A living body feature separation method based on millimeter wave radar signal enhancement

Abstract

This invention discloses a method for separating vital signs based on millimeter-wave radar signal enhancement, comprising the following steps: acquiring the original radar signal of the target being monitored by millimeter-wave radar and preprocessing it; extracting six types of features from the preprocessed signal and classifying the preprocessed signal into four scene categories based on the six types of features; according to the classified scene categories, calling the corresponding signal processing model to enhance or reconstruct the preprocessed signal to obtain the enhanced radar signal; performing signal quality assessment on the enhanced radar signal, and inputting the enhanced signal that meets the quality assessment requirements into a variational mode decomposition module to separate the respiratory and heartbeat signals of the target being monitored. This invention improves the accuracy of vital sign detection in complex environments by classifying millimeter-wave radar vital sign signals in complex scenes, adaptively calling the model to enhance the signal, and using variational mode decomposition to separate respiration and heartbeat.

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and non-contact life recognition technology, and more specifically to a method for separating life features based on millimeter-wave radar signal enhancement. Background Technology

[0002] In recent years, with the acceleration of global population aging, health monitoring of the elderly has become a focus of social attention.

[0003] Existing monitoring devices are mainly divided into two categories: contact and non-contact. While contact devices (such as wearable sensors) offer high accuracy, they require prolonged wear by users, which can cause discomfort and carries the risk of missed wear, leading to poor user compliance. In contrast, non-contact monitoring technologies are favored for their unobtrusive and convenient advantages. Among these, millimeter-wave radar, with its significant characteristics such as insensitivity to light, high anti-interference capabilities, strong penetration, and protection of user privacy (no optical imaging required), is gradually becoming a mainstream emerging product in the field of non-contact vital sign monitoring.

[0004] However, millimeter-wave radar still faces significant challenges in practical applications. First, the complex and variable environments of homes and clinics severely impact signal quality due to multipath effects, static clutter, and dynamic interference. Second, the subtle chest and abdominal movements caused by human respiration and heartbeat are extremely weak, resulting in severe echo signal attenuation and a very low signal-to-noise ratio, especially in medium- or long-range monitoring scenarios. Traditional signal processing methods, such as fixed filters and simple spectral analysis, are often based on ideal signal models and lack adaptability to environmental changes. In noisy environments, these methods struggle to effectively separate weak vital signs, easily leading to feature extraction failures or false detections. Furthermore, individual differences, such as body size and breathing patterns, further increase the difficulty of algorithm generalization.

[0005] The aforementioned technical bottlenecks result in insufficient robustness and reliability of existing millimeter-wave radar products in complex scenarios, severely restricting their large-scale and reliable application in smart elderly care home monitoring and hospital clinical continuous monitoring.

[0006] Therefore, proposing a new method that can adapt to complex scenarios, effectively enhance weak signals, and accurately reconstruct vital signs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, this invention is proposed to provide a life feature separation method based on millimeter-wave radar signal enhancement to overcome or at least partially solve the above problems. The method classifies radar signals collected in complex scenarios by extracting features, and then introduces different neural network models for signal enhancement and life feature extraction.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a method for separating life features based on millimeter-wave radar signal enhancement, comprising the following steps:

[0010] S1. Acquire the raw radar signal of the life monitoring target collected by the millimeter-wave radar and perform preprocessing;

[0011] S2. Extract six types of features from the preprocessed signal, and divide the preprocessed signal into four scene categories based on the six types of features;

[0012] S3. Based on the identified scene categories, the corresponding signal processing model is invoked to enhance or reconstruct the preprocessed signal to obtain the enhanced radar signal; the signal processing model includes: a one-dimensional convolutional neural network model, a temporal coding conditional generative adversarial network model, a Kalman filter processing model, and a combined processing model;

[0013] S4. The enhanced radar signal is evaluated for signal quality. The enhanced signal that meets the quality evaluation requirements is input into the variational mode decomposition module to separate the respiratory signal and heartbeat signal of the life monitoring target.

[0014] Furthermore, in step S1, the preprocessing includes:

[0015] Perform a range-dimensional fast Fourier transform on the original radar signal in complex form to convert the time-domain signal into a range-domain signal;

[0016] The static DC component in the range domain signal is eliminated by subtracting the mean, resulting in a dynamic signal.

[0017] Phase information is extracted from the dynamic signal, and the phase value is restored to a true continuous phase trajectory reflecting the micro-movements of the chest and abdomen of a living organism through a phase unwrapping algorithm.

[0018] Furthermore, in step S2, the six types of features include: spectral flatness, phase standard deviation, estimated signal-to-noise ratio, vital sign band energy ratio, spectral peak amplitude, and signal amplitude stability.

[0019] The four scenario categories include: strong interference-strong signal, weak interference-weak signal, strong interference-weak signal, and weak interference-strong signal.

[0020] Based on the six types of features, the signals are divided into four scene categories, specifically including:

[0021] The spectral flatness, phase standard deviation, and signal amplitude stability are weighted and fused, and then compared with a first threshold. If the value is greater than the first threshold, it is determined to be strong interference; otherwise, it is determined to be weak interference.

[0022] The estimated signal-to-noise ratio, vital sign frequency band energy ratio, and spectral peak amplitude are compared with a preset second threshold. If all of them are greater than the second threshold, they are judged as strong signals; otherwise, they are judged as weak signals.

[0023] Furthermore, step S3 specifically includes:

[0024] For signals classified into strong interference-strong signal scenarios, a one-dimensional convolutional neural network model is used for interference removal and feature extraction to obtain enhanced radar signals.

[0025] For signals classified as weak interference-weak signal scenarios, a temporal coding conditional generative adversarial network model is used to perform generative reconstruction and spectrum restoration of weak signals to obtain enhanced radar signals.

[0026] For signals classified into weak interference-strong signal scenarios, a Kalman filter processing model is used to perform dynamic state estimation and random noise smoothing to obtain the enhanced radar signal.

[0027] For signals classified into strong interference-weak signal scenarios, a combined processing model is used to perform cascaded denoising, temporal modeling, and signal reconstruction to obtain enhanced radar signals. The combined processing model is based on a combination of a one-dimensional convolutional neural network model, a bidirectional simple recurrent unit, a Kalman filter processing model, and a temporal coding conditional generative adversarial network model.

[0028] Furthermore, the one-dimensional convolutional neural network model includes: an encoder and a decoder;

[0029] The encoder includes dual convolutional blocks and four cascaded downsampling modules, used to extract high-dimensional deep features of the signal layer by layer and compress the time dimension;

[0030] The decoder includes four cascaded upsampling modules and a one-dimensional convolutional block, used to recover the temporal resolution of the signal and reconstruct the waveform layer by layer;

[0031] The encoder and decoder perform feature fusion through skip connections between corresponding layers, passing the shallow detail features extracted by the encoder to the decoder to compensate for the spatial information loss during the upsampling process.

[0032] Furthermore, the temporal coding conditional generative adversarial network model uses a weighted sum of L1 loss and spectral consistency loss as the optimization objective during training.

[0033] The L1 loss is used to calculate the absolute error between the reconstructed signal and the true signal in the time domain, and is expressed by the formula:

[0034]

[0035] The spectral consistency loss is used to calculate the mean square error between the reconstructed signal and the true signal in the frequency domain, and is expressed by the formula:

[0036]

[0037] in, B Indicates batch size, L Indicates the number of sampling points. Indicates the first i The first generated signal j One sampling point, Indicates the first i The first real signal j One sampling point; Indicates the first The weights of each frequency component, Indicates the generation signal number 1 The amplitude of each frequency component Represents the true signal. The amplitude of each frequency component.

[0038] Furthermore, the process of using Kalman filtering to perform dynamic state estimation and random noise smoothing specifically includes:

[0039] Construct a linear state-space model; the linear state-space model includes: state transition equations and observation equations;

[0040] Initialize the filter parameters, use the optimal state estimate from the previous time step, calculate the prior state estimate and its error covariance from the state transition equation, and perform time update.

[0041] Calculate the Kalman gain coefficient, and use the actual observation value at the current time to correct the prior state estimate to obtain the posterior optimal state estimate;

[0042] The displacement component is extracted from the posterior optimal state estimate as the enhanced radar signal.

[0043] Furthermore, the construction and data processing of the combined processing model include:

[0044] Load a pre-trained one-dimensional convolutional neural network model and freeze its encoder parameters, then fine-tune only the decoder;

[0045] Load the pre-trained temporal coding conditional generative adversarial network model and freeze the shallow convolution parameters in its conditional encoder and generator.

[0046] A bidirectional simple recurrent unit is inserted between two transfer models to construct a temporal modeling path. The signal output by the one-dimensional convolutional neural network model is compressed and then input into the bidirectional simple recurrent unit for temporal propagation.

[0047] The signal output from the bidirectional simple cyclic unit is input into the Kalman filter processing model for smoothing.

[0048] The smoothed signal is input into the finely tuned temporal coding conditional generative adversarial network model generator for reconstruction.

[0049] Furthermore, in step S4, the enhanced radar signal is evaluated for signal quality, specifically including:

[0050] Based on the enhanced radar signal, three core indicators are calculated; the three core indicators include: peak signal-to-noise ratio, signal amplitude stability, and concentration of vital signs energy.

[0051] The three core indicators are linearly weighted and fused into a comprehensive quality score, which is then compared with a preset threshold.

[0052] Furthermore, in step S4, the variational mode decomposition module separates the respiratory signal and the heartbeat signal, specifically including:

[0053] Based on the spectral characteristics of human vital signs, the total number of modes in the variational mode decomposition module is set to 3, corresponding to low-frequency baseline drift, respiratory component, and heartbeat component, respectively; the center frequency of each mode is initialized.

[0054] The enhanced radar signal that meets the quality assessment requirements is subjected to Fourier transform to obtain its frequency domain representation; a constrained variational model containing data fidelity terms and bandwidth penalty terms is constructed in the frequency domain and iteratively updated using the alternating direction multiplier method; in each iteration, the frequency domain expression of each mode, the corresponding center frequency, and the Lagrange multiplier are updated synchronously until the convergence condition is met or the maximum number of iterations is reached.

[0055] After the iteration is completed, perform an inverse Fourier transform on the frequency domain expression of each mode to obtain the corresponding time domain eigenmode components;

[0056] The dominant frequencies of each intrinsic mode component in the time domain are calculated, and respiratory and heartbeat signals are matched and filtered according to a preset range of human physiological frequencies; wherein the dominant frequency of the respiratory signal is between 0.1Hz and 0.5Hz, and the dominant frequency of the heartbeat signal is between 0.8Hz and 2.0Hz.

[0057] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for separating life features based on millimeter-wave radar signal enhancement, which has the following beneficial effects:

[0058] This invention overcomes the limitations of traditional "one-size-fits-all" methods by innovatively proposing a scene perception and classification mechanism based on six types of features. It accurately identifies four typical scenarios, including "strong interference-strong signal" and "weak interference-weak signal," and adaptively calls the most suitable processing model. This "scenario-specific" strategy effectively solves the problem that a single algorithm cannot simultaneously address both strong noise suppression and weak signal recovery, significantly improving the robustness of radar systems in complex environments such as household clutter interference, long-range detection, and dynamic backgrounds.

[0059] To address the issues of severe signal attenuation and low signal-to-noise ratio in medium- to long-range monitoring, this invention introduces a Conditional Generative Adversarial Network (TECGAN) based on time-series coding and a combined processing model. Leveraging the generative capabilities of deep learning, it not only filters out noise but also "reconstructs" the submerged details of faint breathing and heartbeat waveforms based on prior knowledge, filling in the signal distortion or loss caused by traditional filtering methods. Combined with a closed-loop quality assessment feedback mechanism, it ensures high fidelity of the output signal, significantly reducing the false negative and false positive rates.

[0060] Variational mode decomposition (VMD) is employed in the post-processing stage, combined with prior knowledge of human physiological frequencies (respiration 0.1-0.5 Hz, heartbeat 0.8-2.0 Hz) for mode screening. Compared to traditional empirical mode decomposition, this effectively avoids mode aliasing. Combined with a preset center frequency initialization strategy, it can more accurately separate independent respiratory and heartbeat components from the enhanced signal, maintaining extremely high detection accuracy and physiological consistency even under conditions of large heart rate variability or weak respiration.

[0061] This invention effectively overcomes the bottlenecks of existing millimeter-wave radar technology in terms of anti-interference capability, weak signal extraction, and adaptability to individual differences. It provides a high-precision, high-reliability, and privacy-secure solution for health monitoring, sleep monitoring, and continuous monitoring of clinical vital signs in the elderly, with significant social benefits and broad market application prospects. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0063] Figure 1 This is a flowchart of the life feature separation method based on millimeter-wave radar signal enhancement provided in this embodiment of the invention;

[0064] Figure 2 This is a schematic diagram illustrating the division of six types of features and four scene categories provided in this embodiment of the invention;

[0065] Figure 3 This is a schematic diagram illustrating the four scene categories and their corresponding signal processing model calls provided in this embodiment of the invention;

[0066] Figure 4 This is a schematic diagram of the structure of a one-dimensional convolutional neural network model provided in an embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram of the structure of the temporal coding conditional generative adversarial network model provided in an embodiment of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] This invention discloses a method for separating biological features based on millimeter-wave radar signal enhancement, referring to... Figure 1 As shown, it includes the following steps:

[0070] S1. Acquire the raw radar signal of the life monitoring target collected by the millimeter-wave radar and perform preprocessing;

[0071] S2. Extract six types of features from the preprocessed signal, and divide the preprocessed signal into four scene categories based on the six types of features;

[0072] S3. Based on the identified scene categories, the corresponding signal processing model is invoked to enhance or reconstruct the preprocessed signal to obtain the enhanced radar signal; the signal processing model includes: a one-dimensional convolutional neural network model, a temporal coding conditional generative adversarial network model, a Kalman filter processing model, and a combined processing model;

[0073] S4. The enhanced radar signal is evaluated for signal quality. The enhanced signal that meets the quality evaluation requirements is input into the variational mode decomposition module to separate the respiratory signal and heartbeat signal of the life monitoring target.

[0074] This embodiment is applied to a smart elderly care home sleep monitoring system. It uses a vital sign separation method based on millimeter-wave radar signal enhancement to conduct all-weather, non-intrusive monitoring of vital signs (respiratory rate, heart rate) in the bedroom environment of elderly people living alone or disabled elderly people, so as to promptly detect sudden health risks such as sleep apnea and abnormal heart rate.

[0075] The implementation steps of this embodiment are described in detail below:

[0076] First, following step S1, the raw radar signal acquired by the millimeter-wave radar is read and preprocessed.

[0077] Because the bedroom environment is usually quite complex: the room contains a large number of static objects such as beds, wardrobes, tables and chairs, and there may be interference from the breeze generated by the operation of fans and air conditioners, or vibration interference caused by vehicles passing by outside the window. In addition, the elderly have different body shapes and sleeping positions (side-lying, supine-lying), and some elderly people have weak breathing (such as those with chronic obstructive pulmonary disease), which leads to huge differences in radar echo signals at different times.

[0078] In this embodiment, the data acquired by the millimeter-wave radar are all in complex form and saved in .csv file format. The acquired data may be strong interference-strong signal, i.e., acquired from a scene with significant environmental noise but a close target distance, such as an indoor environment with multiple people in a family; it may be weak interference-weak signal, i.e., acquired from a scene with low environmental interference but the target is located at the edge of radar detection or is obscured by objects, such as a monitoring position at the far end of an indoor space or behind a wall; it may be strong interference-weak signal, i.e., acquired from a scene with strong background noise and a weak target signal, such as an indoor edge area with electrical interference; or it may be weak interference-strong signal, i.e., acquired from an ideal scene with a quiet environment and a clear target signal, such as a specially designed vital sign monitoring environment.

[0079] In this embodiment, the collected data will undergo data preprocessing.

[0080] Specifically, a range-dimensional Fast Fourier Transform (FFT) is performed on the complex signal to convert the time-domain signal to the range-domain signal, thereby separating the target reflection signals at different range cells. Subsequently, the static DC components generated by inherent radar system biases and environmental clutter are eliminated using a mean-subtraction method to highlight dynamic vital signs. Next, the phase information of the complex signal is extracted from the target range cell with the strongest energy or highest signal-to-noise ratio in the dynamic signal, and a phase unwrapping algorithm is used to recover the wrapped phase values ​​from [-π, π] to the true continuous phase trajectory reflecting the subtle movements of the human chest and abdomen.

[0081] Secondly, following step S2, six types of features are extracted from the preprocessed signal data, and scene classification is achieved by combining weighted fusion with individual threshold judgment.

[0082] Reference Figure 2 As shown, based on the two dimensions of environmental interference intensity and target reflected signal intensity, this embodiment divides radar signals into four typical scenarios: strong interference-strong signal, weak interference-weak signal, strong interference-weak signal, and weak interference-strong signal.

[0083] Six features are extracted from the preprocessed radar signal: spectral flatness, phase standard deviation, estimated signal-to-noise ratio, vital sign band energy ratio, spectral peak amplitude, and signal amplitude stability.

[0084] Spectral flatness The calculation formula is:

[0085]

[0086] in, This represents the number of frequency points in the spectrum. For frequency The power spectral density at that location.

[0087] Phase standard deviation The calculation formula is:

[0088]

[0089] in, For time The signal phase value at that location.

[0090] Estimating signal-to-noise ratio The calculation formula is:

[0091]

[0092] in, Signal power within the vital signs frequency band (0.1-3.0Hz). This refers to power noise in other frequency bands.

[0093] Vital signs frequency band energy ratio The calculation formula is:

[0094]

[0095] in, Signal energy within the vital signs frequency band. This represents the total signal energy across the entire frequency band.

[0096] Peak amplitude of spectrum The calculation formula is:

[0097]

[0098] Signal amplitude stability The calculation formula is:

[0099]

[0100] In this embodiment, the preprocessed signal is divided into a 2-second overlapping time window, wherein... For the first The standard deviation of the signal amplitude within a window For the first The average signal amplitude within each window.

[0101] This embodiment uses a weighted fusion of the aforementioned spectral flatness, phase standard deviation, and signal amplitude stability characteristics as the basis for distinguishing between strong and weak interference; expressed by the formula:

[0102]

[0103] 1

[0104] If interference value If the value exceeds the set threshold, it is considered strong interference; otherwise, it is considered weak interference. , and These represent the corresponding weights.

[0105] To determine signal strength, the extracted estimated signal-to-noise ratio, vital sign frequency band energy ratio, and peak amplitude of the spectrum are compared individually with set thresholds. When all three are greater than the threshold, the system determines the signal to be a strong signal; otherwise, it determines the signal to be a weak signal.

[0106] Then, following step S3, the classified radar data is input into the model corresponding to its scene for signal enhancement or reconstruction.

[0107] Reference Figure 3 As shown, the strong interference-strong signal scenario adopts the one-dimensional convolutional neural network (1D U-Net) model, the weak interference-weak signal scenario adopts the temporal coding conditional generative adversarial network (TCGAN) model, the weak interference-strong signal scenario adopts the Kalman filter processing model, and the strong interference-weak signal scenario adopts the combined processing model of 1D U-Net + bidirectional SRU + Kalman filter + TECGAN.

[0108] In this embodiment, the 1D U-Net consists of an encoder and a decoder. The training dataset comprises radar signals collected under strong interference and strong signal conditions, and corresponding vital sign signals. During each training process, a batch of noisy signals is input into the network, and a denoised prediction signal is obtained through forward propagation. The difference between the prediction signal and the actual vital sign signal under the corresponding condition is calculated using the mean squared error loss function. The gradient of the loss value is calculated using the backpropagation algorithm, and the network weights are updated using the Adam optimizer.

[0109] Reference Figure 4 As shown, the classified strong interference-strong signal radar signal is input into the trained 1D U-Net model. First, it enters the encoder, which mainly consists of an initial double convolutional block and four consecutive downsampling blocks. The initial double convolutional block extracts the primary features of the signal, followed by four consecutive downsampling blocks (each consisting of a max-pooling layer and a double convolutional block). Each block halves the signal length through max pooling and then deepens the feature representation through the double convolutional blocks, gradually constructing a hierarchical representation from specific waveforms to abstract features. After the encoder path, the signal enters the decoder path, i.e., gradually restoring the signal resolution through four corresponding upsampling blocks. Each upsampling block first upsamples the feature map to the same size as the corresponding layer of the encoder, and then concatenates the feature map of the same scale as the encoder along the channel dimension through skip connections. Finally, the output convolutional layer maps the high-dimensional features at the end of the decoder back to the single-channel signal space, generating a denoised and enhanced signal with the same length as the input signal.

[0110] In this embodiment, the Temporal Coding Conditional Generative Adversarial Network (TCGAN) model is used to perform generative reconstruction and spectrum restoration of weak signals in weak interference-weak signal scenarios.

[0111] Reference Figure 5 As shown, the input low signal-to-noise ratio radar signal is first converted into a frequency domain representation through a fast Fourier transform, and then input into a conditional encoder for processing. The conditional encoder is composed of a fully connected layer and a ReLU activation function stacked together. Finally, the frequency domain features are compressed into a low-dimensional conditional vector through a fully connected layer. At the same time, a random noise vector sampled from a Gaussian distribution is concatenated with the conditional vector and used as the input of the generator.

[0112] The generator consists of an encoder and a decoder. The encoder comprises four one-dimensional convolutional blocks (encoder 1 to encoder 4), each containing one-dimensional convolution, batch normalization, and a Leaky ReLU activation function, progressively downsampling to extract deep features. The decoder is symmetrically composed of four one-dimensional deconvolutional blocks (decoder 1 to decoder 4), each containing one-dimensional deconvolution, batch normalization, and a ReLU activation function, progressively upsampling to restore signal resolution. Skip connections directly pass the output of each encoder layer to the corresponding layer in the decoder, enabling the decoder to reuse the local features and encoder information extracted by the encoder, improving the overall performance of the network. Finally, the output layer consists of a single one-dimensional convolutional layer.

[0113] In the discriminator section, the generated signal is first concatenated with the conditionally projected features processed by linear projection, and then input into the discriminator. The discriminator consists of four concatenated convolutional blocks, each of which sequentially performs one-dimensional convolution, batch normalization, and LeakyReLU activation. A self-attention mechanism is introduced in the intermediate layer of the discriminator, enabling the network to adaptively focus on key temporal features. Finally, after flattening, the discriminant result is output through a fully connected layer and a sigmoid activation function.

[0114] During training, the TCGAN model employs a dual supervision mechanism of L1 loss and spectral loss. The L1 loss directly constrains the point-by-point differences between the generated signal and the target clean signal in the time domain, ensuring accurate reconstruction of the basic waveform structure. The specific formula is as follows:

[0115]

[0116] in, B Indicates batch size, L Indicates the number of sampling points. Indicates the first i The first generated signal j One sampling point, Indicates the first i The first real signal j One sampling point.

[0117] The spectral loss is achieved by comparing the L1 distance of the amplitude spectra of the two signals using their Fast Fourier Transforms, ensuring that the frequency distribution of the generated signal matches that of real vital signs in the frequency domain. Emphasis is placed on energy matching within the 0.1-3.0 Hz frequency band. The specific formula is as follows:

[0118]

[0119] The above formula is the signal The formula for calculating the spectral amplitude, Indicates signal The orthogonal normalized Fourier transform.

[0120] The specific formula for the spectrum consistency loss is:

[0121]

[0122] Where L represents the number of sampling points, i.e., the signal length. Indicates the first n The weights of each frequency component, Indicates the generation signal n The amplitude of each frequency component Represents the true signal. n The amplitude of each frequency component.

[0123] The weighted sum of the two losses constitutes the overall optimization objective of the generator. During training, backpropagation updates the network parameters simultaneously, enabling the model to recover subtle respiratory and heartbeat waveform features with high fidelity while suppressing noise. The TECGAN module can reconstruct complete vital signs signals from input signals with extremely low signal-to-noise ratios, and its output will be sent to the quality assessment module for reliability verification.

[0124] The formula for calculating the overall optimization objective is:

[0125]

[0126]

[0127] in, The weights represent the magnitudes of the L1 loss function. This indicates the weight of the spectral loss.

[0128] In this embodiment, a Kalman filter processing model is used for dynamic state estimation and random noise smoothing in weak interference-strong signal scenarios.

[0129] Kalman filtering is based on a state-space model and mainly consists of two stages: prediction and update.

[0130] A state-space model of human micro-motion is constructed. This embodiment first defines a state vector containing displacement, velocity, and acceleration components of the human chest and abdomen micro-motions, and establishes a state transition equation describing the periodic motion patterns of vital signs, as well as an observation equation connecting the state vector and the radar observation phase signal; expressed by the formula:

[0131]

[0132]

[0133] in, This represents the state vector at the current sampling time. This is the state vector from the previous sampling time; in this embodiment, the preprocessed signal is used. Taking [k] as an example, its current state vector , For the first k The phase value at each sampling moment is extracted from the complex signal after radar beamforming. Its phase change rate is the first-order difference. F Transition matrix , The actual observed phase value. H Observation matrix . and This corresponds to the noise.

[0134] Initialize filter parameters; in this embodiment, the process noise covariance matrix Q and the observation noise covariance matrix R are adaptively set according to the signal-to-noise ratio estimate of the signal; in weak interference scenarios, a smaller R value is set to trust the observation data, and a moderate Q value is set to allow the state to change naturally with the rhythm of breathing and heartbeat.

[0135] In the prediction phase, this embodiment mainly estimates the current state based on the previous time step. and its error covariance It can be expressed by the formula:

[0136]

[0137]

[0138] in, Indicates based on k -1 time point k The predicted state at time 10:00, i.e., the prior estimate; The error covariance represents the predicted state. express k The optimal estimate at time -1, i.e., the posterior estimate; Let represent the error covariance matrix of the optimal estimate at time k-1. Initial values. = , This is the first actual observation, the initial value. ;

[0139] During the update phase, this embodiment obtains the actual observed values. Then, first calculate the Kalman gain. The calculation formula is:

[0140]

[0141] Using Kalman gain to estimate priors and Compared with actual observation The optimal posterior estimate is obtained by fusion:

[0142]

[0143]

[0144] in, I It is an identity matrix with the same dimensions as the state vector.

[0145] Finally, the posterior state estimate of the Kalman filter output Phase components in This is the smoothed enhanced phase signal. This signal effectively suppresses random noise and high-frequency interference, while fully preserving the macroscopic trends and dynamic details of vital signs, providing a high-quality input for subsequent variational mode decomposition.

[0146] In this embodiment, a combined processing model is used to perform cascaded denoising, timing modeling, and signal reconstruction for strong interference-weak signal scenarios.

[0147] The combined processing model in this embodiment consists of 1D U-Net, bidirectional SRU, Kalman filter, and TECGAN. The 1D U-Net and TECGAN modules are derived through transfer learning from corresponding modules in strong interference-strong signal and weak interference-weak signal scenarios. Specifically, the pre-trained 1D U-Net model (from the strong interference-strong signal scenario) retains its encoder's multi-scale noise feature extraction capability for initial strong interference suppression; the pre-trained TECGAN model (from the weak interference-weak signal scenario) reuses its conditional encoder's spectral understanding capability and the generator's detailed reconstruction capability for final signal recovery.

[0148] In this embodiment, firstly, a 1D U-Net model pre-trained under strong interference-strong signal scenarios is loaded, and the convolutional layer parameters of its encoder are frozen to retain its learned ability to extract strong noise features. Only the decoder is fine-tuned to adapt to the noise distribution differences in the new scenarios. Simultaneously, a TECGAN model pre-trained under weak interference-weak signal scenarios is loaded, and the shallow convolutional parameters in its conditional encoder and generator are frozen to retain its ability to encode vital sign spectral features and reconstruct basic signals.

[0149] Subsequently, a newly initialized bidirectional SRU (Simple Recurrent Unit) module is inserted between the two transfer models to construct the temporal modeling pathway. This module adopts a two-layer stacked bidirectional structure, with its core design targeting the long-range temporal characteristics of vital sign signals: the forward SRU propagates forward in time to capture the historical dependencies of the signal; the backward SRU propagates backward in time to utilize future contextual information. Each SRU unit contains a simplified gating mechanism—a forget gate controls the degree of retention of historical memory, and a reset gate adjusts the fusion ratio of the current input and the historical state. In the specific implementation, the signal output from the 1D U-Net is first compressed to 64 dimensions through a 1×1 convolutional layer, and then the compressed signal is input into the bidirectional SRU model for temporal propagation according to the following formula:

[0150] Candidate state:

[0151]

[0152] Forgotten Gate:

[0153]

[0154] Reset Door:

[0155]

[0156] Status after update:

[0157]

[0158] Output:

[0159]

[0160] in, For the signal after compression s The input vector has time steps, and the total number of time steps is the signal length, i.e., the number of sampling points L. This represents the candidate state after linear transformation. Represents the Gate of Oblivion This indicates that the door is being reset. Indicates the updated status. 、 、 These represent the corresponding weight matrices. 、 、 These represent the corresponding paranoia vectors. This represents element-wise multiplication. The forward SRU propagates forward in time, and the backward SRU propagates backward in time. The bidirectional outputs are concatenated at each time step. Then, it is reduced to 64 dimensions through a fully connected layer.

[0161] In this embodiment, the timing enhancement signal output by the bidirectional SRU is input into the Kalman filter. Finally, the signal smoothed by the Kalman filter is input into the fine-tuned TECGAN generator. By using the joint optimization of L1 reconstruction loss and spectral consistency loss, the high-fidelity reconstruction of weak vital signs signals is completed.

[0162] Finally, following step S4, the radar signal after model enhancement is evaluated for signal quality. The enhanced signal that meets the quality evaluation requirements is input into the variational mode decomposition module to separate the breathing signal and heartbeat signal.

[0163] There are three core metrics for signal quality assessment:

[0164] 1. Peak signal-to-noise ratio (PSNR): This measures the significance of vital signs relative to background noise in the frequency domain.

[0165] In this embodiment, the power spectral density is obtained by Welch power spectrum estimation of the signal. Within the vital signs analysis frequency band (0.1-3.0Hz), the maximum spectral peak is identified in the respiratory sub-band (0.1-0.5Hz) and the heartbeat sub-band (0.8-2.0Hz), respectively, and the larger of the two values ​​is taken as the peak power. And the median power value of all frequency points within the analysis band. The estimation of the background noise level is as follows:

[0166]

[0167] in, This represents the peak signal-to-noise ratio (SNR) of the spectrum. Peak power, This is the median estimate for background noise.

[0168] 2. Signal amplitude stability.

[0169] In this embodiment, the signal is divided into overlapping time windows with a duration of 2 seconds. The signal amplitude stability (SAS) of each window is calculated, and then the median of the SAS of all windows is taken. The smaller the value, the smoother the fluctuation of the signal amplitude over time and the higher the time domain stability. The formula is consistent with the SAS calculation formula in the six types of features in step S2.

[0170] 3. To measure the degree of concentration of vital signs energy in the frequency domain.

[0171] This embodiment extracts the respiratory frequency band. and heart rate band Calculate the maximum spectral peak power within the lifetime frequency band and the average power over the entire lifetime. PThe formula for measuring the concentration of vital energy in the frequency domain is as follows:

[0172]

[0173] In this embodiment, the calculated three indicators are linearly weighted and fused into a comprehensive quality score:

[0174]

[0175] The signal flow direction is determined after comparison with a preset threshold: if the threshold is exceeded, the signal enters the variational mode decomposition module; if the threshold is not exceeded, the signal is fed back to step S2 to re-extract features, classify scenes, and enhance signals, thus forming a closed-loop quality control system.

[0176] In this embodiment, the variational mode decomposition module is a VMD feature separation module. First, the module parameters are initialized, setting the number of modes M=3 to correspond to respiration, heartbeat, and residual components. The center frequencies of each mode are initialized to [0.3, 1.2, 3.0] Hz, i.e., the initial center frequency of the respiration component is 0.3, the initial center frequency of the heartbeat component is 1.2, and the initial center frequency of the residual component is 3. A penalty factor is then set. α =2000 to control modal bandwidth, Lagrange multipliers Set to zero.

[0177] Secondly, the four enhanced signals after quality assessment are collectively referred to as ,right Perform a Fourier transform to obtain its frequency domain representation. A constrained variational model containing data fidelity terms and bandwidth penalty terms is constructed in the frequency domain and iteratively updated using the alternating direction multiplier method. In each iteration, the frequency domain expression, corresponding center frequency, and Lagrange multipliers of each mode are updated synchronously until the convergence condition is met or the maximum number of iterations is reached. The mode update formula is:

[0178]

[0179] in, Indicates the first m The mode in the th ... q Frequency domain estimation after +1 iteration Indicates the first m’ The mode in the th ... q Frequency domain estimation after the next iteration Indicates the first q The Lagrange multipliers in the next iteration are initialized. =0, As a penalty factor, Indicates the first m The mode in the th ... q The center frequency at the next iteration is initialized. =[0.3, 1.2, 3.0], m It can be any one of the modes 1, 2, or 3. m’ To and m Two other different modes; It represents a continuous frequency variable.

[0180] The center frequency update formula is:

[0181]

[0182] in, Indicates the first m The mode in the th ... q The center frequency at +1 iteration, Indicates the first m The mode in the th ... q Power spectral density after +1 iterations.

[0183] The Lagrange multiplier update formula is:

[0184]

[0185] in, Indicates the first q Lagrange multipliers from +1 iteration, Indicates the update step size. M This represents the total number of modes, in this embodiment. M =3.

[0186] The iteration continues until the sum of the changes in the center frequencies of each mode in two consecutive iterations is less than 1. Hz or reaching the maximum number of iterations. After completion, analyze the three frequency domain modes obtained. Perform inverse Fourier transforms on each component to obtain the time-domain modal components. Finally, based on the final center frequencies of each mode... Physiological signal matching: The mode with a center frequency between 0.1-0.5 Hz is selected as the respiratory signal, and the mode with a center frequency between 0.8-2.0 Hz is selected as the heartbeat signal. The remaining modes are classified as residual components containing noise or interference, thus completing the separation of respiration and heartbeat.

[0187] This embodiment extracts six types of features, categorizes the signal into four scene types, and employs a dedicated signal enhancement model for each category, effectively overcoming the limitations of traditional single algorithms. The enhanced signal undergoes rigorous signal quality testing, and finally, life-organization features are extracted using VMD (Virtual Machine Depth Analysis). This embodiment improves signal quality while ensuring the accuracy of feature extraction, and significantly enhances the system's adaptability and reliability in complex environments.

[0188] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0189] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for separating biological features based on millimeter-wave radar signal enhancement, characterized in that, Includes the following steps: S1. Acquire the raw radar signal of the life monitoring target collected by the millimeter-wave radar and perform preprocessing; S2. Extract six types of features from the preprocessed signal, and divide the preprocessed signal into four scene categories based on the six types of features; The six types of features include: spectral flatness, phase standard deviation, estimated signal-to-noise ratio, vital sign band energy ratio, spectral peak amplitude, and signal amplitude stability. The four scenario categories include: strong interference-strong signal, weak interference-weak signal, strong interference-weak signal, and weak interference-strong signal. Based on the six types of features, the signals are divided into four scene categories, specifically including: The spectral flatness, phase standard deviation, and signal amplitude stability are weighted and fused, and then compared with a first threshold. If the value is greater than the first threshold, it is determined to be strong interference; otherwise, it is determined to be weak interference. The estimated signal-to-noise ratio, vital sign frequency band energy ratio, and spectral peak amplitude are compared with a preset second threshold; if all of them are greater than the second threshold, they are judged as strong signals, otherwise they are judged as weak signals. S3. Based on the identified scene categories, the corresponding signal processing model is invoked to enhance or reconstruct the preprocessed signal to obtain the enhanced radar signal; the signal processing model includes: a one-dimensional convolutional neural network model, a temporal coding conditional generative adversarial network model, a Kalman filter processing model, and a combined processing model; S4. The enhanced radar signal is evaluated for signal quality. The enhanced signal that meets the quality evaluation requirements is input into the variational mode decomposition module to separate the respiratory signal and heartbeat signal of the life monitoring target.

2. The method as described in claim 1, characterized in that, In step S1, the preprocessing includes: Perform a range-dimensional fast Fourier transform on the original radar signal in complex form to convert the time-domain signal into a range-domain signal; The static DC component in the range domain signal is eliminated by subtracting the mean, resulting in a dynamic signal. Phase information is extracted from the dynamic signal, and the phase value is restored to a true continuous phase trajectory reflecting the micro-movements of the chest and abdomen of a living organism through a phase unwrapping algorithm.

3. The method as described in claim 1, characterized in that, Step S3 specifically includes: For signals classified into strong interference-strong signal scenarios, a one-dimensional convolutional neural network model is used for interference removal and feature extraction to obtain enhanced radar signals. For signals classified as weak interference-weak signal scenarios, a temporal coding conditional generative adversarial network model is used to perform generative reconstruction and spectrum restoration of weak signals to obtain enhanced radar signals. For signals classified into weak interference-strong signal scenarios, a Kalman filter processing model is used to perform dynamic state estimation and random noise smoothing to obtain the enhanced radar signal. For signals classified into strong interference-weak signal scenarios, a combined processing model is used to perform cascaded denoising, temporal modeling, and signal reconstruction to obtain enhanced radar signals. The combined processing model is based on a combination of a one-dimensional convolutional neural network model, a bidirectional simple recurrent unit, a Kalman filter processing model, and a temporal coding conditional generative adversarial network model.

4. The method as described in claim 3, characterized in that, The one-dimensional convolutional neural network model includes: an encoder and a decoder; The encoder includes dual convolutional blocks and four cascaded downsampling modules, used to extract high-dimensional deep features of the signal layer by layer and compress the time dimension; The decoder includes four cascaded upsampling modules and a one-dimensional convolutional block, used to recover the temporal resolution of the signal and reconstruct the waveform layer by layer; The encoder and decoder perform feature fusion through skip connections between corresponding layers, passing the shallow detail features extracted by the encoder to the decoder to compensate for the spatial information loss during the upsampling process.

5. The method as described in claim 3, characterized in that, The temporal coding conditional generative adversarial network model uses a weighted sum of L1 loss and spectral consistency loss as the optimization objective during training. The L1 loss is used to calculate the absolute error between the reconstructed signal and the true signal in the time domain, and is expressed by the formula: ； The spectral consistency loss is used to calculate the mean square error between the reconstructed signal and the true signal in the frequency domain, and is expressed by the formula: ； in, Indicates batch size, Indicates the number of sampling points. Indicates the first The first generated signal One sampling point, Indicates the first The first real signal One sampling point; Indicates the first The weights of each frequency component, Indicates the generation signal number 1 The amplitude of each frequency component Represents the true signal. The amplitude of each frequency component.

6. The method as described in claim 3, characterized in that, The process of using Kalman filtering to perform dynamic state estimation and random noise smoothing specifically includes: Construct a linear state-space model; the linear state-space model includes: state transition equations and observation equations; Initialize the filter parameters, use the optimal state estimate from the previous time step, calculate the prior state estimate and its error covariance from the state transition equation, and perform time update. Calculate the Kalman gain coefficient, and use the actual observation value at the current time to correct the prior state estimate to obtain the posterior optimal state estimate; The displacement component is extracted from the posterior optimal state estimate as the enhanced radar signal.

7. The method as described in claim 3, characterized in that, The construction and data processing of the combined processing model include: Load a pre-trained one-dimensional convolutional neural network model and freeze its encoder parameters, then fine-tune only the decoder; Load the pre-trained temporal coding conditional generative adversarial network model and freeze the shallow convolution parameters in its conditional encoder and generator. A bidirectional simple recurrent unit is inserted between two transfer models to construct a temporal modeling path. The signal output by the one-dimensional convolutional neural network model is compressed and then input into the bidirectional simple recurrent unit for temporal propagation. The signal output from the bidirectional simple cyclic unit is input into the Kalman filter processing model for smoothing. The smoothed signal is input into the finely tuned temporal coding conditional generative adversarial network model generator for reconstruction.

8. The method as described in claim 1, characterized in that, In step S4, the enhanced radar signal is evaluated for signal quality, specifically including: Based on the enhanced radar signal, three core indicators are calculated; the three core indicators include: peak signal-to-noise ratio, signal amplitude stability, and concentration of vital signs energy. The three core indicators are linearly weighted and fused into a comprehensive quality score, which is then compared with a preset threshold.

9. The method as described in claim 1 or 8, characterized in that, In step S4, the variational mode decomposition module separates the respiratory signal and the heartbeat signal, specifically including: Based on the spectral characteristics of human vital signs, the total number of modes in the variational mode decomposition module is set to 3, corresponding to low-frequency baseline drift, respiratory component, and heartbeat component, respectively; the center frequency of each mode is initialized. The enhanced radar signal that meets the quality assessment requirements is subjected to Fourier transform to obtain its frequency domain representation; a constrained variational model containing data fidelity terms and bandwidth penalty terms is constructed in the frequency domain and iteratively updated using the alternating direction multiplier method; in each iteration, the frequency domain expression of each mode, the corresponding center frequency, and the Lagrange multiplier are updated synchronously until the convergence condition is met or the maximum number of iterations is reached. After the iteration is completed, perform an inverse Fourier transform on the frequency domain expression of each mode to obtain the corresponding time domain eigenmode components; The dominant frequencies of each intrinsic mode component in the time domain are calculated, and respiratory and heartbeat signals are matched and filtered according to a preset range of human physiological frequencies; wherein the dominant frequency of the respiratory signal is between 0.1Hz and 0.5Hz, and the dominant frequency of the heartbeat signal is between 0.8Hz and 2.0Hz.

Images