Method and device for acquiring contactless electrocardiogram signal

Through millimeter wave radar and singular spectrum decomposition combined with CNN-LSTM-Attention model, the problem of low signal-to-noise ratio in contactless electrocardiogram detection is solved, and more accurate electrocardiogram signal acquisition and detection is achieved.

CN118557184BActive Publication Date: 2024-12-06INST OF AUTOMATION CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411048467.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-12-06
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

In the prior art, contactless electrocardiogram detection is less vibration of the chest wall caused by heartbeat and low signal-to-noise ratio, which makes it difficult to accurately detect.

Method used

The contactless detection was performed by millimeter wave radar, and the cardiac vibration signal was extracted through static clutter weakening treatment and singular spectrum decomposition method, and the trained CNN-LSTM-Attention mixed model was used for inference analysis to output the electrocardiogram signal.

Benefits of technology

It improves the detection accuracy and signal-to-noise ratio of the ECG signal, and realizes the acquisition of the ECG signal without the need for conventional instrument attachment, reducing model complexity and resource consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118557184B_ABST
    Figure CN118557184B_ABST
Patent Text Reader

Abstract

The present invention provides a method and device for acquiring a contactless electrocardiogram signal, which relates to the technical field of signal processing. The method comprises: receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter attenuation processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object; after filtering the vibration signal, extracting a cardiac vibration tracing signal from the vibration signal by a singular spectrum decomposition method; and inferring and analyzing the cardiac vibration tracing signal by a trained CNN-LSTM-Attention hybrid model to output the electrocardiogram signal of the target object.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a method and device for acquiring contactless electrocardiogram signals. Background Art

[0002] In the field of medical health monitoring, the non-contact detection method of millimeter wave radar has received widespread attention. For example, in heartbeat detection, the radar can emit and receive electromagnetic waves to sense tiny vibrations of the chest wall, which are related to the beating of the heart. Although millimeter wave radar is effective in detecting breathing and other larger body movements, in the detection of heartbeat signals, the chest wall vibration caused by the heartbeat is relatively small, and its signal-to-noise ratio is usually low, which poses a considerable challenge to accurate detection.

[0003] Therefore, how to perform electrocardiogram detection in a contactless manner has become an urgent problem to be solved in the industry. Summary of the invention

[0004] The present invention provides a method and device for acquiring a contactless electrocardiogram signal, which are used to solve the problem of how to detect an electrocardiogram in a contactless manner in the prior art.

[0005] The present invention provides a method for acquiring a contactless electrocardiogram signal, comprising the following steps.

[0006] receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object;

[0007] After filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method;

[0008] The cardiac vibration tracing signal is inferred and analyzed through the trained CNN-LSTM-Attention hybrid model to output the electrocardiogram signal of the target object.

[0009] According to a method for acquiring a contactless electrocardiogram signal provided by the present invention, before the step of inferring and analyzing the cardiac vibration tracing signal through a trained CNN-LSTM-Attention hybrid model and outputting the electrocardiogram signal of the target object, the method further includes:

[0010] Acquire a plurality of first cardiac vibration tracing sample information carrying systolic period marks and diastolic period marks;

[0011] For any of the first cardiac vibration tracing sample information, the first cardiac vibration tracing sample information is input into an initial CNN-LSTM-Attention hybrid model, and the first cardiac vibration tracing sample information is segmented to obtain systolic sample information and diastolic sample information.

[0012] Calculating a first loss value according to the systolic period mark, the systolic period sample information, the diastolic period mark and the diastolic period sample information;

[0013] If the first loss value is less than a first preset threshold, the first stage training of the initial CNN-LSTM-Attention hybrid model is completed to obtain a first CNN-LSTM-Attention hybrid model.

[0014] According to a method for acquiring a contactless electrocardiogram signal provided by the present invention, after the step of obtaining the first CNN-LSTM-Attention hybrid model, the method further includes:

[0015] Acquire multiple systolic sample information and diastolic sample information carrying reference point peak and trough area labels;

[0016] For any of the systolic sample information or the diastolic sample information, input the systolic sample information or the diastolic sample information into the initial CNN-LSTM-Attention hybrid model to detect the peak and trough areas of the reference points in the systolic sample information or the diastolic sample information;

[0017] Calculating a second loss value according to the reference point peak and trough area and the reference point peak and trough area label;

[0018] If the second loss value is less than the second preset threshold, the second stage training of the first CNN-LSTM-Attention hybrid model is completed to obtain the second CNN-LSTM-Attention hybrid model.

[0019] According to a method for acquiring a contactless electrocardiogram signal provided by the present invention, after the step of obtaining the second CNN-LSTM-Attention hybrid model, the method further includes:

[0020] Acquiring a plurality of third cardiac vibration tracing sample information carrying target reference point labels;

[0021] For any of the third cardiac oscillography sample information, input the third cardiac oscillography sample information into an initial CNN-LSTM-Attention hybrid model to detect a target reference point of the third cardiac oscillography sample information;

[0022] Calculating a third loss value according to the target reference point and the target reference point label;

[0023] If the third loss value is less than the third preset threshold, the third stage training of the second CNN-LSTM-Attention hybrid model is completed to obtain a trained CNN-LSTM-Attention hybrid model.

[0024] According to a method for acquiring a contactless electrocardiogram signal provided by the present invention, the millimeter-wave radar echo signal reflected by the target object is received, and static clutter reduction processing is performed on the millimeter-wave radar echo signal to obtain a vibration signal of the target object, including:

[0025] Obtain the median of the static background signal of the F frame in the millimeter wave radar detection area when there is no target object, and obtain the median background signal;

[0026] Subtracting the median background signal from the millimeter-wave radar echo signal reflected by the target object to implement static clutter attenuation processing and obtain a vibration signal after static clutter attenuation processing;

[0027] The vibration signal after the static clutter attenuation processing is subjected to noise filtering based on a target bandpass filter to obtain the vibration signal of the target object.

[0028] According to a method for acquiring a contactless electrocardiogram signal provided by the present invention, after filtering the vibration signal for noise, a cardiac vibration recording signal in the vibration signal is extracted by a singular spectrum decomposition method, comprising:

[0029] The vibration signal is embedded into a plurality of time series signals, and each of the time series signals is shifted and superimposed on the time axis to form a trajectory matrix;

[0030] Performing singular value analysis on the trajectory matrix to obtain a plurality of singular value components and corresponding left singular vectors and right singular vectors, and constructing a component matrix based on the left singular vectors and the right singular vectors; wherein each of the component matrices corresponds to a singular value and a frequency pattern;

[0031] Perform diagonal averaging on each component matrix to obtain a time series of each component matrix, and then perform fast Fourier transform on each time series to obtain the frequency characteristics of each component matrix;

[0032] Obtaining a target component matrix that matches the frequency characteristic range of the cardiac vibration tracing signal from each of the component matrices, and linearly combining the target component matrices to obtain a cardiac vibration-related time series signal;

[0033] The diagonal averaging algorithm is used to filter and remove noise from the cardiac vibration related time series signal to obtain the cardiac vibration tracing signal.

[0034] According to a method for acquiring a contactless electrocardiogram signal provided by the present invention, the step of inferring and analyzing the cardiac vibration tracing signal through a trained CNN-LSTM-Attention hybrid model and outputting the electrocardiogram signal of the target object includes:

[0035] Inputting the cardiac oscillography signal into the CNN network in the CNN-LSTM-Attention hybrid model to extract the spatial features of the cardiac oscillography signal;

[0036] Inputting the spatial features into the LSTM network in the CNN-LSTM-Attention hybrid model to capture the temporal relationship and dynamic changes between heartbeats and extract seven heartbeat reference points in the cardiac vibration tracing signal;

[0037] Based on the seven heartbeat reference points, the electrocardiogram signal of the target object is reconstructed.

[0038] The present invention also provides a contactless electrocardiogram signal acquisition device, comprising the following modules: a receiving module, an extraction module and an output module.

[0039] A receiving module, used for receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object;

[0040] An extraction module, configured to extract a cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method after filtering out noise from the vibration signal;

[0041] The output module is used to perform inference analysis on the cardiac vibration tracing signal through the trained CNN-LSTM-Attention hybrid model, and output the electrocardiogram signal of the target object.

[0042] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for acquiring a contactless electrocardiogram signal as described above is implemented.

[0043] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for acquiring a contactless electrocardiogram signal as described in any one of the above is implemented.

[0044] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method for acquiring a contactless electrocardiogram signal as described above is implemented.

[0045] The present invention provides a contactless electrocardiogram signal acquisition method and device, which detects the position and speed of the target object by sending and receiving signals by a millimeter-wave radar, and can acquire vibration signals for reconstructing electrocardiogram signals without the need for conventional instruments to be attached, and designs a CNN-LSTM-Attention hybrid deep learning network for ECG signal reconstruction. The CNN network has the ability to extract waveform spatial features; the LSTM can independently learn and identify important time features, and independently realize the positioning of the reference point of the cardiac vibration tracing signal without relying on the electrocardiogram signal as a reference signal. The combination of the two realizes more accurate ECG signal reconstruction. In addition, the CNN-LSTM-Attention hybrid deep learning network is trained by a three-layer positioning algorithm suitable for cardiac vibration signal segmentation to realize the segmentation and positioning of cardiac vibration tracing signals from cardiac contraction and diastole to 7 precise feature points. This method not only improves the processing speed, but also reduces the complexity of the model and resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a flow chart of the method for acquiring contactless electrocardiogram signals provided by the present invention.

[0048] Figure 2 Schematic diagram of the CNN-LSTM-Attention hybrid model provided in an embodiment of the present application.

[0049] Figure 3 A schematic diagram of cardiac vibration recording signal segmentation provided in an embodiment of the present application.

[0050] Figure 4 A schematic diagram of the structure of a contactless electrocardiogram signal acquisition device provided in an embodiment of the present application.

[0051] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Figure 1 FIG. 1 is a flow chart of a method for acquiring a contactless electrocardiogram signal provided by the present invention. Figure 1 As shown, the method includes the following steps 101, 102 and 103.

[0054] Step 101, receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object;

[0055] In the embodiment of the present application, the target object may specifically refer to a user who needs to undergo an electrocardiogram test, or may be an animal or other organism that needs to undergo an electrocardiogram test.

[0056] When detecting human heart vibration recording signals, the millimeter wave radar is generally placed about 30cm in front of the chest of the target object as much as possible, and the radar antenna transceiver array faces the human body. At this time, the displacement fluctuations sensed by the radar are mainly caused by the movement of the chest surface caused by human breathing and heartbeat. Since human breathing and heartbeat can be regarded as periodic movements from a macroscopic perspective.

[0057] The instantaneous distance between the radar antenna array and the human chest cavity is modeled as follows:

[0058]

[0059] in, represents the static average distance from the radar antenna array to the human chest surface, Indicates the amplitude of chest surface vibration caused by human breathing. Indicates the amplitude of chest surface vibration caused by human heartbeat. and Represents the amplitude of breathing and heartbeat respectively. and Represents the frequency of breathing and heartbeat respectively. Indicates slow time.

[0060] The round-trip time from the electromagnetic wave signal emitted by the radar antenna array transmitting antenna to the surface of the human chest cavity is

[0061]

[0062] in, represents the speed of light in air, , , .

[0063] In the embodiment of the present application, the millimeter wave radar echo signal may specifically be a callback signal of the transmission signal of the millimeter wave radar reflected by the chest cavity of the target user.

[0064] For general signal acquisition scenarios, there are static environmental objects such as tables and walls around the human body. Although these objects have different electromagnetic wave reflection coefficients, they will generate echoes to the radar's transmission signal and generate static clutter. In addition, for moving or standing persons being measured, motion artifacts are the biggest obstacle to reliable measurement. Compared with stationary objects such as beds or chairs in the environment, the subjects usually try to remain still during measurement. Swaying posture or walking will produce unnecessary peaks in the measurement signal or cause distortion of the signal waveform. In order to effectively extract the vibration signal reflected by the human body, the static background signal reflected by the surrounding environment needs to be eliminated.

[0065] Optionally, the receiving a millimeter-wave radar echo signal reflected by a target object and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object includes:

[0066] Obtain the median of the static background signal of the F frame in the millimeter wave radar detection area when there is no target object, and obtain the median background signal;

[0067] Subtracting the median background signal from the millimeter-wave radar echo signal reflected by the target object to implement static clutter attenuation processing and obtain a vibration signal after static clutter attenuation processing;

[0068] The vibration signal after the static clutter attenuation processing is subjected to noise filtering based on a target bandpass filter to obtain the vibration signal of the target object.

[0069] In the embodiment of the present application, in the absence of a target object (human body), the millimeter wave radar will detect reflection signals caused by objects in the surrounding environment, which are called static background signals. F frames of static background signals are collected, and the median of these signals is calculated to obtain a median background signal. This median background signal represents the average electromagnetic wave reflection characteristics of the environmental background.

[0070] When a target object enters the radar detection area, the radar receives the millimeter-wave radar echo signal reflected by the target object. The previously acquired median background signal is subtracted from these echo signals to eliminate the static clutter caused by environmental objects, thereby highlighting the reflected signal of the target object.

[0071] The formula is as follows:

[0072]

[0073] in, is the processed signal containing the target human body vibration, is the original radar signal for each frame, is the median of the static background signal in frame F, that is, the background clutter.

[0074] More specifically, since the frequency components of the cardiac oscillography signal are mainly distributed between 5.0 Hz and 30 Hz, in order to associate the characteristics of the heartbeat signal obtained by non-invasive monitoring with the characteristics of the ECG signal, the high-frequency components in the heartbeat signal can reflect the rapid changes of the heart, which is very important for the accurate measurement of heart rate and the identification of potential abnormal heart rhythm.

[0075] Therefore, a bandpass filter (BPF) with a passband of [5.0, 30] Hz is used to filter out noise and unwanted signal components. This filtering process ensures the retention of high-frequency heartbeat components while removing low-frequency and non-heartbeat related high-frequency noise that may mask or confuse important biological signal information, and finally obtains the vibration signal of the target object.

[0076] Step 102, after filtering the vibration signal for noise, extracting the cardiac vibration recording signal from the vibration signal by singular spectrum decomposition method;

[0077] In the embodiment of the present application, the singular spectrum decomposition method includes five steps, namely embedding, decomposition, analysis, combination and reconstruction.

[0078] Optionally, after filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by singular spectrum decomposition method includes:

[0079] The vibration signal is embedded into a plurality of time series signals, and each of the time series signals is shifted and superimposed on the time axis to form a trajectory matrix;

[0080] Performing singular value analysis on the trajectory matrix to obtain a plurality of singular value components and corresponding left singular vectors and right singular vectors, and constructing a component matrix based on the left singular vectors and the right singular vectors; wherein each of the component matrices corresponds to a singular value and a frequency pattern;

[0081] Perform diagonal averaging on each component matrix to obtain a time series of each component matrix, and then perform fast Fourier transform on each time series to obtain the frequency characteristics of each component matrix;

[0082] Obtaining a target component matrix that matches the frequency characteristic range of the cardiac vibration tracing signal from each of the component matrices, and linearly combining the target component matrices to obtain a cardiac vibration-related time series signal;

[0083] The diagonal averaging algorithm is used to filter and remove noise from the cardiac vibration related time series signal to obtain the cardiac vibration tracing signal.

[0084] In the embedding stage, the chest surface vibration signal containing the human heart vibration recording signal is Mapped to a multidimensional lag vector sequence. Select a positive integer L as the window length. The window length L should be neither too long nor too short. Too long will lead to sparse matrix, and too short will fail to capture the structure of the data. Therefore, , time series is a OK, Matrix of columns.

[0085]

[0086] In the decomposition stage, the trajectory matrix obtained in the embedding stage is subjected to singular value decomposition. The singular value decomposition decomposes the trajectory matrix into several singular values ​​and their corresponding left and right singular vector matrices. Among them, the left and right singular vector matrices are called component matrices, and their singular value decomposition can be expressed as:

[0087]

[0088] in, , yes The eigenvalues ​​of a matrix, yes The rank of the matrix, is a second-order matrix of dimensional feature vector, is related to non-zero eigenvalues of -dimensional feature vector.

[0089] In the analysis stage, since the extraction of cardiac vibration waveform from chest vibration signal is mainly based on the frequency characteristics of cardiac vibration tracing signal, it is necessary to solve the frequency of the time series corresponding to each component matrix. The specific method is to perform diagonal averaging on each component matrix to obtain the corresponding time series, and then perform FFT on each time series to obtain its frequency distribution. According to the frequency characteristics of cardiac vibration tracing signal, determine the index i of the corresponding target component.

[0090] In the combination stage, the frequency characteristics of the human heart vibration recording signal and the frequency characteristics of each component matrix are used. According to the experimental results, the frequency range of the human heart vibration recording signal is generally 5-12Hz. Therefore, firstly, according to the frequency analysis of the component matrix, several component matrices whose frequencies fall within the range of 5-12Hz are found and divided into a group. Then, the remaining high-frequency component matrices are divided into several groups again according to their frequency characteristics, and the component matrices with similar frequency characteristics within each group are put together. It is necessary to ensure that the frequency ranges of the component matrices between different groups do not overlap, so that the groups are unrelated. The component matrix can be Combination unrelated groups, and reconstruct each group into a time series. Group contains component matrices, numbered , Then the reconstructed time series of this group Calculation formula, .

[0091] In the reconstruction phase, the operation is repeated until all component matrices are reconstructed into Time series output.

[0092] In the process of extracting cardiac vibration recording signals through singular spectrum decomposition, information related to cardiac activity can be effectively separated from the vibration signal after noise filtering. This method not only improves the quality of signal processing, but also provides high-quality input data for subsequent deep learning model training and ECG signal reconstruction.

[0093] Step 103: Infer and analyze the cardiac vibration recording signal using the trained CNN-LSTM-Attention hybrid model to output the electrocardiogram signal of the target object.

[0094] In the embodiment of the present application, the CNN-LSTM-Attention hybrid model includes the advantages of convolutional neural network (CNN), long short-term memory network (LSTM) and attention mechanism (Attention) to process data with complex spatial and temporal characteristics. This model is particularly suitable for sequence prediction, time series analysis, natural language processing and other tasks that require simultaneous consideration of spatial and temporal information.

[0095] Figure 2 A schematic diagram of the CNN-LSTM-Attention hybrid model provided in the embodiment of the present application is shown in FIG. Figure 2As shown in Figure 1, convolutional neural network (CNN) is a deep learning architecture that can automatically and efficiently learn spatial hierarchical features from image data by imitating biological visual perception mechanisms. It has strong spatial feature extraction capabilities and is particularly good at retaining and extracting spatial attributes from input data during training. They use two very important filters, called convolution and pooling kernels, to create multi-dimensional feature maps from data and analyze these feature maps from various angles to highlight their nonlinear characteristics.

[0096] The gating mechanism of the long short-term memory network (LSTM) enables it to perform a variety of complex time series data processing tasks. In the problem of locating the reference points of the cardiac oscillography signal (also known as the electrocardiogram or ECG), LSTM can independently learn and identify important time features and can independently locate the reference points of the cardiac oscillography signal without relying on the ECG signal as a reference signal. This means that LSTM can automatically identify key time points such as the start and end points of the heartbeat by learning the intrinsic pattern of the ECG signal.

[0097] In electrocardiogram (ECG) signal analysis, CNN can be used to extract spatial features in ECG signals, such as the shape, spacing, and size of the waveform. LSTM can process temporal information and capture the temporal relationship and dynamic changes between heartbeats. Therefore, the method of combining CNN and LSTM takes advantage of CNN's advantages in spatial feature extraction and LSTM's ability in time series analysis, providing a powerful tool for ECG signal analysis. Because CNN can extract useful spatial features from noisy data, and LSTM is responsible for mining the evolution of these features over time. Although it is possible to achieve the conversion of heartbeat signals to ECG signals using only CNN or LSTM, the use of CNN and LSTM makes the model robust to the diversity of heartbeat signals and can generate ECG signals more accurately.

[0098] In an embodiment of the present application, by inferring and analyzing the cardiac oscillography signal through the trained CNN-LSTM-Attention hybrid model, seven target reference points in the cardiac oscillography signal can be obtained, and then the electrocardiogram signal can be effectively reconstructed based on these seven target reference points.

[0099] In an embodiment of the present application, the position and speed of the target object are detected by sending and receiving signals by a millimeter-wave radar, and the vibration signal for reconstructing the electrocardiogram signal can be obtained without the attachment of conventional instruments, and a CNN-LSTM-Attention hybrid deep learning network for ECG signal reconstruction is designed. The CNN network has the ability to extract waveform spatial features; LSTM can independently learn and identify important time features, and independently realize the positioning of the reference point of the cardiac vibration tracing signal without relying on the electrocardiogram signal as a reference signal. The combination of the two realizes more accurate ECG signal reconstruction. In addition, the CNN-LSTM-Attention hybrid deep learning network is trained by a three-layer positioning algorithm suitable for cardiac vibration signal segmentation to achieve segmentation and positioning of cardiac vibration tracing signals from cardiac contraction and diastole to 7 precise feature points. Then the electrocardiogram signal is reconstructed. This method not only improves the processing speed, but also reduces the complexity and resource consumption of the model.

[0100] Optionally, before the step of inferring and analyzing the cardiac vibration tracing signal through the trained CNN-LSTM-Attention hybrid model and outputting the electrocardiogram signal of the target object, the method further includes:

[0101] Acquire a plurality of first cardiac vibration tracing sample information carrying systolic period marks and diastolic period marks;

[0102] For any of the first cardiac vibration tracing sample information, the first cardiac vibration tracing sample information is input into an initial CNN-LSTM-Attention hybrid model, and the first cardiac vibration tracing sample information is segmented to obtain systolic sample information and diastolic sample information.

[0103] Calculating a first loss value according to the systolic period mark, the systolic period sample information, the diastolic period mark and the diastolic period sample information;

[0104] If the first loss value is less than a first preset threshold, the first stage training of the initial CNN-LSTM-Attention hybrid model is completed to obtain a first CNN-LSTM-Attention hybrid model.

[0105] In the embodiment of the present application, during the first stage of training, only the segmentation tasks of the cardiac systole and diastole are focused on.

[0106] First, a plurality of first cardiac vibration tracing sample information is collected, and the sample information carries the marks of the systolic period and the diastolic period, and the marks are used to indicate the exact position of the systolic period and the diastolic period in each sample.

[0107] The labeled cardiac oscillography sample information is input into the initial CNN-LSTM-Attention hybrid model, and the model is used to segment the input cardiac oscillography sample information to distinguish between systolic sample information and diastolic sample information.

[0108] The loss value is calculated based on the known systolic and diastolic markers, as well as the systolic sample information and diastolic sample information obtained by model segmentation. The loss value is an indicator that measures the difference between the model prediction and the actual marker. If the calculated first loss value is less than the preset first threshold, it indicates that the performance of the model has reached an acceptable level.

[0109] After the loss value meets the conditions, the first stage training of the initial CNN-LSTM-Attention hybrid model is completed to obtain the trained first CNN-LSTM-Attention hybrid model. The cardiac vibration tracing signal is inferred and analyzed using the model trained in the first stage to output the electrocardiogram signal of the target object.

[0110] Optionally, after the step of obtaining the first CNN-LSTM-Attention hybrid model, the method further includes:

[0111] Acquire multiple systolic sample information and diastolic sample information carrying reference point peak and trough area labels;

[0112] For any of the systolic sample information or the diastolic sample information, input the systolic sample information or the diastolic sample information into the initial CNN-LSTM-Attention hybrid model to detect the peak and trough areas of the reference points in the systolic sample information or the diastolic sample information;

[0113] Calculating a second loss value according to the reference point peak and trough area and the reference point peak and trough area label;

[0114] If the second loss value is less than the second preset threshold, the second stage training of the first CNN-LSTM-Attention hybrid model is completed to obtain the second CNN-LSTM-Attention hybrid model.

[0115] During the second stage of training, the segmentation tasks of the systolic and diastolic periods have been completed, which allows the current stage to focus on identifying and locating the feature points in these two key periods, namely the peak of the systolic period and the trough of the diastolic period. Since the segmentation has been completed, unnecessary computational burden is avoided, and there is no need to process the entire ECG signal sequence again. Instead, the segmented systolic and diastolic time series segments can be directly input into the deep learning network. This approach not only improves the processing speed, but also reduces the complexity and resource consumption of the model.

[0116] Therefore, in the embodiment of the present application, the model training is performed by using the systolic sample information and the diastolic sample information carrying the reference point peak and trough area labels, and these sample information carry the reference point peak and trough area labels. These labels are used to indicate the exact location of the peak and trough in each sample.

[0117] The second loss value is calculated based on the reference point peak and valley regions detected by the model and the known reference point peak and valley region labels. The loss value is an indicator that measures the difference between the model prediction and the actual label. If the calculated second loss value is less than the preset second threshold, it indicates that the model has achieved an acceptable performance in detecting the reference point peak and valley regions.

[0118] After the loss value meets the conditions, the second stage training of the first stage CNN-LSTM-Attention hybrid model is completed to obtain a second CNN-LSTM-Attention hybrid model that has been further trained.

[0119] At this time, the second CNN-LSTM-Attention hybrid model can effectively identify the peak and trough areas of the benchmark points.

[0120] In an optional embodiment, after the second stage of training, since the peak (trough) area of ​​each reference point has been obtained, the corresponding reference point position can be obtained by finding the extreme value of the area. This step ensures that specific events in the cardiac cycle can be accurately identified while maintaining the time ratio of the original signal. Therefore, even without the third stage of training, the reference point position can be effectively identified and the electrocardiogram signal can be reconstructed.

[0121] Specifically, in order to further improve the accuracy of the model in locating signal feature points and simplify subsequent processing steps, length normalization technology is used. Specifically, the segmented systolic and diastolic time series are adjusted to a unified length standard. This preprocessing step is extremely important for model learning because it eliminates the impact of time series length changes caused by individual differences or changes in measurement conditions, allowing the network to focus more on the morphological characteristics of the signal rather than length changes. The normalized sequence will be input into the network for training and accurate location of feature points.

[0122] Once the model is inferred, these localized fiducials need to be remapped back to the length of the original signal. This step is critical because it ensures that specific events in the cardiac cycle can be accurately identified while maintaining the time scale of the original signal.

[0123] Since the peak (trough) areas of the respective reference points in the systolic and diastolic periods of the second stage have been obtained, the corresponding reference point positions can be obtained by finding the extreme values ​​of the areas.

[0124] In another optional embodiment, the third stage of training may be continued for the second CNN-LSTM-Attention hybrid model.

[0125] Optionally, after the step of obtaining the second CNN-LSTM-Attention hybrid model, the method further includes:

[0126] Acquiring a plurality of third cardiac vibration tracing sample information carrying target reference point labels;

[0127] For any of the third cardiac oscillography sample information, input the third cardiac oscillography sample information into an initial CNN-LSTM-Attention hybrid model to detect a target reference point of the third cardiac oscillography sample information;

[0128] Calculating a third loss value according to the target reference point and the target reference point label;

[0129] If the third loss value is less than the third preset threshold, the third stage training of the second CNN-LSTM-Attention hybrid model is completed to obtain a trained CNN-LSTM-Attention hybrid model.

[0130] In an embodiment of the present application, the third stage of the CNN-LSTM-Attention hybrid model training aims to further refine the performance of the model so that it can accurately detect specific target reference points in the electrocardiogram (ECG) signal.

[0131] A plurality of third cardiac vibration tracing sample information is collected, and the sample information carries labels of target reference points. The labels are used to indicate the exact positions of seven specific reference points in each sample.

[0132] The labeled third heart vibration tracing sample information is input into the CNN-LSTM-Attention hybrid model trained in the second stage. The model is used to analyze the input third heart vibration tracing sample information to detect the target reference point in the signal.

[0133] The third loss value is calculated based on the target reference points detected by the model and the known target reference point labels. The loss value is an indicator of the difference between the model prediction and the actual label. If the calculated third loss value is less than the preset third threshold, it indicates that the performance of the model in detecting the target reference points has reached an acceptable level. After the loss value meets the conditions, the third stage training of the second stage CNN-LSTM-Attention hybrid model is completed to obtain the final trained CNN-LSTM-Attention hybrid model.

[0134] In the embodiment of the present application, the third stage training of the model ensures that the model can accurately identify key electrocardiogram reference points from complex cardiac oscillation tracing signals, effectively ensuring the accurate construction of electrocardiogram signals.

[0135] The model is trained using a three-layer positioning algorithm. The first layer segments the systolic and diastolic periods, the second layer segments the peak (trough) area of ​​the reference points within the period, and the third layer accurately locates the seven reference points. Figure 3 A schematic diagram of segmenting a cardiac vibration recording signal provided in an embodiment of the present application is shown in FIG. Figure 3 As shown in the figure, the systolic period mainly includes four reference points: MC, IVC, AO, and RE, and the diastolic period mainly includes three reference points: AC, MO, and RF. After completing the segmentation tasks of the systolic and diastolic periods of the heart, we focus on identifying and locating the feature points in these two key periods, namely the peak of the systolic period and the trough of the diastolic period. Since the segmentation has been completed, unnecessary computational burden can be avoided, and the entire ECG signal sequence does not need to be processed again. Therefore, the segmented systolic and diastolic time series fragments can be directly input into the deep learning network, which improves the processing speed and reduces the complexity and resource consumption of the model.

[0136] After the first and second layers of positioning are completed, we use length normalization technology to further improve the accuracy of the model in locating signal feature points and simplify subsequent processing steps. Specifically, we adjust the segmented systolic and diastolic time series to a unified length standard. This step can eliminate the impact of time series length changes caused by individual differences or changes in measurement conditions, allowing the network to focus more on the morphological characteristics of the signal rather than length changes. The normalized sequence will be input into the third layer for training and accurate positioning of feature points.

[0137] After the model inference is completed, these located reference points need to be remapped back to the length of the original signal. Since the peak and trough areas of the respective reference points have been obtained, the corresponding reference point positions can be obtained by finding the extreme value of the area. This step ensures that we can accurately identify specific events in the cardiac cycle while maintaining the time ratio of the original signal.

[0138] The three-layer positioning algorithm achieves the effect of extracting 7 reference points without relying on the electrocardiogram signal as a reference signal. Compared with the existing positioning method that relies on electrocardiogram information, it realizes independent detection of reference points, and the number of detected reference points is greater, and the position of key points is more accurate. The following is a description of the non-contact electrocardiogram signal acquisition device provided by the present invention. The non-contact electrocardiogram signal acquisition device described below and the non-contact electrocardiogram signal acquisition method described above can be referred to each other.

[0139] Figure 4 A schematic diagram of the structure of a contactless electrocardiogram signal acquisition device provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, including:

[0140] The receiving module 410 is used to receive the millimeter wave radar echo signal reflected by the target object, and perform static clutter reduction processing on the millimeter wave radar echo signal to obtain the vibration signal of the target object;

[0141] The extraction module 420 is used to extract the cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method after filtering the vibration signal for noise;

[0142] The output module 430 is used to perform inference analysis on the cardiac vibration tracing signal through the trained CNN-LSTM-Attention hybrid model, and output the electrocardiogram signal of the target object.

[0143] In an embodiment of the present application, the position and speed of the target object are detected by transmitting and receiving signals by a millimeter-wave radar, and the vibration signal for reconstructing the electrocardiogram signal can be obtained without the attachment of conventional instruments, and a CNN-LSTM-Attention hybrid deep learning network for ECG signal reconstruction is designed. The CNN network has the ability to extract waveform spatial features; the LSTM can independently learn and identify important time features, and independently locate the reference point of the cardiac vibration tracing signal without relying on the electrocardiogram signal as a reference signal. The combination of the two achieves more accurate ECG signal reconstruction. In addition, the CNN-LSTM-Attention hybrid deep learning network is trained by a three-layer positioning algorithm suitable for cardiac vibration signal segmentation to achieve segmentation and positioning of cardiac vibration tracing signals from cardiac contraction and diastole to 7 precise feature points. This method not only improves the processing speed, but also reduces the complexity of the model and resource consumption.

[0144] Figure 5 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 5 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communication interface 520 and the memory 530 communicate with each other through the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute a method for acquiring a contactless electrocardiogram signal, the method comprising: receiving a millimeter wave radar echo signal reflected by a target object, and performing static clutter attenuation processing on the millimeter wave radar echo signal to obtain a vibration signal of the target object;

[0145] After filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method;

[0146] The cardiac vibration tracing signal is inferred and analyzed by the trained CNN-LSTM-Attention hybrid model, and the electrocardiogram signal of the target object is output. In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0147] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the non-contact electrocardiogram signal acquisition method provided by the above methods, the method comprising: receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object;

[0148] After filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method;

[0149] The cardiac vibration tracing signal is inferred and analyzed through the trained CNN-LSTM-Attention hybrid model to output the electrocardiogram signal of the target object.

[0150] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the non-contact electrocardiogram signal acquisition method provided by the above methods, the method comprising: receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object;

[0151] After filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method;

[0152] The cardiac vibration tracing signal is inferred and analyzed through the trained CNN-LSTM-Attention hybrid model to output the electrocardiogram signal of the target object.

[0153] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0154] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for acquiring a contactless electrocardiogram signal, characterized in that: include: receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object; After filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method; Performing inference analysis on the cardiac vibration tracing signal through a trained CNN-LSTM-Attention hybrid model, and outputting an electrocardiogram signal of the target object; Wherein, before the step of inferring and analyzing the cardiac vibration tracing signal through the trained CNN-LSTM-Attention hybrid model and outputting the electrocardiogram signal of the target object, the method further includes: Acquire a plurality of first cardiac vibration tracing sample information carrying systolic period marks and diastolic period marks; For any of the first cardiac vibration tracing sample information, the first cardiac vibration tracing sample information is input into an initial CNN-LSTM-Attention hybrid model, and the first cardiac vibration tracing sample information is segmented to obtain systolic sample information and diastolic sample information. Calculating a first loss value according to the systolic period mark, the systolic period sample information, the diastolic period mark and the diastolic period sample information; If the first loss value is less than a first preset threshold, the first stage training of the initial CNN-LSTM-Attention hybrid model is completed to obtain a first CNN-LSTM-Attention hybrid model; After obtaining the first CNN-LSTM-Attention hybrid model, the following steps are also included: Acquire multiple systolic sample information and diastolic sample information carrying reference point peak and trough area labels; For any of the systolic sample information or the diastolic sample information, input the systolic sample information or the diastolic sample information into the initial CNN-LSTM-Attention hybrid model to detect the peak and trough areas of the reference points in the systolic sample information or the diastolic sample information; Calculating a second loss value according to the reference point peak and trough area and the reference point peak and trough area label; If the second loss value is less than a second preset threshold, the second stage training of the first CNN-LSTM-Attention hybrid model is completed to obtain a second CNN-LSTM-Attention hybrid model; After obtaining the second CNN-LSTM-Attention hybrid model, the following steps are also included: Acquiring a plurality of third cardiac vibration tracing sample information carrying target reference point labels; For any of the third cardiac oscillography sample information, input the third cardiac oscillography sample information into an initial CNN-LSTM-Attention hybrid model to detect a target reference point of the third cardiac oscillography sample information; Calculating a third loss value according to the target reference point and the target reference point label; If the third loss value is less than the third preset threshold, the third stage training of the second CNN-LSTM-Attention hybrid model is completed to obtain a trained CNN-LSTM-Attention hybrid model.

2. The method for acquiring a contactless electrocardiogram signal according to claim 1, characterized in that: The receiving of the millimeter wave radar echo signal reflected by the target object and performing static clutter reduction processing on the millimeter wave radar echo signal to obtain the vibration signal of the target object includes: Obtain the median of the static background signal of the F frame in the millimeter wave radar detection area when there is no target object, and obtain the median background signal; Subtracting the median background signal from the millimeter-wave radar echo signal reflected by the target object to implement static clutter attenuation processing and obtain a vibration signal after static clutter attenuation processing; The vibration signal after the static clutter attenuation processing is subjected to noise filtering based on a target bandpass filter to obtain the vibration signal of the target object.

3. The method for acquiring a contactless electrocardiogram signal according to claim 1, characterized in that: After filtering the vibration signal for noise removal, extracting the cardiac vibration recording signal from the vibration signal by singular spectrum decomposition method, including: The vibration signal is embedded into a plurality of time series signals, and each of the time series signals is shifted and superimposed on the time axis to form a trajectory matrix; Performing singular value analysis on the trajectory matrix to obtain a plurality of singular value components and corresponding left singular vectors and right singular vectors, and constructing a component matrix based on the left singular vectors and the right singular vectors; wherein each of the component matrices corresponds to a singular value and a frequency pattern; Perform diagonal averaging on each component matrix to obtain a time series of each component matrix, and then perform fast Fourier transform on each time series to obtain the frequency characteristics of each component matrix; Obtaining a target component matrix that matches the frequency characteristic range of the cardiac vibration tracing signal from each of the component matrices, and linearly combining the target component matrices to obtain a cardiac vibration-related time series signal; The diagonal averaging algorithm is used to filter and remove noise from the cardiac vibration related time series signal to obtain the cardiac vibration tracing signal.

4. The method for acquiring a contactless electrocardiogram signal according to claim 1, characterized in that: The step of inferring and analyzing the cardiac vibration tracing signal through the trained CNN-LSTM-Attention hybrid model and outputting the electrocardiogram signal of the target object includes: Inputting the cardiac oscillography signal into the CNN network in the CNN-LSTM-Attention hybrid model to extract the spatial features of the cardiac oscillography signal; Inputting the spatial features into the LSTM network in the CNN-LSTM-Attention hybrid model to capture the temporal relationship and dynamic changes between heartbeats and extract seven heartbeat reference points in the cardiac vibration tracing signal; Based on the seven heartbeat reference points, the electrocardiogram signal of the target object is reconstructed.

5. A contactless electrocardiogram signal acquisition device, characterized in that: include: A receiving module, used for receiving a millimeter-wave radar echo signal reflected by a target object, and performing static clutter reduction processing on the millimeter-wave radar echo signal to obtain a vibration signal of the target object; An extraction module, configured to extract a cardiac vibration recording signal from the vibration signal by using a singular spectrum decomposition method after filtering out noise from the vibration signal; An output module, used to perform inference analysis on the cardiac vibration tracing signal through a trained CNN-LSTM-Attention hybrid model, and output an electrocardiogram signal of the target object; Wherein, the device is also used for: Acquire a plurality of first cardiac vibration tracing sample information carrying systolic period marks and diastolic period marks; For any of the first cardiac vibration tracing sample information, the first cardiac vibration tracing sample information is input into an initial CNN-LSTM-Attention hybrid model, and the first cardiac vibration tracing sample information is segmented to obtain systolic sample information and diastolic sample information. Calculating a first loss value according to the systolic period mark, the systolic period sample information, the diastolic period mark and the diastolic period sample information; If the first loss value is less than a first preset threshold, the first stage training of the initial CNN-LSTM-Attention hybrid model is completed to obtain a first CNN-LSTM-Attention hybrid model; Acquire multiple systolic sample information and diastolic sample information carrying reference point peak and trough area labels; For any of the systolic sample information or the diastolic sample information, input the systolic sample information or the diastolic sample information into the initial CNN-LSTM-Attention hybrid model to detect the peak and trough areas of the reference points in the systolic sample information or the diastolic sample information; Calculating a second loss value according to the reference point peak and trough area and the reference point peak and trough area label; If the second loss value is less than a second preset threshold, the second stage training of the first CNN-LSTM-Attention hybrid model is completed to obtain a second CNN-LSTM-Attention hybrid model; Acquiring a plurality of third cardiac vibration tracing sample information carrying target reference point labels; For any of the third cardiac oscillography sample information, input the third cardiac oscillography sample information into an initial CNN-LSTM-Attention hybrid model to detect a target reference point of the third cardiac oscillography sample information; Calculating a third loss value according to the target reference point and the target reference point label; If the third loss value is less than the third preset threshold, the third stage training of the second CNN-LSTM-Attention hybrid model is completed to obtain a trained CNN-LSTM-Attention hybrid model.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method for acquiring a contactless electrocardiogram signal as described in any one of claims 1 to 4 is implemented.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for acquiring a contactless electrocardiogram signal as claimed in any one of claims 1 to 4 is implemented.