reconstruction method of claim 2, wherein: The radar device adopts a one-transmit-four-receive mode.

4. The method of claim 1 or 3, wherein: The step 3: wireless signal corresponding phase information extraction and preprocessing specifically includes: Step 3.1: target detection, determining the distance interval change range corresponding to the target motion; Step 3.2: extracting the original phase information on the distance interval where the target is located; Step 3.3: eliminating abnormal noise data in the original phase information to obtain clean phase information.

Images