A high-robust non-contact precise electrocardiogram monitoring method based on millimeter wave radar

Abstract

This invention belongs to the fields of wireless sensing and artificial intelligence technology, and discloses a robust, non-contact, and precise electrocardiogram (ECG) monitoring method based on millimeter-wave radar. First, the distance and angle of potential targets are obtained through distance fast Fourier transform and digital beamforming technology, combined with mean filtering and a two-dimensional constant false alarm rate (CFAR) algorithm to remove static background and detect the chest cavity position. Then, continuous phase is extracted using differential cross-multiplication, and a two-step heartbeat-related phase extraction scheme is designed, employing B-spline fitting and differential operations to remove body micro-movements and respiratory interference, obtaining stable heartbeat-related phase signals. A heart rate-guided adaptive wavelet decomposition method is designed to obtain multi-band features, and time-frequency joint features are extracted through a dual-branch attention mechanism and gated fusion. Finally, the time-frequency joint features are input into an ECG signal time-domain reconstruction module based on the TransUNet architecture to achieve high-quality ECG signal reconstruction, featuring non-contact, continuous, and convenient operation.

Description

Technical Field

[0001] This invention belongs to the field of wireless sensing and artificial intelligence technology, specifically relating to a robust non-contact precision electrocardiogram monitoring method based on millimeter-wave radar. Background Technology

[0002] Cardiovascular disease is becoming increasingly prevalent due to population aging and lifestyle changes, posing a significant public health challenge. Currently, cardiac monitoring is primarily achieved through electrocardiographs (ECGs), which acquire ECG signals via adhesive skin electrodes. While this technology provides high-fidelity ECG signals, the contact-based acquisition method makes long-term continuous monitoring difficult, and adhesive electrodes also have limitations in terms of comfort and portability.

[0003] With the continuous advancement of millimeter-wave radar sensing technology, its application in cardiac monitoring is gradually demonstrating its potential. This technology can monitor the heart by capturing the micro-movements in the chest cavity caused by the heartbeat. Compared with traditional adhesive electrode methods, its non-contact nature offers advantages such as comfort and portability, enabling long-term continuous monitoring. Several patents have already reported ECG monitoring methods based on millimeter-wave radar. Patent CN116369934A uses beamforming to locate the heart region and extract the original complex signals from the heart region, mapping the complex signals to corresponding ECG signals through a neural network. Patent CN120130985A extracts the forward and backward temporal features of heartbeat-related signals using a bidirectional long short-term memory network and completes ECG signal conversion through a convolutional structure. Patent CN118319323A utilizes wavelet decomposition to reduce irrelevant component interference and achieves ECG signal conversion through a convolutional encoder and a bidirectional long short-term memory network.

[0004] Although existing millimeter-wave radar-based electrocardiogram (ECG) detection methods can achieve non-contact ECG signal monitoring to a certain extent, they still have several limitations. First, unavoidable minor movements of the human body can interfere with chest cavity micro-vibration signals, and existing methods do not effectively suppress this, affecting the quality of heartbeat signal extraction. Second, when performing wavelet decomposition on radar signals, existing methods do not consider the frequency shift caused by individual differences in heart rate, resulting in a lack of consistent physiological meaning among the decomposed frequency signals, which reduces the reliability of ECG reconstruction. Finally, most existing methods rely primarily on time-domain feature analysis, with insufficient utilization of frequency-domain information, failing to fully leverage the harmonic structure and frequency components inherent in radar signals. These factors collectively limit the robustness and accuracy of existing millimeter-wave radar technology in ECG monitoring. Summary of the Invention

[0005] Considering the limitations of existing methods, this invention provides a robust, non-contact, and precise electrocardiogram (ECG) monitoring method based on millimeter-wave radar. The overall process mainly includes three stages: chest cavity localization, extraction of heartbeat-related phase components, and ECG signal reconstruction. (See attached diagram). Figure 1 .

[0006] In the chest cavity localization phase, the chest cavity is first preliminarily located by performing range-domain Fast Fourier Transform and digital beamforming on the radar signals of each time frame to generate a range-angle map containing potential target location information. Next, mean filtering is used to remove static background, and a constant false alarm rate (CFAR) detector is applied to adaptively extract dynamic targets. To reduce localization jitter caused by system noise, an energy intensity ratio discrimination strategy is designed to update the chest cavity position only when the energy ratio exceeds a threshold.

[0007] In the heartbeat-related phase component extraction stage, to obtain chest cavity vibration information, the mechanism that the phase at the target location is proportional to its vibration amplitude is utilized. At the target location, a differential cross-multiplication method is used to obtain a continuous phase signal. Since this signal is a composite signal caused by respiration, heartbeat, and body micro-movements, B-spline interpolation is further used to model and eliminate micro-movement interference. Then, the heartbeat component is robustly separated using first-order differentiation.

[0008] In the ECG signal reconstruction stage, three modules were designed: a heart rate-guided adaptive wavelet decomposition module, a time-frequency domain feature fusion module, and an ECG signal time-domain reconstruction module. By acquiring paired radar-ECG signals, the three modules learn the nonlinear mapping relationship between the radar heartbeat signal and the ECG signal waveform through a data-driven approach, thereby generating an accurate ECG signal.

[0009] The technical solution of the present invention:

[0010] A robust, non-contact, and precise electrocardiogram monitoring method based on millimeter-wave radar, comprising the following steps:

[0011] Step 1: Locating the chest cavity;

[0012] First, the distance and angle of the chest cavity need to be determined using radar signals, with the chest cavity being a potential target;

[0013] The radar transmitting antenna emits a linear frequency modulated (LFM) wave, and the radar receiving antenna receives the echo signal reflected from a potential target. The echo signal and the LFM wave are mixed to obtain an intermediate frequency (IF) signal. A range-fast Fourier transform (FSF) is performed on the IF signal to obtain its frequency. There is a corresponding relationship between the distance d between the target and the potential target:

[0014] (1)

[0015] in, It is the sweep bandwidth of the linear frequency modulated wave, and c is the speed of light. The duration of the linear frequency modulated wave sweep is given; based on the above correspondence, the distance of the potential target relative to the radar is obtained through the frequency of the intermediate frequency signal;

[0016] Then, digital beamforming technology is used at the receiver to obtain the angle of the potential target: when the echo signal from the angle θ is received, the phase difference between the echo signals of different radar receiving antennas with a spacing of r is r sinθ. The weighting coefficients of each radar receiving antenna at different angles are calculated using this phase difference; the weighting coefficients are obtained by traversing the angles at each range scale and the output power is calculated to obtain the range-angle map of the potential target relative to the radar.

[0017] The distance-angle map of potential targets is processed as follows: (1) In actual scenarios, static structures such as floors or walls will produce strong reflections, interfering with the estimation of the chest cavity position. To solve this problem, the present invention first uses mean filtering to process the distance-angle map. Since the indoor environment remains stable for a short time, the reflection of static objects can be regarded as constant. On the contrary, human body reflection changes due to breathing, heartbeat and subtle body movements, and is manifested as a dynamic target. The mean filtering is used to extract background clutter, and then it is subtracted from the distance-angle map, thus realizing the removal of static background. (2) In order to further deal with the residual non-stationary system noise, a two-dimensional constant false alarm detection algorithm is adopted. Its basic idea is to set the detection threshold adaptively according to the statistics of local background noise under the premise of ensuring a constant false alarm rate. In the specific implementation process, the local background noise level is estimated in real time by selecting reference units in the two dimensions of distance and angle and statistically analyzing their noise power. After obtaining the local background noise level, it is converted into an adaptive detection threshold according to the set false alarm rate, and the detection threshold is used to judge the unit to be detected. When the echo intensity of a certain unit to be detected is higher than the detection threshold, a significant chest cavity reflection point is extracted from the complex and non-uniform background, thereby realizing the location detection of the chest cavity.

[0018] Furthermore, an energy intensity ratio discrimination strategy is designed for the two-dimensional constant false alarm rate (CFAR) detection algorithm, and a discrimination threshold is set. This involves calculating the energy intensity ratio between the current position and the previous position; if the ratio is lower than the threshold, the current detection is considered an outlier, and the original position is retained. Through this process, the distance to the thoracic cavity is obtained. and angle .

[0019] Step 2: Extraction of heartbeat-related phase components;

[0020] Phase at the thoracic cavity Distance from the thoracic cavity in step one Proportional relationship: Phase changes in the position of the thoracic cavity Changes in distance from the thoracic cavity It is also directly proportional: The phase change of the chest cavity position is extracted to obtain the distance change of the chest cavity position caused by the heartbeat. To avoid the problem of phase discontinuity, the differential cross-multiplication technique is used to obtain the phase signal of the chest cavity position. The differential cross-multiplication technique uses the real and imaginary parts of the radar signal and their time change rate. Through the structure of cross-multiplication and subtraction, a quantity proportional to the phase change rate is obtained, and then integrated to obtain a continuous phase curve.

[0021] The phase signal is a composite signal containing heartbeat components, respiratory components, and body micro-movement components. The heartbeat component is separated from the composite signal.

[0022] First, the removal of body micro-movements is crucial. Compared to the quasi-periodic behavior of breathing and heartbeat, body micro-movements introduce irregular and large-amplitude fluctuations, manifesting as trend changes in the phase signal. To suppress this interference, B-spline interpolation is used to model the trend line introduced by body micro-movements. A B-spline can be understood as a smooth, flexible curve that naturally connects several key points to form a continuous, smooth trajectory. It can well fit the slowly changing overall trend in the phase signal without being affected by local noise. Utilizing this characteristic, B-splines can effectively capture the trend of phase signal changes caused by body micro-movements. By subtracting this trend line from the original composite signal, the interference from body micro-movements can be effectively eliminated. Next, to eliminate the respiratory component, the phase signal after removing the trend line is differentiated. Since the heart rate is higher than the respiratory rate, differentiation can enhance the high-frequency heart rate component while suppressing the low-frequency respiratory signal. Although this process may amplify high-frequency noise, the use of first-order differentiation achieves an effective balance between strengthening the heart rate signal and controlling the influence of noise, ultimately resulting in the reliable extraction of the heart rate-related phase signal.

[0023] Step 3: ECG signal reconstruction;

[0024] Through the above processing, a heartbeat-related phase signal can be obtained. However, this heartbeat-related phase signal and the electrophysiological activity depicted by the electrocardiogram (ECG) signal are essentially different physiological processes. The ECG signal reflects the temporal changes of the heart's electrical signals, while the heartbeat-related phase signal reflects the mechanical vibration signal triggered by myocardial contraction. Although the two are significantly different, they are intrinsically linked, namely, the mechanical activity of the heart is triggered and regulated by the ECG signal. Based on the above considerations, this invention adopts a data-driven method, using a neural network to model the mapping relationship between the phase signal of the heartbeat component and the ECG signal, comprising three sequentially connected modules: First, the phase signal of the heartbeat component is decomposed into multi-band signals by a heart rate-guided adaptive wavelet decomposition module; second, the multi-band signals are input into a time-frequency domain feature fusion module to obtain time-frequency joint features; finally, the time-frequency joint features are sent to an ECG signal time-domain reconstruction module to generate the ECG signal.

[0025] The heart rate-guided adaptive wavelet decomposition module is designed to divide the phase signal of the heartbeat component into multi-band signals: a fundamental band and a high-frequency band. Specifically, the heart rate is determined by time-frequency analysis of the phase signal of the heartbeat component. Then, the signal sampling rate is dynamically adjusted according to the heart rate, followed by wavelet decomposition to ensure that the frequency band boundaries of the wavelet decomposition are consistent with the heartbeat rhythm. After the wavelet decomposition is completed, the envelope of the fundamental band is extracted by Hilbert transform, and the fundamental band is divided by the envelope to achieve amplitude normalization, resulting in an updated multi-band signal.

[0026] The design incorporates a time-frequency domain feature fusion module, whose input is a heart rate-guided adaptive wavelet decomposition module that generates updated multi-band signals. This module consists of a shallow encoder, a time-frequency domain feature extraction section, and a feature fusion section. First, the shallow encoder contains four cascaded one-dimensional convolutional blocks. Each one-dimensional convolutional block comprises a convolutional layer, layer normalization, and a PReLU activation function, used for preliminary processing of the updated multi-band signals to extract the basic time structure and stable feature distribution, outputting preliminary multi-band features. The time-frequency domain feature extraction section comprises two parallel branches: a time-domain branch and a frequency-domain branch, both using the preliminary multi-band features as input. The time-domain branch segments the preliminary multi-band features along the time axis and performs time-position encoding, using the responses of different frequency bands at the same time as a token. A Transformer encoder is used to learn the dependencies across time steps to obtain the time-domain features. The frequency-domain branch segments the preliminary multi-band features by frequency band and performs frequency band position encoding, using the complete time series of each frequency band as a token, and a Transformer encoder is used to perform the encoding. The encoder models the correlation between components of different frequency bands, characterizes the harmonic structure of the ECG signal, and obtains frequency domain features. The feature fusion part adopts a gated fusion structure to generate adaptive weight parameters and weights the time domain features and frequency domain features to obtain time-frequency joint features.

[0027] Subsequently, the extracted time-frequency joint features are input into the ECG signal temporal reconstruction module. The ECG signal temporal reconstruction module adopts the TransUNet architecture, based on a U-shaped encoder-decoder structure, and introduces skip connections between the encoder and decoder at the corresponding levels. The encoder consists of two parts: the front end is a multi-level one-dimensional convolutional block, each one-dimensional convolutional block contains a convolutional layer, layer normalization, and PReLU activation function; the end embeds a Transformer encoder to model the global contextual dependency throughout the cardiac cycle; the decoder gradually recovers the ECG signal length through upsampling convolutional blocks and fuses the skip connection features from the encoder, finally outputting the ECG reconstruction result.

[0028] The beneficial effects of this invention are:

[0029] (1) The present invention designs a non-contact ECG signal monitoring method based on millimeter-wave radar, which has high comfort and convenience compared with traditional contact ECG monitoring methods.

[0030] (2) Compared with existing millimeter-wave radar electrocardiogram signal detection methods, this invention specifically suppresses non-cardiac interference such as human micro-movements and respiration, effectively enhancing the ability to separate heartbeat patterns, thereby improving the accuracy of heartbeat-related signal extraction. In addition, this invention designs a heart rate-guided adaptive wavelet decomposition method to achieve cross-individual consistent multi-band partitioning with clear physiological meaning, thereby ensuring that subsequent time-domain and frequency-domain structure extraction are performed under the same physiological reference framework, providing a unified and stable multi-band structural foundation for subsequent feature modeling.

[0031] (3) This invention combines deep learning technology to construct a neural network model of the mapping relationship between cardiac mechanical vibration and electrocardiogram signal. At the same time, it comprehensively considers the dual perspective information of time domain and frequency domain, and comprehensively considers short-term dynamic changes and long-term physiological trends through the TransUNet architecture, thereby improving the accuracy of electrocardiogram signal generation. Attached Figure Description

[0032] Figure 1 This is the overall process of an example of the present invention.

[0033] Figure 2 The test results of the method of the present invention are shown in (a) and (b) respectively. (a) is a comparison of the time waveforms of the reconstructed electrocardiogram and the real electrocardiogram. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0035] A robust, non-contact, and precise electrocardiogram monitoring method based on millimeter-wave radar, comprising the following steps:

[0036] It is divided into three stages: chest cavity localization, extraction of heartbeat-related phase components, and reconstruction of electrocardiogram signals, which can achieve accurate monitoring of human electrocardiogram signals.

[0037] First, a dataset consisting of multiple participants was constructed. Each participant lay supine on a bed, and a millimeter-wave radar module was fixedly installed 0.8m above their chest to collect radar signals. A 12-lead continuous electrocardiogram recorder simultaneously recorded electrocardiogram signals. The collected data were divided into two groups according to the participants: 70% of the data was used as the training set, and 30% as the test set.

[0038] First, following the steps outlined, the radar signal is processed using Fast Fourier Transform (FSFT) and Digital Beamforming to obtain the range and angle of the potential target, enabling preliminary localization of the chest cavity. Then, static background removal is performed using background subtraction, and a two-dimensional constant false alarm rate (CFAR) detection algorithm is used to adaptively extract the chest cavity location. Simultaneously, an energy intensity ratio discrimination strategy is employed to enhance localization stability.

[0039] Step two involves using the differential cross-multiplication technique to obtain the phase at the chest cavity location, which includes heartbeat components, respiratory components, and body micromotion components. Then, B-spline interpolation and first-order differential methods are used to remove body micromotion interference and respiratory interference, respectively, to obtain the heartbeat-related phase signal.

[0040] According to step three, the heart rate-guided adaptive wavelet decomposition module divides the heartbeat-related phase signal into multiple frequency bands, and additionally applies Hilbert transform-based envelope normalization to the fundamental frequency band. The updated multi-frequency band signal is then input into the time-frequency domain feature fusion module to obtain joint time-frequency features. These joint time-frequency features are then input into the ECG signal time-domain reconstruction module to generate the ECG signal.

[0041] In step three, during neural network training, the acquired ECG signal is used as the ground truth, and the root mean square error between the neural network output and the ground truth is used as the loss function. Backpropagation is then used to update the neural network parameters. After training, the model is deployed on a general-purpose computing device to achieve end-to-end mapping from radar phase signals to corresponding ECG signals. Test results are shown below. Figure 2 As shown in the figure. Experimental results demonstrate that this invention can achieve high-precision ECG signal recovery and can be used for routine monitoring, proving its reliability in practical applications.

Claims

1. A robust, non-contact, and precise electrocardiogram (ECG) monitoring method based on millimeter-wave radar, characterized in that, The steps are as follows: Step 1: Locating the chest cavity; First, the distance and angle of the chest cavity need to be determined using radar signals, with the chest cavity being a potential target; Step 2: Extraction of heartbeat-related phase components; Phase at the thoracic cavity Distance from the thoracic cavity in step one Proportional relationship: Phase changes in the position of the thoracic cavity Changes in distance from the thoracic cavity It is also directly proportional: The distance change of the chest cavity position due to heartbeat is obtained by extracting the phase change of the chest cavity position; the phase signal of the chest cavity position is obtained by using the differential cross-multiplication technique. The phase signal is a composite signal containing heartbeat components, respiratory components, and body micro-movement components. The heartbeat component is separated from the composite signal. First, the trend line introduced by body micro-motion is modeled using B-spline interpolation. Then, the trend line is subtracted from the phase signal to eliminate the interference of body micro-motion. To eliminate the respiratory component, the phase signal of the detrend line is differentiated by first order to obtain the phase signal of the heartbeat component. Step 3: ECG signal reconstruction; A data-driven approach is adopted, using neural networks to model the mapping relationship between the phase signal of the heartbeat component and the electrocardiogram (ECG) signal. This approach consists of three sequentially connected modules: First, the phase signal of the heartbeat component is decomposed into multi-band signals by a heart rate-guided adaptive wavelet decomposition module; second, the multi-band signals are input into a time-frequency domain feature fusion module to obtain time-frequency joint features; finally, the time-frequency joint features are fed into an ECG signal time-domain reconstruction module to generate the ECG signal. The specific implementation process of step three is as follows: The heart rate-guided adaptive wavelet decomposition module is designed to divide the phase signal of the heartbeat component into multi-band signals: a fundamental band and a high-frequency band. Specifically, the heart rate is determined by time-frequency analysis of the phase signal of the heartbeat component. Then, the signal sampling rate is dynamically adjusted according to the heart rate, followed by wavelet decomposition to ensure that the frequency band boundaries of the wavelet decomposition are consistent with the heartbeat rhythm. After the wavelet decomposition is completed, the envelope of the fundamental band is extracted by Hilbert transform, and the fundamental band is divided by the envelope to achieve amplitude normalization, resulting in an updated multi-band signal. The design incorporates a time-frequency domain feature fusion module, whose input is a heart rate-guided adaptive wavelet decomposition module that generates updated multi-band signals. This module consists of a shallow encoder, a time-frequency domain feature extraction section, and a feature fusion section. First, the shallow encoder contains four cascaded one-dimensional convolutional blocks. Each one-dimensional convolutional block comprises a convolutional layer, layer normalization, and a PReLU activation function, used for preliminary processing of the updated multi-band signals, outputting preliminary multi-band features. The time-frequency domain feature extraction section comprises two parallel branches: a time-domain branch and a frequency-domain branch, both using the preliminary multi-band features as input. The time-domain branch segments the preliminary multi-band features along the time axis and performs time-position encoding, using the responses of different frequency bands at the same time as a token. A Transformer encoder is used to learn the dependencies across time steps to obtain the time-domain features. The frequency-domain branch segments the preliminary multi-band features by frequency band and performs frequency band position encoding, using the complete time series of each frequency band as a token, and a Transformer encoder is used to further encode these features. The encoder models the correlation between components of different frequency bands, characterizes the harmonic structure of the ECG signal, and obtains frequency domain features. The feature fusion part adopts a gated fusion structure to generate adaptive weight parameters and weights the time domain features and frequency domain features to obtain time-frequency joint features. Subsequently, the extracted time-frequency joint features are input into the ECG signal temporal reconstruction module. The ECG signal temporal reconstruction module adopts the TransUNet architecture, based on a U-shaped encoder-decoder structure, and introduces skip connections between the encoder and decoder at the corresponding levels. The encoder consists of two parts: the front end is a multi-level one-dimensional convolutional block, each one-dimensional convolutional block contains a convolutional layer, layer normalization, and PReLU activation function; the end embeds a Transformer encoder to model the global contextual dependency throughout the cardiac cycle; the decoder gradually recovers the ECG signal length through upsampling convolutional blocks and fuses the skip connection features from the encoder, finally outputting the ECG reconstruction result.

2. The robust non-contact precision electrocardiogram monitoring method based on millimeter-wave radar according to claim 1, characterized in that, The specific implementation process of step one is as follows: The radar transmitting antenna emits a linear frequency modulated (LFM) wave, and the radar receiving antenna receives the echo signal reflected from a potential target. The echo signal and the LFM wave are mixed to obtain an intermediate frequency (IF) signal. A range-fast Fourier transform (FSF) is performed on the IF signal to obtain its frequency. There is a corresponding relationship between the distance d between the target and the potential target: (1) in, It is the sweep bandwidth of the linear frequency modulated wave, and c is the speed of light. The duration of the linear frequency modulated wave sweep is given; based on the above correspondence, the distance of the potential target relative to the radar is obtained through the frequency of the intermediate frequency signal; Then, digital beamforming technology is used at the receiver to obtain the angle of the potential target: when the echo signal from the angle θ is received, the phase difference between the echo signals of different radar receiving antennas with a spacing of r is r sinθ. The weighting coefficients of each radar receiving antenna at different angles are calculated using this phase difference; the weighting coefficients are obtained by traversing the angles at each range scale and the output power is calculated to obtain the range-angle map of the potential target relative to the radar. The distance-angle map of the potential target is processed as follows: first, static background is removed by mean filtering, and then a two-dimensional constant false alarm rate (CFAR) detection algorithm is used for further processing to extract the chest cavity reflection point from the distance-angle map, thereby realizing the location detection of the chest cavity. Furthermore, an energy intensity ratio discrimination strategy is designed for the two-dimensional constant false alarm rate (CFAR) detection algorithm, and a discrimination threshold is set. The position of the thoracic cavity is updated only when the energy intensity ratio exceeds the discrimination threshold, thus obtaining the distance of the thoracic cavity. and angle .

Images