Non-contact electrocardio feature point detection method and system based on millimeter wave radar

Abstract

The invention provides a non-contact electrocardio feature point detection method and system based on a millimeter wave radar. Non-contact real-time electrocardio feature point monitoring can be achieved. The method specifically comprises the steps that radar echo signals and reference electrocardiosignals of the heartbeat of the same subject are synchronously collected; extracting a phase signal in the radar echo signal of each period, and extracting an intrinsic mode component in the phase signal as a radar heartbeat signal component; aligning the reference electrocardiosignal with the radar heartbeat signal component, performing feature point marking on the reference electrocardiosignal, and generating a semantic segmentation tag according to feature points to obtain the reference electrocardiosignal marked with the semantic segmentation tag; constructing a deep learning model, and training the deep learning model by using the reference electrocardiosignal and the heartbeat signal component marked with the semantic segmentation tag; and performing electrocardio feature point detection on a radar heartbeat signal to be detected by using the trained deep learning model.

Description

Technical Field

[0001] The present invention relates to the field of medical detection technology, and in particular to a non-contact electrocardiogram feature point detection method and system based on millimeter wave radar. Background Art

[0002] Cardiovascular disease is the leading cause of death and loss of healthy life expectancy, and ranks first in the global disease burden. Therefore, early detection and prevention of cardiovascular disease are crucial. Electrocardiogram (ECG) is the most commonly used screening and diagnostic tool for heart disease, which can effectively record the electrical activity of the heart. In the electrical activity cycle, the key points that reflect the state of heart contraction include: the starting point of the P wave, the starting point, terminal point and peak of the QRS complex, and the terminal point of the T wave. By detecting the positions of these key points, conventional ECG features such as heart rate, PR interval, QRS duration, heart rate variability and QT interval can be calculated. These ECG features can help doctors screen for early arrhythmias and assist in pre-hospital diagnosis. However, due to the sporadic nature of early arrhythmia diseases, routine physical examinations may miss the diagnosis, and long-term ECG monitoring is required. Traditional ECG monitoring equipment needs to contact the skin through electrodes. Long-term monitoring often causes skin discomfort in patients and is not suitable for newborns, burns and infectious patients. Therefore, it is necessary to study a non-contact ECG monitoring technology and realize real-time ECG feature point detection.

[0003] As an emerging technology in the field of healthcare, low-power millimeter-wave radar can monitor human micro-motion signals and even heartbeat signals without contact. Domestic and foreign radar high-tech development companies have launched a number of low-power, high-resolution millimeter-wave radar chips that can detect submillimeter micro-motions. The millimeter-wave radar system transmits electromagnetic waves, which are modulated by the vibration of the human chest cavity, and receives a mixed signal containing breathing and heartbeat signals in the echo. After a series of radar signal processing such as distance compression, target detection and phase inversion, the mixed micro-motion signals of human breathing, heartbeat and random body movement can be extracted from the echo. In order to further obtain the heartbeat signal, there are currently a variety of signal separation technologies for the separation of breathing and heartbeat signals, but they cannot effectively suppress non-stationary and nonlinear random body motion signals. Summary of the invention

[0004] In order to overcome the defects existing in the above-mentioned prior art, the purpose of the present invention is to provide a non-contact ECG feature point detection method and system based on millimeter-wave radar, which can realize non-contact real-time ECG feature point monitoring and lay a technical foundation for non-contact diagnosis of complex heart diseases.

[0005] In order to achieve the above-mentioned object of the present invention, the present invention provides a non-contact electrocardiogram feature point detection method based on millimeter wave radar, comprising the following steps:

[0006] A millimeter-wave radar is used to collect a radar echo signal of the subject's heartbeat, and a contact-type ECG detection device is used to synchronously collect a reference ECG signal of the subject, wherein the radar echo signal and the reference ECG signal have the same sampling rate;

[0007] Extracting a phase signal from a radar echo signal of each radar transmission cycle, and extracting an intrinsic mode component corresponding to the heartbeat frequency from the phase signal as a radar heartbeat signal component;

[0008] Aligning the reference ECG signal with the radar heartbeat signal component, marking feature points of the reference ECG signal, and generating semantic segmentation labels according to the feature points to obtain a reference ECG signal marked with semantic segmentation labels that corresponds one-to-one to the heart radar heartbeat signal component;

[0009] Constructing a deep learning model and training the deep learning model using reference ECG signals and radar heartbeat signal components marked with semantic segmentation labels in one-to-one correspondence;

[0010] The trained deep learning model is used to detect ECG feature points of the radar echo signal to be detected.

[0011] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar, the step of extracting the radar heartbeat signal component corresponding to the heartbeat frequency in the phase signal is:

[0012] Extract all intrinsic mode components of the phase signal;

[0013] Calculating the spectrogram and instantaneous frequency of all said eigenmode components;

[0014] When the frequency spectrum of the eigenmode component shows that the frequency is concentrated in the heartbeat frequency range and the instantaneous frequency is lower than the instantaneous frequency threshold, the eigenmode component is a radar heartbeat signal component.

[0015] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar, the steps of extracting all eigenmode components of the phase signal are:

[0016] Input the phase signal as a reference signal into the iterative filter and calculate its mask length l(x);

[0017] Constructing a symmetric, non-negative filter matrix according to the mask length l(x), wherein each row in the filter matrix is ​​a normalized and cyclically shifted mask vector, and the sum of the mask vectors over a window length of [-l(x), l(x)] is 1;

[0018] Performing sliding filtering on the phase signal using the filter matrix;

[0019] Find all the maximum and minimum points of the signal after sliding filtering, and extract the intrinsic mode components of the phase signal;

[0020] Subtracting the extracted intrinsic mode component from the phase signal to obtain a new signal;

[0021] If the difference between the extreme point and the zero-crossing point of the new signal does not exceed 1, the extracted intrinsic mode component is added to the intrinsic mode component list; otherwise, the new signal is used as the reference signal and the above steps are repeated until no eigenmode component can be extracted.

[0022] This optional solution can adaptively change the mask length of the local filter according to the characteristics of the signal, and has a good filtering effect on random body motion signals of nonlinear and non-stationary signals, thereby removing breathing and random body motion interference and extracting a relatively pure radar heartbeat signal component.

[0023] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar, the steps of generating semantic segmentation labels are:

[0024] Identifying and marking characteristic points reflecting the cardiac contraction state in the reference electrocardiogram signal;

[0025] The portion between two adjacent characteristic points in an ECG signal of a cardiac cycle is divided into a segment of a category, so that the reference ECG signal is divided into segments of different categories, and a digital label is assigned to each category of the segment;

[0026] Set N groups of empty arrays, where N is equal to the number of digital label categories, and the digital label categories correspond one-to-one to the bits of the empty arrays;

[0027] For the empty array corresponding to each digital label, the bit belonging to the digital label in the empty array is set to 1, and the other bits are set to 0, so as to obtain N kinds of label data.

[0028] In an optional scheme of the non-contact ECG feature point detection method based on millimeter-wave radar, the deep learning model is a U-shaped structure network including a downsampling path and an upsampling path; a deep residual shrinkage convolution module is used in the downsampling path, and the features of the same scale obtained by the downsampling path and the upsampling path are spliced ​​in the channel dimension by using a jump connection.

[0029] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar,

[0030] Performing preliminary convolution on the input data segment through the downsampling path to extract low-level features, reduce data length, and increase the number of channels of the feature map;

[0031] The features of the region of interest in each layer of feature data are extracted through four sets of repeated deep residual shrinkage networks;

[0032] Restoring the image size by upsampling the extracted features of the region of interest;

[0033] The feature map of the downsampling path is fused with the feature map of the upsampling path through the skip connection operation;

[0034] The number of output channels of the last layer of deconvolution is set according to the number of classifications, and the feature map output by the upsampling path is deconvolved to obtain the recognition and positioning results.

[0035] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar, the specific steps of extracting the features of the region of interest by the deep residual shrinkage network are as follows:

[0036] First, perform a convolution operation on the input data;

[0037] Then take the absolute value of the convolution result, and then perform a global average pooling operation to reduce the data length to 1;

[0038] The result of global average pooling is subjected to two full connection operations and one sigmoid operation to obtain γ, and then the result of global average pooling is multiplied by γ to obtain the threshold τ. Finally, the threshold operation is performed on the original input data to extract the features of the region of interest.

[0039] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar, the step of extracting the phase signal in the radar heartbeat signal is:

[0040] The radar echo signal of each radar cycle is subjected to fast Fourier transform to obtain a one-dimensional range image, and the one-dimensional range images of multiple cycles are rearranged along the slow time direction to obtain a range-time matrix; the distance unit where the human target is located in the radar echo signal is analyzed; in the corresponding distance unit of the range-time matrix, the phase of the distance unit signal is unwrapped along the slow time axis direction to obtain the phase signal corresponding to the potential human target.

[0041] In an optional solution of the non-contact ECG feature point detection method based on millimeter wave radar, the prediction result output by the deep learning model is transformed into a feature point position result through a function that detects 0 to 1 transformation.

[0042] The present application also proposes a non-contact ECG feature point detection system based on millimeter wave radar, comprising a millimeter wave radar for collecting a heartbeat radar echo signal to be detected, a data acquisition unit, a processing unit and a storage unit;

[0043] The data acquisition unit acquires the radar echo signal of the subject's heartbeat used for model training and the reference ECG signal, as well as the heartbeat radar echo signal to be detected; the data acquisition unit is communicatively connected to the processing unit, the processing unit is communicatively connected to the storage unit, the storage unit is used to store at least one executable instruction, the executable instruction enables the processing unit to perform operations corresponding to the above-mentioned millimeter-wave radar-based non-contact ECG feature point detection method based on the radar echo signal of the subject's heartbeat used for model training, the reference ECG signal and the heartbeat radar echo signal to be detected, and perform ECG feature point detection on the heartbeat radar echo signal to be detected.

[0044] The beneficial effects of the present invention are as follows: the present invention can effectively suppress random body motion interference signals in radar echoes and extract relatively pure heartbeat signals, and use the signals and reference ECG signals with semantic segmentation labels to train a deep learning model, which can automatically identify and locate ECG key points in radar heartbeat signal components, and automatically calculate ECG indicators (such as PR interval, QT interval, and RR interval, etc.), thereby improving the efficiency and accuracy of non-contact ECG monitoring and assisting doctors in diagnosis.

[0045] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0047] Figure 1 It is a schematic diagram of the process of embodiment 1;

[0048] Figure 2 It is a schematic diagram of semantic segmentation label data;

[0049] Figure 3 It is a schematic diagram of the network structure of the deep learning model. DETAILED DESCRIPTION

[0050] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0051] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0052] Embodiment 1

[0053] like Figure 1 As shown, the present invention provides a non-contact ECG feature point detection method based on millimeter wave radar, which can realize non-contact real-time ECG feature point monitoring and lay a technical foundation for non-contact complex heart disease diagnosis.

[0054] Specifically, the method includes the following steps: using millimeter wave radar and contact ECG detection equipment to synchronously collect the radar echo signal and reference ECG signal of the subject's heartbeat, and the radar echo signal and the reference ECG signal have the same sampling rate and the same length.

[0055] When the millimeter-wave radar is used to collect the radar heartbeat signal of the subject, in this embodiment, the subject lies flat on the bed and maintains stable breathing, and the low-power millimeter-wave radar is placed 50 cm above the subject's xiphoid process. The radar emits multiple cycles of electromagnetic waves. After being reflected by the subject's chest cavity, the radar echo signal of each radar emission cycle contains the heartbeat signal and clutter (clutter mainly includes breathing and random body motion interference). The receiving antenna receives the radar echo signals of multiple radar emission cycles after micro-motion modulation.

[0056] Then, the phase signal in the radar echo signal of each radar transmission cycle is extracted, and the intrinsic mode component corresponding to the heartbeat frequency in the phase signal is extracted as the heartbeat signal component. Specifically,

[0057] The process of extracting the phase signal in this embodiment is as follows: first, the radar echo signal of each radar transmission cycle is subjected to a fast Fourier transform to obtain a one-dimensional range image, and then the one-dimensional range images of multiple cycles are rearranged along the slow time direction to obtain a range-time matrix. Then, the static interference is removed by moving target detection, and then the distance unit where the human target is located in the radar echo signal is obtained by using a constant false alarm rate detection algorithm. In the distance unit corresponding to the distance-time matrix, the distance unit signal is phase unwrapped along the slow time direction to obtain a phase signal S corresponding to the potential human target, and the phase signal S contains a heartbeat signal, a breathing signal, and a random body motion signal.

[0058] In this embodiment, the phase signal S is decomposed by an adaptive local iterative filtering algorithm to suppress the random body motion signal and decompose the phase signal S(x) into the eigenmodal components and the remaining trend term Sα , extract the intrinsic mode component as the heartbeat signal component. Specifically:

[0059] 1. Initialization

[0060] a) Input phase signal S(x), denoted as S 0 (x).

[0061] b) Initialize the list of intrinsic mode components IMFs to be empty.

[0062] 2. Extract intrinsic mode components IMF

[0063] a) Calculate the mask length: Substitute the reference signal S n (x) Input the iterative filter and calculate its mask length l n (x), n refers to the nth iteration. Initially, the reference signal S n (x) is the phase signal S(x), that is, S 0 (x).

[0064] b) Construct the filter matrix: according to the mask length l n (x) Construct a symmetric, non-negative filter matrix W, and each row in the filter matrix W is a normalized and cyclically shifted mask vector w(x), and satisfies the condition: the sum of the mask vector w(x) over the window length [-l(x), l(x)] is 1, that is, t refers to time.

[0065] c) Filtering operation: Apply the filter matrix W to the signal S n (x), and get the signal h after sliding filtering n (x).

[0066] d) Calculate the extreme point: find the signal h n All the maximum and minimum points of (x) are used to extract the eigenmode components

[0067] e) Update signal: From the original reference signal S n Subtract (x) Get a new signal

[0068] 3. Check the stop condition

[0069] a) Stop condition: current eigenmode component When the difference between the extreme point and the zero crossing point does not exceed 1, the current eigenmode component Add to the list of intrinsic mode components IMFs.

[0070] b) If the stop condition is not met, use the new signal S n+1(x) is used as the reference signal S n (x) Repeat step 2 until the remaining signal S n+1 (x) The eigenmode components can no longer be extracted And the remaining signal is recorded as the trend term S α .

[0071] 4. Extract heartbeat signal

[0072] a) Calculate each eigenmode component Spectrum plot and Hilbert transform instantaneous frequency.

[0073] b) Set the instantaneous frequency threshold. When the instantaneous frequency exceeds the instantaneous frequency threshold, the component is considered to be a random body motion signal. Considering that the random body motion signal is a non-stationary mutation signal, the instantaneous frequency threshold in this embodiment is preferably but not limited to 20 Hz.

[0074] c) When the frequency spectrum shows that the frequency is concentrated in the heart rate range (0.7-2 Hz), and the instantaneous frequency of the Hilbert transform is lower than the instantaneous frequency threshold, it is considered that the intrinsic mode component The heart rate.

[0075] Through the above processing, the interference of breathing and random body movement can be removed, and a relatively pure radar heartbeat signal component can be extracted.

[0076] This embodiment also aligns the reference ECG signal with the radar heartbeat signal component, marks feature points of the reference ECG signal, and generates semantic segmentation labels based on the feature points. In this embodiment, the steps of generating semantic segmentation labels are:

[0077] a) Align the radar heartbeat signal with the reference ECG signal and ensure that the two signals have the same sampling rate and length.

[0078] b) Identify and mark the characteristic points in the reference ECG signal that reflect the heart's contraction state, including: the starting point of the P wave, the starting point of the QRS complex, the peak of the R wave, the end point of the QRS complex, and the end point of the T wave.

[0079] c) The portion between two adjacent characteristic points in the ECG signal of a cardiac cycle is divided into a segment of a category, so that the reference ECG signal is divided into segments of different categories, and a digital label is assigned to each category of the segment. Specifically, in the ECG signal of a cardiac cycle, the signal segment from the starting point of the P wave to the starting point of the QRS complex is marked as 1, representing the P wave; the signal segment from the starting point of the QRS complex to the peak of the R wave is marked as 2, representing the rising part of the QRS complex; the signal segment from the peak of the R wave to the end point of the QRS complex is marked as 3, representing the falling part of the QRS complex; the signal segment from the end point of the QRS complex to the end point of the T wave is marked as 4, representing the ST segment and T wave; the signal segment from the end point of the T wave to the starting point of the next P wave is marked as 5, representing the heart recovery period after the T wave.

[0080] d) Set N groups of empty arrays, where N is equal to the number of digital label categories. Each digital label corresponds to a group of empty arrays. The number of empty arrays is the same as the number of samples of the corresponding ECG signal, and the digital label categories correspond one-to-one to the bits of the empty arrays. For the empty array corresponding to each digital label, set the bit belonging to the digital label in the empty array to 1, and the other bits to 0, to obtain N types of label data. In this way, a set of label data sets in binary form can be obtained, in which the change position from 0 to 1 corresponds to the position of each ECG feature point. From the above, it can be seen that there are 5 categories of digital labels in this embodiment, so 5 groups of empty arrays are set to obtain five types of label data, such as Figure 2 shown.

[0081] Through the above operations, a reference ECG signal marked with semantic segmentation labels corresponding to the heart radar heartbeat signal components is obtained.

[0082] A deep learning model is constructed and trained using reference ECG and heartbeat signal components marked with semantic segmentation labels to obtain the optimal parameter configuration.

[0083] In this embodiment, in order to realize the automatic semantic segmentation detection of ECG key points, a one-dimensional convolutional neural network for feature point recognition and positioning is designed as a deep learning model. The network is a U-shaped structure network consisting of a downsampling path (encoder structure) and an upsampling path (decoder structure). The network uses a deep residual shrinkage convolution module in the downsampling path, and also uses a jump connection method to splice the features of the same scale obtained by downsampling and upsampling in the channel dimension. Figure 3 As shown, take the data segment with an input length of 5000 as an example:

[0084] 1. Perform preliminary convolution on the input data segment through the downsampling path to extract low-level features, reduce the data length, and increase the number of channels of the feature map.

[0085] 2. Through four sets of repeated deep residual shrinkage networks, the features of the region of interest in each layer of feature data are extracted and the interfering data point information is discarded. The specific steps of the deep residual shrinkage network are as follows:

[0086] a) The module first performs a convolution operation on the input data.

[0087] b) Take the absolute value of the convolution result and perform a global average pooling operation to reduce the data length to 1.

[0088] c) The result of global average pooling is obtained through two full connection operations and one sigmoid operation to obtain γ, and then the result of global average pooling is multiplied by γ to obtain the threshold τ. Finally, the threshold operation is performed with the original input data to extract the features of the region of interest. The formula for the threshold operation is as follows:

[0089]

[0090] A represents the input feature, B represents the output feature, and τ represents the threshold that can be adjusted adaptively. The thresholding operation sets the features in the range of [-τ,τ] to 0, which is a more flexible way to remove redundant information.

[0091] 3. Restore the image size by upsampling the extracted features of the region of interest (deconvolution operation).

[0092] 4. The feature map of the downsampling path (encoder stage) is fused with the feature map of the upsampling path (decoder stage) through the skip connection operation to retain more detail information.

[0093] 5. According to the number of classifications (i.e. the number of key point types), the number of output channels of the last layer of deconvolution is set, and the feature map output by the upsampling path (decoder) is deconvolved to obtain the recognition and positioning results.

[0094] In this embodiment, the training and testing steps for the deep learning model are as follows:

[0095] The reference ECG signal and radar heartbeat signal component data marked with semantic segmentation labels are divided into training sets and test sets (in pairs) one by one, the data are normalized, and the data are segmented into segments of 10 seconds each.

[0096] In the deep learning model designed by inputting the radar heartbeat signal in the training set, five types of predicted semantic segmentation data are output.

[0097] The Adam optimizer was used during training, with the learning rate set to 0.005, the batch size to 512, and the number of iterations to 200.

[0098] The root mean square error loss function is used to calculate the difference between the predicted result and the true label. By optimizing the loss function, the network parameters are continuously adjusted to improve the accuracy of recognition and positioning.

[0099] The prediction results output by the network are transformed into feature point position results through a function that detects 0 to 1 transformation.

[0100] The evaluation index calculation method during the test is based on the standards set by the American Association for the Advancement of Medical Instrumentation (AAMI), and the 75ms before and after the ECG reference point position is selected as the tolerance range. Taking the P wave starting point position discrimination as an example, referring to the position of the P wave starting point in the ECG, if the P wave starting point is correctly detected within the tolerance range, it is counted as a true positive (TP) of the P wave starting point classification; if the corresponding P wave starting point is not detected within this tolerance range, it is counted as a false negative (FN) of the P wave starting point classification; if the P wave starting point is detected outside this tolerance range, it is counted as a false positive (FP) of the P wave starting point classification.

[0101] This example uses precision, recall, and F1-score values ​​to compare the performance of different methods. The calculation formula for the evaluation index used is as follows:

[0102]

[0103]

[0104] Among them, TP is the number of positive samples judged as positive samples, FP is the number of negative samples judged as positive samples, and FN is the number of positive samples judged as negative samples.

[0105] Finally, the radar heartbeat signal to be detected is input into the trained deep learning model for ECG feature point detection, which can automatically identify the positions of five ECG key points: the starting point of the P wave, the starting point, end point and peak point of the QRS complex, and the end point of the T wave.

[0106] Embodiment 2

[0107] The present invention also provides a non-contact electrocardiogram feature point detection system based on millimeter-wave radar, comprising a millimeter-wave radar for collecting a heartbeat radar echo signal to be detected, a data acquisition unit, a processing unit and a storage unit.

[0108] The data acquisition unit acquires the radar echo signal of the subject's heartbeat and the reference electrocardiogram signal for model training, as well as the heartbeat radar echo signal to be detected; the data acquisition unit is communicatively connected to the processing unit, the processing unit is communicatively connected to the storage unit, the storage unit is used to store at least one executable instruction, the executable instruction enables the processing unit to perform operations corresponding to the non-contact electrocardiogram feature point detection method based on millimeter-wave radar described in Example 1 based on the radar echo signal of the subject's heartbeat for model training, the reference electrocardiogram signal and the heartbeat radar echo signal to be detected, and perform electrocardiogram feature point detection on the heartbeat radar echo signal to be detected.

[0109] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0110] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A non-contact electrocardiogram feature point detection method based on millimeter wave radar, characterized in that: The following steps are involved: A millimeter-wave radar is used to collect a radar echo signal of the subject's heartbeat, and a contact-type ECG detection device is used to synchronously collect a reference ECG signal of the subject, wherein the radar echo signal and the reference ECG signal have the same sampling rate; Extracting a phase signal from a radar echo signal of each radar transmission cycle, and extracting an intrinsic mode component corresponding to the heartbeat frequency from the phase signal as a radar heartbeat signal component; Aligning the reference ECG signal with the radar heartbeat signal component, marking feature points of the reference ECG signal, and generating semantic segmentation labels according to the feature points to obtain a reference ECG signal marked with semantic segmentation labels that corresponds one-to-one to the radar heartbeat signal component; Constructing a deep learning model and training the deep learning model using reference ECG signals and radar heartbeat signal components marked with semantic segmentation labels in one-to-one correspondence; The trained deep learning model is used to detect ECG feature points of the radar echo signal to be detected.

2. The non-contact ECG feature point detection method based on millimeter wave radar according to claim 1 is characterized in that: The steps of extracting the radar heartbeat signal component corresponding to the heartbeat frequency in the phase signal are: Extract all intrinsic mode components of the phase signal; Calculating the spectrogram and instantaneous frequency of all said eigenmode components; When the frequency spectrum of the eigenmode component shows that the frequency is concentrated in the heartbeat frequency range and the instantaneous frequency is lower than the instantaneous frequency threshold, the eigenmode component is a radar heartbeat signal component.

3. The non-contact electrocardiogram feature point detection method based on millimeter wave radar according to claim 1 or 2, characterized in that: The steps to extract all the eigenmode components of the phase signal are: Input the phase signal as a reference signal into the iterative filter and calculate its mask length l(x); Constructing a symmetric, non-negative filter matrix according to the mask length l(x), wherein each row in the filter matrix is ​​a normalized and cyclically shifted mask vector, and the sum of the mask vectors over a window length of [-l(x), l(x)] is 1; Performing sliding filtering on the phase signal using the filter matrix; Find all the maximum and minimum points of the signal after sliding filtering, and extract the intrinsic mode components of the phase signal; Subtracting the extracted intrinsic mode component from the phase signal to obtain a new signal; If the difference between the extreme point and the zero-crossing point of the new signal does not exceed 1, the extracted intrinsic mode component is added to the intrinsic mode component list; otherwise, the new signal is used as the reference signal and the above steps are repeated until no eigenmode component can be extracted.

4. The non-contact electrocardiogram feature point detection method based on millimeter wave radar according to claim 1 is characterized in that: The steps to generate semantic segmentation labels are: Identifying and marking characteristic points reflecting the cardiac contraction state in the reference electrocardiogram signal; The portion between two adjacent characteristic points in an ECG signal of a cardiac cycle is classified as a segment of a category, so that the reference ECG signal is divided into segments of different categories, and a digital label is assigned to each category of the segment; Set N groups of empty arrays, where N is equal to the number of digital label categories, and the digital label categories correspond one-to-one to the bits of the empty arrays; For the empty array corresponding to each digital label, the bit belonging to the digital label in the empty array is set to 1, and the other bits are set to 0, so as to obtain N kinds of label data.

5. The non-contact ECG feature point detection method based on millimeter wave radar according to claim 1, characterized in that: The deep learning model is a U-shaped structure network including a downsampling path and an upsampling path; a deep residual shrinkage convolution module is used in the downsampling path, and features of the same scale obtained by the downsampling path and the upsampling path are spliced ​​in the channel dimension by using a jump connection.

6. The non-contact ECG feature point detection method based on millimeter wave radar according to claim 5 is characterized in that: Performing preliminary convolution on the input data segment through the downsampling path to extract low-level features, reduce data length, and increase the number of channels of the feature map; The features of the region of interest in each layer of feature data are extracted through four sets of repeated deep residual shrinkage networks; Restoring the image size by upsampling the extracted features of the region of interest; The feature map of the downsampling path is fused with the feature map of the upsampling path through the skip connection operation; The number of output channels of the last layer of deconvolution is set according to the number of classifications, and the deconvolution operation is performed on the feature map output by the upsampling path to obtain the recognition and positioning results.

7. The non-contact ECG feature point detection method based on millimeter wave radar according to claim 6 is characterized in that: The specific steps of deep residual shrinkage network to extract features of the region of interest are as follows: First, perform a convolution operation on the input data; Then take the absolute value of the convolution result, and then perform a global average pooling operation to reduce the data length to 1; The result of global average pooling is subjected to two full connection operations and one sigmoid operation to obtain γ, and then the result of global average pooling is multiplied by γ to obtain the threshold τ. Finally, the threshold operation is performed on the original input data to extract the features of the region of interest.

8. The non-contact ECG feature point detection method based on millimeter wave radar according to claim 1, characterized in that: The steps to extract the phase signal from the radar echo signal are: Perform fast Fourier transform on the radar echo signal of each radar cycle to obtain a one-dimensional range image, and rearrange the one-dimensional range images of multiple cycles along the slow time direction to obtain a range-time matrix; analyze the distance unit where the human target is located in the radar echo signal; In the distance unit corresponding to the distance-time matrix, the distance unit signal is phase unwrapped along the slow time axis to obtain the phase signal corresponding to the potential human target.

9. The non-contact ECG feature point detection method based on millimeter wave radar according to claim 4, characterized in that: The prediction results output by the deep learning model are transformed into feature point position results through a function that detects 0 to 1 transformation.

10. A non-contact ECG feature point detection system based on millimeter wave radar, characterized in that: It includes a millimeter wave radar for collecting the heartbeat radar echo signal to be detected, a data acquisition unit, a processing unit and a storage unit; The data acquisition unit acquires the radar echo signal of the subject's heartbeat used for model training and the reference ECG signal, as well as the heartbeat radar echo signal to be detected; the data acquisition unit is communicatively connected to the processing unit, the processing unit is communicatively connected to the storage unit, the storage unit is used to store at least one executable instruction, the executable instruction enables the processing unit to perform operations corresponding to the non-contact ECG feature point detection method based on millimeter-wave radar as described in any one of claims 1 to 9 based on the radar echo signal of the subject's heartbeat used for model training, the reference ECG signal and the heartbeat radar echo signal to be detected, and perform ECG feature point detection on the heartbeat radar echo signal to be detected.

Images