Millimeter wave vital sign monitoring method based on Narx neural network

Abstract

The invention discloses a millimeter wave vital sign monitoring method based on a Narx neural network, and the method comprises the following steps: 1, initializing parameters of a monitoring system, and configuring the working frequency of a millimeter wave radar sensor to be 77GHz and the sampling frequency to be 100MHz; step 2, executing signal acquisition, and transmitting and receiving millimeter wave signals to and from a monitoring area through a millimeter wave radar sensor assembly; 3, judging whether an abnormal condition exists after the acquisition is completed, and if not, executing the step 4; and step 4, executing signal processing and analysis. The system has the following advantages: real-time and accurate monitoring and deep analysis of respiration and heartbeat signals are realized; and comprehensive monitoring requirements in the fields of medical monitoring, smart home, security and protection and the like are met.

Description

Technical Field

[0001] This invention relates to the field of vital sign monitoring, specifically to a method for real-time monitoring and in-depth analysis of vital signs using millimeter-wave radar based on a NARX neural network. Background Technology

[0002] With increasing public awareness of health monitoring and healthcare, vital sign monitoring technology is widely used in fields such as medicine, smart homes, and security. Traditional vital sign monitoring methods mainly rely on contact sensors, such as electrocardiogram electrodes and chest strap respiratory sensors. While these devices can achieve high-precision monitoring, prolonged wear can cause discomfort for users and is unsuitable for certain groups, such as infants and burn patients. Furthermore, contact sensors have limitations in scenarios involving multiple people and long-distance monitoring, making it difficult to meet the needs of smart homes and public spaces for seamless monitoring.

[0003] In recent years, non-contact vital sign monitoring technologies have gradually developed, including traditional methods such as optics and ultrasound, as well as emerging technologies based on millimeter-wave radar. As an emerging technology for non-contact monitoring, millimeter-wave radar has attracted attention due to its advantages such as strong signal penetration, no risk of privacy leakage, and susceptibility to external environmental factors. However, existing related systems still have significant drawbacks: First, the measurement results are not stable enough due to limitations in hardware performance and frequency estimation algorithms; second, the signal acquisition time is relatively long, resulting in delays in detection results and affecting real-time monitoring effectiveness; third, the systems and algorithms largely rely on PCs for implementation, resulting in large and costly equipment that is difficult to miniaturize and deploy at low cost.

[0004] To address these issues, researchers have begun exploring new technological solutions. Among them, deep learning technology based on neural networks has proven to better solve nonlinear problems and can predict heart rate with high accuracy and generalization. Although deep learning technology based on neural networks has been proven useful for heart rate prediction, and some solutions attempt to combine millimeter-wave radar with lightweight neural networks, key problems remain: lightweight neural network models have low complexity and weak ability to process time-series information. When faced with highly fluctuating data sequences, the stability of heart rate prediction is insufficient, failing to guarantee monitoring accuracy. How to achieve real-time, accurate monitoring and in-depth analysis of respiratory and heartbeat signals using high-frequency sampling millimeter-wave radar sensors, combined with advanced signal processing techniques and deep learning algorithms, has become a pressing technical problem to be solved.

[0005] On the one hand, existing technologies still have room for improvement in the speed and accuracy of signal processing modules, making it difficult to achieve real-time, accurate monitoring and in-depth analysis of respiratory and heartbeat signals; on the other hand, existing systems lack the ability to integrate multiple sensors and intelligently analyze data, failing to meet the comprehensive monitoring needs of different application scenarios and limiting the promotion and application of the technology. Summary of the Invention

[0006] The technical problem this invention aims to solve is to address the above-mentioned shortcomings by providing a millimeter-wave vital sign monitoring method based on Narx neural networks. This method addresses many pain points of existing vital sign monitoring technologies, offering an efficient, accurate, and non-invasive monitoring solution that overcomes the limitations of traditional contact monitoring, avoids discomfort from prolonged wear, and is suitable for special populations such as infants and burn patients, as well as non-invasive monitoring scenarios. It also solves the problems of unstable measurement, signal acquisition delay, large size, and high cost of existing millimeter-wave radar monitoring systems; compensates for the weak ability of lightweight neural networks to process temporal information and the instability of heart rate prediction; improves the speed and accuracy of signal processing, enabling real-time, accurate monitoring and in-depth analysis of respiratory and heartbeat signals; and meets the comprehensive monitoring needs of multiple fields such as medical monitoring, smart homes, and security.

[0007] To solve the above technical problems, the present invention adopts the following technical solution:

[0008] A millimeter-wave vital sign monitoring method based on Narx neural networks includes the following steps:

[0009] Step 1: Initialize the monitoring system parameters and configure the millimeter-wave radar sensor to operate at a frequency of 77 GHz and a sampling frequency of 100 MHz.

[0010] Step 2: Perform signal acquisition by transmitting and receiving millimeter-wave signals to the monitoring area through the millimeter-wave radar sensor assembly;

[0011] Step 3: Determine if there are any abnormalities after the data collection is completed. If there are no abnormalities, proceed to Step 4: Through three-layer verification and hierarchical processing, gradually screen the data integrity, signal validity, and target existence.

[0012] Step 4: Perform signal processing and analysis. The high-quality acquired signals that have been screened are precisely processed and deeply analyzed to purify vital sign-related signals. The NARX neural network is used to achieve real-time and accurate prediction of heartbeat signals and heart rate, while also supporting abnormal alarms and historical data storage.

[0013] Furthermore, the specific steps of step 1 are as follows:

[0014] Step 1.1, Monitoring system configuration;

[0015] The monitoring system includes a millimeter-wave radar sensor assembly, a signal processing module, a main control module, and auxiliary function modules. The main control module connects the radar sensor assembly, the signal processing module, the main control module, and the auxiliary function modules.

[0016] The millimeter-wave radar sensor assembly, as the core of non-contact signal acquisition, directly determines the frequency, power, and reliability of signal acquisition, including:

[0017] The frequency synthesizer, configured with a 77GHz operating frequency, generates precise millimeter-wave carrier signals and supports 77GHz band output.

[0018] A power amplifier amplifies the millimeter-wave signal generated by the frequency synthesizer to a suitable power level of 18dBm.

[0019] The antenna array uses a 2-transmit, 4-receive antenna array for transmitting millimeter-wave signals and receiving echo signals.

[0020] The analog-to-digital conversion module converts the analog echo signal received by the antenna array into a digital signal, and performs time-domain acquisition at a sampling frequency of 100MHz to provide raw data for subsequent digital signal processing.

[0021] The signal processing module is used to filter out interference and extract features from the acquired digital signals, and to run the NARX neural network to predict heartbeat signals, ensuring monitoring accuracy. This includes:

[0022] Digital Signal Processing (DSP) Unit: Iteratively filters out strong signals while retaining weak signals related to respiration and heartbeat; performs spectrum analysis, performs short-time Fourier transform on intermediate frequency signals to identify respiratory signals and harmonic frequencies; and performs linear filtering to remove harmonics from respiratory signals and purify the effective signal.

[0023] Neural network computing unit: runs the NARX neural network model, including model training, weight adjustment, and heartbeat signal prediction;

[0024] The auxiliary function module includes:

[0025] The power management module provides a stable power supply to all modules and supports the voltage requirements of different modules;

[0026] The storage module caches the raw digital signals collected by the analog-to-digital conversion module and the intermediate data preprocessed by the digital signal processing (DSP) unit; it stores the parameters of the trained NARX neural network model so that it does not need to be retrained every time it starts; it can optionally store the final respiratory / heartbeat monitoring results for later review.

[0027] The main control module connects and coordinates the working timing of each module, sends configuration commands for the 77GHz working frequency and 100MHz sampling frequency to the frequency synthesizer and analog-to-digital converter module; triggers the timing coordination of signal acquisition, anomaly detection, signal processing and neural network calculation; reads the working status of each module and confirms that the parameter configuration is effective;

[0028] Step 1.2, Configure the operating frequency of the millimeter-wave radar sensor;

[0029] Step 1.3, Sampling frequency configuration;

[0030] Step 1.4, System parameter verification and status confirmation.

[0031] Furthermore, step 2 includes the following steps:

[0032] Step 2.1, signal transmission;

[0033] The 77 GHz millimeter-wave signal, amplified by a power amplifier, is radiated directionally towards the monitoring area through a transmitting antenna array. The arrangement of the antenna array must be adapted to the wavelength of the 77 GHz signal, where wavelength λ = c / f, and c is the speed of light ≈ 3 × 10⁻⁶. 8 m / s, f=77×109Hz, a uniform linear array is used, the element spacing is set to λ / 2, the array is arranged horizontally and fits the center line of the monitoring area;

[0034] Step 2.2, signal reception;

[0035] When a millimeter-wave signal irradiates a human body, the slight displacement of the chest cavity caused by the human body's breathing and heartbeat will cause Doppler frequency shift and phase change in the reflected signal. These reflected signals carrying vital signs information are captured by the receiving antenna array. During the reception process, the antenna array focuses the target signal through beamforming technology and suppresses interference signals such as environmental reflections.

[0036] Step 2.3, Time Domain Acquisition;

[0037] The analog signal output from the receiving antenna is transmitted to the analog-to-digital converter module; the analog-to-digital converter module discretizes the intermediate frequency signal at a set sampling frequency of 100MHz and converts the analog signal into a digital signal;

[0038] Step 2.4, data storage;

[0039] The sampled time-domain digital signals are stored in frame format with a frame period of 10ms. The number of data points collected in a single frame is 1024, and a timestamp is added. The signals are stored in 16-bit binary format and organized into triplets of 'frame number-timestamp-signal amplitude' for easy timing alignment and tracing during subsequent signal processing.

[0040] Furthermore, step 3 includes the following steps:

[0041] Step 3.1, Data integrity verification, check whether there are any missing or lost packets in the collected data, and ensure that the data sequence is continuous and unbroken;

[0042] Read the acquisition data of the current frame from the buffer, count the number of valid data points. Valid data points are data points with complete amplitude information and no null / invalid values. Set the integrity threshold to the percentage of valid data points ≥ 95%, that is, the missing rate ≤ 5%. If the percentage of valid data points < 95%, it is determined that the data integrity is abnormal. Clear the abnormal data of the current frame, return to step 2 and re-execute the signal acquisition. If it is still abnormal after 3 retries, output a hardware alarm.

[0043] Step 3.2, verify the validity of the signal amplitude, check for problems such as signal saturation and abnormal amplitude, and ensure that the collected signal does not exceed the hardware processing range and retains weak vital signs information;

[0044] Iterate through the amplitude of all valid data points in the current frame, extract the maximum and minimum values, and set the amplitude threshold to ±1.8V. If any data point has an amplitude exceeding ±1.8V, it is determined to be a signal saturation abnormality. Clear the abnormal data of the current frame and return to step 2 to re-execute the signal acquisition. If it is still abnormal after 3 retries, output a hardware alarm.

[0045] Furthermore, step 3 also includes the following steps:

[0046] Step 3.3, Target Existence Verification: Confirm whether there are human targets in the monitoring area to avoid ineffective processing of environmental noise without targets;

[0047] Perform a simple energy detection on the current frame data: calculate the sum of squared amplitudes of all valid data points, and set the target determination threshold to be ≥ 5 times the ambient noise energy. If the sum of squared amplitudes is lower than the threshold, it means that there are no human targets in the monitoring area and only ambient noise exists. It is determined to be an abnormality of missing targets, the acquisition process is paused, and a prompt of no monitored targets is output. If it is higher than the threshold, it means that the data contains valid signals reflected by human bodies, and it is determined to be normal.

[0048] Furthermore, step 4 includes the following steps:

[0049] Step 4.1: The acquired intermediate frequency signal is processed by iteratively filtering out strong interference signals and retaining weak signals carrying vital signs information, laying the foundation for subsequent respiratory signal identification and heartbeat signal prediction.

[0050] Step 4.2: Identify the respiratory signal and its harmonics based on spectral characteristics;

[0051] Step 4.3: Use a linear filter factor to filter out the harmonics of the respiratory signal from the intermediate frequency signal to avoid frequency overlap and interference with the heartbeat signal, thereby purifying the characteristics of the heartbeat signal.

[0052] Step 4.4: Use the NARX neural network to perform in-depth analysis on the processed signal to predict the heartbeat signal.

[0053] Furthermore, step 4.1 includes the following steps:

[0054] Step 4.1.1: Set the iterative filtering threshold and adopt a three-level screening to achieve layered interference filtering from strong to weak, so as to avoid the loss of weak vital signs signals caused by a one-time strong threshold filtering.

[0055] Normalize the input single-frame intermediate frequency signal: calculate the maximum amplitude of all data points as A_max, and convert the amplitude A_i of each data point into the relative intensity S_i=A_i / A_max, S_i∈[0,1];

[0056] Fixed three-level thresholds: Level 1 threshold T1=0.6, Level 2 threshold T2=0.4, Level 3 threshold T3=0.15; T1 is used to filter out the strongest fixed interference signals, such as hard environmental reflections, T2 filters out secondary interference signals, and T3 retains weak effective signals. The relative intensity of the heartbeat signal is usually between 0.15 and 0.4.

[0057] Step 4.1.2, first-stage filtering, preliminary screening of strong interference signals;

[0058] Traverse the normalized signal sequence (S_1,S_2,...,S_m), which has a length of 1024 sampling points, where m is the number of data points per frame; retain all data points that satisfy S_i>T1, and remove data points that S_i≤0.6; obtain the signal sequence S1 after the first level of filtering, which contains only strong signal components, mainly fixed interference signals, and does not contain effective vital signs signals for the time being;

[0059] Step 4.1.3, the second-level filtering, further eliminates secondary interference based on the first-level screening, and gradually focuses on the effective signal;

[0060] Taking the signal sequence S1 after the first-level filtering as the processing object, the maximum relative intensity of the signal sequence S1 is recalculated and denoted as S1_max. Since weak signals have been removed, S1_max is still close to 1. The second-level filtering is performed according to the threshold T2=0.4, retaining the data points in S1 that satisfy S_i>0.4 and removing the data points that S_i≤0.4. The signal sequence S2 after the second-level filtering is obtained. The interference signal is further reduced, but strong interference signals still remain.

[0061] Step 4.1.4, the third stage of filtering, removes the remaining interference and retains the weak signals corresponding to breathing and heartbeat, thus completing the process of removing strong signals and retaining weak signals;

[0062] Taking the signal sequence S2 after the second-level filtering as the processing object, the maximum relative intensity S2_max of S2 is calculated; the third-level threshold T3=0.15 is used to perform filtering, retaining the data points in the signal sequence S2 that satisfy S_i>0.15 and removing the data points that S_i≤0.15; the effective data points retained after the three filterings are arranged according to the original time sequence, and the signal at the missing position is filled in by linear interpolation to avoid time sequence breakage, forming a pure signal sequence x(m) after filtering out strong interference, with a length of 1024 sampling points.

[0063] Furthermore, step 4.2 includes the following steps:

[0064] Step 4.2.1: Calculate the short-time Fourier transform of the intermediate frequency signal to convert the time-domain signal into a two-dimensional time-frequency spectrum, thereby enabling time-series tracking of frequency components and avoiding the loss of time information in the traditional Fourier transform.

[0065] Signal framing: The input time-domain signal is divided into continuous frames with a fixed window length of 2s. The 2s window length can completely cover one breathing cycle, taking into account both frequency resolution and time resolution. A 50% overlap rate is set between frames to avoid spectral distortion caused by the loss of signal at frame boundaries.

[0066] Window function selection: Hanning window is used to suppress spectral leakage;

[0067] Fourier Transform (FFT) point configuration: The FFT point count for each frame of signal is set to 2048 points;

[0068] Perform STFT: After applying a Hanning window to each frame of signal, perform a Fast Fourier Transform (FFT) to output a three-dimensional spectrum of time, frequency, and amplitude.

[0069] Step 4.2.2: Perform spectral analysis on the Fourier transform results, extract the effective frequency components from the spectrum output by STFT, eliminate noise interference, and focus on respiratory-related frequencies;

[0070] Spectral amplitude normalization: The spectral amplitude of each frame is normalized, and the maximum amplitude within a single frame is recorded as 1, which facilitates the comparison of the intensity of frequency components across frames.

[0071] Noise threshold filtering: Calculate the mean background noise in the spectrum, take the mean amplitude of the low frequency band of 0-0.05Hz, which has no vital signs. Set the noise threshold = noise mean × 3, and remove frequency components in the spectrum with amplitudes lower than the threshold to reduce noise interference.

[0072] Frequency component extraction: retain frequency points whose amplitude is greater than the threshold after filtering, record the amplitude and corresponding time frame of each frequency point, and form an effective frequency-amplitude-time list;

[0073] Cross-frame consistency verification: Since the respiratory signal frequency is stable and fluctuates ≤ ±0.1Hz in the resting state, the extracted effective frequency points are verified across frames. Frequency components that appear in 3 or more consecutive frames are retained, and false frequencies caused by instantaneous noise are eliminated to obtain a candidate frequency set.

[0074] Step 4.2.3: Identify the respiratory signal and its harmonic frequencies, distinguish the respiratory fundamental frequency from the candidate frequency set, and determine their frequency values;

[0075] Respiratory fundamental frequency identification: Select frequency points in the range of 0.1-1Hz from the candidate frequency set; the frequency point with the largest amplitude in this range is the respiratory fundamental frequency f0, because the respiratory fundamental frequency signal has the strongest energy and the amplitude is significantly higher than that of harmonics and surrounding noise;

[0076] Respiratory harmonic identification: Based on the fundamental respiratory frequency f0, select frequency points in the candidate frequency set that satisfy f=n×f0 (n=2,3,...);

[0077] Verify harmonic characteristics: The amplitude of the harmonics must meet the requirement of ≤30%×f0, the energy attenuation characteristics of respiratory harmonics, and the frequency deviation ≤±0.05Hz, to avoid misjudging similar heartbeat signals as harmonics;

[0078] Record the respiratory fundamental frequency f0 and all harmonic frequencies that meet the conditions.

[0079] Furthermore, step 4.3 includes the following steps:

[0080] Step 4.3.1: Determine the type of linear filter factor;

[0081] The linear filtering factor corresponds to the unit impulse response of the linear IIR notch filter. The respiratory harmonics are discrete frequency points and are close to the frequency of the heartbeat signal. Therefore, a narrow bandwidth and high attenuation notch filter is required to accurately suppress specific discrete frequencies and avoid over-filtering that could distort the heartbeat signal. This meets the subsequent NARX neural network's requirements for analyzing time-series signals.

[0082] Step 4.3.2: Configure notch filter parameters based on breathing harmonic frequency;

[0083] The center frequency fn is consistent with the breathing harmonic frequency, precisely targeting the harmonic frequency to be suppressed, ensuring that the filtering target is clear;

[0084] The bandwidth BW is a fixed narrow bandwidth of 0.1Hz, which only covers a narrow range near the harmonic frequency to avoid affecting adjacent heartbeat signals;

[0085] The attenuation A is ≥40dB to ensure that the harmonic amplitude is attenuated to less than 1% of the original amplitude, completely suppressing harmonic interference and avoiding residual harmonics from affecting the extraction of heartbeat signals.

[0086] The filter is of order 4, which balances the filtering effect with the computational load. Order 4 can achieve high attenuation and low computational load, making it suitable for real-time monitoring requirements.

[0087] If multiple breathing harmonics are identified, an independent notch filter is designed for each harmonic frequency and cascaded in the order of high frequency harmonics → low frequency harmonics to avoid mutual interference between multiple filters, ensure that each harmonic is fully suppressed, and reduce the cumulative distortion of the overall signal phase.

[0088] Step 4.3.3: Generate the linear filter factor;

[0089] Based on the selected IIR notch filter and parameters, the transfer function is calculated using the filter design formula:

[0090] Transfer function of IIR notch filter: ;

[0091] in:

[0092] Normalized center angular frequency f n The frequency is the breathing harmonic frequency, and the actual sampling frequency is Fs'=1kHz;

[0093] Attenuation coefficient r≈0.9997, ensuring narrow bandwidth and sufficient attenuation;

[0094] The inverse Z-transform of the transfer function H(z) yields the unit impulse response h(n), which is the linear filtering factor. The impulse response h(n) of the 4th-order IIR filter is a finite-length sequence with a length of 30 sampling points, balancing filtering effect and computational efficiency.

[0095] Standardization of the filter factor: h(n) is normalized by adjusting the sum of all elements to 1 to avoid the overall amplitude of the signal being attenuated after filtering;

[0096] Verify the frequency response of the filter factor: Analyze the frequency response of h(n) using Fast Fourier Transform (FFT) to ensure that f n The amplitude attenuation is ≥40dB at the frequency of the signal and ≤0.5dB at the frequency of the heartbeat signal.

[0097] Step 4.3.4: Apply a filtering factor to filter out respiratory harmonics. Combine the filtering factor with the input signal through time-domain convolution to accurately suppress respiratory harmonics.

[0098] Convolution operation: Performing a linear convolution between the pure signal sequence x(m) and the filter factor h(n) yields the filtered signal y(n). ;

[0099] Where N is the length of the filter factor h(n), and n is the index of the signal sampling point;

[0100] Pure signal sequence x(m) and f n Consistent breathing harmonic components are suppressed by the frequency response of h(n), while other frequency components are preserved with almost no distortion;

[0101] If multiple harmonics exist, perform multi-harmonic cascade filtering:

[0102] First-stage filtering: Convolve x(m) with the filter factor for 3f0 to obtain the signal x1(m) after filtering out 3f0;

[0103] Second-stage filtering: Convolve x1(m) with the filter factor for 2f0 to obtain the signal y(n) after filtering out all harmonics;

[0104] After each filtering stage, the following verification is required: calculate the amplitude of the filtered signal at the harmonic frequency to ensure that the attenuation is ≥40dB. If the standard is not met, adjust the r value of the filtering factor.

[0105] Signal edge processing: Convolution operation can cause edge distortion at both ends of the signal. Zero padding is used to process it: N / 2 zeros are added before and after the input signal x(m). After convolution, the effective signal segment in the middle is truncated to ensure that the length of the output signal y(n) is consistent with the input signal and the timing is continuous.

[0106] Furthermore, step 4.4 includes the following steps:

[0107] Step 4.4.1: Construct the NARX neural network model;

[0108] The NARX neural network model structure is as follows:

[0109] The input layer is 1 layer with an input dimension of 100. The input is time series data extracted by a sliding time window: the window length is 100 sampling points, that is, the input vector dimension is 100×1, which is suitable for capturing time series features.

[0110] The hidden layers consist of 5 layers with 80 nodes each. The activation function is ReLU, which solves the gradient vanishing problem and has low computational cost, making it suitable for embedded hardware. Batch normalization is added to each layer to accelerate training convergence and improve the model's generalization feedback.

[0111] The connection layer is 1 layer, and output feedback is introduced: the predicted heartbeat signal amplitude output by the model at the previous time step is fed back to the input layer at the current time step, and after being concatenated with the original signal, it is input into the hidden layer to strengthen the temporal dependency relationship and fit the continuous characteristics of the heartbeat signal.

[0112] The output layer is 1 layer with an output dimension of 1. The output is the amplitude of the heartbeat signal at the current moment. The activation function is Sigmoid, which is mapped to the [0,1] interval. The true amplitude can be restored by inverse normalization.

[0113] Weight initialization: He normal distribution is used to avoid training failure due to initial weights being too large or too small;

[0114] Bias initialization: All layer bias terms are initialized to 0.01 to accelerate convergence during the initial training phase;

[0115] Step 4.4.2: Use the Adam optimization algorithm to train the model and adjust the network parameters using training data to enable the model to predict heartbeat signals.

[0116] Training dataset preparation:

[0117] Data source: Synchronous data from 100 subjects, pure signals acquired by millimeter-wave radar + real heartbeat signals acquired by ECG;

[0118] Data partitioning: The data is divided into training, validation, and test sets in a 7:2:1 ratio; the training set is used for parameter updates, the validation set is used to monitor overfitting, and the test set is used for final performance evaluation.

[0119] Data augmentation: Add small noise and time stretching to the training set signal. The amplitude of the small noise is ≤5% and the length of the time stretching is ±10%, which improves the model's anti-interference ability and adapts it to the environmental noise in actual monitoring.

[0120] Training parameter configuration: The learning rate of the Adam optimization algorithm is fixed at 0.0005 to avoid oscillations caused by an excessively large learning rate and slow convergence caused by an excessively small learning rate;

[0121] Loss function: Mean Squared Error (MSE) formula is as follows B represents the batch size. For real labels, The model predicts values ​​to adapt to continuous value regression tasks;

[0122] Training batch and iteration: Batch size = 32, balancing training speed and memory usage; Number of iterations = 5000, if the validation set MSE does not decrease for 50 consecutive iterations, the training will stop early to avoid overfitting;

[0123] Training process:

[0124] Forward propagation: The training set input vector is fed into the network, and the signal is fused through hidden layer calculation and feedback connection to output the predicted heartbeat signal amplitude;

[0125] Backpropagation: Based on the MSE loss value, the gradient of each layer's parameters is calculated using the chain rule, and the weights and biases are updated using the Adam algorithm.

[0126] Momentum parameters: First-order momentum β1=0.9, second-order momentum β2=0.9, default values ​​balance gradient stability and convergence speed;

[0127] Weight update formula: η is the learning rate. , These are the first- and second-order momentum correction values. Avoid having a denominator of 0;

[0128] Every 100 iterations, the MSE is calculated using the validation set. If the MSE continues to rise, the learning rate decay is triggered with a decay coefficient of 0.9 to suppress overfitting.

[0129] Step 4.4.3: Adjust the model weights after training is complete;

[0130] Weighting importance assessment:

[0131] Calculate the importance values ​​of each weight parameter during training. T is the number of training iterations. Let i be the weight;

[0132] The higher the importance value, the greater the contribution of the weight to fitting the historical training data, and the more it needs to be protected; conversely, a lower importance value indicates a non-critical weight that can be adjusted flexibly.

[0133] Weighting adjustment rules:

[0134] Importance value The weights, with a threshold θ=0.001: only small adjustments are allowed, with the adjustment magnitude ≤ 5% of the initial value, to avoid compromising the model's ability to fit existing subject data;

[0135] Importance value The weights can be freely adjusted to adapt to the signal characteristics of new monitoring scenarios;

[0136] Step 4.4.3: Use the adjusted model to predict the heartbeat signal, apply the trained and optimized model to the actual monitored signal, and output the final heartbeat signal and heart rate value.

[0137] Preprocessing before prediction:

[0138] Input signal interception: For the real-time signal output in step 4.3, the input vector is intercepted according to the sliding time window (). The length of the sliding time window is 100 sampling points and the step size is 1 sampling point to ensure the continuity of timing.

[0139] Normalization: The input vector is normalized using the normalization parameters of the training set, the mean μ and the standard deviation σ. To avoid prediction distortion caused by inconsistent distribution;

[0140] Real-time prediction process:

[0141] Initialization: The first 100 sampling points are used as initial input, and the model outputs the first predicted value. ;

[0142] Time series iteration: At time t, the input vector = real-time signal from time t to t+99 + predicted value at time t-1. (Feedback connection), Model output ;

[0143] Inverse normalization: transforming the predicted values Restored to the amplitude of the actual heartbeat signal: ;

[0144] Heart rate calculation: Perform spectral analysis (FFT) on a 20-second continuous predicted heartbeat signal to extract the peak frequency. (Unit: Hz), Heart rate value = ×60 (unit: times / minute), and perform moving average smoothing, with a window length of 5 seconds to reduce instantaneous fluctuations;

[0145] Prediction results output:

[0146] Real-time output: The heart rate value is updated once per second, and the time-domain waveform of the heartbeat signal is output synchronously for display on the monitoring terminal;

[0147] Abnormal alarm: If the heart rate value exceeds the normal range and continues for 3 seconds, an alarm prompt will be output;

[0148] Data storage: Historical data is stored in the format of timestamp-heart rate value-heartbeat signal amplitude for easy subsequent traceability and analysis.

[0149] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0150] 1. Overcoming the limitations of contact-based monitoring, adapting to diverse scenario needs;

[0151] The 77GHz millimeter-wave radar sensor enables non-contact signal acquisition, eliminating the need for direct contact with the human body. This solves the discomfort caused by prolonged wear of traditional contact sensors and is particularly suitable for monitoring the needs of special populations such as infants and burn patients. It also meets the core requirements of seamless monitoring in smart homes, public spaces, and other scenarios.

[0152] 2. Improve monitoring accuracy and stability, and address key technical pain points;

[0153] By integrating cutting-edge signal processing technology with NARX neural networks, the limitations of hardware and frequency estimation algorithms in existing systems are effectively overcome, significantly improving the problem of unstable measurement results and enhancing the accuracy and reliability of signal acquisition.

[0154] The NARX neural network has powerful time-series information processing capabilities. Through the output feedback mechanism, it strengthens the capture of the dependence on continuous signals, improves the adaptability to data sequences with large fluctuations, and greatly improves the instability of heart rate prediction in traditional lightweight neural networks.

[0155] 3. Achieve system miniaturization and cost reduction, and optimize user experience;

[0156] High-frequency sampling technology (100MHz sampling frequency) is used to reduce signal acquisition time and avoid detection result delay. At the same time, through the compact system architecture (radar sensor components + signal processing module + main control module + auxiliary function module), the overall system size and cost are reduced, achieving the goals of miniaturization and low cost, and eliminating the dependence on PC in existing technologies.

[0157] The storage module caches the parameters of the trained neural network model, eliminating the need for retraining on each startup, thus improving system startup and running efficiency and further optimizing the user experience.

[0158] 4. Enhance signal processing capabilities to achieve real-time and accurate monitoring;

[0159] By employing a series of signal processing procedures, including three-level threshold layering to filter out strong interference signals, short-time Fourier transform (STFT) to accurately identify respiratory fundamental frequency and harmonics, and cascaded fourth-order IIR narrowband notch filters to suppress harmonic interference, the speed and accuracy of signal processing are greatly improved, successfully achieving real-time, accurate monitoring and in-depth analysis of respiratory and heartbeat signals.

[0160] 5. Expanding application scenarios, possessing high practical value and promising prospects for promotion;

[0161] It provides a complete solution integrating real-time heart rate output, heartbeat signal time-domain waveform display, abnormal alarm (triggered when heart rate exceeds the normal range for 3 seconds), and historical data storage and traceability. It is applicable to many fields such as medical monitoring, smart home, and security, and provides an efficient and reliable new technical path for non-invasive vital sign monitoring. It has important practical value and broad application prospects. Detailed Implementation

[0162] An example of a millimeter-wave vital sign monitoring method based on Narx neural networks includes the following steps:

[0163] Step 1: Initialize the monitoring system parameters and configure the millimeter-wave radar sensor to operate at a frequency of 77 GHz and a sampling frequency of 100 MHz.

[0164] Step 1.1, Monitoring system configuration;

[0165] The monitoring system includes a millimeter-wave radar sensor assembly, a signal processing module, a main control module, and auxiliary function modules. The main control module connects the radar sensor assembly, the signal processing module, the main control module, and the auxiliary function modules.

[0166] The millimeter-wave radar sensor assembly, as the core of non-contact signal acquisition, directly determines the frequency, power, and reliability of signal acquisition, including:

[0167] The frequency synthesizer, configured with a 77GHz operating frequency, generates precise millimeter-wave carrier signals, supports 77GHz band output, and has high frequency stability (avoiding signal offset that leads to acquisition distortion). Frequency parameters can be configured via commands from the main control module.

[0168] The power amplifier amplifies the millimeter-wave signal generated by the frequency synthesizer to a suitable power level of 18dBm, ensuring that the signal has sufficient transmission distance and penetration capability to meet the coverage requirements of the monitoring area.

[0169] The antenna array, employing a 2-transmit, 4-receive antenna array, enhances signal reception stability. It is used for transmitting millimeter-wave signals and receiving echo signals, sending amplified 77GHz signals to the monitoring area and capturing reflected signals modulated by human breathing and heartbeat.

[0170] The analog-to-digital converter module converts the analog echo signal received by the antenna array into a digital signal, and performs time-domain acquisition at a sampling frequency of 100MHz to provide raw data for subsequent digital signal processing.

[0171] The signal processing module is used to filter out interference and extract features from the acquired digital signals, and to run the NARX neural network to predict heartbeat signals, ensuring monitoring accuracy. This includes:

[0172] The digital signal processing (DSP) unit iteratively filters out strong signals while retaining weak signals related to respiration and heartbeat; spectrum analysis performs short-time Fourier transform on the intermediate frequency signal to identify respiratory signals and harmonic frequencies; and linear filtering removes harmonics from the respiratory signal and purifies the effective signal.

[0173] Neural network computing unit: runs NARX neural network models, including model training, weight adjustment, and heartbeat signal prediction, and can be equipped with FPGA, GPU or dedicated AI acceleration chip.

[0174] The auxiliary function module includes:

[0175] The power management module provides a stable power supply to all modules and supports the voltage requirements of different modules.

[0176] The storage module caches the raw digital signals collected by the analog-to-digital conversion module and the intermediate data preprocessed by the digital signal processing (DSP) unit; it stores the parameters of the trained NARX neural network model so that it does not need to be retrained every time it starts; it can optionally store the final respiratory / heartbeat monitoring results (such as heart rate value and monitoring timestamp) for later traceability.

[0177] The main control module connects and coordinates the working timing of each module, sends configuration commands for a 77GHz working frequency and a 100MHz sampling frequency to the frequency synthesizer and analog-to-digital converter module; triggers the timing coordination of signal acquisition, anomaly detection, signal processing and neural network calculation; reads the working status of each module, confirms that the parameter configuration is effective, and avoids abnormal processes. An embedded processor (such as ARM Cortex-M series), FPGA or MCU can be selected, and it needs to have a multi-module communication interface.

[0178] Step 1.2, Configure the operating frequency of the millimeter-wave radar sensor;

[0179] The main control module sends a frequency configuration command to the frequency synthesizer module of the millimeter-wave radar sensor through the communication link, specifying that the target operating frequency is 77GHz. This frequency belongs to the millimeter-wave band and has both signal penetration and resolution, which meets the requirements of non-contact signal acquisition for vital sign monitoring.

[0180] The frequency synthesizer responds to commands and generates a precise 77 GHz millimeter-wave carrier signal.

[0181] Step 1.3, Sampling frequency configuration;

[0182] The main control module sends a sampling frequency configuration command to the analog-to-digital conversion module of the millimeter-wave radar sensor, setting the sampling frequency to 100MHz. High-frequency sampling can reduce the time interval of signal acquisition, improve the ability to capture details of timing signals, and provide a guarantee for subsequent real-time monitoring and reduced latency.

[0183] Configure the sampling clock of the analog-to-digital converter module: Based on the 100MHz sampling frequency requirement, calibrate the sampling clock source of the ADC to ensure a stable sampling period, i.e., a sampling interval of 1 / 100MHz = 10ns, to avoid sampling point misalignment due to clock jitter.

[0184] Set the sampling accuracy of the ADC: In view of the detection requirements of vital signs signals (breathing and heartbeat signals are low-frequency and weak signals), the resolution of the ADC is configured to 16 bits to balance sampling accuracy and data transmission efficiency.

[0185] Step 1.4, System parameter verification and status confirmation;

[0186] The main control module reads the configuration confirmation information returned by the sensor, including the current values ​​of the actual operating frequency and sampling frequency, and compares them with the preset parameters (77GHz, 100MHz) to confirm whether the configuration is effective.

[0187] Check the sensor's operating status indicator light or status register to verify whether the sensor has entered the data acquisition state.

[0188] Step 1 clearly configures the millimeter-wave radar with a 77GHz operating frequency and a 100MHz high-frequency sampling frequency. Combined with the 16-bit ADC sampling accuracy, this ensures both signal penetration and resolution, while also capturing timing signal details through high-frequency sampling, providing hardware support for real-time monitoring and reducing latency.

[0189] A compact architecture consisting of radar sensor components, signal processing modules, main control modules, and auxiliary function modules is constructed. The main control module coordinates timing and parameter verification to ensure that the configuration of each module is effective and to avoid abnormal processes, thus laying a structural foundation for the miniaturization and cost reduction of the system.

[0190] The storage module caches raw data and trained model parameters, eliminating the need for repeated training and improving system startup and operation efficiency.

[0191] Step 2: Perform signal acquisition by transmitting and receiving millimeter-wave signals to the monitoring area through the millimeter-wave radar sensor assembly;

[0192] Step 2.1, signal transmission;

[0193] The 77 GHz millimeter-wave signal, amplified by a power amplifier, is radiated directionally towards the monitoring area through a transmitting antenna array. The arrangement of the antenna array must be adapted to the wavelength of the 77 GHz signal, where wavelength λ = c / f, and c is the speed of light ≈ 3 × 10⁻⁶. 8 m / s, f=77×10 9The signal strength is 1 Hz, and a uniform linear array is used with an element spacing of λ / 2. The array is arranged horizontally and fits the center line of the monitoring area to ensure that the signal covers the target monitoring area (such as a range with a diameter of 1-3m) and the energy is concentrated to improve the signal-to-noise ratio.

[0194] Step 2.2, signal reception;

[0195] When a millimeter-wave signal irradiates a human body, the slight displacement of the chest cavity caused by the human body's breathing and heartbeat will cause Doppler frequency shift and phase change in the reflected signal. These reflected signals carrying vital signs information are captured by the receiving antenna array. During the reception process, the antenna array focuses the target signal through beamforming technology and suppresses interference signals such as environmental reflections.

[0196] Step 2.3, Time Domain Acquisition;

[0197] The analog signal output from the receiving antenna is transmitted to the analog-to-digital converter module; the analog-to-digital converter module discretizes the intermediate frequency signal at a set sampling frequency of 100MHz and converts the analog signal into a digital signal.

[0198] Step 2.4, data storage;

[0199] The sampled time-domain digital signals are stored in frame format with a frame period of 10ms. The number of data points collected in a single frame is 1024, and a timestamp is added. The signals are stored in 16-bit binary format and organized into triplets of 'frame number-timestamp-signal amplitude' for easy timing alignment and tracing during subsequent signal processing.

[0200] Step 2 employs a 2-transmit, 4-receive antenna array and beamforming technology to directionally radiate 77GHz millimeter-wave signals, accurately capturing the frequency shift and phase change of reflected signals caused by human breathing and heartbeat. This method requires no contact with the human body, breaking through the limitations of traditional contact monitoring and making it suitable for special populations and non-contact monitoring scenarios.

[0201] Data is stored in a format with a 10ms frame period and 1024 data points / frame, with timestamps added to ensure data timing integrity and provide support for timing alignment and tracing in subsequent signal processing.

[0202] Step 3: Determine if there are any abnormalities after the data collection is completed. If there are no abnormalities, proceed to Step 4. Through three-layer verification and hierarchical processing, the data integrity, signal validity, and target existence are gradually screened to ensure that the original data input to Step 4 is of high quality and valuable. At the same time, the monitoring efficiency and reliability are balanced through flexible process jumps.

[0203] Step 3.1, Data integrity verification, check whether there are any missing or lost packets in the collected data, and ensure that the data sequence is continuous and unbroken;

[0204] Read the current frame's acquisition data from the buffer, count the number of valid data points. Valid data points are data points with complete amplitude information and no null / invalid values. Set the integrity threshold to the percentage of valid data points ≥ 95%, i.e., the missing rate ≤ 5%. If the percentage of valid data points < 95%, it is determined to be a data integrity abnormality. Clear the abnormal data of the current frame, return to step 2, and re-execute the signal acquisition. If it is still abnormal after 3 retries, output a hardware alarm.

[0205] Step 3.2, verify the validity of the signal amplitude, check for problems such as signal saturation and abnormal amplitude, and ensure that the collected signal does not exceed the hardware processing range and retains weak vital signs information;

[0206] Iterate through the amplitude of all valid data points in the current frame, extract the maximum and minimum values, and set the amplitude threshold to ±1.8V. If any data point has an amplitude exceeding ±1.8V, it is determined to be a signal saturation abnormality. Clear the abnormal data of the current frame and return to step 2 to re-execute the signal acquisition. If it is still abnormal after 3 retries, output a hardware alarm.

[0207] Step 3.3, Target Existence Verification: Confirm whether there are human targets in the monitoring area to avoid ineffective processing of environmental noise without targets;

[0208] Perform a simple energy detection on the current frame data: calculate the sum of squared amplitudes of all valid data points, and set the target determination threshold to be ≥ 5 times the ambient noise energy. If the sum of squared amplitudes is lower than the threshold, it means that there are no human targets in the monitoring area and only ambient noise exists. It is determined to be an abnormality of missing targets, the acquisition process is paused, and a prompt of no monitored targets is output. If it is higher than the threshold, it means that the data contains valid signals reflected by human bodies, and it is determined to be normal.

[0209] Through a three-level verification process—data integrity (valid data point ratio ≥ 95%), signal amplitude validity (±1.8V threshold), and target existence (sum of squared amplitudes ≥ 5 times the ambient noise)—high-quality data is screened layer by layer, and invalid data such as missing, saturated, and targetless data are eliminated to avoid subsequent invalid processing.

[0210] A tiered processing mechanism of 3 retries + alarm / pause is adopted to ensure data reliability while avoiding excessive retries from affecting monitoring efficiency.

[0211] Step 4: Perform signal processing and analysis. The high-quality acquired signals are precisely processed and deeply analyzed to purify vital sign-related signals. The NARX neural network is used to achieve real-time and accurate prediction of heartbeat signals and heart rate. At the same time, it supports abnormal alarms and historical data storage, providing reliable data support for multi-scenario monitoring.

[0212] Step 4.1: The acquired intermediate frequency signal is processed by iteratively filtering out strong interference signals and retaining weak signals carrying vital signs information, laying the foundation for subsequent respiratory signal identification and heartbeat signal prediction.

[0213] Step 4.1.1: Set the iterative filtering threshold and adopt a three-level screening to achieve layered interference filtering from strong to weak, so as to avoid the loss of weak vital signs signals caused by a one-time strong threshold filtering.

[0214] Normalize the input single-frame intermediate frequency signal: calculate the maximum amplitude of all data points as A_max, and convert the amplitude A_i of each data point into the relative intensity S_i=A_i / A_max, S_i∈[0,1];

[0215] The three fixed threshold levels are: Level 1 threshold T1=0.6, Level 2 threshold T2=0.4, and Level 3 threshold T3=0.15. T1 is used to filter out the strongest fixed interference signals, such as hard environmental reflections; T2 filters out secondary interference signals; and T3 retains weak effective signals. The relative intensity of the heartbeat signal is usually between 0.15 and 0.4.

[0216] Step 4.1.2, first-stage filtering, preliminary screening of strong interference signals;

[0217] Traverse the normalized signal sequence (S_1,S_2,...,S_m), where m is the number of data points in a single frame; retain all data points that satisfy S_i>T1 and remove data points that S_i≤0.6; obtain the signal sequence S1 after the first level of filtering, which contains only strong signal components, mainly fixed interference signals, and does not contain effective vital signs signals for the time being.

[0218] Step 4.1.3, the second-level filtering, further eliminates secondary interference based on the first-level screening, and gradually focuses on the effective signal;

[0219] Taking the signal sequence S1 after the first-level screening as the processing object, the maximum relative intensity of the signal sequence S1 is recalculated and denoted as S1_max. Since weak signals have been removed, S1_max is still close to 1. The second-level threshold T2=0.4 is used to perform screening, retaining the data points in S1 that satisfy S_i>0.4 and removing the data points that S_i≤0.4. The signal sequence S2 after the second-level screening is obtained. The interference signal is further reduced, but strong interference signals still remain.

[0220] Step 4.1.4, the third stage of filtering, removes the remaining interference and retains the weak signals corresponding to breathing and heartbeat, thus completing the process of removing strong signals and retaining weak signals;

[0221] Taking the signal sequence S2 after the second-level screening as the processing object, the maximum relative intensity S2_max of S2 is calculated; the third-level threshold T3=0.15 is used to perform screening, retaining the data points in the signal sequence S2 that satisfy S_i>0.15 and removing the data points that S_i≤0.15; the effective data points retained after the three screenings are arranged according to the original time sequence, and the signal at the missing position is filled in by linear interpolation to avoid time sequence breakage, forming a pure signal sequence x(m) after filtering out strong interference.

[0222] Step 4.1 uses a three-level threshold layer to filter out strong interference such as hard environmental reflections, retains weak respiratory and heartbeat signals, and uses linear interpolation to complete the timing sequence to avoid loss of effective signals and provide clean input for subsequent signal purification.

[0223] Step 4.2: Identify the respiratory signal and its harmonics based on spectral characteristics;

[0224] Step 4.2.1: Calculate the short-time Fourier transform of the intermediate frequency signal to convert the time-domain signal into a two-dimensional time-frequency spectrum, thereby enabling time-series tracking of frequency components and avoiding the loss of time information in the traditional Fourier transform.

[0225] Signal framing: The input time-domain signal is divided into continuous frames with a fixed window length of 2s. The 2s window length can completely cover one breathing cycle, taking into account both frequency resolution and time resolution. A 50% overlap rate is set between frames to avoid spectral distortion caused by the loss of signal at frame boundaries.

[0226] Window function selection: Hanning window is used to suppress spectral leakage;

[0227] Fourier Transform (FFT) point configuration: The FFT point count for each frame of signal is set to 2048 points;

[0228] Perform STFT: After applying a Hanning window to each frame of signal, perform a Fast Fourier Transform (FFT) to output a three-dimensional spectrum of time, frequency, and amplitude.

[0229] Step 4.2.2: Perform spectral analysis on the Fourier transform results, extract the effective frequency components from the spectrum output by STFT, eliminate noise interference, and focus on respiratory-related frequencies;

[0230] Spectral amplitude normalization: The spectral amplitude of each frame is normalized, and the maximum amplitude within a single frame is recorded as 1, which facilitates the comparison of the intensity of frequency components across frames.

[0231] Noise threshold filtering: Calculate the mean background noise in the spectrum, take the mean amplitude of the low frequency band of 0-0.05Hz, which has no vital signs. Set the noise threshold = noise mean × 3, and remove frequency components in the spectrum with amplitudes lower than the threshold to reduce noise interference.

[0232] Frequency component extraction: retain frequency points whose amplitude is greater than the threshold after filtering, record the amplitude and corresponding time frame of each frequency point, and form an effective frequency-amplitude-time list;

[0233] Cross-frame consistency verification: Since the respiratory signal frequency is stable and fluctuates ≤ ±0.1Hz in the resting state, the extracted effective frequency points are verified across frames. Frequency components that appear in 3 or more consecutive frames are retained, and false frequencies caused by instantaneous noise are eliminated to obtain a candidate frequency set.

[0234] Step 4.2.3: Identify the respiratory signal and its harmonic frequencies, distinguish the respiratory fundamental frequency from the candidate frequency set, and determine their frequency values;

[0235] Respiratory fundamental frequency identification: Select frequency points in the range of 0.1-1Hz from the candidate frequency set; the frequency point with the largest amplitude in this range is the respiratory fundamental frequency f0, because the respiratory fundamental frequency signal has the strongest energy and the amplitude is significantly higher than that of harmonics and surrounding noise;

[0236] Respiratory harmonic identification: Based on the fundamental respiratory frequency f0, select frequency points in the candidate frequency set that satisfy f=n×f0 (n=2,3,...);

[0237] Verify harmonic characteristics: The amplitude of the harmonics must meet the requirement of ≤30%×f0, the energy attenuation characteristics of respiratory harmonics, and the frequency deviation ≤±0.05Hz, to avoid misjudging similar heartbeat signals as harmonics;

[0238] Record the respiratory fundamental frequency f0 and all harmonic frequencies that meet the conditions.

[0239] Step 4.3: Use a linear filter factor to filter out the harmonics of the respiratory signal from the intermediate frequency signal to avoid frequency overlap and interference with the heartbeat signal, thereby purifying the characteristics of the heartbeat signal.

[0240] Step 4.3.1: Determine the type of linear filter factor;

[0241] The linear filtering factor corresponds to the unit impulse response of the linear IIR notch filter. The breathing harmonics are discrete frequency points and are close to the frequency of the heartbeat signal. Therefore, a narrow bandwidth and high attenuation notch filter is required to accurately suppress specific discrete frequencies (breathing harmonics) and avoid over-filtering that could distort the heartbeat signal. This meets the subsequent NARX neural network's requirements for analyzing time-series signals.

[0242] Step 4.3.2: Configure notch filter parameters based on breathing harmonic frequency;

[0243] The center frequency fn is consistent with the breathing harmonic frequency, precisely targeting the harmonic frequency to be suppressed, ensuring that the filtering target is clear;

[0244] The bandwidth BW is a fixed narrow bandwidth of 0.1Hz, which only covers a narrow range near the harmonic frequency to avoid affecting adjacent heartbeat signals;

[0245] The attenuation A is ≥40dB to ensure that the harmonic amplitude is attenuated to less than 1% of the original amplitude, completely suppressing harmonic interference and avoiding residual harmonics from affecting the extraction of heartbeat signals.

[0246] The filter is of order 4, which balances the filtering effect with the computational load. Order 4 can achieve high attenuation and low computational load, making it suitable for real-time monitoring requirements.

[0247] If multiple breathing harmonics are identified, an independent notch filter is designed for each harmonic frequency and cascaded in the order of high-frequency harmonics → low-frequency harmonics to avoid mutual interference between multiple filters, ensure that each harmonic is fully suppressed, and reduce the cumulative distortion of the overall signal phase.

[0248] Step 4.3.3: Generate the linear filter factor;

[0249] Based on the selected IIR notch filter and parameters, the transfer function is calculated using the filter design formula:

[0250] Transfer function of IIR notch filter: ;

[0251] in:

[0252] Normalized center angular frequency f n The frequency is the breathing harmonic frequency, and the actual sampling frequency is Fs'=1kHz;

[0253] Attenuation coefficient r≈0.9997, ensuring narrow bandwidth and sufficient attenuation.

[0254] The inverse Z-transform of the transfer function H(z) yields the unit impulse response h(n), which is the linear filtering factor. The impulse response h(n) of the 4th-order IIR filter is a finite-length sequence with a length of 30 sampling points, balancing filtering effect and computational efficiency.

[0255] Standardization of the filter factor: h(n) is normalized by adjusting the sum of all elements to 1 to avoid attenuation of the overall signal amplitude after filtering.

[0256] Verify the frequency response of the filter factor: Analyze the frequency response of h(n) using Fast Fourier Transform (FFT) to ensure that f n The amplitude attenuation is ≥40dB at the high frequency and ≤0.5dB at the low frequency of the heartbeat signal.

[0257] Step 4.3.4: Apply a filtering factor to filter out respiratory harmonics. Combine the filtering factor with the input signal through time-domain convolution to accurately suppress respiratory harmonics.

[0258] Convolution operation: Performing a linear convolution between the pure signal sequence x(m) and the filter factor h(n) yields the filtered signal y(n). ;

[0259] Where N is the length of the filter factor h(n), and n is the index of the signal sampling point;

[0260] Pure signal sequence x(m) and f n Consistent breathing harmonic components are suppressed by the frequency response of h(n), while other frequency components are preserved with almost no distortion.

[0261] If multiple harmonics exist, perform multi-harmonic cascade filtering:

[0262] First-stage filtering: Convolve x(m) with the filter factor for 3f0 to obtain the signal x1(m) after filtering out 3f0;

[0263] Second-stage filtering: Convolve x1(m) with the filter factor for 2f0 to obtain the signal y(n) after filtering out all harmonics;

[0264] After each filtering stage, verification is required: calculate the amplitude of the filtered signal at the harmonic frequency to ensure attenuation ≥ 40dB. If the standard is not met, adjust the r value of the filtering factor.

[0265] Signal edge processing: Convolution operation can cause edge distortion at both ends of the signal. Zero padding is used to process it: N / 2 zeros are added before and after the input signal x(m). After convolution, the effective signal segment in the middle is truncated to ensure that the length of the output signal y(n) is consistent with the input signal and the timing is continuous.

[0266] Steps 4.2-4.3 use Short Time Fourier Transform (STFT) to accurately identify the respiratory fundamental frequency and harmonics, and use cascaded fourth-order IIR narrowband notch filters to suppress harmonics, avoid their overlap with the heartbeat signal frequency, purify the heartbeat signal characteristics, and solve the frequency interference problem.

[0267] Step 4.4: Use the NARX neural network to perform in-depth analysis on the processed signal to predict the heartbeat signal;

[0268] Step 4.4.1: Construct the NARX neural network model;

[0269] The NARX neural network model structure is as follows:

[0270] The input layer is 1 layer with an input dimension of 100. The input is time series data extracted by a sliding time window: the window length is 100 sampling points, that is, the input vector dimension is 100×1, which is suitable for capturing time series features.

[0271] The hidden layers consist of 5 layers with 80 nodes each. The activation function is ReLU, which solves the gradient vanishing problem and has low computational cost, making it suitable for embedded hardware. Batch normalization is added to each layer to accelerate training convergence and improve the model's generalization feedback.

[0272] The connection layer is 1 layer, and output feedback is introduced: the predicted heartbeat signal amplitude output by the model at the previous time step is fed back to the input layer at the current time step, and after being concatenated with the original signal, it is input into the hidden layer to strengthen the temporal dependency relationship and fit the continuous characteristics of the heartbeat signal.

[0273] The output layer is 1 layer with an output dimension of 1. The output is the amplitude of the heartbeat signal at the current moment (a continuous value). The activation function is Sigmoid, which is mapped to the [0,1] interval. The true amplitude can be restored by inverse normalization.

[0274] Weight initialization: He normal distribution is used to avoid training failure due to initial weights being too large or too small;

[0275] Bias initialization: All layer bias terms are initialized to 0.01 to accelerate convergence during the initial training phase.

[0276] Step 4.4.2: Use the Adam optimization algorithm to train the model and adjust the network parameters using training data to enable the model to predict heartbeat signals.

[0277] Training dataset preparation:

[0278] Data source: Synchronous data from 100 subjects (aged 20-60 years, covering different body types and weights), pure signals acquired by millimeter-wave radar + real heartbeat signals acquired by electrocardiogram (ECG);

[0279] Data partitioning: The data is divided into training, validation, and test sets in a 7:2:1 ratio; the training set is used for parameter updates, the validation set is used to monitor overfitting, and the test set is used for final performance evaluation.

[0280] Data augmentation: Add small noise and time stretching to the training set signal. The amplitude of the small noise is ≤5% and the length of the time stretching is ±10%, which improves the model's anti-interference ability and adapts it to the environmental noise in actual monitoring.

[0281] Training parameter configuration: The learning rate of the Adam optimization algorithm is fixed at 0.0005 to avoid oscillations caused by an excessively large learning rate and slow convergence caused by an excessively small learning rate;

[0282] Loss function: Mean Squared Error (MSE) formula is as follows B represents the batch size. For real labels, The model predicts values ​​to adapt to continuous value regression tasks;

[0283] Training batch and iteration: Batch size = 32, balancing training speed and memory usage; Number of iterations = 5000, if the validation set MSE does not decrease for 50 consecutive iterations, the training will stop early to avoid overfitting.

[0284] Training process:

[0285] Forward propagation: The training set input vector is fed into the network, and the signal is fused through hidden layer calculation and feedback connection to output the predicted heartbeat signal amplitude;

[0286] Backpropagation: Based on the MSE loss value, the gradient of each layer's parameters is calculated using the chain rule, and the weights and biases are updated using the Adam algorithm.

[0287] Momentum parameters: First-order momentum β1=0.9, second-order momentum β2=0.9, default values ​​balance gradient stability and convergence speed;

[0288] Weight update formula: η is the learning rate. , These are the first- and second-order momentum correction values. Avoid having a denominator of 0;

[0289] Every 100 iterations, the MSE is calculated using the validation set. If the MSE continues to rise, the learning rate decay is triggered with a decay coefficient of 0.9 to suppress overfitting.

[0290] Step 4.4.3: Adjust the model weights after training is complete;

[0291] Weighting importance assessment:

[0292] Calculate the importance values ​​of each weight parameter during training. T is the number of training iterations. Let i be the weight;

[0293] The higher the importance value, the greater the contribution of the weight to fitting the historical training data, and the more it needs to be protected; conversely, a lower importance value indicates a non-critical weight that can be adjusted flexibly.

[0294] Weighting adjustment rules:

[0295] Importance value The weights, with a threshold θ=0.001: only small adjustments are allowed, with the adjustment magnitude ≤ 5% of the initial value, to avoid compromising the model's ability to fit existing subject data;

[0296] Importance value The weights can be freely adjusted to adapt to the signal characteristics of new monitoring scenarios;

[0297] While maintaining the original prediction accuracy, the model's adaptability to new scenarios is improved, catastrophic forgetting is reduced, and the original capabilities are avoided after training on new data.

[0298] Step 4.4.3: Use the adjusted model to predict the heartbeat signal, apply the trained and optimized model to the actual monitored signal, and output the final heartbeat signal and heart rate value.

[0299] Preprocessing before prediction:

[0300] Input signal interception: For the real-time signal output in step 4.3, the input vector is intercepted according to the sliding time window (). The length of the sliding time window is 100 sampling points and the step size is 1 sampling point to ensure the continuity of timing.

[0301] Normalization: The input vector is normalized using the normalization parameters of the training set, the mean μ and the standard deviation σ. This helps avoid prediction distortion caused by inconsistent distribution.

[0302] Real-time prediction process:

[0303] Initialization: The first 100 sampling points are used as initial input, and the model outputs the first predicted value. ;

[0304] Time series iteration: At time t, the input vector = real-time signal from time t to t+99 + predicted value at time t-1. (Feedback connection), Model output ;

[0305] Inverse normalization: transforming the predicted values Restored to the amplitude of the actual heartbeat signal: ;

[0306] Heart rate calculation: Perform spectral analysis (FFT) on a 20-second continuous predicted heartbeat signal to extract the peak frequency. (Unit: Hz), Heart rate value = ×60 (unit: times / minute), and perform moving average smoothing, with a window length of 5 seconds to reduce instantaneous fluctuations.

[0307] Prediction results output:

[0308] Real-time output: The heart rate value is updated once per second, and the time-domain waveform of the heartbeat signal is output synchronously for display on the monitoring terminal;

[0309] Abnormal alarm: If the heart rate value exceeds the normal range and continues for 3 seconds, an alarm prompt will be output;

[0310] Data storage: Historical data is stored in the format of timestamp-heart rate value-heartbeat signal amplitude for easy subsequent traceability and analysis.

[0311] Step 4.4 Based on the model structure of input layer + 5 hidden layers + feedback connection + output layer, the previous time step output feedback is introduced to enhance the time-dependent processing capability, improve the adaptability to fluctuating data sequences, and solve the problems of weak time-series processing and unstable heart rate prediction in traditional lightweight neural networks.

[0312] The model was trained with data from 100 subjects and optimized using the Adam algorithm. Combined with weight importance assessment and flexible adjustment rules, the model has strong generalization ability and high prediction accuracy. Real-time monitoring is achieved by inputting heart rate values ​​in real time through a sliding window and updating the heart rate value every second, thus avoiding detection delay.

[0313] It outputs heart rate values ​​and time-domain waveforms, supports abnormal alarms (exceeding the normal range for 3 seconds) and historical data storage, and provides a complete monitoring solution for applications in multiple fields.

[0314] The description of this invention is given for illustrative and descriptive purposes only and is not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention and to enable those skilled in the art to understand the invention and design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A Narx neural network based millimeter wave vital sign monitoring method, characterized in that: The method comprises the following steps: Step 1, initialize the monitoring system parameters, configure the millimeter wave radar sensor working frequency to be 77GHz, and the sampling frequency to be 100MHz; Step 2, perform signal acquisition, emit and receive millimeter wave signals to the monitoring area through the millimeter wave radar sensor assembly; Step 3, judge whether there is an abnormal situation after the acquisition is completed, if not, execute Step 4, through three-layer verification and hierarchical processing, gradually screen the data integrity, signal validity and target existence; Step 4, perform signal processing and analysis, accurately process and deeply analyze the screened high-quality acquisition signals, purify the vital sign related signals, realize real-time and accurate prediction of heartbeat signals and heart rate by means of NARX neural network, and support abnormal alarm and historical data storage.

2. The Narx neural network based millimeter wave vital sign monitoring method of claim 1, wherein: The specific steps of Step 1 are as follows: Step 1.1, monitoring system configuration; The monitoring system comprises a millimeter wave radar sensor assembly, a signal processing module, a main control module and an auxiliary function module, and the main control module is connected with the radar sensor assembly, the signal processing module, the main control module and the auxiliary function module; The millimeter wave radar sensor assembly is the core of non-contact signal acquisition, directly determines the frequency, power and reliability of signal acquisition, and comprises: A frequency synthesizer configured with a 77GHz working frequency generates accurate millimeter wave carrier signals and supports 77GHz band output; A power amplifier amplifies the millimeter wave signals generated by the frequency synthesizer to a suitable power level of 18dBm; An antenna array adopts a 2-transmit 4-receive antenna array for transmitting millimeter wave signals and receiving echo signals; An analog-to-digital conversion module converts the analog echo signals received by the antenna array into digital signals, performs time domain acquisition at a sampling frequency of 100MHz, and provides original data for subsequent digital signal processing; The signal processing module is used for filtering interference, feature extraction and running NARX neural network to realize heartbeat signal prediction, and guarantees monitoring accuracy, and comprises: A digital signal processing DSP unit: iteratively filters strong signals, and retains weak signals related to breathing and heartbeat; spectrum analysis is performed on intermediate frequency signals to identify breathing signals and harmonic frequencies; linear filtering is performed to filter out harmonics of breathing signals and purify effective signals; A neural network calculation unit: running NARX neural network model, including model training, weight adjustment and heartbeat signal prediction; The auxiliary function module comprises: A power management module provides stable power supply for all modules and supports voltage requirements of different modules; A storage module caches original digital signals collected by the analog-to-digital conversion module and intermediate data preprocessed by the digital signal processing DSP unit; stores the NARX neural network model parameters after training, without retraining every time; and can optionally store the final breathing / heartbeat monitoring results for subsequent tracing. The master module connects the working time sequence of each module, and issues configuration instructions of 77GHz working frequency and 100MHz sampling frequency to the frequency synthesizer and the analog-to-digital conversion module; the time sequence of signal acquisition, abnormality judgment, signal processing and neural network calculation is triggered; the working state of each module is read, and the parameter configuration is confirmed to take effect; Step 1.2, millimeter wave radar sensor working frequency configuration; Step 1.3, sampling frequency configuration; Step 1.4, system parameter verification and state confirmation.

3. The Narx neural network based millimeter wave vital sign monitoring method of claim 1, wherein: The step 2 includes the following steps: Step 2.1, signal transmission; The 77GHz millimeter wave signal amplified by the power amplifier is radiated to the monitoring area through the transmitting antenna array. The arrangement of the antenna array needs to adapt to the wavelength of the 77GHz signal. The wavelength λ=c / f, where c is the speed of light≈3×10 8 m / s, f=77×109Hz. A uniform linear array is adopted, the element spacing is set to λ / 2, and the array as a whole is arranged along the horizontal direction and adheres to the center line of the monitoring area. Step 2.2, signal reception; After the millimeter wave signal irradiates the human body, the micro displacement of the chest caused by the breathing and heartbeat of the human body will cause the reflected signal to produce Doppler shift and phase change. These reflected signals carrying vital sign information are captured by the receiving antenna array. During the receiving process, the antenna array focuses the target signal through beamforming technology and suppresses the interference signals such as environmental reflection; Step 2.3, time domain acquisition; The analog signal output by the receiving antenna is transmitted to the analog-to-digital conversion module; the analog-to-digital conversion module discretely samples the intermediate frequency signal at a set sampling frequency of 100MHz, and converts the analog signal into a digital signal; Step 2.4, data storage; The time domain digital signal obtained by sampling is stored in frame format, the frame period is 10ms, the number of single frame acquisition data points is 1024, and a time stamp is attached. The data is stored in 16-bit binary format and organized in a 'frame number-time stamp-signal amplitude' triplet for time sequence alignment and tracing during subsequent signal processing.

4. The Narx neural network based millimeter wave vital sign monitoring method of claim 1, wherein: The step 3 includes the following steps: Step 3.1, data integrity verification, checking whether the acquisition data is missing or packet loss, ensuring continuous data sequence without interruption; Read the acquisition data of the current frame in the cache area, count the number of valid data points, and the valid data points are the data points with complete amplitude information and no null / invalid values. Set the integrity threshold to be that the valid data point ratio is greater than or equal to 95%, that is, the missing rate is less than or equal to 5%. If the valid data point ratio is less than 95%, it is determined that the data integrity is abnormal, the current frame abnormal data is emptied, and the signal acquisition is re-executed in step 2. If it is still abnormal after retrying 3 times, output hardware alarm; Step 3.2, signal amplitude validity verification, checking signal saturation, amplitude abnormality and other problems, ensuring that the acquired signal does not exceed the hardware processing range, and retaining weak vital sign information; Step 3.2, signal amplitude validity verification, checking signal saturation, amplitude abnormality and other problems, ensuring that the acquired signal does not exceed the hardware processing range, and retaining weak vital sign information; 5. The Narx neural network based millimeter wave vital sign monitoring method of claim 1, wherein: The step 3 also includes the following steps: Step 3.3, target existence verification, confirming whether there is a human target in the monitoring area to avoid invalid processing of environmental noise without target; Simple energy detection on current frame data: calculate the sum of amplitude squares of all valid data points, set the target decision threshold to be 5 times the amplitude square sum greater than the ambient noise energy, if the amplitude square sum is lower than the threshold, it means that there is no human target in the monitoring area, only environmental noise, and it is determined to be a target missing anomaly, the collection process is suspended, and a no monitoring target prompt is output; if it is higher than the threshold, it means that the data contains effective signals reflected by the human body, and it is determined to be no anomaly.

6. The Narx neural network based millimeter wave vital sign monitoring method of claim 1, wherein: The step 4 comprises the following steps: Step 4.1, sequentially iteratively filtering out strong interference signals to the collected intermediate frequency signal, and retaining weak signals carrying vital sign information, laying a foundation for subsequent respiratory signal identification and heartbeat signal prediction; Step 4.2, identifying the respiratory signal and the harmonic of the respiratory signal according to the spectral characteristics; Step 4.3, using a linear filter factor to filter out the harmonic of the respiratory signal from the intermediate frequency signal, avoiding frequency overlap interference with the heartbeat signal, thereby purifying the characteristics of the heartbeat signal; Step 4.4, using NARX neural network to deeply analyze the processed signal and predict the heartbeat signal.

7. The Narx neural network based millimeter wave vital sign monitoring method of claim 6, wherein: The step 4.1 comprises the following steps: Step 4.1.1, setting an iterative filtering threshold, using a three-level screening to realize hierarchical interference filtering from strong to weak, and avoiding the loss of weak vital sign signals caused by one-time strong threshold filtering; Normalizing the input single-frame intermediate frequency signal: calculate the maximum amplitude A_max of all data points, and convert the amplitude A_i of each data point to relative intensity S_i=A_i / A_max, S_i∈[0,1]; Fixing three-level thresholds: first-level threshold T1=0.6, second-level threshold T2=0.4, and third-level threshold T3=0.15; T1 is used to filter out the strongest fixed interference signals such as environmental hard reflections, T2 filters out secondary interference signals, and T3 retains weak effective signals, and the relative intensity of the heartbeat signal is usually between 0.15 and 0.4; Step 4.1.2, first-level filtering, strong interference signal preliminary screening; Traverse the normalized signal sequence (S_1, S_2,..., S_m), m is the number of single-frame data points; retain all data points satisfying S_i>T1, and remove data points with S_i≤0.6; obtain the first-level screened signal sequence S1, which only contains strong signal components, mainly fixed interference signals, and does not contain effective vital sign signals; Step 4.1.3, second-level filtering, further removing secondary interference on the basis of first-level screening, and gradually focusing on effective signals; Taking the first-level screened signal sequence S1 as the processing object, re-calculating the maximum relative intensity of the signal sequence S1, denoted as S1_max, since weak signals have been removed, S1_max is still close to 1; performing screening according to the second-level threshold T2=0.4, retaining data points in S1 that satisfy S_i>0.4, and removing data points that satisfy S_i≤0.4; obtaining the second-level screened signal sequence S2, the interference signals are further reduced, and the strong interference signal residues are still dominant; Step 4.1.4, third level filtering, remove the remaining interference, keep the weak signal corresponding to breathing and heartbeat, complete the strong and weak; Taking the signal sequence S2 filtered in the second stage as the processing object, the maximum relative intensity S2_max of S2 is calculated; filtering is performed according to the third level threshold T3 = 0.15, the data points in the signal sequence S2 that satisfy S_i > 0.15 are kept, and the data points that satisfy S_i ≤ 0.15 are removed; the valid data points kept after the three filtering are arranged according to the original time sequence, the linear interpolation method is used to complete the signal in the missing position to avoid time sequence breakage, and a pure signal sequence x(m) after filtering out strong interference is formed.

8. The Narx neural network based millimeter wave vital sign monitoring method of claim 6, wherein: The step 4.2 includes the following steps: Step 4.2.1, calculating the short-time Fourier transform of the intermediate frequency signal, converting the time domain signal into a time-frequency two-dimensional spectrum, realizing the time sequence tracking of the frequency component, and avoiding the loss of time information in the traditional Fourier transform; Signal frame processing: the input time domain signal is divided into continuous frames according to a fixed window length, the window length is set to 2s, the 2s window length can completely cover 1 breathing cycle, and the frequency resolution and time resolution are considered; a 50% overlap rate is set between frames to avoid spectral distortion caused by frame boundary signal loss; Window function selection: Hanning window is used to suppress spectral leakage; Fourier transform FFT point number configuration: the FFT point number of each frame signal is set to 2048 points; Performing STFT: after applying the Hanning window to each frame signal, performing fast Fourier transform FFT, and outputting a time-frequency-amplitude three-dimensional spectrum; Step 4.2.2, performing spectrum analysis on the Fourier transform result, extracting the effective frequency component from the spectrum output by STFT, excluding noise interference, and focusing on the breathing related frequency; Spectrum amplitude normalization: the spectrum amplitude of each frame is normalized, and the maximum amplitude in a single frame is recorded as 1, which facilitates cross-frame comparison of the intensity of the frequency component; Noise threshold filtering: calculating the background noise mean value in the spectrum, taking the amplitude mean value of the 0-0.05Hz low frequency band, which has no vital sign signal, setting the noise threshold = noise mean value x 3, removing the frequency components in the spectrum whose amplitude is lower than the threshold, and reducing noise interference; Frequency component extraction: keeping the frequency points whose amplitudes are greater than the threshold after filtering, recording the amplitude and corresponding time frame of each frequency point, and forming an effective frequency-amplitude-time list; Cross-frame consistency verification: because the frequency of the breathing signal is stable, the fluctuation under the resting state is ≤±0.1Hz, the effective frequency points extracted are verified cross-frame, the frequency components that appear in continuous 3 frames and above are kept, the false frequencies caused by transient noise are excluded, and a candidate frequency set is obtained; Step 4.2.3, identifying the breathing signal and the harmonic frequency of the breathing signal, distinguishing the breathing fundamental frequency from the harmonics from the candidate frequency set, and determining the frequency values of the two; Breathing fundamental frequency identification: selecting the frequency points in the range of 0.1-1Hz from the candidate frequency set; the frequency point with the maximum amplitude in this range is the breathing fundamental frequency f0, because the breathing fundamental frequency signal has the strongest energy, the amplitude is significantly higher than the harmonics and the surrounding noise; Breathing harmonic recognition: based on the breathing fundamental frequency f0, the candidate frequency set is screened to meet f = n x f0 (n = 2, 3, …); Verify harmonic characteristics: the amplitude of the harmonic needs to meet ≤30% x f0, the energy attenuation characteristics of the breathing harmonic, and the frequency deviation ≤±0.05 Hz, to avoid misjudging the similar heartbeat signal as a harmonic; Record the breathing fundamental frequency f0 and all harmonic frequencies that meet the conditions.

9. The Narx neural network based millimeter wave vital sign monitoring method of claim 6, wherein: The step 4.3 includes the following steps: Step 4.3.1, determine the type of linear filter factor; The linear filter factor corresponds to the unit impulse response of the linear IIR notch filter. The breathing harmonic is a discrete frequency point, and is close to the frequency of the heartbeat signal. A narrow-band and high-attenuation notch filter needs to be used to accurately suppress specific discrete frequencies and avoid excessive filtering that distorts the heartbeat signal. This meets the analysis needs of the subsequent NARX neural network for time series signals. Step 4.3.2, configure the notch filter parameters based on the breathing harmonic frequency; The center frequency fn is consistent with the breathing harmonic frequency, accurately aligning with the harmonic frequency that needs to be suppressed, ensuring that the filtering target is clear; The bandwidth BW is a fixed narrow bandwidth of 0.1 Hz, covering only a narrow range around the harmonic frequency, avoiding affecting adjacent heartbeat signals; The attenuation A is ≥40 dB, ensuring that the harmonic amplitude is attenuated to less than 1% of the original amplitude, completely suppressing harmonic interference and avoiding residual harmonic affecting heartbeat signal extraction; The filter order is 4, balancing the filtering effect and computational load. A 4th order can achieve high attenuation with small computational load, suitable for real-time monitoring needs; If multiple breathing harmonics are identified, an independent notch filter is designed for each harmonic frequency, cascaded in the order of high-frequency harmonic → low-frequency harmonic, to avoid mutual interference between multiple filters and ensure that each harmonic is fully suppressed, while reducing the cumulative distortion of the overall signal phase; Step 4.3.3, generate a linear filter factor; Based on the selected IIR notch filter and parameters, the transfer function is calculated through the filter design formula: Transfer function of the IIR notch filter: ; Where: Normalized center angle frequency , f n is the respiratory harmonic frequency, the actual sampling frequency Fs'= 1 kHz; Attenuation coefficient r ~ 0.9997, ensuring a narrow bandwidth and sufficient attenuation; Perform inverse Z-transform on the transfer function H(z) to get the unit impulse response h(n), which is the linear filter factor: The impulse response h(n) of the 4th order IIR filter is a finite length sequence, with a length of 30 sampling points, taking into account the filtering effect and computational efficiency; Standardization of the filter factor: normalize h(n) to adjust the sum of all elements to 1, to avoid overall amplitude attenuation of the filtered signal; Verify the frequency response of the filter factor: analyze the frequency response of h(n) by fast Fourier transform (FFT) to ensure that the amplitude attenuation is ≥ 40 dB at f n and the amplitude attenuation is ≤ 0.5 dB in the heart beat signal frequency band. Step 4.3.4, apply the filter factor to remove the breathing harmonic, combine the filter factor with the input signal through time domain convolution to accurately suppress the breathing harmonic; Convolution operation: linearly convolve the pure signal sequence x(m) with the filter factor h(n) to obtain the filtered signal y(n): ; Where N is the length of the filter factor h(n) and n is the signal sampling point index; The pure signal sequence x(m) and f n The consistent breathing harmonic components are suppressed by the frequency response of h(n), and other frequency components are almost kept without distortion; If there are multiple harmonics, perform multi-harmonic cascaded filtering: First stage filtering: convolve the filter factor for 3f0 with x(m) to get the signal x1(m) after filtering out 3f0; Second stage filtering: convolve the filter factor for 2f0 with x1(m) to get the signal y(n) after filtering out all harmonics; After each stage of filtering, verify the amplitude of the filtered signal at the harmonic frequency to ensure that the attenuation is ≥40 dB. If not, adjust the r value of the filter factor. Signal edge processing: Convolution operation will cause boundary distortion at both ends of the signal, which is handled by zero padding: fill N / 2 zeros before and after the input signal x(m), and after convolution, the effective signal segment in the middle is intercepted to ensure that the length of the output signal y(n) is consistent with the input signal, and the timing is continuous.

10. The Narx neural network based millimeter wave vital sign monitoring method of claim 6, wherein: The step 4.4 includes the following steps: Step 4.4.1, constructing a NARX neural network model; The NARX neural network model structure is as follows: The input layer is 1 layer, the input dimension = 100, and the input is the time series data intercepted by the sliding time window: the window length = 100 sampling points, that is, the input vector dimension is 100x1, which is suitable for capturing time series characteristics; The hidden layer is 5 layers, each layer has 80 nodes, the activation function uses ReLU to solve the gradient disappearance problem, and the calculation amount is small, which is suitable for embedded hardware; batch normalization is added to each layer to speed up training convergence and improve model generalization feedback; The connection layer is 1 layer, and the output feedback is introduced: the predicted heartbeat signal amplitude of the model output at the previous time is fed back to the input layer at the current time, and the original signal is concatenated and input to the hidden layer, which strengthens the time series dependence and fits the continuity characteristics of the heartbeat signal; The output layer is 1 layer, the output dimension = 1, and the output is the heartbeat signal amplitude at the current time, and the activation function uses Sigmoid to map to the [0, 1] interval, which can be restored to the true amplitude through de-normalization in the future; Weight initialization: He normal distribution is used to avoid training failure caused by too large or too small initial weights; Bias initialization: all layer bias items are initialized to 0.01 to speed up the convergence in the initial training stage; Step 4.4.2, training the model using Adam optimization algorithm Adjust the network parameters through the training data to make the model have the ability to predict the heartbeat signal; Training data set preparation: Data source: 100 subjects' synchronous data, pure signal collected by millimeter wave radar + real heartbeat signal collected by ECG; Data division: divided into training set, validation set and test set according to 7:2:1; the training set is used for parameter update, the validation set is used for overfitting monitoring, and the test set is used for final performance evaluation; Data enhancement: add small noise and time stretching to the training set signal, the amplitude of the small noise is ≤5%, and the time stretching length is ±10%, which improves the model's anti-interference ability and adapts to the environmental noise in actual monitoring; Training parameter configuration: the learning rate of Adam optimization algorithm is fixed at 0.0005 to avoid oscillation caused by too large learning rate and slow convergence caused by too small learning rate; Loss function: Mean Squared Error (MSE) formula is , B is the batch size, is the true label, is the model prediction value, which is suitable for continuous value regression tasks; Training batch and iteration: batch size = 32 to balance training speed and memory occupation; iteration number = 5000 times, stop early when the validation set MSE does not decrease for 50 consecutive iterations to avoid overfitting; Training process: Forward propagation: input the training set input vector into the network, calculate through the hidden layer, and fuse the feedback connection to output the predicted heartbeat signal amplitude; Back propagation: based on the MSE loss value, calculate the gradient of each layer parameter using the chain rule, and update the weight and bias through the Adam algorithm: Momentum parameter: first order momentum β1=0.9, second order momentum β2=0.9, the default value balances the gradient stability and convergence speed; Weight update formula: η is a learning rate, , is a first-order and second-order momentum correction value, Avoiding denominator 0; MSE is calculated with the validation set every 100 iterations. If MSE continues to rise, trigger learning rate decay, decay coefficient = 0.9, to suppress overfitting. Step 4.4.3, adjust model weights after training is complete. Weight importance evaluation: computing importance values for each weight parameter in a training process , T is the number of training iterations, is the i-th weight; The greater the importance value, the greater the contribution of the weight to fitting historical training data, and it needs to be protected; otherwise, it is a non-critical weight and can be adjusted flexibly. Weight adjustment rules: weight of importance value threshold θ = 0.001: only allow minor adjustment, adjustment amplitude ≤ 5% of initial value, avoid destroying model fitting ability to existing subject data The weight of the importance value can be freely adjusted to adapt to the signal characteristics of a new monitoring scene. Step 4.4.3, use the adjusted model to predict the heartbeat signal, apply the optimized model to the actual monitoring signal, and output the final heartbeat signal and heart rate value. Pre-processing before prediction: Input signal interception: intercept the input vector according to the sliding time window () from the real-time signal output in step 4.3, the length of the sliding time window = 100 sampling points, the step = 1 sampling point, and ensure the time sequence continuity. Normalization: The input vector is normalized using the normalization parameters, mean μ and standard deviation σ, of the training set: , to avoid distribution inconsistency leading to prediction distortion; Real-time prediction process: Initialization: the first 100 samples are used as initial input and the model outputs the first prediction ; Timing iteration: at time t, input vector = real-time signal of sample points t to t+99 + predicted value at time t-1 (model output, feedback connection) ; De-normalization: Predicted values are reduced to real heartbeat signal amplitudes: ; Heart rate calculation: Perform spectral analysis (FFT) on the predicted heartbeat signal for 20 consecutive seconds, extract the peak frequency (unit: Hz), heart rate value = 60 × (unit: beats / minute), and perform moving average smoothing with a window length of 5 seconds to reduce transient fluctuations; Prediction result output: Real-time output: update the heart rate value every second, and output the time-domain waveform of the heartbeat signal synchronously, which is suitable for monitoring terminal display. Abnormal alarm: if the heart rate value exceeds the normal range and lasts for 3 seconds, output an alarm prompt. Data storage: store historical data in the format of timestamp-heart rate value-heartbeat signal amplitude for subsequent traceability analysis.

Images