Non-contact ECG signal monitoring method based on millimeter wave radar

Abstract

The invention discloses a non-contact ECG signal monitoring method based on a millimeter wave radar, belongs to the field of ECG signal monitoring, realizes high-precision electrocardiogram reconstruction by simultaneously sensing chest vibration caused by heartbeat and carotid artery pulsation from the neck and fusing spatial-temporal characteristics, and solves the problems that a traditional scheme depends on electrodes and an existing radar technology cannot recover complete ECG. According to the method, non-contact ECG signal monitoring is carried out in a daily scene, and good privacy and comfort are achieved.

Description

Technical Field

[0001] This invention belongs to the field of millimeter-wave radar vital sign monitoring technology, specifically relating to a non-contact electrocardiogram (ECG) signal monitoring method based on millimeter-wave radar. Background Technology

[0002] Cardiovascular disease is the leading cause of death worldwide, and electrocardiogram (ECG) is the most commonly used method for diagnosing cardiovascular disease. Traditional ECG monitoring mainly relies on wearable devices, which suffer from limited battery life and electrode detachment, affecting the continuity of monitoring and data reliability. Among non-contact ECG monitoring technologies, ultra-wideband (UWB) and millimeter-wave radar are two main solutions. However, UWBB suffers from high power consumption and susceptibility to multipath interference. Millimeter-wave radar offers good privacy, is penetrable through clothing, and is not absorbed by the skin, while also being less expensive, making it suitable for ECG monitoring in daily life.

[0003] Existing millimeter-wave radar solutions can achieve non-contact heartbeat monitoring by detecting subtle vibrations on the human body surface. However, existing solutions often only extract heart rate or partial heartbeat features (such as for identity authentication) and cannot fully reconstruct the ECG waveform, limiting their application in medical-grade monitoring. Therefore, researching how to capture effective heartbeat signals and achieve high-fidelity ECG waveform reconstruction is of great significance in the field of vital sign monitoring. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a non-contact ECG signal monitoring method based on millimeter-wave radar. By simultaneously sensing chest vibrations caused by heartbeats and carotid artery pulsations from the neck, it fuses spatiotemporal features to achieve high-precision ECG reconstruction, solving the problems of electrode dependence in traditional solutions and the inability of existing radar technology to recover a complete ECG.

[0005] The technical solution adopted by this invention to solve its technical problem is: A non-contact ECG signal monitoring method based on millimeter-wave radar includes the following steps: S1. The millimeter-wave radar emits a linear frequency modulated continuous wave and receives the reflected signal from the human body. It generates an intermediate frequency signal through a mixer and performs a Fourier transform on the signal in the range dimension to construct a two-dimensional data matrix. S2. By analyzing the distribution of respiratory frequency band energy in a two-dimensional data matrix at different distances, static and dynamic interference is eliminated from the matrix, and the location of the human chest cavity is located using the energy intensity of the matrix after interference elimination. S3. Based on the relative spatial relationship between the chest cavity heart and neck, further locate the chest cavity heart and neck regions, and use the second-order difference method to eliminate the interference of body and respiratory activities in the target region to obtain a set of candidate signals for chest cavity heartbeat and carotid pulse. S4. A variational mode decomposition-based segmentation algorithm is used on the candidate signals to obtain a series of heartbeat cycle signals. The similarity between the cycle signals is used to perform an evaluation algorithm based on dynamic time programming to screen out high-quality chest heartbeat signals and carotid pulse signals. S5. The ECG data is processed using the equal-frequency binning method to balance the amplitude distribution of the ECG waveform. The selected chest heartbeat signal and carotid pulse signal are combined with the synchronously acquired ECG signal, and a time-dependent dataset is constructed using the sliding window sampling method. S6. Input the generated dataset into the ECG prediction model. The model captures the periodicity and waveform details of the heartbeat signal by extracting the temporal and spatial features between signals, and predicts the ECG value.

[0006] Furthermore, in step S1, the millimeter-wave radar is placed in front of the person to be detected, and the relevant parameters for transmitting linear frequency modulated electromagnetic waves are set. A mixer mixes the transmitted and received signals to generate an intermediate frequency (IF) signal. The IF signal data is stored according to the received time sequence, and its dimension is... ,in This represents the number of sampling points for a single chirp. This indicates the number of chirps contained in each frame of data. This indicates the number of antennas for receiving signals. The length of each frame is given; a range-dimension Fast Fourier Transform is performed on the data received by a single antenna to generate a two-dimensional matrix containing signal amplitude and phase information, also with dimensions of 1. .

[0007] Furthermore, in step S2, a two-dimensional data matrix is ​​extracted. Phase signals from different distance chambers are de-wrapped and then passed through a bandpass filter of the classical respiratory band to obtain sub-band signals. The instantaneous energy of these sub-band signals is calculated and reconstructed into a respiratory energy matrix. Its dimensions are the same as the original matrix. Same, for The matrix is ​​obtained by normalizing the distance dimension. ; then With the original matrix Obtain the matrix by performing the Hadamard product operation. By searching The energy extremum point of the matrix in the distance dimension locates the center of the human chest cavity.

[0008] Furthermore, in step S3, based on the fact that the radiation diameter of the thoracic cavity region is approximately 15 to 25 centimeters, and the carotid artery region is located 15 to 25 centimeters above the thoracic cavity with a radiation range of approximately 8 to 12 centimeters; combined with the range resolution of millimeter-wave radar of approximately 4 centimeters, five consecutive range chambers within the thoracic cavity region and three range chambers near the carotid artery region are selected as target signal extraction areas; the phase signal of each range chamber is subjected to second-order difference processing to suppress interference from the body and respiration, resulting in multiple candidate heartbeat signals, which respectively constitute candidate signal sets for thoracic heartbeat and carotid pulse.

[0009] In step S4, the phase signal is decomposed using a variational mode decomposition algorithm to extract the main frequency signal of the heartbeat; and a peak detection algorithm is performed based on the amplitude threshold to segment the heartbeat cycle according to the location of the peak, resulting in a series of cycle segments; the similarity between cycle segments is calculated using dynamic time warping (DTW), and hierarchical clustering is performed using the average linking method. The optimal number of clusters is determined according to the elbow rule, and the cluster center with the largest number of samples is selected as the template signal of this candidate signal. The DTW average of the template signal and all cycle segments is calculated as the score P, and the signal with the lowest score in the candidate signal set is selected as the high-quality chest heartbeat signal and carotid artery signal.

[0010] In step S5, to avoid model training bias, the synchronously acquired ECG signal is preprocessed using an equal-frequency binning method. The sampling points of the original ECG signal are arranged in ascending order of voltage value and divided into several bins to ensure that each bin contains the same number of data points. The ECG data in the bins are normalized to ensure that they are distributed within the interval [0,1] while preserving the trend of the original waveform. The chest heartbeat signal, carotid pulse signal and ECG signal are aligned with the time and sampling rate and stored in a joint data matrix of size 3*n according to the timestamp, where n is the length of the signal. The matrix is ​​divided using a sliding window method, and the ECG voltage value data of the next time step of the current window is used as the label to construct the dataset.

[0011] In step S6, the ECG prediction model employs a parallel network architecture. A temporal convolutional network combined with channel attention mechanism (TCN-SE) extracts spatial features and dynamically allocates weights for chest heartbeat signals and carotid pulse signals. A bidirectional gated circulation unit combined with global attention (BiGRU-GA) extracts temporal features, focusing on capturing the periodicity of the P-QRS-T complex in the electrocardiogram.

[0012] The beneficial effects of this invention are mainly reflected in: 1. It can perform non-contact ECG signal monitoring in everyday scenarios, offering excellent privacy and comfort; 2. Interference cancellation based on respiratory frequency band energy can effectively suppress the influence of dynamic and static interference in the environment, thereby accurately locating the human body; 3. Segmentation methods based on variational mode decomposition and evaluation methods based on dynamic time programming can achieve high-quality extraction of chest cavity heartbeat signals and carotid pulse signals; 4. The spatiotemporal fusion model based on TCN and BiGRU effectively captures and reconstructs the local details and global periodicity of ECG signals through channel attention mechanism and global attention mechanism. Attached Figure Description

[0013] Figure 1 This is a flowchart of the non-contact ECG signal monitoring method based on millimeter-wave radar according to the present invention; Figure 2 This is a schematic diagram of the network structure for spatiotemporal feature fusion according to the present invention. Detailed Implementation

[0014] The present invention will now be further described with reference to the accompanying drawings.

[0015] Reference Figure 1 and Figure 2 A non-contact ECG signal monitoring method based on millimeter-wave radar includes the following steps: S1. The millimeter-wave radar emits a linear frequency modulated continuous wave and receives the reflected signal from the human body. It generates an intermediate frequency signal through a mixer and performs a Fourier transform on the signal in the range dimension to construct a two-dimensional data matrix. The process of step S1 is as follows: Place the millimeter-wave radar in front of the person to be detected, set the relevant parameters for transmitting linear frequency modulated electromagnetic waves, and use a mixer to mix the transmitted and received signals to generate an intermediate frequency signal: ; in, Indicates the amplitude of the signal. , , This indicates the bandwidth, scan period, and wavelength of the transmitted signal. Indicates the speed of electromagnetic waves. Indicates the distance of the object from the radar; Intermediate frequency signal data is stored according to the received timing sequence, and its dimension is... ,in This represents the number of sampling points for a single chirp. This indicates the number of chirps contained in each frame of data. This indicates the number of antennas for receiving signals. The length of each frame of data; Perform a distance-dimensional Fast Fourier Transform on the data received by a single antenna to generate a two-dimensional matrix containing signal amplitude and phase information.

[0016] S2. By analyzing the distribution of respiratory frequency band energy in a two-dimensional data matrix at different distances, static and dynamic interference is eliminated from the matrix, and the location of the human chest cavity is located using the energy intensity of the matrix after interference elimination. The process of step S2 is as follows: First extract the two-dimensional data matrix Phase signals from different distance chambers are analyzed, and phase unwrapping is performed to recover the phase signals. The phase signals are then passed through a 0.1-0.5Hz bandpass filter to obtain the corresponding sub-band signals. The instantaneous energy of each sub-band signal is calculated and reconstructed into a respiratory energy matrix. Its dimensions are the same as the original matrix. Same. (Regarding) The matrix is ​​obtained by normalizing the distance dimension. Subsequently With the original matrix Obtain the matrix by performing the Hadamard product operation. ( = ⊙ ). By searching The energy extremum point of the matrix in the distance dimension locates the center of the human chest cavity.

[0017] S3. Based on the relative spatial relationship between the thoracic cavity, heart, and neck, further locate the thoracic cavity, heart, and neck regions. Use the second-order difference method on the target region to eliminate interference from body and respiratory activities, and obtain a set of candidate signals for thoracic heartbeat and carotid pulse; The process of step S3 is as follows: To obtain a more comprehensive heartbeat signal from the target area, this example requires extracting all signals within that area. Since the radiation diameter of the thoracic cavity region is approximately 15 to 25 centimeters, and the carotid artery region is located 15 to 25 centimeters above the thoracic cavity with a radiation range of approximately 8 to 12 centimeters, and considering the range resolution of millimeter-wave radar (approximately 4 centimeters), five consecutive range chambers within the thoracic cavity region (covering an area of ​​approximately 20 centimeters) and three range chambers near the carotid artery region (covering an area of ​​approximately 12 centimeters) are selected as the target signal extraction areas. Next, the phase signal of each distance cell is processed using second-order differential processing: ; in, Indicates the first One sampling point, This represents the time interval between sampling points; second-order differential processing can effectively suppress low-frequency interference from the body and respiration, thereby separating the heartbeat signal. Performing second-order differential processing on all phase signals in the target area yields multiple candidate heartbeat signals, which respectively constitute candidate signal sets for chest heartbeat and carotid pulse.

[0018] S4. A variational mode decomposition-based segmentation algorithm is used on the candidate signals to obtain a series of heartbeat cycle signals. The similarity between the cycle signals is used to perform an evaluation algorithm based on dynamic time programming to screen out high-quality chest heartbeat signals and carotid pulse signals. The process of step S4 is as follows: The phase signal is decomposed using a variational mode decomposition algorithm to extract the dominant frequency signal of the heartbeat. In this example, 15 decomposition layers are selected, and the heartbeat range is set to 1.0~3.0Hz. A peak detection algorithm based on amplitude thresholding is used to segment the heartbeat period according to the location of the peak, resulting in a series of periodic segments. ; Calculating the similarity between periodic segments using Dynamic Time Warping (DTW) Hierarchical clustering is performed using the average linking method, and the optimal number of clusters is determined according to the elbow rule. The cluster center with the largest number of samples is selected as the template signal for this candidate signal. The DTW average of the template signal and all periodic segments is calculated as the score. : ; The lower the average DTW of the candidate signals, the higher the similarity between the periodic segments, the less interference the signal is subjected to, and the better the signal quality. The signal with the lowest score in the candidate signal set is selected as the high-quality chest heartbeat signal and carotid artery signal for reconstructing the ECG signal.

[0019] S5. The ECG data is processed using the equal-frequency binning method to balance the amplitude distribution of the ECG waveform. The selected chest heartbeat signal and carotid pulse signal are combined with the synchronously acquired ECG signal, and a time-dependent dataset is constructed using the sliding window sampling method. The process of step S5 is as follows: To avoid model training bias, the equal-frequency binning method is used to preprocess the synchronously acquired ECG signal. Specifically, the sampling points of the original ECG signal are arranged in ascending order of voltage value and divided into 256 bins. Each bin contains the same number of data points. The ECG data in the bins are normalized to ensure that they are distributed in the range of [0,1] while preserving the trend of the original waveform. The selected chest cavity heartbeat signal, carotid pulse signal, and ECG signal were time-aligned using a cross-correlation function and downsampled to the same sampling rate of 100Hz. They were stored in a joint data matrix of size 3*n, where n is the signal length, and each column represents the chest cavity vibration amplitude, carotid pulse amplitude, and ECG voltage value at the same time. A sliding window method was used to segment the matrix, and the dataset was constructed using the ECG voltage value data of the next time step of the current window as the label. In this example, the window length is 80 sampling points, and the step size is 10 sampling points.

[0020] S6. Input the generated dataset into the ECG prediction model. The model captures the periodicity and waveform details of the heartbeat signal by extracting the temporal and spatial features between signals, and predicts the ECG value. The process of step S6 is as follows: like Figure 2 This example demonstrates a parallel network architecture for reconstructing ECG signals. A temporal convolutional network (TCN-SE) is used to extract spatial features. The TCN uses dilated causal convolution to acquire current and past information, making it suitable for prediction tasks; the SE mechanism dynamically allocates weights between chest heartbeat and carotid pulse signals, obtaining more favorable local detail features. A bidirectional gated recurrent unit (BiGRU-GA) is used to extract temporal features. Two BiGRU layers use bidirectional gated units to acquire forward and backward temporal dependencies; the GA mechanism acquires global temporal dependencies, focusing on capturing the periodicity of the P-QRS-T complex in the ECG. After concatenating and fusing the temporal and spatial features, the ECG value for the next time step is output through a fully connected layer.

[0021] The solution presented in this embodiment achieves high-precision non-contact ECG signal reconstruction, exhibiting excellent performance in both waveform and timing. Regarding waveform morphology, the reconstructed signal shows a high degree of consistency with the actual ECG signal, with an average Pearson correlation coefficient (PCC) of 94.1% and an average root mean square error (RMSE) as low as 0.068 mV. In terms of timing characteristics, the average time interval errors for key physiological parameters such as the RR interval, QRS duration, PR interval, and QT interval are 0.18%, 0.99%, 2.60%, and 1.34%, respectively. The extremely low error in the RR interval indicates that this example demonstrates exceptionally high temporal accuracy in cardiac cycle recognition.

[0022] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A non-contact ECG signal monitoring method based on millimeter-wave radar, characterized in that, The method includes the following steps: S1. The millimeter-wave radar emits a linear frequency modulated continuous wave and receives the reflected signal from the human body. It generates an intermediate frequency signal through a mixer and performs a Fourier transform on the signal in the range dimension to construct a two-dimensional data matrix. S2. By analyzing the distribution of respiratory frequency band energy in a two-dimensional data matrix at different distances, static and dynamic interference is eliminated from the matrix, and the location of the human chest cavity is located using the energy intensity of the matrix after interference elimination. S3. Based on the relative spatial relationship between the chest cavity heart and neck, further locate the chest cavity heart and neck regions, and use the second-order difference method to eliminate the interference of body and respiratory activities in the target region to obtain a set of candidate signals for chest cavity heartbeat and carotid pulse. S4. A variational mode decomposition-based segmentation algorithm is used on the candidate signals to obtain a series of heartbeat cycle signals. The similarity between the cycle signals is used to perform an evaluation algorithm based on dynamic time programming to screen out high-quality chest heartbeat signals and carotid pulse signals. S5. The ECG data is processed using the equal-frequency binning method to balance the amplitude distribution of the ECG waveform. The selected chest heartbeat signal and carotid pulse signal are combined with the synchronously acquired ECG signal, and a time-dependent dataset is constructed using the sliding window sampling method. S6. Input the generated dataset into the ECG prediction model. The model captures the periodicity and waveform details of the heartbeat signal by extracting the temporal and spatial features between signals, and predicts the ECG value.

2. The non-contact ECG signal monitoring method based on millimeter-wave radar as described in claim 1, characterized in that, In step S1, the millimeter-wave radar is placed in front of the person to be detected, and the relevant parameters for transmitting linear frequency modulated electromagnetic waves are set. A mixer mixes the transmitted and received signals to generate an intermediate frequency (IF) signal. The IF signal data is stored according to the received time sequence, and its dimension is... ,in This represents the number of sampling points for a single chirp. This indicates the number of chirps contained in each frame of data. This indicates the number of antennas for receiving signals. The length of each frame is given; a range-dimension Fast Fourier Transform is performed on the data received by a single antenna to generate a two-dimensional matrix containing signal amplitude and phase information, also with dimensions of 1. .

3. A non-contact ECG signal monitoring method based on millimeter-wave radar as described in claim 1 or 2, characterized in that, In step S2, the two-dimensional data matrix is ​​extracted. The phase signals from different distance chambers are de-wound and then passed through a bandpass filter of the classical breathing frequency band to obtain sub-band signals. Calculate the instantaneous energy of its sub-band signal and reconstruct it into a breathing energy matrix. Its dimensions are the same as the original matrix. Same, for The matrix is ​​obtained by normalizing the distance dimension. ; then With the original matrix Obtain the matrix by performing the Hadamard product operation. By searching The energy extremum point of the matrix in the distance dimension locates the center of the human chest cavity.

4. A non-contact ECG signal monitoring method based on millimeter-wave radar as described in claim 1 or 2, characterized in that, In step S3, based on the radiation diameter of the thoracic cavity region being 15 to 25 cm and the carotid artery region being located 15 to 25 cm above the thoracic cavity with a radiation range of 8 to 12 cm, and considering the range resolution of the millimeter-wave radar being 4 cm, five consecutive range chambers within the thoracic cavity region and three range chambers near the carotid artery region are selected as target signal extraction areas. The phase signal of each range chamber is subjected to second-order difference processing to suppress interference from the body and respiration, resulting in multiple candidate heartbeat signals, which respectively constitute candidate signal sets for thoracic heartbeat and carotid pulse.

5. A non-contact ECG signal monitoring method based on millimeter-wave radar as described in claim 1 or 2, characterized in that, In step S4, the phase signal is decomposed using a variational mode decomposition algorithm to extract the main frequency signal of the heartbeat; and a peak detection algorithm is performed based on the amplitude threshold to segment the heartbeat cycle according to the location of the peak, resulting in a series of cycle segments; the similarity between cycle segments is calculated using dynamic time warping (DTW), and hierarchical clustering is performed using the average linking method. The optimal number of clusters is determined according to the elbow rule, and the cluster center with the largest number of samples is selected as the template signal of this candidate signal. The DTW average of the template signal and all cycle segments is calculated as the score P, and the signal with the lowest score in the candidate signal set is selected as the high-quality chest heartbeat signal and carotid artery signal.

6. A non-contact ECG signal monitoring method based on millimeter-wave radar as described in claim 1 or 2, characterized in that, In step S5, to avoid model training bias, the synchronously acquired ECG signal is preprocessed using an equal-frequency binning method. The sampling points of the original ECG signal are arranged in ascending order of voltage value and divided into several bins to ensure that each bin contains the same number of data points. The ECG data in the bins are normalized to ensure that they are distributed within the interval [0,1] while preserving the trend of the original waveform. The chest heartbeat signal, carotid pulse signal and ECG signal are aligned with the time and sampling rate and stored in a joint data matrix of size 3*n according to the timestamp, where n is the length of the signal. The matrix is ​​divided using a sliding window method, and the ECG voltage value data of the next time step of the current window is used as the label to construct the dataset.

7. A non-contact ECG signal monitoring method based on millimeter-wave radar as described in claim 1 or 2, characterized in that, In step S6, the ECG prediction model adopts a parallel network architecture. The temporal convolutional network is combined with the channel attention mechanism to extract spatial features and dynamically allocate the weights of the chest heartbeat signal and the carotid pulse signal. The bidirectional gated circulation unit is combined with global attention to extract temporal features and focuses on capturing the periodicity of the P-QRS-T wave group in the electrocardiogram.

Images