Non-contact heartbeat monitoring method and system based on millimeter wave signal

Abstract

The invention belongs to the technical field of remote health monitoring, medical diagnosis and intelligent health equipment, and discloses a non-contact heartbeat monitoring method and system based on millimeter wave signals. According to the non-contact heartbeat monitoring system, heartbeat monitoring is achieved by transmitting and receiving millimeter wave signals reflected from the chest of the human body through the millimeter wave radar technology. Tiny movement generated by cardiac pulsation can cause tiny changes of the chest of the human body, and the changes are shown through reflected millimeter wave signals. By processing and analyzing the reflected signals in real time, the influence of heartbeat on the signals can be accurately extracted, so that the heartbeat frequency and physiological indexes related to the electrocardiogram can be deduced.

Description

Technical Field

[0001] The present invention relates to the technical fields of remote health monitoring, medical diagnosis, and intelligent health devices, and particularly to a non-contact heart rate monitoring method and system based on millimeter-wave signals. Background Art

[0002] With the rapid development of medical technology and the increasing demand for health management, non-contact heart rate monitoring systems have gradually become an important research direction in the modern medical field. Traditional heart rate monitoring methods, such as electrocardiogram (ECG) measurement, usually rely on devices that need to be in direct contact with the patient's skin (such as electrodes, chest patches, etc.). This contact-based monitoring is not only complex to operate but may also cause discomfort to the patient. Especially during an ECG test, the patient needs to keep their upper body bare and lie still, which limits the flexibility of monitoring and may cause inconvenience and psychological burden.

[0003] Efficient heart rate monitoring using millimeter-wave radar technology can solve these problems. Due to its good penetration and high resolution, millimeter-wave signals can effectively sense the minute changes in the human body, especially the minute movements related to the heartbeat. Specifically, the system emits millimeter-wave signals and receives the signals reflected from the human chest. The minute chest movements caused by the heartbeat will lead to subtle changes in the reflected signals. These changes reflect the rhythm and frequency of the heartbeat, and by precisely analyzing and processing these reflected signals, physiological information related to the heartbeat can be inferred.

[0004] Compared with the traditional electrocardiogram measurement method, this non-contact monitoring system does not require any physical contact during the monitoring process and can provide real-time, comfortable, and efficient heart rate monitoring without disturbing the patient's daily activities. The advantages of this technology are not only reflected in the simplicity and comfort of operation but also in its ability to perform long-term and continuous health monitoring, especially suitable for groups that require long-term heart health monitoring, such as the elderly, chronic disease patients, and high-risk populations. In addition, the non-contact feature of this system provides the feasibility for remote health monitoring, supports the integration of intelligent health devices and remote medical platforms, and further improves the intelligence and accuracy of personalized medical services. Summary of the Invention

[0005] With the development and iteration of medical technology, non-contact electrocardiogram monitoring has become a new focus. Compared with previous contact-based electrocardiogram monitoring and wireless sensing technologies, the main technical problems to be solved in the implementation of this system are as follows.

[0006] The first technical problem is the wireless capture of tiny electrocardiogram signals. During the human heartbeat activity, there are various changes in electrical signals. These changes in electrical signals will lead to myocardial contraction and relaxation. Further, this contraction and relaxation action is transmitted to the human chest cavity, causing fluctuations. This tiny mechanical movement is extremely vulnerable to various factors, including respiratory movement, irregular human movement, etc. It becomes very difficult to successfully capture the displacement changes of the human chest cavity using wireless signals and further isolate the more subtle heartbeat activity from them.

[0007] The second technical problem is the accuracy and comprehensiveness of the results of electrocardiogram monitoring. Traditional electrocardiogram monitoring is based on the 12-lead data of the human body from an electrocardiograph, so as to analyze cardiac activities from multiple angles. However, the non-contact method cannot directly capture the changes in human electrical signals and can only deduce the results by analyzing the influence of cardiac physical and mechanical activities on signal characteristics. Therefore, how to achieve the feature extraction of wireless signals and electrocardiogram feature analysis is challenging.

[0008] To solve the first problem, we have carefully designed a series of signal processing methods. We select specific frequency signals to ensure fine-grained motion capture. Based on an array antenna, signal focusing and directivity are achieved, thus avoiding multipath effects from the environment. Based on the special regularity of heartbeat activity, the separation of heartbeat signals from interference signals such as respiration is realized. Based on filtering technology, the extraction of periodic heartbeat signals is achieved.

[0009] To solve the second problem, we have designed a specific machine learning model to enhance the learning of the physical signs of wireless signals, thereby constructing an electrocardiogram and inferring the corresponding electrocardiogram indicators. First, a method for segmenting periodic heartbeat segments is designed. Segmenting by a single heartbeat cycle is convenient for fully learning the activity details of each heartbeat and the signal differences between different heartbeats. Finally, combined with medical theory, a model is designed to realize wavelet recognition and indicator inference for the constructed electrocardiogram.

[0010] The technical solution of the present invention is as follows: A non-contact heartbeat monitoring method based on millimeter-wave signals, the steps are as follows:

[0011] Step 1: Heartbeat signal acquisition;

[0012] The heartbeat signal acquisition device consists of a millimeter-wave radar transmitting board and a data acquisition board. The two are connected by a data transmission interface and fixed together; during the acquisition process, the target to be measured maintains a sitting or lying position and remains relatively stationary. The heartbeat signal acquisition device is placed in front of the human chest cavity, and the distance is set between 30 cm and 100 cm. The millimeter-wave radar antenna is aligned with the position of the heart; the millimeter-wave radar transmitting board emits millimeter-wave signals. After the receiving antenna of the millimeter-wave radar transmitting board receives the human echo data, it is transmitted to the data acquisition board for post-data processing;

[0013] Step 2: Signal separation and processing;

[0014] Perform target location positioning, multipath clutter suppression and heartbeat signal separation in sequence to separate clear heartbeat activity information from the captured millimeter wave signal;

[0015] Step 3: Construct an electrocardiogram based on clear heartbeat activity information;

[0016] Step 4: Perform wavelet identification and index inference based on the constructed electrocardiogram.

[0017] Furthermore, the target position positioning is specifically:

[0018] The millimeter wave radar transmitting board works in a frequency modulated continuous wave (FMCW) mode; the frequency modulated continuous wave (FMCW) working mode is that the millimeter wave signal has a fixed starting frequency f c In the subsequent millimeter wave signal transmission process, the frequency of the transmitted millimeter wave signal increases with a certain slope k, that is, the transmission signal frequency at the tth moment is f t =f c +kt; therefore, the signals at different times have their own frequencies; the millimeter wave signal in a transmission cycle is converted into a frequency domain signal by calculating FFT to obtain the frequency of the millimeter wave reflected signal at each moment; based on the FMCW working mode, it is inferred that the signal is transformed from the starting frequency to a certain frequency f t The time t is obtained, and the distance of the target to be measured is d=t×c, where c represents the speed of light;

[0019] There are multiple peaks in the frequency domain signal, each of which represents the reflection signal of a reflector in the scene. The human body is the nearest strong reflector within the line of sight of the device. The highest peak in the frequency domain signal is identified as the distance bin where the human body is located, and the human body reflection signal at this location is processed;

[0020]

[0021] Formula (1) is the phase calculation formula of the millimeter wave signal; d represents the propagation distance of the millimeter wave signal, and λ represents the wavelength of the millimeter wave signal. During the acquisition of human chest cavity signals, human breathing and heartbeat movements will cause the chest cavity to expand and contract. The chest cavity fluctuations affect the propagation distance d of the millimeter wave signal, thereby affecting the phase of the reflected signal.

[0022] Furthermore, the multipath clutter suppression is specifically:

[0023] According to the physical arrangement of the multi-antenna receiving array, the weight formula for beamforming phase adjustment is constructed:

[0024]

[0025] Among them, n represents the number of antennas, i represents the i-th millimeter-wave radar antenna, θ is the angle where the target is located, and l is the antenna spacing; traverse θ = (-π, π) to find the signal with the strongest reflection within the range of 180° directly in front of the millimeter-wave radar transmitting board; obtain the angle where the reflection is the strongest according to the signal peak; perform subsequent processing on the reflection signal at this angle.

[0026] Further, the separation of the heartbeat signal is specifically as follows: perform phase unwrapping and heartbeat signal separation on the echo signal at the currently determined distance and angle.

[0027] Phase unwrapping: The process of phase unwrapping is to ensure the continuity of the signal by gradually adjusting the phase difference; first, calculate the phase difference between the current time point and the previous time point; if the phase difference is greater than π, it means that the phase has jumped, and the phase at the current moment needs to be subtracted by a period of 2π; if the phase difference is less than -π, it indicates that the direction of the phase jump is reverse, and at this time, a period of 2π needs to be added to the phase at the current moment.

[0028] Heartbeat signal separation: Obtain the waveform with periodic changes from the result of phase unwrapping; the undulation of the waveform corresponds to the undulation of the chest cavity during human breathing, and perform the separation of the heartbeat signal.

[0029] The acceleration of the millimeter-wave signal sequence after phase unwrapping is calculated by the numerical method to highlight the characteristics of the heartbeat:

[0030]

[0031] Among them, S 0 is the value of a certain point in the sequence, S -i and S i represent the values of the i-th point before and the i-th point after the current point respectively, and h represents the spatial step, that is, the distance between adjacent points; after performing a low-pass filter on the acceleration result to eliminate irrelevant high-frequency noise, calculate the envelope signal of the filtered signal; convert the envelope signal to the frequency domain, and find the frequency corresponding to the heartbeat through spectrum analysis; after determining the frequency corresponding to the heartbeat, use the method of frequency truncation to extract the signal within the frequency corresponding to the heartbeat to obtain the heartbeat signal sequence.

[0032] Further, the specific content of step three is as follows: perform periodic heartbeat segmentation, cut the clear heartbeat activity information for a period of time to be processed according to individual heartbeat cycles and then construct an electrocardiogram.

[0033] The periodic heartbeat segmentation is performed using a template matching algorithm for segmentation:

[0034] First, perform preliminary segmentation on the heartbeat signal sequence obtained by frequency truncation through the heartbeat frequency; according to the clear heartbeat activity information and device parameters, calculate the number of sampling points corresponding to one heartbeat cycle, and use this number of sampling points as the initial size of a sliding window; use the sliding window to slide on the heartbeat signal sequence, record each selected segmentation point and move the sliding window to continue searching; the selection of the segmentation point is as follows: through theoretical analysis, the highest point of the processed heartbeat signal sequence corresponds to the R wave in the electrocardiogram, so the peak value in the heartbeat signal sequence is used as the segmentation point, and the RR interval is used to represent one heartbeat cycle;

[0035] Next, perform template matching segmentation; first, construct a template. After adjusting the initial segmentation segments obtained in the previous step to the same length, calculate the mean value to obtain the initial template. Template matching will be iterated multiple rounds, and a new template will be recalculated according to the segmentation segments after each round of matching; during the template matching process, use a sliding window with variable start and end points to move backward; starting from the start point, the start point of the sliding window is fixed, and the end point varies within an interval; during the change of the end point, calculate the similarity between the current window and the new template obtained in the previous round each time, and select the end point of the window with the highest similarity as the segmentation point, which is also the start point of the next window; after traversing from the start to the end of the heartbeat signal sequence with such a segmentation point selection method, it is considered the end of one iteration to obtain the latest segmentation segment group according to template matching; calculate a new template in the same way based on the latest segmentation segment group, and then repeat the above process iteratively. After setting the number of iteration rounds, obtain the most accurate segmentation result;

[0036] Based on the segmentation results, use machine learning methods to learn the potential relationship between heartbeat activity and thoracic reflection millimeter-wave signals; before training the network model, perform data preprocessing. Recombine the previous heartbeat signal segmentation segments as training data, and recombine the training data at the head and tail in different proportions; transform the recombined data to the same size through interpolation or decimation, and then perform normalization; use the simultaneously collected real electrocardiogram as the real label to train the corresponding radio frequency signal.

[0037] Furthermore, the network model is a hybrid neural network that combines a Convolutional Neural Network (CNN) and a Long Short-Term Memory Network (LSTM). A single-channel sequence data with a length of at least 300 is input and passed into the network model through the sequenceInputLayer. First, the input data goes through a 1D convolutional layer, using 3 convolutional kernels to output 16 feature maps, and is processed through a batch normalization layer and a ReLU activation function. Next, the data is reduced in dimension through a max pooling layer and then enters the second 1D convolutional layer, using 5 convolutional kernels to output 32 feature maps, and is processed again through a batch normalization layer and a ReLU activation function. The pooling operation continues to reduce the time dimension of the data and further extract features. Subsequently, the model uses a dropout layer with a 25% dropout rate to reduce the risk of overfitting. Next, the LSTM layer is used to model the temporal information. The LSTM layer has 200 units and can capture the temporal dependencies in the input sequence. Then, the network uses two transposed convolutional (deconvolution) layers for upsampling to restore the time dimension and generate higher-resolution feature maps. Next, the data passes through two fully connected layers. The first fully connected layer outputs 8 neurons, and the second fully connected layer outputs a single predicted value as the result of the regression task. Finally, the output layer uses a regression layer for optimization and outputs a scalar value. The data stream starts from the input sequence, goes through the processing of convolutional, pooling, LSTM, deconvolutional, and fully connected layers, and finally outputs a continuous value for regression. The training of the network is carried out through the Adam optimizer with a learning rate of 0.001, a batch size of 32, and is trained for at most 100 epochs, and is trained on the GPU to improve the computational efficiency.

[0038] Furthermore, the weighted cosine similarity is used to calculate the similarity. Based on the calculation method of cosine similarity, a higher weight is assigned to the electrocardiogram wavelet position corresponding to the current segmentation segment.

[0039] Furthermore, a classification network is used for wavelet identification; the electrocardiogram judgment criteria include the RR interval duration to judge whether the patient's heart rate is too slow, normal, or too fast, the QRS waveform hypertrophy, the T wave inversion, and the P wave inversion, which are used to judge different cardiac indicators; the features of the QRS complex, T wave, and P wave in the electrocardiogram are identified for the derivation of various cardiac indicators;

[0040] The label construction method of the electrocardiogram is as follows:

[0041] The generated electrocardiogram is one-dimensional data, and different numerical representations are set for different sub-waves; construct a one-dimensional data of equal length, uniformly assign the corresponding numerical representation to the index interval where the corresponding sub-wave is located, and assign 0 to the remaining non-sub-wave parts; during the training process, use the label array to supervise the training of electrocardiogram data to complete classification; finally, transform the corresponding numerical values into the corresponding sub-waves in the generated results to infer the index intervals occupied by different indicators and infer the sub-wave duration.

[0042] Derivation of cardiac indicators: Transformed into code logic according to electrocardiogram knowledge; restore the electrocardiogram results constructed by the classification network to the original size; splice together the electrocardiogram signals corresponding to each heartbeat and recalculate various indicators, including: P-wave duration, T-wave duration, QRS-wave duration, PR interval, QT interval; judge different physiological index situations according to medical standards.

[0043] A non-contact heartbeat monitoring system for a non-contact heartbeat monitoring method based on millimeter-wave signals, including:

[0044] A heartbeat signal acquisition module for acquiring heartbeat signals;

[0045] A signal separation and processing module that sequentially performs target position positioning, multipath clutter suppression, and heartbeat signal separation to separate clear heartbeat activity information from the captured millimeter-wave reflection signals;

[0046] An electrocardiogram construction module for performing periodic heartbeat segmentation and electrocardiogram construction;

[0047] A sub-wave identification and index inference module for identifying corresponding sub-waves and judging physiological index situations.

[0048] The beneficial effects of the present invention:

[0049] 1. Adopt non-contact electrocardiogram detection, which does not require complex equipment installation and operation, greatly improving the efficiency of the electrocardiogram detection process. It does not make any contact with the human skin, avoiding problems such as allergies and cross-infections, and is safe and healthy.

[0050] 2. The system can achieve rapid detection in a short time and display the real-time results of synchronous cardiac indicators. The index classification is comprehensive and diverse, and the results are accurate.

[0051] 3. Skillfully capture the micro-movement changes caused by the heartbeat based on the short wavelength characteristics of millimeter waves. Utilize the small phase shift caused by the movement to design a signal processing method to make the heartbeat the dominant signal and eliminate the influence of noise, and extract the clean heartbeat corresponding phase information from the complex environment echo signals.

[0052] 4. Combine medical theory knowledge with machine learning, identify the sub-wave characteristics in the electrocardiogram through machine learning and calculate the corresponding heartbeat indicators to achieve the derivation of cardiac indicators.

[0053] 5. Design a network learning model, use traditional professional electrocardiograms for supervision, and construct an electrocardiogram based on the regular phase changes of the signals caused by heartbeat activities.

[0054] 6. The device is small and compact, facilitating deployment and installation. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the human body target positioning result of the present invention.

[0056] Figure 2 It is a schematic diagram of the beam after beamforming of the antenna array of the present invention.

[0057] Figure 3 It is a schematic diagram for comparing signal phase unwrapping of the present invention. (a) is before phase unwrapping, and (b) is after phase unwrapping.

[0058] Figure 4 It is a schematic diagram of the heartbeat signal separation result of the present invention.

[0059] Figure 5 It is a schematic diagram of the periodic heartbeat segmentation result of the present invention.

[0060] Figure 6 It is a schematic diagram of the data preprocessing result of the present invention. (a) The R wave is at the 1 / 3 position, (b) the R wave is at the 2 / 3 position, and (c) the R wave is at both ends.

[0061] Figure 7 It is a schematic diagram of constructing an electrocardiogram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0062] The technical solution of the present invention will be further described below in conjunction with the specific embodiments and the accompanying drawings of the specification.

[0063] With the development of wireless signal sensing technology, millimeter-wave radar technology has gradually become an important means of non-contact physiological monitoring due to its excellent penetration, resolution, and high precision. Existing electrocardiogram monitoring methods usually rely on contact devices, such as electrode patches or electrocardiogram suction balls, which not only cause certain discomfort to patients but also limit the flexibility and convenience of monitoring. The non-contact heartbeat monitoring system proposed by the present invention uses millimeter-wave radar technology to monitor the heartbeat by transmitting and receiving millimeter-wave signals reflected from the human chest. The minute movements generated by the heartbeat will cause minute changes in the human chest, which are manifested through the reflected millimeter-wave signals. By real-time processing and analysis of the reflected signals, the influence of the heartbeat on the signals can be accurately extracted, thereby inferring the heartbeat frequency and physiological indicators related to the electrocardiogram.

[0064] We conducted a series of field tests using radar equipment to verify the feasibility of the system. The present invention uses a combination of an IWR1642 millimeter-wave radar board and a DWM1000 data acquisition board.

[0065] Step 1: Heartbeat signal acquisition:

[0066] The equipment placement and human body posture during the heartbeat signal acquisition process. The equipment is placed at a position 30 cm to 100 cm away from the human body, and the equipment is directly facing the human chest position. The person can take a sitting or lying posture in front of the equipment and remain relatively stationary during the acquisition process.

[0067] The acquisition device hardware is developed based on a commercial IWR1642 millimeter-wave radar development board combined with a DWM1000 data acquisition board. For the non-contact heartbeat monitoring scenario implemented by the millimeter-wave radar, the high-frequency radar signal can generate an extremely short millimeter wavelength, which is also the basis for the radar to perceive millimeter-level heartbeat displacement.

[0068] After the device is started, it will send and receive signals at a specific frequency, and the collected human reflection signals will be processed step by step by the system.

[0069]

[0070] Formula (1) is the calculation formula for the phase of the radio frequency signal, where d represents the signal propagation distance and λ represents the signal wavelength. Assuming that the millimeter wavelength is 4 mm and the chest displacement generated by the heartbeat is 1 mm, the influence of the heartbeat movement on the phase of the reflected signal is Therefore, the heartbeat change can be deduced by the phase change. However, for a larger displacement movement, such as body shaking, assuming the shaking displacement is 1 cm, the influence of the body shaking on the phase of the reflected signal is It is easy to generate phase wrapping, and the number of phase wrapping generated by irregular movement is unknown. At this time, the influence generated by the heartbeat signal will be submerged and difficult to be extracted. Therefore, during the acquisition process, the target to be measured remains sitting or lying and remains relatively stationary, and the radar antenna is aligned with the position of the heart, which will further improve the signal acquisition quality.

[0071] Step 2: Signal separation processing;

[0072] The signal separation processing is mainly to separate clear heartbeat activity information from the captured signals, which involves target position positioning, multipath clutter suppression, and heartbeat signal separation;

[0073] (2.1) Target position localization: The radar device emits signals forward in a fan-shaped manner with specified parameters. Reflective objects in the scene will change the signal propagation direction, and finally, a part of the signals will reach the receiving antenna through different reflection paths with different time delays. During the signal acquisition process, the reflected signals of the environment are received at all times. Locating the reflected signals of the body part of our interest from them is the key to the system.

[0074] The parameter settings are as follows:

[0075] Data channel: LVDS, ADC bit number: 16, ADC sampling number: 256, starting frequency: 77.0 GHz, frequency slope: 59.963 MHz / us, idle time: 250.0 us, sampling rate: 5000 ksps, number of transmitting antennas: 2, number of receiving antennas: 4, number of transmitted frames: 2999, frame period: 10.02 ms, number of chirps per frame: 32.

[0076] The millimeter-wave radar operates in the frequency-modulated continuous-wave (FMCW) mode, which also provides us with a way to separate the reflected signals of different targets by frequency. The FMCW operating mode is that there is a fixed starting frequency f c . During the subsequent signal transmission process, the frequency of the transmitted signal will increase at a certain slope k, that is, the transmitted signal frequency at the t-th moment is f t = f c + kt. Therefore, the signals at different moments have their own individual frequencies. When we convert the signals within a transmission period to the frequency domain by calculating the FFT, we will obtain the frequencies of the reflected signals at each moment. Based on the FMCW operating mode, we can infer the time t when the signal changes from the starting frequency to a certain frequency f t . Therefore, we can infer the distance d of the target = t × c.

[0077] Finally, a part of the signals will reach the receiving antenna through different reflection paths with different time delays. The intensity of the reflected signal is affected by the reflection coefficient of the target on the one hand. The larger the reflection coefficient of the target, the stronger the reflected signal intensity. On the other hand, it is affected by the propagation distance. According to the Friis free-space path loss model, the received signal intensity is inversely proportional to the signal propagation distance. Therefore, in the calculated frequency-domain signal, the clearly visible peaks correspond to the reflected signals of strong reflectors at a certain distance (bin). Referring to Figure 1 , the earliest and highest peak that appears in the calculation result can be identified as the bin where the human body is located. Subsequently, the human body reflected signals at this bin will be processed.

[0078] (2.2) Multipath clutter suppression: Millimeter-wave signals are greatly affected by multipath. As described above, the signal is emitted forward in a fan shape and receives echo signals that have undergone one or more reflections from the environment. Among them, there will be two or more signals from different propagation paths arriving at the receiving end at the same time, so that the signal received at the current moment is the result of the linear superposition of multiple different signals. Therefore, we need to solve the clutter influence of multipath first.

[0079] With the help of the multi-antenna receiving array of the device. When the target distance is relatively large compared to the antenna spacing in the antenna array, we can regard the reflected signal from the target as arriving at the array antennas in parallel. If there is a certain angle between the direction of the target and the positive direction of the antenna array, the parallel signals will have different degrees of delay when arriving at different receiving antennas, and this delay is manifested as a phase difference.

[0080] If we can control the phase of the signal received by each receiving antenna, make up this phase difference and adjust it to the same level, then we can consider that the reflected signals arrive at each antenna simultaneously. At this time, when we superimpose the antenna array, we will get the strongest reflected signal from the target (without peak shift caused by phase offset, so that the superposition of all peaks cannot be obtained). So far, we can construct the weight formula for beamforming phase adjustment according to the physical arrangement of the array:

[0081]

[0082] Where n represents the number of antennas, i represents the i-th antenna, θ is the angle where the target is located, and l is the antenna spacing. By traversing θ = (-π, π), we can reflect the strongest signal within the range of 180° in the front direction. For example, assume that there is a fixed reflector only in the direction 30° to the left front in the environment. Taking the first antenna in the antenna array as a reference (the antenna array is Rx1, Rx2, Rx3, Rx4 from left to right), the strongest reflected signal is obtained at the receiving end when the phase is π. Then for the remaining receiving antennas, the strongest reflected signals may fall at phases π + 0.5, π + 1, π + 1.5, and this phase difference is jointly determined by the antenna spacing and the target angle. Next, scan within the range of (-π, π). If the value of θ is not the true angle of the target, then substitute the current θ into the weight formula to calculate the adjustment amount of different antennas. Due to inaccurate angles, for each antenna, the current phase position is not the point where the reflected signal is the strongest. Therefore, the result obtained by superimposing the array at this time is not the maximum. Only when the value of θ is When the phases of the four receiving antennas after adjustment all correspond to the highest peak values, the superimposed result is the largest at this time. Therefore, by searching for the peak within the range of (-π, π), the angle where the reflection is the strongest can be obtained. Therefore, in the following, we only need to focus on the reflection signal at the current angle to suppress the echoes in other directions because we are not concerned about them.

[0083] We select the first antenna as the reference antenna, calculate the phase differences of the remaining receiving antennas compared to the reference antenna according to the beamforming formula, and after compensating for this phase difference, adjust them to the same level. In this way, it can be considered that the reflection signals arrive at each antenna simultaneously. At this time, when we superimpose the antenna array, we will obtain the strongest reflection signal from the target. So far, we can traverse θ = (-π, π) to find the signal with the strongest reflection within the 180° range directly in front. Refer to Figure 2 , which shows that after beamforming, the signal has a stronger beam directivity pointing in the direction of the target.

[0084] (2.3) Heartbeat signal separation: After the aforementioned operations, we can now obtain the echo signal at a specific distance and angle in space, which is the reflection signal containing heartbeat information that we are concerned about. However, the current signal cannot be directly used for heartbeat feature extraction because the current signal contains respiration, body vibrations, and some irrelevant environmental noises. These signals are linearly mixed together.

[0085] (2.3.1) Phase unwrapping: The phase values extracted for complex signals generally fall within the range of (-π, π). When motion continuously affects the phase, such as when the current phase is 3.10, and the phase will continue to increase due to the influence of motion. Due to the range limitation, the next phase may become -3.11, which results in phase wrapping. The phase wrapping problem can be simply solved by comparing the phases at two adjacent moments. Specific method: Calculate the phase difference between the phases at two adjacent time points. If the difference is greater than π, it indicates a cliff-like change in the phase. According to the magnitude relationship between the two adjacent phases, add or subtract a period to the latter phase and all subsequent phases, so as to achieve phase unwrapping.

[0086] By calculating the phase difference between the signals collected at two adjacent time points within the time series, according to the magnitude relationship between the two adjacent phases, add or subtract a period to the latter phase and all subsequent phases, so as to achieve phase unwrapping. Refer to Figure 3 , which shows the change in the phase of the sequence signal before and after unwrapping.

[0087] (2.3.2) Heartbeat signal separation: From the unwrapped result, it can be clearly seen that there are waveforms with periodic changes, and this set of waveform features is mainly dominated by respiratory movement. The respiratory rate is low and the amplitude is large, showing obvious undulations in the right figure. The heartbeat rate is higher than that of respiration and the amplitude is smaller, so it appears as serrations on the model. In principle, since the general respiratory rate is 0.2 - 0.4 Hz and the heartbeat rate is 1 - 2 Hz, the respiratory and heartbeat signals can be easily separated by a band-pass filter. However, the movement pattern of the chest cavity undulation is not regular. On the one hand, the contraction and relaxation of the lungs during respiration drive the slow undulation of the chest cavity skeleton; on the other hand, the contraction and relaxation of the heartbeat drive the movement of the surrounding fat and muscles and transmit it to the skin surface. These two different mechanical movement methods are mixed in a complex way, and moreover, the undulation caused by the two movements is not the undulation of a single point, but the vibration of the whole body. Therefore, in a broad filtering within a small frequency range (0.2, 0.4)(1, 2), it is impossible to handle all situations of different heart rates. For example, in medicine, the normal heart rate range is 60 - 100 beats per minute (corresponding to a frequency of 1 - 1.67), and the normal respiratory range is 12 - 20 breaths per minute (corresponding to a frequency of 0.2 - 0.33). However, in reality, some abnormal physiological activities will cause abnormal respiratory and heart rates. The respiratory rate can reach 50 - 60 breaths per minute (corresponding to a frequency of 0.83 - 1), and the heart rate can be higher than 240 beats per minute (corresponding to a frequency of 4). So, in order for the system to handle abnormal heartbeat situations, the frequency of the filter needs to be set at 1 - 5. Then, the situation that occurs at this time is that for the lower frequency bound, there may be an overlap with the respiratory rate. Therefore, a different method is needed.

[0088] First, the waveform result of the current phase unwrapping is dominated by respiration, and the respiratory pattern can be clearly seen. Then, from another perspective, as mentioned before, respiration is the slow undulation of the chest cavity driven by the lungs, while the heartbeat is the chest cavity vibration caused by a short-time beat. Therefore, the accelerations of the two movements are different. Acceleration can be expressed as the second derivative of displacement with respect to time. Due to the functional relationship between phase and displacement in formula (1), acceleration can be expressed as the second derivative of phase with respect to time. Therefore, the numerical method is used to calculate the acceleration of the time signal sequence:

[0089]

[0090] Our aim is to be able to design a method for extracting heartbeat signals, which can adaptively adjust to appropriate frequencies for different acquisition situations, that is, slow heart rate, normal heart rate, or fast heart rate, rather than having fixed frequency upper and lower limits. To achieve this, we use the method of spectrum analysis. First, we need to perform a rough filtering on the acceleration results obtained above to eliminate irrelevant high-frequency noise. This can be achieved by using a low-pass filter. And then plot the envelope. Next, convert the envelope signal to the frequency domain and determine the accurate heart rate by analyzing the frequency. After determining the precise heartbeat frequency, we use the method of frequency truncation to extract signals within a small range of the precise heartbeat frequency. In this way, our filtering will be more flexible and no longer limited to a fixed frequency range.

[0091] From the unwrapped results, it can be clearly seen that the waveform with periodic changes is mainly dominated by respiratory movements. We further calculate the acceleration results based on the processed signals according to formula (3), thus transforming the waveform to be dominated by heartbeat movements.

[0092] Then perform a low-pass filter on the acceleration results according to the heartbeat frequency to eliminate the interference of high-frequency signals therein. Refer to Figure 4 , which shows the waveform results of the periodic heartbeat signals after separation.

[0093] Step 3: Electrocardiogram construction;

[0094] (3.1) Periodic heartbeat segmentation: The purpose of heartbeat segmentation is to cut the heartbeat signals collected over a period of time into individual heartbeat cycles. In this way, a more accurate mapping between the radio frequency signal and the electrocardiogram can be learned, and the corresponding relationship between different moments of a heartbeat cycle and the radio frequency signal can be learned more accurately.

[0095] We use the template matching algorithm for segmentation:

[0096] First, in the first step, the heartbeat sequence is preliminarily segmented. In the previous step, we were able to obtain the exact heartbeat frequency. Combining with the device parameters, we can calculate the number of sampling points corresponding to a heartbeat cycle, and use this sample number as the initial size of a sliding window. Then we will use the sliding window to slide on the signal sequence, record each selected segmentation point and move the window to continue searching. The principle for selecting the segmentation point is as follows: As mentioned before, the chest displacement will affect the phase of the signal. Calculating the second derivative of the phase with respect to time can reflect the acceleration of the heartbeat. During the heartbeat cycle, the beating is most intense when the ventricle contracts to pump blood into the body, which means the acceleration is the largest at this time. In the electrocardiogram, the QRS complex corresponds to ventricular depolarization, and the ventricle contraction transports blood to the aorta and pulmonary artery. Therefore, we align the peak in the acceleration result with the R wave in the electrocardiogram as the segmentation point, and use the RR interval to represent a heartbeat cycle.

[0097] Next, template matching segmentation is performed. First, construct a template. After adjusting the initial segmentation segments obtained in the previous step to the same length, calculate the mean to obtain the initial template. Next, at the starting position of the sequence, use a window with both the starting point and the ending point variable to calculate a similarity with the template to find a window with the maximum similarity, and use the starting point of the window as the starting point of the subsequent template matching algorithm.

[0098] During the template matching process, starting from the starting point, the starting point of the window is fixed, and the ending point varies within an interval. During the change of the ending point, calculate the similarity between the current window and the template each time, and select the ending point of the window with the highest similarity as the segmentation point, which is also the starting point of the next window. Since the template is calculated based on the previous segmentation result, it may contain the influence of abnormal data or unsatisfactory segmentation segments, resulting in a difference between the template and the real template. For the calculation of similarity, we use the method of weighted cosine similarity. On the basis of the original cosine similarity calculation method, a higher weight is given to the position of the electrocardiogram sub-wave corresponding to the current segment. For example, the two ends of the ideal segmentation segment are R waves, so there should be two high peaks, and we give a higher weight to this part. When the values at both ends of a certain segment are not high, the calculated similarity will naturally not be high under the action of the weight. After traversing from the starting point to the end of the sequence with such a segmentation point selection method, it is considered as the end of one iteration. At this time, the latest segmentation segment group according to template matching is obtained. Use the current segment group to calculate a new template in the same way, and then repeat the above process iteratively. After setting the number of iteration rounds, the most accurate segmentation result can be obtained.

[0099] Refer to Figure 5 , which shows the situation of the segments after segmentation by the template method. The heartbeat sequence is segmented by the dotted line according to the heartbeat segments.

[0100] (3.2) Electrocardiogram construction:

[0101] Use machine learning methods to learn the potential relationship between heartbeat activities and reflection signals. First is the data preprocessing part before network training. In the data preprocessing part before network training. The first step is to recombine the training data. The original training data is individual heartbeat cycles represented by RR intervals. To increase the diversity of samples and prevent overfitting, we recombine the training data at the head and tail in different proportions so that the QRS complex and the like may appear at any position in a training segment. The second step is to transform the data to the same size by interpolation or decimation to facilitate unified learning by the network; the third step is to normalize the data to avoid inconsistent signal amplitudes caused during signal processing. Finally, the simultaneously acquired real electrocardiogram is used as a template to train the corresponding radio frequency signal. The network model is designed as follows:

[0102]

[0103] Refer to Figure 6 , the first step is to recombine the training data. The original training data is individual heartbeat cycles represented by RR intervals. To increase the diversity of samples and prevent overfitting, we recombine the training data at the head and tail in different proportions; the second step is to transform the data to the same size by interpolation or decimation to facilitate unified learning by the network; the third step is to normalize the data to avoid inconsistent signal amplitudes caused during signal processing. Refer to Figure 7 , which shows the electrocardiogram learned by the model through wireless signals.

[0104] Step four: Wavelet identification and index inference;

[0105] (4.1) Wavelet identification: The purpose of wavelet identification is that in clinical electrocardiogram examinations, doctors will judge various indicators of patients based on medical knowledge and experience for the generated electrocardiogram. The judgment criteria include that the RR interval duration can be used to judge whether the patient's heartbeat is too slow, normal or too fast, QRS waveform hypertrophy, T-wave inversion, P-wave inversion, etc. can all be used to judge different cardiac indicators. To realize the function of machine analysis of electrocardiograms, we identify the characteristics of QRS complexes, T waves, and P waves among them for the derivation of various cardiac indicators.

[0106] Wavelet identification still uses machine learning methods and adopts a classification network. The label construction method is as follows:

[0107] The generated electrocardiogram is one-dimensional data. Different numerical representations are set for different sub-waves. For example, 1 corresponds to the T wave and 2 corresponds to the P wave. Then, a one-dimensional data of the same length is constructed. The index intervals corresponding to the sub-waves are uniformly assigned the corresponding numerical representations, and the non-sub-wave parts are assigned 0. During the training process, the electrocardiogram data is supervised by a label array to complete the classification. Finally, in the generated results, the corresponding numerical values are transformed into the corresponding sub-waves, and the index intervals occupied by different indicators can be inferred, and the duration of the sub-waves can also be inferred therefrom.

[0108] The network model is implemented using a basic LSTM network.

[0109] (4.2) Deduction of cardiac indicators: The deduction of indicators is based on electrocardiogram knowledge and transformed into code logic. First, the electrocardiogram result obtained from the previous operation is restored to the original size. Next, the signals are spliced together and various indicators (P wave duration, T wave duration, QRS wave duration, PR interval, QT interval) are calculated and recalculated, and different physiological indicator conditions are judged according to medical standards.

[0110] Refer to the following table, which shows the monitoring results of the system's heartbeat indicators.

[0111] Table 1 Basic Indicator Results

[0112]

[0113] Table 2 Pathological Inference Results

[0114]

Claims

1. A non-contact heartbeat monitoring method based on millimeter wave signals, characterized in that: Here are the steps: Step 1: Heartbeat signal collection; The heartbeat signal acquisition device consists of a millimeter-wave radar transmitting board and a data acquisition board, which are connected and fixed by a data transmission interface. During the acquisition process, the target to be measured remains sitting or lying flat and relatively still, and the heartbeat signal acquisition device is placed in front of the human chest cavity, with a distance set between 30cm and 100cm, and the millimeter-wave radar antenna is aimed at the position of the heart. The millimeter-wave radar transmitting board transmits millimeter-wave signals, and the millimeter-wave radar transmitting board receiving antenna receives the human echo data and transmits it to the data acquisition board for data post-processing. Step 2: signal separation processing; Perform target location positioning, multipath clutter suppression and heartbeat signal separation in sequence to separate clear heartbeat activity information from the captured millimeter wave signal; Step 3: Construct an electrocardiogram based on clear heartbeat activity information; Step 4: Perform wavelet identification and index inference based on the constructed electrocardiogram.

2. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 1, characterized in that: The target position positioning is specifically: The millimeter wave radar transmitting board works in a frequency modulated continuous wave (FMCW) mode; the frequency modulated continuous wave (FMCW) working mode is that the millimeter wave signal has a fixed starting frequency f cc In the subsequent millimeter wave signal transmission process, the frequency of the transmitted millimeter wave signal increases with a certain slope k, that is, the transmission signal frequency at the tth moment is f tt =f cc +kt, so the signals at different times have their own frequencies; the millimeter wave signal in a transmission cycle is converted into a frequency domain signal by calculating FFT to obtain the frequency of the millimeter wave reflected signal at each moment; based on the FMCW working mode, it is inferred that the signal is transformed from the starting frequency to a certain frequency f tt The time t is obtained, and the distance of the target to be measured is d=t×c, where c represents the speed of light; There are multiple peaks in the frequency domain signal, each of which represents the reflection signal of a reflector in the scene. The human body is the nearest strong reflector within the line of sight of the device. The highest peak in the frequency domain signal is identified as the distance bin where the human body is located, and the human body reflection signal at this location is processed; Formula (1) is the phase calculation formula of the millimeter wave signal; d represents the propagation distance of the millimeter wave signal, and λ represents the wavelength of the millimeter wave signal; During the process of collecting human chest signals, human breathing and heartbeat movements will cause the chest to expand and contract. The chest fluctuations affect the propagation distance d of the millimeter wave signal, thereby affecting the phase of the reflected signal.

3. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 2, characterized in that: The multipath clutter suppression is specifically as follows: According to the physical arrangement of the multi-antenna receiving array, the weight formula for beamforming phase adjustment is constructed: Where n represents the number of antennas, i represents the i-th millimeter-wave radar antenna, θ represents the angle of the target, and l represents the antenna spacing; traverse θ=(-π,π) to find the signal with the strongest reflection within a range of 180° directly in front of the millimeter-wave radar transmitting plate; obtain the angle with the strongest reflection based on the signal peak; and perform subsequent processing on the reflected signal at this angle.

4. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 3 is characterized in that: The heartbeat signal separation is specifically: performing phase unwrapping and heartbeat signal separation on the echo signal at the current determined distance and determined angle; Phase unwrapping: The process of phase unwrapping is to ensure the continuity of the signal by gradually adjusting the phase difference. First, calculate the phase difference between the current time point and the previous time point. If the phase difference is greater than π, it means that the phase jumps, and the phase at the current moment needs to be subtracted by one period 2π. If the phase difference is less than -π, it means that the direction of the phase jump is reversed. In this case, a period of 2π needs to be added to the current phase. Heartbeat signal separation: A periodically changing waveform is obtained from the phase unwrapping result; the fluctuation of the waveform corresponds to the fluctuation of the chest cavity during human breathing, and the heartbeat signal is separated; The acceleration of the phase-unwrapped millimeter-wave signal sequence is calculated numerically to highlight the characteristics of the heartbeat: Among them, S0 is the value of a certain point in the sequence, S -ii and S ii It represents the value of the i-th point before the current point and the i-th point after the current point, and h represents the spatial step size, that is, the distance between adjacent points; after performing a low-pass filtering on the acceleration result to eliminate irrelevant high-frequency noise, the envelope signal of the filtered signal is calculated; the envelope signal is converted to the frequency domain, and the frequency corresponding to the heartbeat is found through spectrum analysis; after determining the frequency corresponding to the heartbeat, the frequency truncation method is used to extract the signal within the frequency corresponding to the heartbeat to obtain the heartbeat signal sequence.

5. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 1, characterized in that: The step three is specifically: performing periodic heartbeat segmentation, cutting the processed clear heartbeat activity information for a period of time into separate heartbeat cycles and then constructing an electrocardiogram; The periodic heartbeat segmentation is performed using a template matching algorithm: First, the heartbeat signal sequence obtained by frequency interception through the heartbeat frequency is preliminarily cut; according to the clear heartbeat activity information and equipment parameters, the number of sampling points corresponding to a heartbeat cycle is calculated, and the number of sampling points is used as the initial size of a sliding window; the sliding window is used to slide on the heartbeat signal sequence, and each time a segmentation point is selected, it is recorded and the sliding window is moved to continue searching; the segmentation point is selected as follows: through theoretical analysis, the highest point of the processed heartbeat signal sequence corresponds to the R wave in the electrocardiogram, so the peak value in the heartbeat signal sequence is used as the segmentation point, and the RR interval is used to represent a heartbeat cycle; Next, perform template matching segmentation. First, construct a template, adjust the initial segmented segments obtained in the previous step to the same length, and then calculate the average value to obtain the initial template. Template matching will be iterated for multiple rounds, and a new template will be recalculated based on the segmented segments after each round of matching. During the template matching process, a sliding window with a variable starting point and ending point is used to move backward. Starting from the starting point, the starting point of the sliding window is fixed, and the ending point varies within a range. During the changing of the ending point, the similarity between the current window and the new template obtained in the previous round is calculated each time, and the ending point of the window with the highest similarity is selected as the segmentation point, which is also the starting point of the next window. Using this segmentation point selection method, traversing from the beginning to the end of the heartbeat signal sequence is counted as one iteration to obtain the latest segmented segment group according to template matching. Calculate a new template in the same way based on the latest segmented segment group, and then repeat the above process to get the most accurate segmentation result after setting the number of iterations; Based on the segmentation results, a machine learning method is used to learn the potential connection between heartbeat activity and millimeter wave signals reflected from the chest cavity; Before the network model is trained, data preprocessing is performed. The previous segmented segments of the heartbeat signal are reassembled as training data, and the training data are reassembled head to tail in different proportions. The reassembled data is transformed to the same size by interpolation or extraction, and then normalized. The real electrocardiogram collected synchronously is used as the real label to train the corresponding RF signal.

6. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 5, characterized in that: The network model is a hybrid neural network that combines a convolutional neural network (CNN) and a long short-term memory network.

7. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 5, characterized in that: The similarity is calculated using weighted cosine similarity. In the cosine similarity calculation method, a higher weight is assigned to the electrocardiogram wavelet position corresponding to the current segmentation segment.

8. The non-contact heartbeat monitoring method based on millimeter wave signals according to claim 1, characterized in that: The sub-wave identification adopts a classification network; the electrocardiogram judgment criteria include the RR period duration to judge whether the patient's heartbeat is too slow, normal or too fast, QRS waveform hypertrophy, T wave inversion, P wave inversion, which are used to judge different heart indicators; identify the QRS complex, T wave, P wave characteristics in the electrocardiogram, and use them to derive various heart indicators; The label for the ECG is constructed as follows: The generated ECG is one-dimensional data, and different numerical representations are set for different sub-waves; a one-dimensional data of equal length is constructed, and the index intervals where the corresponding sub-waves are located are uniformly assigned to the corresponding numerical representations, and the rest of the non-wavelet parts are assigned to 0; during the training process, the label array is used to supervise the training of the ECG data to complete the classification; finally, in the generated results, the corresponding numerical values ​​are transformed into the corresponding sub-waves to infer the index intervals occupied by different indicators and infer the duration of the sub-waves; Derivation of cardiac indicators: convert ECG knowledge into code logic; restore the ECG results constructed by the classification network to their original size; splice the ECG signals corresponding to each heartbeat together and recalculate various indicators, including: P wave duration, T wave duration, QRS wave duration, PR interval, QT interval; judge different physiological indicators according to medical standards.

9. A non-contact heartbeat monitoring system according to any one of claims 1 to 8, characterized in that: include: A heartbeat signal acquisition module, used for acquiring heartbeat signals; The signal separation processing module performs target location positioning, multipath clutter suppression and heartbeat signal separation in sequence, separating clear heartbeat activity information from the captured millimeter wave reflection signal; ECG construction module, used for periodic heartbeat segmentation and ECG construction; The wavelet identification and index inference module is used to identify the corresponding wavelets and judge the physiological index conditions.

Images