Non-contact respiration monitoring method based on Doppler radar

Abstract

The invention discloses a non-contact respiration monitoring method based on Doppler radar, which relates to the technical field of respiration monitoring, and comprises the following steps: constructing a dual-band radar system containing 24-26GHz and 77-81GHz, the signal acquisition included angle of two modules being 30-60 degrees; the method comprises the following steps: preprocessing an original signal, eliminating power frequency interference by self-adaptive notch filtering, denoising a wavelet threshold, and smoothing a sliding window mean value; establishing a dual-band fusion model, taking a 24-26GHz signal period as an initial reference, extracting an instantaneous frequency from 77-81GHz signals through Hilbert transform, and obtaining a precise respiratory period through Kalman filtering; 5-10 periodic waveform features are extracted by using a convolutional neural network, and the respiratory rate, depth and rhythm stability are output in combination with a threshold value; and sound-light alarm and wireless data transmission terminal are realized in case of abnormity. The system is widely applied to the scenes of family health, hospital nursing, old-age care and the like, and has technical innovation and practical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of respiratory monitoring, in particular to a non-contact respiratory monitoring method based on Doppler radar. BACKGROUND

[0002] Respiratory monitoring is an important means to evaluate human vital signs. Traditional respiratory monitoring methods are mainly divided into two types: contact type and single-frequency radar non-contact type. However, there are obvious defects: Contact monitoring (such as chest strap sensor, nasal cannula): it needs to be in direct contact with the human body, which can easily cause skin irritation and discomfort, especially for infants, burn patients and people sleeping. Moreover, it is affected by clothing obstruction and body movement, and the monitoring data is easy to be distorted. Long-term wearing compliance is poor. Single-frequency radar non-contact monitoring: it usually uses 24GHz or 77GHz single frequency band, which can only capture a single signal dimension in chest displacement or airflow vibration. It is easily affected by environmental noise (such as power frequency interference and furniture reflection) and changes in monitoring distance, resulting in large measurement error of respiratory frequency (usually >10%), and unable to accurately distinguish respiratory depth and rhythm stability.

[0003] Poor scene and population adaptability: the existing technology does not optimize parameters for different scenes such as low noise at home and high interference in hospital, nor does it adjust monitoring strategies for special populations such as infants and the elderly. This results in low monitoring accuracy (<85%) in complex scenes or special populations, and lack of historical data trend analysis and early warning capability, making it difficult to meet the long-term health management needs. Therefore, there is an urgent need for a non-contact respiratory monitoring method with high precision, strong adaptability, full population coverage and perfect function to solve the pain points of existing technology. SUMMARY

[0004] The present application aims to at least partially solve one of the above technical problems.

[0005] To achieve the above-mentioned purpose, the first aspect of the present application proposes a non-contact respiratory monitoring method based on Doppler radar, comprising the following steps: S1: constructing a multi-frequency Doppler radar monitoring system, the monitoring system comprising at least two radar modules with different working frequencies, wherein the first radar module works in the 24GHz-26GHz frequency band to capture the chest displacement signal caused by breathing, and the second radar module works in the 77GHz-81GHz frequency band to collect the small vibration signal of respiratory airflow, and the signal collection directions of the two radar modules are arranged at an angle of 30°-60°; S2: preprocessing the original signals collected by the two radar modules respectively, the preprocessing including: eliminating 50Hz / 60Hz power frequency interference by adaptive notch filter, removing environmental noise by wavelet threshold denoising algorithm, and smoothing the signal based on sliding window mean method; S3: Establish a dual-band signal fusion model. Use the preprocessed signal from the first radar module as the base signal, extract its signal periodic features as the initial reference value for the respiratory cycle, perform Hilbert transform on the signal from the second radar module to obtain the analytical signal, extract the instantaneous frequency features of the analytical signal, and fuse the periodic features and instantaneous frequency features using the Kalman filter algorithm to obtain accurate respiratory cycle parameters. S4: Based on the fused respiratory cycle parameters, a respiratory status assessment model is constructed. The assessment model extracts features from the signal waveforms of 5-10 consecutive respiratory cycles through a convolutional neural network. Combined with a preset respiratory rate threshold range (8-20 breaths / minute for adults in a resting state and 20-40 breaths / minute for adults in an exercise state), the model outputs assessment results of respiratory rate, respiratory depth, and respiratory rhythm stability. S5: When the respiratory rate is detected to exceed the corresponding state threshold range or the respiratory rhythm stability coefficient is lower than 0.8, the audible and visual alarm module is triggered, and the real-time monitoring data and alarm information are transmitted to the terminal device through the wireless communication module.

[0006] As an improvement, in step S1, the angle between the signal acquisition directions of the two radar modules is adjusted in real time by an electric adjustment bracket. The electric adjustment bracket dynamically adjusts the angle according to the distance between the monitored object and the radar module. When the distance is less than 1m, the angle is set to 30°-40°, and when the distance is 1m-3m, the angle is set to 40°-60°.

[0007] As an improvement, the wavelet thresholding denoising algorithm in step S2 adopts the db4 wavelet basis function, sets the number of decomposition layers to 3, and adopts the Birgé-Massart strategy for threshold selection. The signal is reconstructed after thresholding the high-frequency wavelet coefficients.

[0008] As an improvement, the state equation of the Kalman filter algorithm in step S3 is constructed as follows: X(k)=AX(k-1)+BU(k-1)+W(k-1), where X(k) is the state vector at time k, which includes the breathing cycle and instantaneous frequency parameters, A is the state transition matrix, which takes the value of a diagonal matrix diag(0.95, 0.98), B is the control matrix, U(k-1) is the input vector, and W(k-1) is the process noise vector, whose covariance matrix is ​​set to diag(0.01, 0.005).

[0009] As an improvement, the convolutional neural network in step S4 includes 3 convolutional layers, 2 pooling layers, and 2 fully connected layers. The first convolutional layer uses 16 3×3 kernels with a stride of 1 and the same padding method. The second convolutional layer uses 32 3×3 kernels with a stride of 1 and the same padding method. The third convolutional layer uses 64 3×3 kernels with a stride of 1 and the same padding method. All pooling layers use 2×2 max pooling kernels with a stride of 2. The number of neurons in the fully connected layers are 256 and 3, respectively. The output layer uses the softmax activation function to output the respiratory state assessment result.

[0010] As an improvement, the wireless communication module in step S5 supports both Bluetooth 5.0 and LoRa communication protocols. When the monitoring distance is less than 10m, the Bluetooth 5.0 protocol is used to transmit data, and when the monitoring distance is between 10m and 100m, the LoRa protocol is used to transmit data. In addition, the monitoring data is encrypted using the AES-128 encryption algorithm during data transmission.

[0011] As an improvement, step S6 is also included: periodically calibrating the monitoring system by generating respiratory signals with known frequency (5 breaths / min - 50 breaths / min) and depth (0.1L - 1.5L) using a standard respiratory simulator, comparing the measurement results of the monitoring system with the standard signal, calculating the error value, and automatically adjusting the filtering parameters in step S2 and the fusion model parameters in step S3 when the error value is greater than 5% to achieve system calibration.

[0012] As an improvement, step S1 also includes adding an infrared thermal imaging module to the monitoring system. The infrared thermal imaging module collects data synchronously with the two radar modules, determines the outline of the chest cavity region of the monitored object through the infrared thermal imaging image, feeds the outline information back to the signal acquisition unit of the radar module, and adjusts the signal focusing range of the radar module so that the signal acquisition area accurately covers the chest cavity outline and reduces signal interference in non-target areas.

[0013] As an improvement, step S2 also includes noise adaptation processing for different monitoring environments: when the monitoring environment is a home bedroom (ambient noise ≤ 40dB), the notch bandwidth of the adaptive notch filter is set to 2Hz; when the monitoring environment is a hospital ward (ambient noise 40dB-60dB), the notch bandwidth is adjusted to 4Hz. At the same time, after wavelet threshold denoising, adaptive Wiener filtering is added. The window function size of the Wiener filter is dynamically set according to the ambient noise intensity. For every 10dB increase in noise intensity, the window function size increases by 2 sampling points.

[0014] As an improvement, step S4 also includes a threshold adaptation strategy for special populations: for infants (0-3 years old), the preset respiratory rate threshold range is set to 20-40 breaths / minute; for the elderly (≥65 years old), the respiratory rate threshold range in a resting state is set to 12-24 breaths / minute. At the same time, during the training of the convolutional neural network, respiratory signal sample sets of infants and the elderly (≥5000 samples each) are added, and the network parameters are optimized through transfer learning to improve the accuracy of respiratory status assessment for special populations.

[0015] As an improvement, step S5 also includes a multi-level alarm mechanism: when the respiratory rate exceeds the threshold range by less than 10% or the respiratory rhythm stability coefficient is 0.7-0.8, a level 1 alarm is triggered (only a pop-up notification is displayed on the terminal device); when the respiratory rate exceeds the threshold range by 10%-20% or the respiratory rhythm stability coefficient is 0.6-0.7, a level 2 alarm is triggered (the audible and visual alarm module emits a low-frequency alert sound, and the terminal device pushes an SMS notification); when the respiratory rate exceeds the threshold range by more than 20% or the respiratory rhythm stability coefficient is <0.6, a level 3 alarm is triggered (the audible and visual alarm module emits a high-frequency alarm sound, and the terminal device simultaneously sends alarm information to the preset emergency contacts and medical institution platforms, including the location of the monitored object, real-time respiratory data, and historical data trend charts).

[0016] As an improvement, step S7 is also included: constructing a respiratory data historical analysis model, storing 30 consecutive days of respiratory monitoring data (including respiratory rate, depth, and rhythm stability) in a local database, performing trend analysis on the historical data through a long short-term memory network (LSTM), predicting the respiratory status change trend for the next 7 days, and pushing early warning prompts to terminal devices in advance when the respiratory rate is predicted to show a continuous upward / downward trend (daily average change rate ≥5%) or the rhythm stability is predicted to continuously decrease (daily average decrease rate ≥3%), assisting users or medical staff in early intervention.

[0017] As an improvement, the synchronous acquisition of the infrared thermal imaging module and the radar module is achieved through timestamp calibration: a synchronization timestamp is generated every 1 second to mark the acquisition time of the infrared thermal imaging image and the radar signal respectively. When the timestamp deviation between the two exceeds 50ms, time synchronization correction is initiated, and the sampling clock frequency of the radar module is adjusted to keep the timestamp deviation between the two within 20ms, ensuring the time consistency of multimodal data.

[0018] Beneficial effects By employing dual-band radar to collaboratively acquire respiratory-related signals, combining data fusion algorithms to improve signal quality, and then eliminating interference through infrared calibration, the overall monitoring results are more accurate and reliable.

[0019] The parameters can be dynamically adjusted according to the noise level and monitoring distance of different scenarios, and it can work stably in various environments such as homes and hospitals.

[0020] It can optimize monitoring strategies for special groups such as infants and the elderly, and its non-contact design avoids discomfort when wearing it, making it suitable for the needs of various groups.

[0021] It has a tiered alarm mechanism to respond to anomalies in a timely manner, and can also predict respiratory status trends through historical data, while ensuring data security, reducing maintenance costs, and facilitating practical application.

[0022] In summary, this solution has significant advantages in accuracy, adaptability, and practicality, and can be widely used in scenarios such as family health, hospital care, and elderly care, combining technological innovation with practical application value. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a control system diagram of a non-contact respiratory monitoring method based on Doppler radar according to an embodiment of the present invention. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] The following describes a non-contact respiratory monitoring method based on Doppler radar according to an embodiment of the present invention, with reference to the accompanying drawings.

[0026] like Figure 1 As shown in the figure, a non-contact respiratory monitoring method based on Doppler radar according to an embodiment of the present invention is characterized by comprising the following steps: S1: Construct a monitoring system that integrates multi-band Doppler radar and infrared thermal imaging. The system comprises at least two radar modules operating at different frequencies, an infrared thermal imaging module, and an electrically adjustable support. The first radar module uses a TI AWR1642 millimeter-wave radar chip (operating in the 24GHz-26GHz band) to capture macroscopic displacement signals of the thoracic cavity caused by respiration. The second radar module uses an NXP MR24H251 radar chip (operating in the 77GHz-81GHz band) to collect minute vibration signals from respiratory airflow. The infrared thermal imaging module uses a Hikvision DS-2TD2617B-6 / PA sensor (320×240 resolution, 15fps frame rate) to synchronously acquire data with the radar modules to locate the thoracic cavity contour of the monitored object. The signal acquisition directions of the two radar modules are set at an angle of 30°-60°. The electrically adjustable support is an ST-M100 type support driven by a stepper motor (adjustment angle range 0-90°, control accuracy ±0.5°). S2: Adaptive preprocessing is performed on the raw signals acquired by the two radar modules. The preprocessing includes: eliminating 50Hz / 60Hz power frequency interference using an adaptive notch filter based on the LMS algorithm (implemented using the AD8233 chip from ADI); removing environmental noise using a wavelet threshold denoising algorithm; smoothing the signal based on the sliding window mean method (window size of 5 sampling points); and dynamically adjusting the filtering parameters according to the intensity of the monitored environmental noise (determined by the noise detection module associated with the SensirionSHT31 temperature and humidity sensor built into the system). S3: Establish a multimodal fusion model of dual-band signals and infrared image features. The preprocessed signal from the first radar module is used as the base signal, and periodic features are extracted using a peak detection algorithm (implemented using a hardware acceleration module based on the STM32H743 microprocessor). The signal from the second radar module undergoes Hilbert transform (operated using a TMS320C6748 DSP chip) to extract instantaneous frequency features. The chest contour image output by the infrared thermal imaging module is processed using a multi-feature fusion algorithm (temperature threshold segmentation + Canny edge detection + MobileNetV2 pose estimation) to extract dynamic contour change features. The periodic features, instantaneous frequency features, and dynamic contour change features are fused using a Kalman filter algorithm to obtain accurate respiratory cycle, respiratory depth, and chest movement amplitude parameters. S4: Based on the fused respiratory parameters, a multi-scenario adapted respiratory status assessment model is constructed. The assessment model uses a hybrid network structure of convolutional neural network (CNN) and long short-term memory network (LSTM) (equipped on an NVIDIA Jetson Nano embedded AI module) to jointly extract signal waveforms and chest movement features over 5-10 consecutive respiratory cycles. Combined with preset respiratory frequency threshold ranges for different scenarios (8-20 breaths / minute for adults in a calm state, 20-40 breaths / minute for exercise, and 12-20 breaths / minute for sleep), the model outputs assessment results for respiratory frequency, respiratory depth, respiratory rhythm stability, and chest movement coordination. S5: When the respiratory rate exceeds the corresponding scene threshold range, the respiratory rhythm stability coefficient is lower than 0.8, or the chest movement coordination coefficient is lower than 0.75, a multi-level sound and light alarm module is triggered (using NeoPixel RGB LED beads and Maxim MAX98357 audio amplifier); and the real-time monitoring data and alarm information are transmitted to the terminal device through a multi-protocol wireless communication module (Bluetooth 5.0 using TI CC2640R2F chip, LoRa using Semtech SX1278 chip, NB-IoT using Huawei MH5000-31 module); S6: Periodically perform automatic calibration of the monitoring system. Generate a respiratory signal with known parameters using a standard respiratory simulator (TSI 4080, which can simulate a frequency of 5-50 breaths / minute and a depth of 0.1L-1.5L). Compare the measurement results of the monitoring system with the standard signal and calculate the error value. When the error value is greater than 5%, automatically adjust the filtering parameters (notch bandwidth, wavelet decomposition level) in step S2 and the fusion model parameters (Kalman filter process noise covariance matrix) in step S3 to achieve system calibration.

[0027] It should be noted that in step S1, the electric adjustable bracket (ST-M100 type) dynamically adjusts the included angle according to the distance between the monitored object and the radar module (obtained in real time through the ranging function of the first radar module) and the monitoring scenario (manually selected through the terminal APP or automatically determined by the system): when the distance is less than 1m and it is a sleep monitoring scenario, the included angle is set to 30°-35°; when the distance is 1m-3m and it is a calm state monitoring scenario, the included angle is set to 40°-50°; when the distance is 1m-2m and it is a motion state monitoring scenario, the included angle is set to 50°-60°; at the same time, the sampling rate of the radar module is adjusted synchronously. In the motion state, the sampling rate of the first radar module (AWR1642) is increased to 300Hz, and the sampling rate of the second radar module (MR24H251) is increased to 400Hz, which is achieved by configuring the sampling clock register through the module's SPI interface.

[0028] It should be noted that the adaptive preprocessing in step S2 also includes a motion artifact elimination step: when the monitored scene is in motion, the body motion acceleration signal of the monitored object is collected by an accelerometer (Bosch BMA280, range ±2g, resolution 16-bit); the acceleration signal and radar signal are subjected to real-time correlation analysis (using the Pearson correlation coefficient algorithm) through the DMA transfer channel of the STM32H743 microprocessor to filter out motion-related interference signal segments; the interference segments are reconstructed using a cubic spline interpolation algorithm (based on embedded code generated by MATLAB Coder), with the interpolation node interval set to 2 sampling points to ensure that the continuity deviation between the reconstructed signal and the original signal is ≤3%.

[0029] It should be noted that the wavelet thresholding denoising algorithm in step S2 uses the db4 wavelet basis function, with a decomposition layer of 3, and adopts the Birgé-Massart strategy for threshold selection. Furthermore, the threshold coefficient is dynamically adjusted according to the ambient noise intensity (the Sensirion SHT31 correlation module's detection range is 0dB-120dB): when the ambient noise is ≤40dB, the threshold coefficient is set to 0.8; when the ambient noise is 40dB-60dB, the threshold coefficient is set to 1.0; and when the ambient noise is >60dB, the threshold coefficient is set to 1.2. After soft thresholding of the high-frequency wavelet coefficients (threshold calculation formula: λ = coefficient × σ√(2lnN), where σ is the noise standard deviation and N is the signal length), the signal is reconstructed using the TI TMS320C6748 DSP chip, resulting in a signal-to-noise ratio improvement of ≥20dB.

[0030] It should be noted that in step S3, the state equation of the Kalman filter algorithm is constructed as: X(k) = AX(k-1) + BU(k-1) + W(k-1), and the observation equation is constructed as: Z(k) = H*X(k) + V(k); where X(k) is the state vector at time k, including the respiratory cycle, instantaneous frequency, and chest movement amplitude parameters; A is the state transition matrix, with values ​​of diagonal matrix diag (0.95, 0.98, 0.96); B is the control matrix, with values ​​of 3×2 matrix ([0.1, 0.05; 0.08, 0.04; 0.06, 0.03]); U(k-1) is the infrared image contour feature input vector (including contour area change rate and contour center offset); W(k-1) is the process noise vector, with its covariance matrix set as diag (0.01, 0.005, 0.008); H The observation matrix is ​​a 3×3 identity matrix; V(k) is the observation noise vector, and its covariance matrix is ​​set as diag(0.008, 0.004, 0.006); the filtering algorithm is accelerated by the CUDA core of NVIDIA Jetson Nano, and the single filtering iteration time is ≤1ms.

[0031] It should be noted that the hybrid network structure in step S4 includes 3 convolutional layers, 2 pooling layers, 1 LSTM layer, and 2 fully connected layers: the first convolutional layer uses 16 3×3 kernels (stride 1, same padding) and selects the ReLU activation function; the second convolutional layer uses 32 3×3 kernels (stride 1, same padding) and selects the ReLU activation function; the third convolutional layer uses 64 3×3 kernels (stride 1, same padding) and selects the ReLU activation function; all pooling layers use 2×2 max pooling kernels (stride 2); the LSTM layer contains 128 neurons (using the tanh activation function, dropout rate 0.2); the number of neurons in the fully connected layers are 256 (ReLU activation) and 4 (softmax activation), respectively; the network model is deployed on an NVIDIA Jetson Nano based on the TensorFlow Lite framework, with a single evaluation inference time ≤50ms and an evaluation accuracy ≥96%.

[0032] It should be noted that step S4 also includes scenario adaptation strategies for specific populations: for infants (0-3 years old), the preset respiratory rate threshold range is set to 20-40 breaths / minute during sleep and 30-50 breaths / minute during exercise; for the elderly (≥65 years old), the respiratory rate threshold range is set to 12-24 breaths / minute during rest and 14-22 breaths / minute during sleep; at the same time, during the training of the hybrid network, respiratory signal sample sets of infants and the elderly in different scenarios are added (≥8000 samples for each scenario, samples are from the clinical database of Shenzhen Third People's Hospital), and the network parameters are optimized through transfer learning (fine-tuning based on the pre-trained ResNet50 model), so that the accuracy of respiratory status assessment for specific populations in different scenarios is improved to ≥97%.

[0033] It should be noted that the multi-level audio-visual alarm module in step S5 works as follows: At level one, the NeoPixel RGB LED emits a flashing green light (1Hz), and a pop-up notification appears on the terminal device (smartphone app); at level two, the Maxim MAX98357 audio amplifier outputs a low-frequency alert tone (60dB volume) of 500Hz-800Hz, the NeoPixel LED remains solid yellow, and the terminal device pushes an SMS notification via the SMS gateway (using Alibaba Cloud SMS service API); at level three, the audio amplifier outputs a high-frequency alarm tone (80dB volume) of 1000Hz-1500Hz, the NeoPixel LED flashes red (2Hz), and the terminal device simultaneously sends alarm information to preset emergency contacts (stored in an encrypted local database) and the medical institution platform (via the HL7 FHIR medical data interface). This information includes the location of the monitored object (obtained via the terminal device's GPS module, accuracy ±5m), real-time respiratory data, and a historical data trend chart for the past 24 hours (based on ECharts). (Charts generated from the library) and images of thoracic motion (frames captured by the infrared thermal imaging module).

[0034] It should be noted that in step S5, the multi-protocol wireless communication modules are as follows: Bluetooth 5.0 (TI CC2640R2F chip) uses the GATT protocol when the transmission distance is less than 10m, with a transmission rate of 2Mbps and a single data transmission volume ≤1024 bytes; LoRa (Semtech SX1278 chip) uses the LoRaWAN protocol when the transmission distance is 10m-100m, with a spreading factor of SF12, a transmission rate of 50kbps, and a communication interval ≤10s; NB-IoT (Huawei MH5000-31 module) connects to the operator's NB-IoT network when the transmission distance is >100m, using the CoAP protocol, and the transmission cycle is adjustable from 1min to 5min according to monitoring requirements (configurable via AT commands). The data transmission process under all three protocols uses the AES-128 encryption algorithm (implemented based on the STM32H743's built-in hardware encryption engine) to encrypt the monitoring data. The encryption key is obtained through the ECC during the initial pairing of the terminal device and the monitoring system. Elliptic curve cryptography (256-bit key length) is dynamically generated and the key is automatically updated every 7 days via Bluetooth 5.0. The update process log is stored in the module's Flash chip (4MB capacity).

[0035] It should be noted that the standard breathing simulator (TSI 4080 model) in step S6 can also simulate standard signals for different breathing modes, including thoracic breathing (chest movement amplitude ≥70%), abdominal breathing (abdominal movement amplitude ≥70%), and mixed breathing mode (chest and abdominal movement amplitude 30%-70%). The standard signal frequency range for each breathing mode is 5 breaths / minute to 50 breaths / minute, and the depth range is 0.1L-1.5L. During system calibration, the initial 24-hour monitoring data (stored in 8GB eMMC flash memory) is used to determine the regular breathing mode of the monitored object using the K-means clustering algorithm (cluster number k=3). The standard signal of the corresponding mode is automatically selected for comparison and calibration, reducing the monitoring error caused by breathing mode differences to ≤3%.

[0036] It should be noted that step S7 is also included: constructing a correlation analysis model between respiratory data and physiological parameters. Heart rate data is collected using a heart rate sensor (Maxim MAX30102, sampling rate 100Hz, measurement range 30-250 beats / minute) built into the monitoring system, and body temperature data is collected using a temperature sensor (Sensirion SHT31, measurement range -40℃-125℃, accuracy ±0.3℃). Respiratory monitoring data (respiratory rate, depth, rhythm stability, chest wall movement coordination) from 30 consecutive days, along with heart rate and body temperature data, are transmitted to the correlation analysis model via a data fusion interface (based on an MQTT message queue). The model uses the XGBoost algorithm (deployed on a cloud server, trained using GPU acceleration) to construct a mapping relationship between respiratory parameters and physiological states (such as the risk of colds or asthma attacks). When the model predicts an abnormal physiological state (such as a heart rate to respiratory rate ratio exceeding the normal range of 1.5), the model will analyze the data. When the ratio of abnormal indicators to normal (4-5:1 in a calm adult state) is more than twice that of normal, the NB-IoT module can push physiological status warnings (including abnormal indicators, risk levels, and suggested interventions) to the terminal device to assist users or medical staff in health intervention.

[0037] It should be noted that the monitoring system in step S1 also has an energy consumption optimization function: when the monitored object is in a sleep state (determined by linkage with the sleep monitoring APP on the terminal device or by human activity detection of the infrared thermal imaging module) and the respiratory parameters are stable (respiratory rate fluctuation ≤5% for 30 consecutive minutes, rhythm stability coefficient ≥0.9), the system automatically enters a low-power mode through the PMIC power management chip (TI TPS65987DDH), reducing the sampling rate of the first radar module (AWR1642) to 50Hz (shutting down some transmission channels of the module), reducing the sampling rate of the second radar module (MR24H251) to 100Hz (reducing the sampling clock cycle), and reducing the sampling frame rate of the infrared thermal imaging module (DS-2TD2617B-6 / PA) to 5fps (using an inter-frame compression algorithm); at this time, the overall power consumption of the system is ≤50mA (3.7V). Power consumption is reduced by 60% compared to normal mode. When respiratory parameter fluctuations exceed the stable range (frequency fluctuation > 5%) or the monitored object awakens (infrared module detects human activity > 30%), the system triggers a return to normal working mode via an interrupt signal, with a response time ≤ 1s, ensuring a balance between energy consumption and monitoring accuracy.

[0038] It should be noted that the chest cavity contour localization of the infrared thermal imaging module in step S3 adopts a multi-feature fusion algorithm: First, the human body area is initially screened by temperature threshold segmentation (the threshold is set to 32℃-37℃, dynamically adjusted based on the module's temperature calibration function), eliminating interference areas in the environment that are below 32℃ or above 37℃; then, the human body contour is extracted by the Canny edge detection algorithm (the threshold is determined by the Otsu adaptive algorithm, with a high threshold of 200 and a low threshold of 100), removing small noise (areas with an area < 50 pixels) from the contour; finally, the chest cavity position is determined by combining the human pose estimation model (a lightweight model based on MobileNetV2, with a model size ≤ 5MB and an inference time ≤ 20ms) (the line connecting the midpoint of the clavicle and the xiphoid process is used as the boundary of the chest cavity area), with a positioning accuracy error ≤ 2cm; and when the monitored object undergoes a change in body posture (such as turning over or sitting up), the contour offset of 3 consecutive frames is > 10cm. When determining the location, the positioning algorithm improves the response time to ≤1s by increasing the interrupt priority (set to the highest priority of the system), ensuring that the signal acquisition area of ​​the radar module (AWR1642, MR24H251) always accurately covers the chest cavity, and the signal acquisition efficiency is improved by ≥40%.

[0039] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0040] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0041] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A Doppler radar based non-contact respiration monitoring method, characterized in that, The method comprises the following steps: S1: constructing a multi-frequency Doppler radar monitoring system, the monitoring system comprising at least two radar modules different in working frequency band, wherein a first radar module works in a 24-26 GHz frequency band to capture a chest macro-displacement signal caused by respiration, and a second radar module works in a 77-81 GHz frequency band to collect a slight vibration signal of respiratory airflow, and the signal collection directions of the two radar modules are arranged at an angle of 30-60°; S2: respectively pre-processing the original signals collected by the two radar modules, the pre-processing comprising: eliminating 50 / 60 Hz power frequency interference through an adaptive notch filter, removing environmental noise through a wavelet threshold denoising algorithm, and smoothing the signals based on a sliding window mean method; S3: establishing a dual-frequency signal fusion model, taking the pre-processed first radar module signal as a basic signal, extracting a signal period feature thereof as an initial reference value of a respiratory period, performing Hilbert transform on the second radar module signal to obtain an analytical signal, extracting an instantaneous frequency feature of the analytical signal, and fusing the period feature and the instantaneous frequency feature through a Kalman filtering algorithm to obtain accurate respiratory period parameters; S4: based on the fused respiratory period parameters, constructing a respiratory state evaluation model, the evaluation model extracting features of signal waveforms of 5-10 continuous respiratory periods through a convolutional neural network, combining a preset respiratory frequency threshold range, and outputting respiratory frequency, respiratory depth and respiratory rhythm stability evaluation results; S5: when it is monitored that the respiratory frequency exceeds the corresponding state threshold range or the respiratory rhythm stability coefficient is lower than 0.8, triggering an audible and light alarm module, and transmitting real-time monitoring data and alarm information to a terminal device through a wireless communication module.

2. The Doppler radar-based contactless respiration monitoring method according to claim 1, characterized in that, In step S1, the angle between the signal collection directions of the two radar modules is adjusted in real time through an electrically adjustable support, which dynamically adjusts the angle according to the distance between the monitoring object and the radar module, and the angle is set to 30-40° when the distance is less than 1 m, and the angle is set to 40-60° when the distance is 1-3 m.

3. The Doppler radar-based contactless respiration monitoring method according to claim 1, characterized in that, In step S2, the wavelet threshold denoising algorithm adopts a db4 wavelet basis function, the decomposition layer number is set to 3 layers, and the Birgé-Massart strategy is adopted for threshold selection, and the signal is reconstructed after threshold processing of high-frequency wavelet coefficients.

4. The Doppler radar-based contactless respiration monitoring method according to claim 1, characterized in that, In step S3, the state equation of the Kalman filtering algorithm is constructed as: X(k)=AX(k-1)+BU(k-1)+W(k-1), wherein X(k) is a state vector at time k, containing respiratory period and instantaneous frequency parameters, A is a state transition matrix, taking a diagonal matrix diag(0.95, 0.98), B is a control matrix, U(k-1) is an input vector, and W(k-1) is a process noise vector, and the covariance matrix thereof is set to diag(0.01, 0.005).

5. The Doppler radar-based contactless respiration monitoring method according to claim 1, characterized in that, The convolutional neural network in step S4 includes 3 convolutional layers, 2 pooling layers and 2 fully connected layers, the first convolutional layer uses 16 3*3 size convolution kernels, the step is 1, and the padding mode is same; the second convolutional layer uses 32 3*3 size convolution kernels, the step is 1, and the padding mode is same; the third convolutional layer uses 64 3*3 size convolution kernels, the step is 1, and the padding mode is same; the pooling layer uses 2*2 size maximum pooling kernel, the step is 2; the number of neurons of the fully connected layer is 256 and 3 respectively, the output layer uses the softmax activation function, and the output layer uses the softmax activation function.

6. The Doppler radar-based contactless respiration monitoring method according to claim 1, characterized in that, The wireless communication module in step S5 supports two communication protocols of Bluetooth 5.0 and LoRa, when the monitoring distance is less than 10m, the Bluetooth 5.0 protocol is used to transmit data, when the monitoring distance is 10m-100m, the LoRa protocol is used to transmit data, and in the data transmission process, the monitoring data is encrypted by using the AES-128 encryption algorithm.

7. The Doppler radar-based contactless respiration monitoring method according to claim 1, characterized in that, Further comprising: Periodically calibrate the monitoring system, generate a known frequency and depth respiratory signal through a standard respiratory simulator, compare the measurement results of the monitoring system with the standard signal, calculate the error value, when the error value is greater than 5%, automatically adjust the filtering parameters in step S2 and the fusion model parameters in step S3, realize system calibration.

Images