A vital sign monitoring method, system and apparatus

Abstract

This invention discloses a method, system, and device for monitoring vital signs, relating to the field of biomedical engineering technology. The method includes acquiring complex scattering parameters of electromagnetic wave signals from the thoracic region of a human subject after being modulated by a metasurface based on a quasi-continuous domain bound state of a ring dipole. The metasurface is positioned between a microwave transceiver unit and the human subject to focus or shape the electromagnetic beam transmitted to the subject and receive electromagnetic wave signals reflected or transmitted through the thoracic region. Multidimensional feature calculations are performed on the complex scattering parameters to generate feature vectors. These feature vectors are then input into a pre-trained deep learning neural network model to infer and output electrocardiogram waveforms, respiratory waveforms, heart rate values, and respiratory rate values. This method achieves a system-level breakthrough from physical signal enhancement to intelligent analysis by deeply integrating a high-quality factor Q-BIC metasurface with multidimensional feature extraction and deep learning inference algorithms.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering, electromagnetic field and microwave technology, and artificial intelligence application technology, specifically to a method, system, and device for monitoring vital signs. Background Technology

[0002] Remote non-contact vital sign monitoring, as a cutting-edge technology, enables the non-intrusive monitoring of heart rate, respiratory rate, and even electrocardiogram waveforms by remotely detecting the slight displacement of the chest cavity caused by heartbeat and respiration. It shows great application prospects in fields such as neonatal monitoring, home-based elderly care, and burn patient care. The core challenge of this technology lies in how to accurately and stably capture and analyze extremely weak physiological signals from a noisy and complex environment.

[0003] In the early stages of non-contact vital sign monitoring technology development, researchers mainly used systems based on traditional radar (such as continuous wave Doppler radar) or vector network analyzers (VNAs). By analyzing the phase or amplitude changes of microwave signals reflected by the human body, respiratory and heartbeat information could be indirectly extracted. However, these traditional radio frequency sensing systems generally face the bottleneck of signal-to-noise ratio (SNR). Because the physiological modulation signal reflected by the human body has extremely low energy, it is easily overwhelmed by large-scale body movements, multipath effects in the environment, and noise from the device itself. More importantly, most existing signal processing methods only utilize a single dimension of the signal, severely underutilizing the rich information contained in the complex scattering parameters (S-parameters) (such as real part, imaginary part, group delay, etc.), resulting in a need to improve the accuracy and robustness of signal analysis.

[0004] In recent years, metasurface technology has attracted attention due to its ability to precisely control the propagation of electromagnetic waves. As an artificially designed two-dimensional electromagnetic structure, metasurfaces can control the propagation of electromagnetic waves with unprecedented precision with extremely low profile and design degrees of freedom. Through careful design of metasurface unit structures, precise focusing or field enhancement of microwave energy can be achieved, providing a new physical approach to improve the sensitivity to weak physiological signals. Preliminary theoretical and simulation studies have verified the feasibility of using metasurfaces to enhance biosensing signals.

[0005] However, existing metasurface research generally remains at the level of unit design and physical simulation, lacking systematic integration with back-end intelligent signal processing algorithms. This disconnect between the "physical layer" and the "algorithm layer" means that the signal enhancement advantages brought by metasurfaces cannot be efficiently utilized. Summary of the Invention

[0006] To address the shortcomings of existing technologies where the signal enhancement advantages of metasurfaces cannot be efficiently utilized due to the disconnect between the "physical layer" and the "algorithm layer," this invention proposes a vital sign monitoring method, system, and device. By deeply integrating high-quality factor Q-BIC metasurfaces with multi-dimensional feature extraction and deep learning inference algorithms, a highly robust monitoring system with pattern adaptation and intelligent state perception capabilities is constructed. This achieves a system-level breakthrough from physical signal enhancement to intelligent analysis, thereby solving the problems existing in the prior art.

[0007] A method for monitoring vital signs includes the following steps: The complex scattering parameters of electromagnetic wave signals from the thoracic region of the human body under test are collected after being modulated by a metasurface based on the quasi-continuous domain bound state of a ring dipole. The metasurface is disposed between the microwave transceiver unit and the human body under test, and is used to focus or shape the electromagnetic beam transmitted from the microwave transceiver unit to the human body under test, and to receive electromagnetic wave signals reflected or transmitted through the thoracic region of the human body. After resampling the complex scattering parameters to a preset fixed number of frequency points, multidimensional feature calculation is performed to generate a feature vector set containing the real part, imaginary part, amplitude logarithm, unwrapped phase, group delay, and the difference between adjacent frequency points of the real part, the difference between adjacent frequency points of the imaginary part, and the difference between adjacent frequency points of the amplitude. The feature vector set is input into the pre-trained neural network model VNA2VitalNet. The feature vector set is subjected to hierarchical feature extraction by a shared encoder composed of multiple stacked one-dimensional convolutional neural networks to obtain abstract features. The abstract features are then decoded by a multi-head decoder containing multiple parallel fully connected layer branches to output electrocardiogram waveforms, respiratory waveforms, heart rate values, and respiratory rate values, respectively.

[0008] Furthermore, the calculation process for the unwrapped phase specifically includes the following steps: By resampling the complex scattering parameters to a preset fixed number of frequency points N ; The original phase is obtained by performing arctangent calculation on the real and imaginary parts of each frequency point; The phase jumps between adjacent frequency points in the original phase that exceed a preset threshold are corrected to obtain a continuous phase sequence; the correction process specifically eliminates phase winding by accumulating ±2π offsets point by point.

[0009] Furthermore, the group delay is calculated by performing adjacent frequency point differential on the unwrapped phase and taking the negative value.

[0010] Furthermore, it also includes displaying waveforms using the logarithm of amplitude as the interface, and using the peak value of the real part or the energy of a specified frequency band as the presence index; based on the presence index, the mean of the presence index is calculated in real time through a sliding window, and the monitoring area is judged to be in a state of being occupied or unoccupied based on the comparison result of the mean of the presence index and a preset threshold; when the state changes from occupied to unoccupied, the ECG and respiratory waveforms are cleared, and the heart rate and respiratory rate values ​​are set to invalid.

[0011] Furthermore, the training process of the VNA2VitalNet model includes the following steps: Simultaneously acquire complex scattering parameter data output from the metasurface sensor front end and real electrocardiogram waveforms and real respiratory waveforms output from standard medical equipment; The continuous complex scattering parameter data is divided into sample frames, and each frame corresponds to a segment of real electrocardiogram waveform and real respiratory waveform within a synchronous time period. The true heart rate is calculated by detecting the R-peak of the true electrocardiogram waveform; the true respiratory rate is calculated by detecting the peak value of the true respiratory waveform. Construct a set of data pairs containing feature vectors, actual electrocardiogram waveforms, actual respiratory waveforms, actual heart rate values, and actual respiratory rate values; Based on the data set, a weighted combined loss function is used to perform multi-task joint training of the model; wherein, the weighted combined loss function is expressed as: Loss total =w ecg *Loss ecg +w resp *Loss resp +w hr *Loss hr +w rr *Loss rr ; Among them, w ecg w resp w hr w rr These are preset weighting coefficients; Loss ecg Used to calculate the error between the model-predicted ECG waveform and the actual ECG waveform in the data pair, Loss resp Used to calculate the error between the model-predicted respiratory waveform and the actual respiratory waveform in the data pair; Loss hr and Loss rr Calculate the prediction errors for heart rate and respiratory rate values ​​separately.

[0012] The present invention also includes a vital signs monitoring system, comprising: The radio frequency signal acquisition module is used to acquire the complex scattering parameters of electromagnetic wave signals from the chest cavity region of the human body under test after being modulated by a metasurface based on the quasi-continuous domain bound state of a ring dipole. The metasurface is disposed between the microwave transceiver unit and the human body under test, and is used to focus or shape the electromagnetic beam transmitted from the microwave transceiver unit to the human body under test, and to receive electromagnetic wave signals reflected or transmitted through the chest cavity region of the human body. The feature extraction module is used to resample the complex scattering parameters to a preset fixed number of frequency points and then perform multi-dimensional feature calculations to generate a feature vector set containing the real part, imaginary part, amplitude logarithm, unwrapped phase, group delay, and the difference between adjacent frequency points of the real part, the difference between adjacent frequency points of the imaginary part, and the difference between adjacent frequency points of the amplitude. The deep learning vital sign inference module is used to input the feature vector set into the pre-trained neural network model VNA2VitalNet, and extract the feature vector set hierarchically through a shared encoder composed of multiple stacked one-dimensional convolutional neural networks to obtain abstract features; the abstract features are decoded by a multi-head decoder containing multiple parallel fully connected layer branches, and the electrocardiogram waveform, respiratory waveform, heart rate value and respiratory rate value are output respectively.

[0013] Furthermore, the deep learning vital sign inference module also includes a dual-mode working mechanism: when the deep learning model is detected to be unavailable or fails to load, it automatically switches to the fallback prediction mode; the fallback prediction mode is used to directly read the pre-stored vital sign values ​​from the data packet or output the preset reference values.

[0014] Furthermore, it also includes an automatic data source switching module; the automatic data source switching module is used to detect the availability of the radio frequency signal acquisition module when the system starts up; when the radio frequency signal acquisition module is available, the system enters the real-time monitoring mode; when the radio frequency signal acquisition module is unavailable, the system automatically switches to the pre-stored demonstration mode to replay the continuous data sequence containing both manned and unmanned states.

[0015] The present invention also includes a vital signs monitoring computer device, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the vital signs monitoring method.

[0016] The present invention also includes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the vital signs monitoring method.

[0017] This invention provides a method for monitoring vital signs, which has the following beneficial effects: This invention is the first to deeply integrate the physical control capabilities of Q-BIC metasurfaces with high quality factors with the data-driven inference capabilities of deep learning. By focusing electromagnetic energy onto the human chest cavity through a specially designed Q-BIC metasurface, the weak phase and amplitude modulations caused by heartbeats and respiration are greatly amplified. Combined with an eight-dimensional feature set extracted from complex scattering parameters, including real parts, imaginary parts, and group delays, an unprecedented information richness is provided for deep learning models. This collaborative method of "physical enhancement + intelligent processing" achieves high-fidelity capture and precise separation of weak vital signs at the non-contact sensing level for the first time, significantly improving the reconstruction accuracy of ECG and respiratory waveforms and the calculation stability of heart rate and respiratory rate, while maintaining high robustness even in low signal-to-noise ratio environments. This method fills the engineering technology gap from metasurface fundamental physics to vital sign monitoring applications, providing a core technical solution and system-level implementation for a new generation of high-precision, high-robust non-contact medical monitoring equipment. It has significant application value and broad market prospects in fields such as smart elderly care, neonatal monitoring, and special environment operation monitoring. Attached Figure Description

[0018] Figure 1 This is a block diagram of the vital signs monitoring system based on q-BIC metasurface enhanced sensing and deep learning in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the physical principle and experimental verification of the sensing front end in an embodiment of the present invention; Figure 3 This is a flowchart of the complex S21 multidimensional feature extraction module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the architecture of the VNA2VitalNet deep learning model in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the training process and performance verification results of the VNA2VitalNet model in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the robustness test results of the system in various scenarios, including free space and through walls, in an embodiment of the present invention. Figure 7 The above is a metasurface sensing structure based on Q-BIC in an embodiment of the present invention; in the figure: 1-circular arc resonator; 2-f4b plate; 3-metallic copper backplate. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] This invention proposes a vital sign monitoring system based on q-BIC metasurface enhanced sensing and deep learning. By deeply integrating a specially designed q-BIC metasurface with a high quality factor (high Q value) with a systematic multidimensional feature extraction and deep learning inference algorithm, a highly robust monitoring system with pattern adaptation and intelligent state perception capabilities is constructed for the first time. This system works synergistically on two levels: First, at the physical front end, it uses the q-BIC metasurface to finely control the spatial electromagnetic field, greatly amplifying the weak signal modulation generated by the micro-movement of the human chest cavity, thereby capturing complex scattering parameters with high signal-to-noise ratio and high information richness. Second, at the intelligent back end, it pioneers a set of processing and inference methods that can fully mine this high-dimensional information and ensure that the system still operates stably under abnormal conditions such as equipment failure and model failure. This invention successfully advances metasurface technology from device-level signal enhancement research to a new stage of robust application and intelligent analysis at the system level.

[0021] like Figure 1 As shown, the system specifically includes: (1) Acquisition Module: A metasurface sensing front-end based on Q-BIC is designed using a specific microwave frequency band (e.g., 2-4 GHz) sensitive to biological signals. This metasurface, based on the quasi-continuous bound state of the ring dipole (Q-BIC), forms a highly sensitive electromagnetic field environment in the monitoring area through its high quality factor characteristics. It can efficiently convert the small displacements of the chest cavity caused by human heartbeat and respiration into significant phase and amplitude changes in the complex scattering parameter (S21), serving as a high-fidelity input source for subsequent signal processing.

[0022] This embodiment provides a high-Q metasurface sensing structure based on quasi-continuum bound states (q-BIC), such as Figure 7 As shown in (a); the metasurface consists of periodically arranged unit structures, each periodic unit comprising two asymmetric arc-shaped resonators, an f4b plate, and a copper backplate, as shown in (a). Figure 7As shown in (b) of the diagram; in the ideal case of perfect symmetry, the two circular arc structures have identical geometric parameters (such as arc radius, central angle, and opening direction). In this case, the system supports a strictly continuum bound state (BIC), and the mode is completely confined within the metasurface plane, unable to couple with the free-space radiation field. Therefore, the quality factor theoretically tends towards infinity. To obtain a high-Q resonance that can be excited by an external light field, this invention transforms the BIC into a quasi-BIC (q-BIC) by introducing structural asymmetry. Specifically, the central angles of the two arcs are different, and the asymmetry parameter is defined as the absolute value of the difference between the two central angles. When this asymmetry is greater than zero, the originally confined mode acquires finite radiation loss, thus enabling it to be excited by far-field incident light. Furthermore, the quality factor is inversely proportional to the square of the asymmetry; that is, the smaller the asymmetry, the higher the Q value. To further enhance the local field enhancement effect of the resonance mode and its sensitivity to changes in the surrounding refractive index, this embodiment further combines the basic units composed of the two asymmetric arcs into a 2×2 supramolecular structure. By specially designing the relative orientation and spatial arrangement of these four units, such as using a rotationally symmetric configuration around a central axis, the entire 2×2 subarray primarily produces a toroidal dipole radiation mode under excitation. The toroidal dipole is essentially composed of a closed magnetically polarized current loop, and its far-field radiation is extremely weak, further suppressing radiation loss. This, combined with the q-BIC effect, results in experimentally extremely high quality factors, typically reaching hundreds or even tens of thousands.

[0023] (2) Complex S21 feature extraction module: Receives a frame of data containing frequency points and complex S21 (real part + imaginary part), and performs a set of feature engineering aimed at maximizing information utilization. Specifically, it calculates the following eight types of feature vectors, each with a length of N, which is the number of fixed frequency points after resampling: real part, imaginary part, amplitude logarithm, unwrapped phase, group delay, and the difference between adjacent frequency points of the real part, the difference between adjacent frequency points of the imaginary part, and the difference between adjacent frequency points of the amplitude.

[0024] The generation of unwrapped phase includes: first, performing arctangent operation on the real and imaginary parts of each frequency point to obtain the original phase; then, correcting the phase jumps between adjacent frequency points in the original phase that exceed a preset threshold; and eliminating phase entanglement by accumulating ±2π offsets point by point to obtain a continuous phase sequence.

[0025] The group delay is calculated by taking the negative value of the difference between adjacent frequency points of the unwrapped phase.

[0026] Finally, all feature vectors are concatenated into a high-dimensional total feature vector of 8N dimensions, which is used as the input to the deep learning inference module.

[0027] (3) Intelligent Detection Module for Occupied / Unoccupied Status: This module separates one or more presence indicators from the features extracted from each frame. It maintains a sliding window of length L and calculates the average value of the presence indicators in real time. If the average value is greater than a preset threshold, the current status is determined to be "occupied"; otherwise, it is "unoccupied". This module is linked to the interface display. When the status changes from "occupied" to "unoccupied", it automatically clears the ECG and respiratory waveforms and sets the heart rate and respiratory rate values ​​to invalid indicators such as "--" to avoid outputting false information.

[0028] (4) Dual-mode vital signs deep learning inference module: The main and backup dual-mode design is adopted to ensure that the system is never interrupted.

[0029] The main mode is deep learning inference: the pre-trained neural network model VNA2VitalNet is loaded, the above 8N-dimensional feature vectors are received, the feature vectors are extracted hierarchically by a multi-layer one-dimensional convolutional neural network with a shared encoder, and the abstract features are decoded by four independent fully connected layer branches of the multi-head decoder. The parallel inference outputs high-resolution electrocardiogram waveforms, respiratory waveforms, heart rate values ​​and respiratory rate values.

[0030] The backup mode is a fallback prediction mode: when the system detects a missing model file, an unavailable deep learning library, or a loading failure, it automatically and seamlessly switches to this mode. This mode directly reads pre-stored vital sign values ​​from the current data packet or outputs preset reference values ​​(sine wave and 75 bpm heart rate), ensuring that the system can run completely and display effective output in any environment.

[0031] To directly and in parallel extract four heterogeneous indicators—ECG waveform, respiratory waveform, heart rate, and respiratory rate—from a single, metasurface-enhanced complex S21 data frame, this invention proposes an end-to-end training and inference method based on multi-task deep learning. A specific neural network structure, VNA2VitalNet, is designed, and a complete offline training process is established, specifically including the following steps: S1. Synchronous Data Acquisition and Labeling: To train the neural network model, a large-scale training dataset with time-synchronized labels is first needed. The acquisition process is as follows:

[0032] The test subjects were invited to be within the effective monitoring range of the Q-BIC metasurface sensor front end. While the system acquired complex S21 data through VNA, the test subjects' real electrocardiogram (ECG) waveform and real respiratory waveform were simultaneously recorded using medical standard equipment as ground truth. The acquired VNA data stream and standard physiological signal stream were strictly time-aligned. The continuous VNA data was divided into frame-by-frame data packets, with each frame of VNA data corresponding to a real ECG waveform segment and a respiratory waveform segment within a synchronous time period. In addition, the real heart rate (HR) was calculated by detecting the R-peak of the real ECG waveform, and the real respiratory rate (RR) was calculated by detecting the peak value of the real respiratory waveform. Finally, a set of data pairs containing VNA feature vectors, real ECG waveforms, real respiratory waveforms, real HR, and real RR was formed as training samples for the model.

[0033] S2, VNA2VitalNet multi-task neural network model architecture: This model adopts a "shared encoder-multi-head decoder" architecture to efficiently learn the complex mapping relationship between S21 features and various vital signs.

[0034] It receives an 8N-dimensional high-dimensional feature vector generated by the "complex S21 feature extraction module"; the shared encoder is composed of multiple layers of one-dimensional convolutional neural networks stacked together. Each convolutional network is used to automatically learn and extract different levels of abstract features from the input 8N-dimensional feature vector. The shallow convolutional network focuses on local correlations in the frequency domain, while the deep network can capture more global and complex combination patterns. These patterns are deeply related to electromagnetic field disturbances caused by heartbeat and respiration. The output of the encoder is a highly condensed feature representation that contains all the information needed to decouple all vital signs.

[0035] The multi-head decoder starts from the output of the shared encoder, and the network structure is divided into four independent branches, each responsible for a specific task: the ECG consists of several fully connected layers, and its task is to decode the shared feature representation into a ECG time series waveform; the respiratory head has a similar structure to the ECG head, but its goal is to decode the respiratory waveform time series; the heart rate usually consists of fewer fully connected layers, and finally outputs a single value representing the predicted heart rate value; the respiratory rate has a similar structure to the heart rate head, and outputs a single value representing the predicted respiratory rate.

[0036] S3. Multi-task Joint Training and Loss Function: The training objective of the model is to make the outputs of the four heads simultaneously approximate their respective "ground truth" values; this is achieved through a weighted combined loss function, the total loss function of which is defined as: Loss total =w ecg ×Loss ecg +w resp×Loss resp +w hr ×Loss hr +w rr ×Loss rr ; Among them, w ecg w resp w hr w rr These are preset weighting coefficients used to balance the importance of the four different tasks and the magnitude of their loss values; Loss ecg and Loss resp Mean squared error or smoothed L1 loss is used to calculate the difference between the predicted waveform and the true waveform at each point; Loss hr and Loss rr The mean squared error (MSE) or mean absolute error is used to calculate the difference between the predicted and the actual values.

[0037] Training Process: In training iterations, a batch of training data is input into the model, and the model performs forward propagation to obtain four predicted outputs. Then, the total error is calculated based on the total loss function described above. Finally, the gradient of the loss with respect to all parameters in the network is calculated using the backpropagation algorithm, and an optimizer is used to update the network parameters so that the total loss function value decreases. This process is repeated until the model reaches optimal performance on an independent validation set.

[0038] Based on the same inventive concept, this invention also proposes a complete and highly robust non-contact vital sign monitoring and demonstration method, which includes the following steps: ① System self-test and mode switching: Upon system startup, the system first detects the connection status and availability of the RF sensing hardware. If the hardware is available and communication is normal, the system automatically enters real-time monitoring mode. If the detection fails, it seamlessly switches to pre-stored demonstration mode, ensuring that the system can still function as a complete demonstration platform even without devices.

[0039] ② Real-time high-fidelity signal processing flow: In real-time monitoring mode, the system periodically acquires complex S21 data from the sensing front end, sends it sequentially to the feature extraction module to generate an 8N-dimensional feature vector, then the manned / unmanned state detection module judges the scene state, and finally inputs the feature vector into the deep learning inference module to obtain the final vital signs result and push it to the interface for display.

[0040] ③ Realistic alternating unmanned / manned demonstration process: In demonstration mode, the system reads data from a preset demonstration sequence file. This sequence consists of alternating continuous data blocks of real "manned" states and "unmanned" state data blocks generated by adding random noise to a reference waveform library of real unmanned environments. This design simulates the scenario of people entering and leaving in real monitoring, making the demonstration process more coherent and convincing than random frame sampling.

[0041] ④ Information Fusion and Multidimensional Visualization: The multidimensional information obtained from processing and reasoning is fused and displayed. The interface simultaneously displays: a) the VNA amplitude-frequency curve of the current frame; b) the historical amplitude spectrum heatmap of the most recent M frames, used to observe the dynamic changes of the signal; c) real-time scrolling electrocardiogram (ECG) waveform; d) real-time scrolling respiratory waveform; e) accurate heart rate and respiratory rate values ​​in the presence of a person.

[0042] The overall system architecture and workflow of this invention are as follows: Figure 1 As shown, the system is logically and rigorously divided into a data source layer, a core processing layer, and a user interface layer. The data source layer includes a real-time vector network analyzer (VNA) device and a demonstration sequence management module, whose functions are scheduled by an automatic data source switching module. This switching module automatically detects the availability of the VNA hardware upon system startup. If the device connection is normal, the system enters real-time monitoring mode; otherwise, it seamlessly switches to demonstration mode, thus ensuring the system's operability in any environment. Regardless of the data source, all data is sent to the core processing layer. This layer is the system's technical hub, containing, in sequence, a complex S21 feature extraction module, a manned / unmanned status detection module, and a dual-mode vital sign inference module.

[0043] The physical basis and core innovation of this invention lie in its unique sensing front end, the principle of which and experimental setup are as follows: Figure 2 As shown; the slight displacement of the thoracic cavity caused by heart contractions and respiratory movements (such as...) Figure 2 (as shown in (a)) modulates the surrounding electromagnetic field.

[0044] To efficiently capture this weak modulation, this invention employs a specially designed metasurface based on quasi-continuous bound states of the ring dipole (Q-BIC) (such as...). Figure 2 As shown in (h) in the figure, the key to its design lies in the fact that, near the medical safety frequency band of 2.45 GHz, this metasurface has a sharp resonance peak with an extremely high quality factor (Q value), which is reflected in the measured and simulated S-parameter curves ( Figure 2This has been fully verified in (e), (f), (g), and (i) of the model. This high-Q resonance characteristic makes the electromagnetic field near the metasurface exceptionally sensitive to minute perturbations in dielectric constant (i.e., human movement), enabling the efficient conversion of submillimeter-level chest cavity displacement into significant frequency and amplitude changes in the S21 parameters; in an anechoic chamber (… Figure 2 Experiments conducted in (b), (c), and (d) demonstrate that by continuously measuring the S21 parameter using a VNA, a high signal-to-noise ratio vital signs signal that has been significantly amplified at the physical level can be obtained.

[0045] After capturing the high-fidelity signal, the system enters the intelligent processing stage: the first step is the multi-dimensional feature extraction stage, the detailed process of which is as follows: Figure 3 As shown, when a frame of original complex S21 data enters the module, the system will first resample it to a uniform frequency point. N Then, it performs parallel computation of eight types of features designed to maximize information utilization, specifically including the real part, imaginary part, logarithm of amplitude, unwrapped phase, group delay, and the difference between adjacent frequency points of these three. These eight types of features together constitute an 8N-dimensional high-dimensional feature vector, which comprehensively describes the information of the original signal in different dimensions and serves as the input to the subsequent deep learning model. Simultaneously, this module also separates the logarithm of amplitude curve for interface waveform display and extracts the peak-to-peak value of the real part signal as an "existence indicator" to determine the presence of the target, which is then sent to the state detection module.

[0046] These 8N-dimensional feature vectors constitute the input to the core algorithm of this invention—the VNA2VitalNet neural network model, whose ingenious architecture is as follows: Figure 4 As shown, this model employs an advanced "shared encoder-multi-head decoder" paradigm. Input data is first processed by a one-dimensional convolutional network to extract local features in the frequency domain, which are then fed into a Transformer encoder. The Transformer, leveraging its powerful self-attention mechanism, captures long-range dependencies between features across the entire frequency domain, resulting in a highly condensed deep feature representation with a global perspective. This powerful shared feature representation is then fed into four independent decoding parts, each a multilayer perceptron (MLP), responsible for decoding the shared features into four heterogeneous vital signs: electrocardiogram (ECG) waveforms, respiratory waveforms, heart rate values, and respiratory rate values. This enables end-to-end, parallel, and efficient inference from a single radio frequency signal to multi-dimensional physiological information.

[0047] The model's superior performance has been fully demonstrated through training and validation on a large self-collected dataset, and its key performance metrics are summarized in [database name missing]. Figure 5 .from Figure 5As shown in the loss function curve (a) in the figure, the losses on both the training and validation sets steadily decreased and converged, indicating that the model was sufficiently trained and there was no serious overfitting. Regarding numerical prediction accuracy, the mean absolute error (MAE) of both heart rate and respiratory rate converged to below 1 bpm / rpm. Figure 5 In (b) of the equation, the coefficient of determination R² is close to a perfect 1.0. Figure 5 (c) in the middle), a scatter plot of predicted values ​​versus actual values ​​( Figure 5 (g) and (h) are closely aligned with the diagonal, and the error histogram ( Figure 5 The values ​​(i) and (j) in the figure also exhibit an ideal normal distribution, jointly demonstrating its extremely high numerical prediction accuracy. In the more challenging waveform reconstruction task, the correlation coefficient between the predicted waveform and the actual waveform is also as high as 0.95 or higher (5(e) and (f) in the figure), demonstrating direct waveform comparison ( Figure 5 The (k) and (l) in the model also show a high degree of consistency between the two in terms of morphology, period, and phase, proving the model's ability to accurately recover details of physiological signals.

[0048] Finally, to verify the practicality of this invention in real-world complex environments, we conducted a series of rigorous robustness tests, the results of which are as follows: Figure 6 As shown. The test scenarios covered unobstructed free space, through-wall monitoring with walls as obstructions, and different postures of the test subject facing away from or towards the sensor under through-wall conditions, as well as conditions with different incident angles. The experimental results clearly show that, regardless of whether in ideal free space ( Figure 6 (a) and (e) in the above are still under the scenario of severe signal attenuation through walls ( Figure 6 As shown in (b), (c), (f), and (g)), this system consistently and successfully recovers high-quality ECG waveforms from weak signals. These successful tests under various challenging conditions strongly demonstrate that this invention, by combining a Q-BIC metasurface-enhanced front-end with a powerful back-end intelligent algorithm, possesses the ability to overcome practical interference such as obstacle occlusion and changes in human posture, showcasing its outstanding robustness and broad application potential as a complete solution.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring vital signs, characterized in that, Includes the following steps: The complex scattering parameters of electromagnetic wave signals from the thoracic region of the human body under test are collected after being modulated by a metasurface based on the quasi-continuous domain bound state of a ring dipole. The metasurface is disposed between the microwave transceiver unit and the human body under test, and is used to focus or shape the electromagnetic beam transmitted from the microwave transceiver unit to the human body under test, and to receive electromagnetic wave signals reflected or transmitted through the thoracic region of the human body. After resampling the complex scattering parameters to a preset fixed number of frequency points, multidimensional feature calculation is performed to generate a feature vector set containing the real part, imaginary part, amplitude logarithm, unwrapped phase, group delay, and the difference between adjacent frequency points of the real part, the difference between adjacent frequency points of the imaginary part, and the difference between adjacent frequency points of the amplitude. The feature vector set is input into the pre-trained neural network model VNA2VitalNet. The feature vector set is subjected to hierarchical feature extraction by a shared encoder composed of multiple stacked one-dimensional convolutional neural networks to obtain abstract features. The abstract features are then decoded by a multi-head decoder containing multiple parallel fully connected layer branches to output electrocardiogram waveforms, respiratory waveforms, heart rate values, and respiratory rate values, respectively.

2. The method for monitoring vital signs according to claim 1, characterized in that, The calculation process for the unwrapped phase specifically includes the following steps: By resampling the complex scattering parameters to a preset fixed number of frequency points N ; The original phase is obtained by performing arctangent calculation on the real and imaginary parts of each frequency point; The phase jumps between adjacent frequency points in the original phase that exceed a preset threshold are corrected to obtain a continuous phase sequence; the correction process specifically eliminates phase winding by accumulating ±2π offsets point by point.

3. The method for monitoring vital signs according to claim 1, characterized in that, The group delay is calculated by taking the negative value of the difference between adjacent frequency points of the unwrapped phase.

4. The method for monitoring vital signs according to claim 1, characterized in that, It also includes displaying waveforms using the logarithm of amplitude as the interface, and using the peak value of the real part or the energy of a specified frequency band as the presence index; based on the presence index, the mean of the presence index is calculated in real time through a sliding window, and the presence index is compared with a preset threshold to determine whether the monitoring area is occupied or unoccupied. When the status changes from occupied to unoccupied, the ECG and respiratory waveforms are cleared, and the heart rate and respiratory rate values ​​are set to invalid.

5. The method for monitoring vital signs according to claim 1, characterized in that, The training process of the VNA2VitalNet model includes the following steps: Simultaneously acquire complex scattering parameter data output from the metasurface sensor front end and real electrocardiogram waveforms and real respiratory waveforms output from standard medical equipment; The continuous complex scattering parameter data is divided into sample frames, and each frame corresponds to a segment of real electrocardiogram waveform and real respiratory waveform within a synchronous time period. The true heart rate is calculated by detecting the R-peak of the true electrocardiogram waveform; the true respiratory rate is calculated by detecting the peak value of the true respiratory waveform. Construct a set of data pairs containing feature vectors, actual electrocardiogram waveforms, actual respiratory waveforms, actual heart rate values, and actual respiratory rate values; Based on the data set, a weighted combined loss function is used to perform multi-task joint training of the model; wherein, the weighted combined loss function is expressed as: Loss total =w ecg *Loss ecg +w resp *Loss resp +w hr *Loss hr +w rr *Loss rr ; Among them, w ecg w resp w hr w rr These are preset weighting coefficients; Loss ecg Used to calculate the error between the model-predicted ECG waveform and the actual ECG waveform in the data pair, Loss resp Used to calculate the error between the model-predicted respiratory waveform and the actual respiratory waveform in the data pair; Loss hr and Loss rr Calculate the prediction errors for heart rate and respiratory rate values ​​separately.

6. A vital signs monitoring system, characterized in that, include: The radio frequency signal acquisition module is used to acquire the complex scattering parameters of electromagnetic wave signals from the chest cavity region of the human body under test after being modulated by a metasurface based on the quasi-continuous domain bound state of a ring dipole. The metasurface is disposed between the microwave transceiver unit and the human body under test, and is used to focus or shape the electromagnetic beam transmitted from the microwave transceiver unit to the human body under test, and to receive electromagnetic wave signals reflected or transmitted through the chest cavity region of the human body. The feature extraction module is used to resample the complex scattering parameters to a preset fixed number of frequency points and then perform multi-dimensional feature calculations to generate a feature vector set containing the real part, imaginary part, amplitude logarithm, unwrapped phase, group delay, and the difference between adjacent frequency points of the real part, the difference between adjacent frequency points of the imaginary part, and the difference between adjacent frequency points of the amplitude. The deep learning vital sign inference module is used to input the feature vector set into the pre-trained neural network model VNA2VitalNet, and extract the feature vector set hierarchically through a shared encoder composed of multiple stacked one-dimensional convolutional neural networks to obtain abstract features; the abstract features are decoded by a multi-head decoder containing multiple parallel fully connected layer branches, and the electrocardiogram waveform, respiratory waveform, heart rate value and respiratory rate value are output respectively.

7. A vital signs monitoring system according to claim 6, characterized in that, The deep learning vital sign inference module also includes a dual-mode working mechanism: when the deep learning model is detected to be unavailable or fails to load, it automatically switches to the fallback prediction mode; the fallback prediction mode is used to directly read the pre-stored vital sign values ​​from the data packet or output the preset reference values.

8. A vital signs monitoring system according to claim 6, characterized in that, It also includes an automatic data source switching module; the automatic data source switching module is used to detect the availability of the radio frequency signal acquisition module when the system starts up; when the radio frequency signal acquisition module is available, the system enters the real-time monitoring mode; when the radio frequency signal acquisition module is unavailable, the system automatically switches to the pre-stored demonstration mode to replay the continuous data sequence containing both manned and unmanned states.

9. A computer device for monitoring vital signs, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the vital signs monitoring method according to any one of claims 1-5.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the vital signs monitoring method according to any one of claims 1-5.

Images