Radar breath monitoring controllable data enhancement method based on parameterized environment modeling

Abstract

The present application relates to the technical field of digital health, wireless sensing, signal processing and artificial intelligence, in particular to a radar respiratory monitoring controllable data enhancement method based on parameterized environment modeling. The data enhancement method first acquires a pure radar respiratory signal, constructs four types of models with physical constraint coupling and corresponding parameter space, and calculates the true value of each interference. Then, according to the physical mechanism, the noise signal is synthesized by convolution / time-varying filtering, phase modulation and additive superposition operators to form a four-tuple data containing the pure radar respiratory signal, the noise signal, the true value of the interference and the model parameters. After generating a batch of data sets, multi-level physical fidelity verification is carried out, and if it does not meet the standard, the parameters are adjusted: scene fingerprints are extracted from real monitoring data to obtain parameter adjustment and update the parameter space. The present application solves the problems of missing physical modeling and unreasonable synthesis mechanism, effectively improves the simulation data precision, model generalization ability and respiratory monitoring robustness.

Description

Technical Field

[0001] This invention relates to the fields of digital healthcare, wireless sensing, signal processing and artificial intelligence, specifically the field of controllable data enhancement technology for radar respiratory monitoring. Background Technology

[0002] Millimeter-wave radar, with its non-contact, anti-interference, and all-weather operation characteristics, is widely used in fields such as human respiratory monitoring and health surveillance. However, in actual deployment scenarios, respiratory monitoring signals are easily affected by multiple complex factors such as multipath propagation reflection, human motion artifacts, physiological activity interference, electromagnetic noise, and narrowband interference, leading to a significant decrease in monitoring accuracy. Existing technologies typically rely on collecting large amounts of real-world labeled data for model training, but acquiring real-world data is costly, difficult to label, and has limited coverage. Therefore, data augmentation has become a core means to improve model robustness.

[0003] Current mainstream radar respiration monitoring data augmentation methods suffer from several inherent flaws: First, they often employ simple noise superposition and signal translation / flipping, failing to construct a parameterized modeling system for multipath propagation, human motion artifacts, physiological interference, electromagnetic interference, and noise. Furthermore, they lack physical constraints and coupling relationships for these four types of interference, resulting in simulation data that is disconnected from the physical laws of the real scene and distorted interference characteristics. Second, the synthesis of interference signals does not follow actual physical mechanisms, failing to differentiate between multipath propagation, motion artifacts, physiological interference, and electromagnetic noise using convolution / time-varying filtering, phase modulation, additive superposition, and other differentiated operators, leading to insufficient physical plausibility in the synthesized signals. Third, the generated datasets only contain raw signal samples, lacking a pure radar respiration profile. The structured annotation information, such as signal, interference ground truth, and model parameters, cannot support the model's physically decoupled learning of pathological signals and environmental interference, and the model reasoning lacks interpretability. Fourth, there is a lack of a multi-level physical fidelity verification mechanism for simulation data, the data quality cannot be quantitatively evaluated, and substandard data is difficult to automatically iterate and optimize. Fifth, the model parameter space is mostly statically preset, and it is impossible to extract scene fingerprints from real monitoring data to achieve parameter space mapping adjustment and closed-loop adaptive evolution. This makes it difficult for augmented data to adapt to the personalized interference characteristics of real scenes, the model has weak generalization ability in unfamiliar scenes, and the incremental optimization efficiency is low. Ultimately, it is difficult to meet the high-precision and high-robustness millimeter-wave radar respiratory monitoring requirements in complex real environments.

[0004] To overcome the above problems, there is an urgent need for a controllable data augmentation method based on parametric environment modeling, with physical constraint coupling, capable of physical fidelity verification and closed-loop evolution of parameter space, so as to achieve high-quality data generation that fits the real scene and improve the scene adaptability and monitoring accuracy of respiratory monitoring models. Summary of the Invention

[0005] To address the technical problems of existing millimeter-wave radar respiratory monitoring data augmentation methods, which suffer from simulation data distortion, insufficient model generalization, and inadequate monitoring robustness due to a lack of physically constrained coupling modeling, unreasonable data synthesis mechanisms, lack of physical fidelity verification, and inability of the parameter space to adaptively evolve, this invention provides a controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling. Furthermore, this invention also provides a controllable data augmentation system for millimeter-wave radar respiratory monitoring based on parametric environment modeling for implementing this method, and a millimeter-wave radar for respiratory monitoring.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling, characterized by the following steps:

[0008] Acquire clean radar breathing signals;

[0009] We constructed a multipath propagation model with physical constraints and coupling, a human motion artifact model, a physiological interference model, and an electromagnetic interference and noise model, and determined the parameter spaces corresponding to the four types of models.

[0010] Based on the four types of models and the values ​​of their parameters in the parameter space, the true values ​​of each disturbance are calculated.

[0011] Based on the pure radar breathing signal and the true values ​​of each interference, the noise-added signal is calculated sequentially through convolution / time-varying filtering operator, phase modulation operator, and additive superposition operator according to the physical mechanism.

[0012] The data is integrated into a four-tuple data structure, which includes the pure radar breathing signal, the noisy signal, the true values ​​of each interference, and the parameter values ​​of each model.

[0013] Repeat the above steps to generate augmented datasets consisting of quadruple data structures in batches, and perform multi-level physical fidelity verification on the augmented datasets. If the verification fails, trigger an iterative parameter tuning process until the standard is met: extract scene fingerprints from real monitoring data, map scene fingerprints to corresponding parameter space adjustment values, update the parameter space based on the parameter space adjustment values, and realize closed-loop adaptive evolution of the parameter space.

[0014] As a further improvement to the above scheme, the parameters of the four types of models with physical constraint coupling are:

[0015] The parameters of the multipath propagation model include path delay, path propagation distance, and path complex amplitude, which satisfy the physical constraint coupling of the multipath propagation model, which is composed of delay-distance constraints, amplitude-distance constraints, and path number-delay spread constraints.

[0016] The parameters of the human motion artifact model include motion amplitude, motion duration, motion velocity, and motion acceleration, which satisfy the physical constraint coupling of the human motion artifact model, which is composed of amplitude-duration constraints, rise-fall time constraints, and velocity-acceleration constraints.

[0017] The parameters of the physiological disturbance model include heart rate, heart rate variability, and heart rate amplitude, which satisfy the physical constraint coupling of the physiological disturbance model, which is composed of heart rate-variability constraints and heart rate-amplitude constraints.

[0018] The parameters of the electromagnetic interference and noise model include thermal noise power, interference frequency, interference bandwidth, and sampling frequency, which satisfy the physical constraint coupling of the electromagnetic interference and noise model, which is composed of thermal noise power-bandwidth constraints and narrowband interference frequency-sampling rate constraints.

[0019] As a further improvement to the above scheme, the delay-distance constraint is:

[0020] ;

[0021] in, Let d be the time delay of the i-th path at time t. i (t) represents the propagation distance of the i-th path at time t; c represents the radar wave speed;

[0022] The amplitude-distance constraint is:

[0023] ;

[0024] Where, α i (t) represents the complex amplitude of the i-th path at time t; ∝ represents the proportionality sign; |Γ| represents the absolute value of the reflection coefficient Γ;

[0025] Path number-delay spread constraint: Number of effective multipath paths N mp With multipath delay spread There is a positive statistical correlation between them;

[0026] Amplitude-duration constraint: A during rolling over movement roll ·T roll =constant, A roll For the range of motion of rolling over, T roll The duration of the rolling motion;

[0027] Ascent-descent time constraint: Coughing movement T rise <T fall T rise T represents the duration of thoracic displacement from resting state to peak value. fall The duration of the movement in the thoracic cavity as it recovers from its peak displacement to its resting state;

[0028] Velocity-acceleration constraint: the velocity v of the k-th event. k The acceleration a of the motion of the kth event k Meeting the physiological safety threshold, v max ≤2m / s, a max ≤5m / s 2 v max a is the maximum speed of the event. max This represents the maximum acceleration of the event's motion.

[0029] Heart rate variability constraint: HRV = 0.1·f heart HRV stands for heart rate variability, f heart Heart rate;

[0030] Heart rate-amplitude constraint: A heart ∝1 / f heart A heart This refers to the amplitude of the heartbeat.

[0031] Thermal noise power-bandwidth constraint: , Where B is the thermal noise power, and B is the noise bandwidth.

[0032] Narrowband interference frequency-sampling rate constraint: f q <F s / 2,f q For narrowband interference frequency, F s The sampling frequency.

[0033] As a further improvement to the above scheme, the calculation process for the noisy signal is as follows:

[0034] First, multipath propagation interference is synthesized using a convolution / time-varying filter operator:

[0035] ;

[0036] Among them, s multipath (t) represents the convolution / time-varying filter operator at time t, i.e., the multipath propagation interference at time t; The impulse response of a time-varying multipath channel is given by the expression representing the time delay of the multipath channel at time t. The response of the signal components; for The pure radar breathing signal generated only by the target's minute breathing movements. This indicates that time delay has elapsed at time t. The moment after; For time delay The differential;

[0037] Next, a phase modulation operator is used to synthesize human motion artifact interference and physiological interference:

[0038] ;

[0039] Among them, s phase (t) represents the phase modulation operator at time t, i.e., the human motion artifact interference and physiological interference at time t; exp(·) is an exponential function with the natural constant as the base; j is the imaginary unit; π is pi; λ is the carrier wavelength of the radar signal; Δd total (t) represents the total change in the round-trip path caused by the slight movement of the target at time t;

[0040] Finally, an additive superposition operator is used to synthesize electromagnetic interference and thermal noise:

[0041] ;

[0042] Among them, sIF output (t) is the additive superposition operator at time t, i.e., the noisy signal at time t; G m Let i be the gain coefficient for the m-th type of electromagnetic interference. EMI,m (t) represents the m-th type of electromagnetic interference base signal, i EMI,m (t)=i NBI (t)+i WBI (t), i NBI (t) represents the narrowband interference signal at time t, i WBI (t) represents the broadband interference signal at time t; w th (t) represents the thermal noise at time t.

[0043] As a further improvement to the above scheme: multi-level physical fidelity verification refers to the verification of data generated by four types of models with physical constraint coupling. When any physical fidelity verification index fails to reach the preset threshold, an iterative parameter tuning process is triggered until the threshold is met. The specific verification content is as follows:

[0044] Multipath power delay spectrum KS statistic DKS < 0.15, delay-distance mapping error Eego < 5%, phase change rate correlation with motion velocity Rvel > 0.95, motion amplitude A - duration T product A·T ∈ [physiological range], heart rate estimation error EHR < 3%, heart rate variability HRV and heart rate f heart The correlation RHR > 0.9, the spectral structure similarity SSIMspec > 0.85, and the thermal noise power The consistency error with the noise bandwidth B is <5%, and the F1 score of the downstream respiratory monitoring task is >0.85.

[0045] As a further improvement to the above scheme: mapping the scene fingerprint to the corresponding parameter space adjustment amount refers to performing physically constrained directional adjustments within the physically constrained parameter space based on the deviation between the scene fingerprint and the corresponding parameters in the current parameter space. The specific directional adjustment rules are as follows:

[0046] When the actual multipath delay spread is greater than the multipath delay spread value in the current parameter space, increase the upper limit of the multipath delay spread sampling in the current parameter space to the target sampling upper limit:

[0047] ;

[0048] in, , These are the upper limits for the multipath delay spread in the k-th and k+1th iterations, respectively. This is the amplification factor for multipath delay spread;

[0049] When the actual rolling event frequency is greater than the rolling event frequency value in the current parameter space, increase the rolling event sampling probability in the current parameter space to the target sampling probability:

[0050] ;

[0051] in, , The sampling probabilities of the overturning event in the k-th and k+1th iterations are respectively; δ f p is the amplification factor for the sampling probability of the overturning event. max The physical upper limit of the sampling probability for a comeback event;

[0052] When the true heart rate deviates from the heart rate value within the current parameter space, adjust the heart rate parameter sampling center in the current parameter space to the target sampling center:

[0053] ;

[0054] ;

[0055] in, , Let be the mean and standard deviation of the heart rate distribution in the (k+1)th iteration, respectively. Let f be the standard deviation of the heart rate distribution in the k-th iteration. heart,real For true heart rate, δ h This is the adjustment factor for the standard deviation of heart rate;

[0056] When the actual background noise is greater than the background noise value in the current parameter space, reduce the lower limit of the signal-to-noise ratio sampling in the current parameter space to the target sampling lower limit:

[0057] ;

[0058] in, , The lower sampling limits of the signal-to-noise ratio (SNR) for the k-th and k+1-th iterations are δ. NF This is the adjustment step size for the lower limit of signal-to-noise ratio (SNR) sampling.

[0059] As a further improvement to the above scheme: extracting scene fingerprints from real monitoring data is based on noisy signals and a four-tuple data structure. This involves extracting scene fingerprints from real monitoring data that correspond one-to-one with the four types of models, specifically including:

[0060] Multipath fingerprinting: estimating power delay spectrum, multipath delay spread, and principal path percentage;

[0061] Motion fingerprints: frequency of rolling over events, distribution of rolling over amplitude, and duration of rolling over;

[0062] Physiological fingerprints: heart rate, heart rate variability, intestinal motility;

[0063] Electromagnetic fingerprint: background noise spectrum, narrowband interference frequency points.

[0064] As a further improvement to the above scheme, the quadruple data structure is a data structure that supports physically decoupled learning of pathological signals and environmental interference. Specifically, it includes: a pure radar respiratory signal as the carrier of pathological signals, a noisy signal simulating a real monitoring scenario, the true value of interference containing all components of multipath, motion, physiological and electromagnetic interference, and parameter value labels for scenario inversion and personalized adaptation.

[0065] A controllable data augmentation system for millimeter-wave radar respiratory monitoring based on parametric environment modeling is provided for executing a controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling, comprising:

[0066] A clean radar breathing signal acquisition module is used to acquire clean radar breathing signals.

[0067] The parameterized environment model library construction module is used to construct multipath propagation models with physical constraints and coupling, human motion artifact models, physiological interference models, and electromagnetic interference and noise models, and to determine the parameter spaces corresponding to the four types of models.

[0068] The interference truth value calculation module is used to calculate the true values ​​of each interference based on the four types of models and the values ​​of their parameters in the parameter space.

[0069] The physical mechanism differentiation synthesis module is used to calculate the noise-added signal based on the pure radar breathing signal and the true values ​​of each interference according to the physical mechanism, through convolution / time-varying filtering operator, phase modulation operator, and additive superposition operator.

[0070] The quadruple data encapsulation module is used to integrate and form a quadruple data structure containing the pure radar breathing signal, the noise-added signal, the true values ​​of each interference, and the values ​​of each model parameter.

[0071] The dataset generation and physical fidelity verification module is used to generate augmented datasets in batches and perform multi-level physical fidelity verification on the augmented datasets. If the verification fails, parameter tuning is automatically triggered until the verification is successful.

[0072] The closed-loop adaptive evolution module is used to extract scene fingerprints from real monitoring data, map scene fingerprints to corresponding parameter space adjustment values, update the parameter space based on the adjustment values, and realize closed-loop adaptive evolution of the parameter space.

[0073] A millimeter-wave radar for respiratory monitoring includes a radar radio frequency transceiver unit and a controllable data enhancement system for millimeter-wave radar respiratory monitoring based on parametric environment modeling.

[0074] Compared with the prior art, the beneficial effects of the present invention are:

[0075] 1. In the data augmentation method, a clean radar breathing signal is first obtained as the basic signal source. Then, four parameterized models coupled with physical constraints are constructed for multipath propagation, human motion artifacts, physiological interference, and electromagnetic interference and noise, and corresponding parameter spaces are defined. Based on the model and parameters, the accurate ground truth of the interference is calculated. Subsequently, strictly following the physical transmission mechanism of the signal, convolution / time-varying filtering, phase modulation, and additive superposition operators are used to differentiate and synthesize the noise-added signal. At the same time, the clean radar breathing signal, the noise-added signal, the ground truth of the interference, and the model parameters are integrated to form a quadruple that can support physically decoupled learning. The data structure involves batch generating datasets and then verifying data quality through multi-level physical fidelity. Data that fails to meet the standards is automatically adjusted and optimized. Finally, multi-dimensional scene fingerprints are extracted from real monitoring data, mapped to directional adjustment values ​​in the parameter space, and the parameter space is updated. This achieves closed-loop adaptive evolution of the parameter space, fundamentally solving the problems of disconnect between simulation data and the physical laws of real scenes, unreasonable interference synthesis, lack of structured data labeling, and static solidification of the parameter space. Ultimately, this improves the authenticity and adaptability of the enhanced data and optimizes the generalization ability and monitoring robustness of downstream respiratory monitoring models.

[0076] 2. This invention provides a unified modeling of multiple physical mechanism parameters. Based on a parametric environment modeling framework, it integrates four core physical mechanism parameters: multipath propagation, human motion, physiological interference, and electromagnetic interference. Each parameter has a clear physical meaning: time delay is related to room space size, motion amplitude corresponds to the range of motion of human joints, heart rate reflects the state of the cardiovascular system, and signal-to-noise ratio characterizes the noise characteristics of the receiver. This enables the simulation-generated data to directly match the characteristics of the real physical world, ensuring the physical relevance and authenticity of the data from the source.

[0077] 3. This invention constructs a parameter space with physical constraint coupling characteristics. This parameter space is not a simple combination of independent parameters, but forms a high-dimensional space containing multi-dimensional physical constraint coupling such as time delay-distance, amplitude-duration, heart rate-heart rate variability, and thermal noise power-bandwidth. By constraining the parameter value logic through built-in physical coupling rules, it ensures that each generated sample strictly conforms to objective physical laws, effectively avoids invalid samples that violate physical common sense, and greatly improves the rationality and reliability of simulation data.

[0078] 4. This invention realizes a physical mechanism-driven parameter-signal mapping relationship. The conversion of parameters to signals strictly follows the corresponding physical eigenvalue equations. Convolution is used to address multipath interference, phase modulation is used to address motion artifacts, and additive superposition is used to address electromagnetic interference to complete signal synthesis. By relying on physical mechanisms to construct the inherent coupling relationship between parameters and signals, not only is the physical authenticity of the generated signal guaranteed, but the entire data generation process also has clear interpretability.

[0079] 5. This invention establishes a physical correspondence binding mechanism between parameter types and synthesis operators. Based on the physical nature of different interferences, it matches exclusive signal synthesis operators to various parameters, fundamentally solving the problem of physical logic inconsistency caused by the decoupling of parameters and synthesis operators in the prior art. Among them, multipath interference is bound to the convolution operator, motion and physiological interference is bound to the phase modulation operator, and electromagnetic interference is bound to the additive superposition operator. This binding relationship is determined by physical mechanism rather than being arbitrarily set by humans, further ensuring the physical rigor of signal synthesis.

[0080] 6. This invention achieves closed-loop adaptive evolution of the parameter space. The parameter space is not statically fixed, but is continuously iteratively optimized based on feedback from real scene data. A complete closed loop is formed by scene fingerprint extraction, parameter space mapping adjustment and incremental iterative optimization, so that the parameter distribution continuously approaches the real physical characteristics of the target scene, reducing the system's personalized adaptation time from several weeks to several hours, while significantly reducing the cost of scene adaptation.

[0081] 7. This invention can effectively support the physical decoupled representation learning of pathological signals and environmental interference. By constructing a four-tuple data structure of "pure radar respiratory signal - noisy signal - interference truth value - physical parameters", it provides a training basis for downstream monitoring models that can be decoupled for learning. After the model is deployed, it can synchronously output pathological event results and estimates of various interference components, realizing end-to-end interpretable respiratory monitoring, effectively filling the technical gap in AI interpretability in the field of medical millimeter-wave radar.

[0082] 8. This invention enables the model to perform accurate inference in zero-sample scenarios. The model internalizes the physical laws of the parameterized environment model during the pre-training stage. During the cold start stage, it can distinguish whether the signal change comes from a pathological event or environmental interference by relying on physical cognition without relying on real data of the target scene. Even in a low signal-to-noise ratio and strong interference environment of 0dB, the model performance retention rate is still higher than 85%, and the misjudgment rate of pathological events is significantly reduced in zero-sample scenarios.

[0083] 9. This invention has significant advantages in data generation efficiency and application cost. A single server can generate tens of thousands of samples per day, which greatly improves data supply efficiency, significantly reduces data collection costs, and effectively shortens the R&D cycle of medical radar monitoring systems. It provides key technical support for the large-scale implementation and commercial application of millimeter-wave radar in the field of respiratory health monitoring. Attached Figure Description

[0084] Figure 1 Flowchart for a method to enhance controllable data.

[0085] Figure 2 This is a schematic diagram of a quadruple data structure.

[0086] Figure 3 The graph shows a comparison of the power delay spectrum between real-world and simulation-generated data.

[0087] Figure 4 The graph shows the F1 scores of the downstream respiratory monitoring task for the two models under different signal-to-noise ratio conditions. Detailed Implementation

[0088] 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.

[0089] This invention addresses the core shortcomings of existing millimeter-wave radar respiratory monitoring data augmentation methods, such as parameters lacking physical meaning, unconstrained parameter space, signal mapping violating physical laws, decoupling of parameters and synthesis operators, and poor scene adaptability. It proposes a controllable data augmentation scheme based on parameterized environment modeling. This scheme constructs four types of parameterized models coupled with physical constraints: multipath propagation, human motion artifacts, physiological interference, and electromagnetic interference and noise. It then forcibly binds three types of synthesis operators—convolution, phase modulation, and additive superposition—according to the physical nature of the interference to generate high-fidelity noise-added signals. The output is a four-tuple data structure consisting of a clean radar respiratory signal, the noise-added signal, the ground truth of the interference, and the physical parameters. Simultaneously, relying on multi-level physical fidelity verification and a closed-loop adaptive evolution mechanism based on real-world data scene fingerprints, it continuously optimizes the parameter space to fit the physical characteristics of the target scene. This supports downstream models in achieving physically decoupled learning of pathological respiratory signals and environmental interference, significantly improving the detection accuracy of respiratory monitoring models in zero-sample and complex interference scenarios, and significantly reducing the cost and cycle of personalized scene adaptation. This provides key technical support for the large-scale deployment of millimeter-wave radar in the field of medical and health monitoring.

[0090] I. Controllable Data Augmentation Methods

[0091] like Figure 1 As shown, this data augmentation method strictly follows the physical mechanisms and physical constraints of multipath propagation, human motion artifacts, physiological interference, electromagnetic interference, and noise. It is implemented step-by-step according to a complete process: acquiring clean radar breathing signals, constructing parameterized models, physical mechanism-driven synthesis, quadruple data encapsulation, physical fidelity verification, and closed-loop adaptive evolution. The specific details are as follows:

[0092] (a) Acquiring pure respiratory radar signals

[0093] Acquiring pure breathing radar intermediate frequency signal sIF clean (t), this signal is generated solely by the respiratory micro-motion excitation of the monitored target's thoracic cavity, and contains no environmental multipath propagation, human motion artifacts, physiological crosstalk, electromagnetic noise, or other external interference components. sIF clean (t) can be obtained in two ways: by real-world collection in a shielded, controlled experimental environment, or by numerical simulation based on the respiratory microphysical mechanism of millimeter-wave radar.

[0094] (ii) Constructing a parametric environment model library

[0095] The parameterized environment model library constructed in this invention completes unified parameterized modeling for four core interference factors in millimeter-wave radar respiratory monitoring scenarios: multipath propagation, human motion artifacts, physiological interference, and electromagnetic interference and noise. The core characteristics of this model library are: each type of parameter has a clear and explicit physical meaning; the overall parameter space is constrained by physical rules to form a coupled relationship; and the mapping relationship between parameters and radar signals strictly follows the physical eigenvalue equations, ensuring the physical authenticity and logical rigor of the subsequent data augmentation process from the bottom layer of modeling.

[0096] 1. Multipath propagation model: Physical parameters of electromagnetic propagation

[0097] (1) The physical nature of multipath propagation

[0098] Electromagnetic waves propagate through space via reflection, refraction, and diffraction, forming multiple paths, each with specific time delay, attenuation, and phase. These parameters directly correspond to room geometry (size, layout) and material properties (reflection coefficient, absorptivity).

[0099] (2) Parametric model

[0100] The multipath propagation model uses the equivalent time-varying multipath channel to affect the echo, rather than a simple additive superposition. The real-time (time t) multipath channel impulse response h... mp (t) can be written as:

[0101] ;

[0102] (3) Physical mechanism parameters

[0103] N mp : Number of effective multipath paths, representing the complexity of the scenario.

[0104] The time delay of the i-th path at time t is determined by the propagation distance. It directly reflects the room size, where d i (t) represents the propagation distance of the i-th path at time t, and c represents the radar wave speed.

[0105] α i (t): The complex amplitude of the i-th path at time t, determined jointly by free space loss and reflection coefficient, α i (t)∝|Γ| / (d i (t)) 2 Γ reflects the material properties, where Γ is the reflection coefficient.

[0106] The phase of the i-th path at time t is determined by the propagation distance and the reflection phase shift.

[0107] δ(·): Impulse function.

[0108] (4) Physical constraint coupling

[0109] Delay-distance constraints: The time delay is strictly coupled with the propagation distance.

[0110] Amplitude-Distance Constraint: α i (t)∝|Γ| / (d i (t)) 2 The amplitude is inversely proportional to the square of the distance.

[0111] Path number - delay spread constraint: N mp With multipath delay spread There is a positive statistical correlation between them; the larger the room, the more paths there are, and the greater the latency spread.

[0112] (5) Parameter space

[0113] ;

[0114] Among them, Ω mp N represents the parameter space of the multipath propagation model. min N max These are the minimum and maximum values ​​of the number of effective multipath paths, respectively. , These represent the minimum and maximum path delays, respectively; |α i | min 、|α i | max Let be the minimum and maximum values ​​of the complex amplitude of the i-th path, respectively; ∠α i (t) is α i The principal argument of (t).

[0115] Parameter space Ω mp All parameters strictly satisfy the physical constraint coupling conditions preset by the corresponding model, ensuring that the parameter values ​​and combinations completely conform to the inherent laws of the real physical scene.

[0116] (6) Summary

[0117] The multipath propagation model, a core component of the parametric environment model library, is built upon the physical mechanism of electromagnetic propagation. Its physical essence lies in the fact that electromagnetic waves, within the monitoring space, form multiple propagation paths through reflection, refraction, and diffraction. The time delay, attenuation, and phase parameters corresponding to each path can be directly mapped to the room's geometric dimensions, layout, and the reflection and absorption properties of materials such as walls and floors. This model employs an equivalent time-varying multipath channel acting on the radar echo, abandoning the simple additive superposition method. It achieves parametric characterization through the multipath channel impulse response. Core physical mechanism parameters include the number of effective multipath paths, path length, and other parameters. Path delay, complex amplitude, and phase are parameters that characterize physical features such as scene complexity, room size, and material properties, respectively. The model establishes three types of strong physical constraint coupling relationships: delay-distance, amplitude-squared distance, and path number-delay spread, which constrain the parameter value logic from the perspective of physical laws. At the same time, a standardized parameter space including path number, delay, complex amplitude, and argument is defined. All parameter values ​​and combinations strictly adhere to the preset physical constraint coupling conditions to ensure that the multipath propagation modeling fully conforms to the real indoor electromagnetic propagation law, providing underlying support for subsequent data enhancement with physical fidelity.

[0118] 2. Human motion artifact model: biomechanical physical parameters

[0119] (1) The principle of human motion artifacts

[0120] The physical essence of human motion artifacts is that human trunk movements such as turning over and coughing follow inherent biomechanical laws. Key parameters such as movement speed, acceleration, duration, and displacement amplitude are constrained by human physiological structure and motor ability. These parameters can directly characterize the type, intensity, and temporal characteristics of the corresponding human movement.

[0121] (2) Parametric model

[0122] The human motion artifact model mainly includes event-driven displacement perturbations such as turning over, limb twitching, and cough clusters, which are mapped to real-time (time t) phase abrupt change / diffusion Δd. motion (t):

[0123] ;

[0124] (3) Physical mechanism parameters

[0125] K event : Number of events; Event type: Turning over (Sigmoid function simulates smooth rotation), coughing (double exponential decaying sine function simulates rapid contraction and slow recovery).

[0126] A kThe range of motion for the kth event; the range of motion for rolling over is limited by the range of joint movement (trunk rotation angle 0-90°), and the range of motion for coughing is limited by the elasticity of the thoracic cage (displacement 5-20mm).

[0127] t k : The time when the k-th event occurs.

[0128] g k (tt k ): for tt k The prototype function of the k-th event at time (e.g., Gaussian impulse / exponential decay / piecewise linear).

[0129] (4) Physical constraint coupling (biomechanical constraint coupling)

[0130] Amplitude-duration constraint: A during rolling over movement roll ·T roll =Constant (angular velocity is limited by the speed of muscle contraction and is a set value), A roll For the range of motion of rolling over, T roll The duration of the rolling motion.

[0131] Rise-fall time constraint: Coughing movement T rise <T fall (The biomechanical principle of rapid contraction and slow recovery of the thoracic cavity), T rise T represents the duration of the rapid increase in thoracic displacement from the resting state to its peak value. fall The duration of the movement in which the thoracic cavity slowly returns to its resting state from its peak displacement.

[0132] Velocity-acceleration constraint: the velocity v of the k-th event. k The acceleration a of the motion of the kth event k Meeting the physiological safety threshold, v max ≤2m / s, a max ≤5m / s 2 It is limited by the speed of muscle contraction.

[0133] (5) Parameter space

[0134] ;

[0135] Among them, Ω motion For the parameter space of the human motion artifact model; A min A max These are the minimum and maximum values ​​of the event's motion amplitude, respectively; T dur The duration of the movement is 0.5-2 seconds for rolling over and 0.2-0.5 seconds for coughing; T min T maxThese are the minimum and maximum values ​​of the event's motion duration, respectively; v min v max These are the minimum and maximum values ​​of the event's movement speed, respectively.

[0136] Parameter space Ω motion All parameters strictly satisfy the physical constraint coupling conditions preset by the corresponding model, ensuring that the parameter values ​​and combinations completely conform to the inherent laws of the real physical scene.

[0137] (6) Summary

[0138] The occurrence time and prototype function are the core physical mechanism parameters. Setting up physiological human motion artifact models for different motion types is a key module in the parameterized environment model library for characterizing non-respiratory trunk interference. Based on the biomechanical laws of the real human body, it primarily reflects the physiological constraints of event-driven movements such as turning over, coughing, and limb twitching. It is modeled equivalently through event-driven displacement perturbations and mapped to the phase changes of radar echoes. The model constructs three types of biomechanical constraint coupling relationships: amplitude-duration coupling, coughing movement rise-fall time sequence constraints, and physiological safety thresholds for movement speed and acceleration, based on the number of events, motion amplitude, and actual parameter value range. This ensures that the simulated perturbations conform to the laws of real human motion from the perspective of motion mechanism. The model further defines a standardized parameter space including motion amplitude, duration, and speed. All parameter values ​​and combinations strictly adhere to preset physical constraint coupling conditions, ensuring that the human motion artifact modeling highly matches the characteristics of real physiological motion, providing reliable motion interference modeling support for subsequent physically accurate respiratory monitoring data enhancement.

[0139] 3. Physiological interference model: medical physiological and physical parameters

[0140] (1) The physical nature of physiological interference

[0141] The physical nature of physiological interference stems from the inherent rhythmic characteristics of human autonomous physiological activities: heartbeat interference has stable periodicity, typical P-QRS-T waveform morphology and heart rate variability; intestinal peristalsis interference exhibits low-frequency, intermittent and non-stationary motion characteristics. Both follow the laws of human medical physiology and constitute an endogenous physiological interference component that cannot be ignored in respiratory monitoring.

[0142] (2) Heartbeat displacement model

[0143] A real-time (time t) heart rate displacement model based on ECG morphology and cardiac cycle generation. heart (t) is represented as follows:

[0144] ;

[0145] (3) Physical mechanism parameters of the heartbeat displacement model

[0146] K hconv ECG-displacement conversion coefficient.

[0147] s ECG (t): ECG waveform (P-QRS-T wave) at time t.

[0148] g kernel (t): The electrocardiogram-displacement transformation kernel function at time t (simulating the delay and smoothing of cardiac mechanical contraction).

[0149] (4) Intestinal peristalsis disturbance model

[0150] The intestinal peristalsis disturbance model is characterized by low-frequency, non-stationary, and intermittently enhanced real-time (time t) displacement perturbations d. peri (t) (Combination of colored noise and intermittent bursts):

[0151] ;

[0152] (5) Physical mechanism parameters of the intestinal peristalsis disturbance model

[0153] u peri (t): Low-pass / bandpass kernel at time t (forming low-frequency colored noise).

[0154] η white (t): White noise at time t.

[0155] N burst Total number of intermittent sudden events.

[0156] b n Strength coefficient.

[0157] : A low-frequency envelope that rises and falls slowly.

[0158] : The start time of the nth intermittent sudden event.

[0159] (6) Physical constraint coupling (physiological constraint coupling)

[0160] Heart rate variability constraint: HRV = 0.1·f heart (Heart rate-variability coupling in physiology); HRV stands for heart rate variability, reflecting the regulatory capacity of the autonomic nervous system; f heart Heart rate, with a normal range of 60-100 beats per minute, is affected by age and health status.

[0161] Heart rate-amplitude constraint: A heart ∝1 / f heart (The faster the heart rate, the shorter the myocardial contraction time, and the smaller the stroke volume); A heartThis refers to the amplitude of the heartbeat, which is related to the myocardial contractility and is affected by the heart rate (the faster the heart rate, the shorter the contraction time and the smaller the amplitude).

[0162] (7) Parameter space

[0163] ;

[0164] Among them, Ω phys For the parameter space of the physiological interference model; f min f max These are the minimum and maximum heart rate, respectively; HRV min HRV max These are the minimum and maximum values ​​of heart rate variability, respectively; A min A max These are the minimum and maximum values ​​of the heart rate amplitude, respectively; f peri This refers to the frequency of intestinal peristalsis, with a normal range of 0.05-0.2 Hz, occurring intermittently; f pmin f pmax These represent the minimum and maximum values ​​of intestinal peristalsis frequency, respectively.

[0165] Parameter space Ω phys All parameters strictly satisfy the physical constraint coupling conditions preset by the corresponding model, ensuring that the parameter values ​​and combinations completely conform to the inherent laws of the real physical scene.

[0166] (8) Summary

[0167] As a core component of the parametric environment model library representing endogenous physiological disturbances in the human body, the physiological disturbance model is constructed based on the laws of human medical physiology. Its physical essence is reflected in the periodic rhythm of heartbeat activity, typical ECG waveform characteristics and heart rate variability, as well as the low-frequency, intermittent, and non-stationary disturbance characteristics of intestinal peristalsis. The model separately builds a heartbeat displacement model and an intestinal peristalsis disturbance model. Through physical mechanism parameters such as ECG-displacement conversion coefficient, ECG waveform, number of intermittent bursts, and intensity coefficient, it accurately describes the temporal variation characteristics of the two types of endogenous disturbances. At the same time, it establishes two types of medical physiological constraint relationships: heart rate-heart rate variability coupling and heart rate-heartbeat amplitude coupling, so that the modeling process strictly follows the real physiological regulatory mechanism of the human body. In addition, the model also delineates a standardized parameter space including heart rate, heart rate variability, heartbeat amplitude, and intestinal peristalsis frequency. All parameter values ​​and combinations strictly adhere to the preset physical constraint coupling conditions, ensuring that the physiological disturbance modeling closely matches the laws of real human physiological activities. This provides an accurate modeling foundation for endogenous physiological disturbances to achieve physically faithful millimeter-wave radar respiratory monitoring data enhancement.

[0168] 4. Electromagnetic Interference and Noise Models: Statistical Physics and Communication Physics Parameters

[0169] (1) Physical nature of electromagnetic interference

[0170] The physical nature of electromagnetic interference: thermal noise originates from the thermal motion of electronic devices and satisfies Boltzmann's law; narrowband / broadband electromagnetic interference originates from external radiation sources and has specific spectral structure and time-domain gating characteristics.

[0171] (2) Thermal noise model

[0172] Typical thermal noise power Modeled as real-time (time t) complex white Gaussian noise w th (t):

[0173] ;

[0174] Where CN(·,·) is a complex normal distribution; k B =1.38×10 -23 J / K (Boltzmann constant), T is the absolute temperature (K), and B is the noise bandwidth (Hz).

[0175] (3) Narrowband electromagnetic interference

[0176] Real-time (time t) narrowband electromagnetic interference i NBI (t) (single / multiple notes) is represented as follows:

[0177] ;

[0178] Where Q represents the number of tones in the narrowband interference (i.e., the number of superimposed single tones); A q f is the amplitude of the q-th pitch (a real number that determines the intensity of the interference); q Let θ be the frequency of the q-th tone. q The initial phase of the qth tone; m gate (t) is the gating function at time t, which takes the value 0 or 1 and is used to describe the time duty cycle of the interference.

[0179] (4) Broadband electromagnetic interference

[0180] Real-time (time t) broadband electromagnetic interference i WBI (t) (band-limited noise) is expressed as follows:

[0181] ;

[0182] Among them, b base (t) represents the baseband white noise at time t; u filter (t) is the impulse response of the band-limited filter at time t, used to limit white noise within a specified bandwidth.

[0183] (5) Physical mechanism parameters of electromagnetic interference and noise model

[0184] SNR: Signal-to-noise ratio, determined by the receiver noise figure and the target echo intensity.

[0185] f q Narrowband interference frequencies are constrained by the spectrum occupancy of external radiation sources (WiFi, Bluetooth, etc.).

[0186] B emi Broadband interference bandwidth, conforming to communication standard bandwidth (such as WiFi 20 / 40 / 80MHz).

[0187] D: Duty cycle, the time gating characteristic of the interference, reflecting the suddenness of the interference.

[0188] (6) Physical constraint coupling of electromagnetic interference and noise model

[0189] Thermal noise power-bandwidth constraint: (Boltzmann's Law).

[0190] Narrowband interference frequency-sampling rate constraint: f q <F s / 2 (Nyquist sampling theorem), where F s The sampling frequency.

[0191] (7) Parameter space of electromagnetic interference and noise model

[0192] ;

[0193] Among them, Ω EMI For the parameter space of the electromagnetic interference and noise model; SNR min SNR max For the minimum and maximum values ​​of the signal-to-noise ratio; f min f max For the minimum and maximum values ​​of the narrowband interference frequency; B min B max These represent the minimum and maximum values ​​of the broadband interference bandwidth.

[0194] Parameter space Ω EMI All parameters strictly satisfy the physical constraint coupling conditions preset by the corresponding model, ensuring that the parameter values ​​and combinations completely conform to the inherent laws of the real physical scene.

[0195] (8) Summary

[0196] Electromagnetic interference and noise models, as core components of the parametric environment model library representing inherent system noise and external electromagnetic interference, are constructed based on statistical physics and communication physics. Their physical essence manifests as thermal noise originating from the thermal motion of devices, and narrowband / broadband interference originating from external radiation sources, exhibiting specific spectral and time-domain gating characteristics. The model establishes parametric expressions for three types of noise / interference: thermal noise uses a complex Gaussian white noise model based on Boltzmann's law; narrowband interference employs multi-tone superposition combined with time-domain gating; and broadband interference is achieved through white noise band-limited filtering combined with gating. Simultaneously... Key physical mechanism parameters such as signal-to-noise ratio, interference frequency, bandwidth, and duty cycle were defined, and physical relationships such as thermal noise power-bandwidth coupling and narrowband interference frequency-sampling rate constraints were constructed to ensure that the modeling process strictly follows the physical laws of the real scene. The model further delineates a standardized parameter space including signal-to-noise ratio, interference frequency, bandwidth, and duty cycle. All parameter values ​​and combinations strictly meet the preset physical constraints, ensuring that the electromagnetic interference and noise modeling closely matches the real scene of the radar receiver. This provides accurate modeling support for system noise and electromagnetic interference for subsequent physically faithful radar signal data enhancement.

[0197] (III) Differentiated Synthesis and Controlled Data Augmentation Driven by Physical Mechanisms

[0198] The core innovation and uniqueness of this step lies in: based on the physical nature of various interference parameters, multipath propagation, human motion artifacts, physiological interference, and electromagnetic noise are forcibly bound to convolution / time-varying filtering, phase modulation, and additive superposition operators, respectively. This ensures that the mapping process from parameters to radar signals strictly follows the physical eigenvalue equations of electromagnetic propagation and physiological micro-motion, guaranteeing the physical authenticity of signal synthesis from a mechanistic perspective, and ultimately achieving highly controllable and high-fidelity respiratory monitoring data enhancement.

[0199] To clearly illustrate the differentiated signal synthesis logic of this invention for different types of interference, Table 1 systematically breaks down the core innovations from five dimensions: parameter type, physical nature, synthesis operator, physical basis, and essential differences from existing technologies. Existing radar signal enhancement technologies generally use a uniform additive superposition method to synthesize various types of interference, ignoring the fundamental differences in the physical generation mechanisms of multipath propagation, human motion, and electromagnetic interference, resulting in poor physical consistency between the generated data and the real scene. This invention, however, matches specific synthesis operators to each type of interference based on its physical nature, achieving a precise mapping from physical mechanism to signal generation.

[0200] Table 1. Physical mechanisms, synthesis operators, and differences from existing technologies for different parameter types.

[0201]

[0202] As shown in Table 1, this invention constructs a differentiated synthesis system deeply bound to physical mechanisms to address three core types of interference: multipath propagation, motion, and electromagnetic interference. Multipath propagation follows the characteristics of a linear time-varying system, employing convolution / time-varying filtering operators to ensure the physical consistency between the signal and the channel impulse response. Human motion is based on the principle of displacement-modulated phase, using phase modulation operators to realistically simulate echo phase changes. Electromagnetic interference, as an independent signal source, uses additive superposition operators to restore its linear superposition characteristics. This forced binding of "mechanism-operator" completely solves the physical distortion problem caused by the "one-size-fits-all" additive superposition of existing technologies, ensuring that the data enhancement process fully follows the generation law of radar echoes, providing crucial support for subsequent high-fidelity and controllable respiratory monitoring data enhancement.

[0203] 1. Multipath propagation: Convolution / Time-varying filter operators

[0204] Real-time (time t) convolution / time-varying filter operator s multipath (t) (multipath propagation interference at time t) is represented as:

[0205] ;

[0206] in, The impulse response of a time-varying multipath channel is given by the expression representing the time delay of the multipath channel at time t. The response of the signal components; This is a pure radar breathing signal generated solely by the target's minute breathing movements. This indicates that time delay has elapsed at time t. The moment after; For time delay The differential.

[0207] Physical Derivation: Impulse Response of Multipath Channel It is the impulse response of a linear time-varying system. The received signal is the convolution of the transmitted signal and the channel impulse response; this is a fundamental physical law of electromagnetic wave propagation. Using a convolution / time-varying filtering operator instead of additive superposition is a necessary condition to ensure the physical reality of multipath propagation.

[0208] 2. Motion artifacts and physiological disturbances: Phase modulation operator

[0209] Real-time (time t) phase modulation operator s phase (t) (human motion artifacts and physiological interference at time t) is represented as:

[0210] ;

[0211] Where exp(·) is an exponential function with the natural constant as its base; j is the imaginary unit; λ is the carrier wavelength of the radar signal; Δd total(t) represents the total change in the round-trip path caused by the target's slight movement at time t, which is a function of time. It includes the sum of displacement changes caused by all motion artifacts and physiological slight movements such as breathing, heartbeat, turning over, and coughing, and directly reflects the change in distance between the target and the radar. π is the mathematical constant pi.

[0212] Δd total (t)=Δd motion (t)+d physio (t), d physio (t)=d heart (t)+d peri (t), where Δd motion (t) represents the phase perturbation caused by the motion artifact at time t, and d physio (t) represents the phase perturbation caused by physiological micro-motion, and d heart (t) represents the displacement disturbance caused by the heartbeat at time t, and d peri (t) represents the displacement disturbance caused by intestinal peristalsis at time t.

[0213] Physical derivation: Real-time (time t) phase of radar echo The distance R(t) between (t) and the target satisfies (t) = (4πR(t)) / λ, R(t) = R0 + Δd(t), where R0 is the static reference distance and Δd(t) is the change in target displacement at time t. Therefore, the change in target displacement Δd(t) is directly converted into the phase change Δt. (t)=(4πΔd(t) / λ. Using a phase modulation operator instead of additive superposition is a necessary condition to ensure the motion-phase physical mapping.

[0214] 3. Electromagnetic interference and thermal noise: Additive superposition operator

[0215] Real-time (time t) additive superposition operator sIF output (t) (the noisy signal at time t) is represented as:

[0216] ;

[0217] Among them, G m Let i be the gain coefficient for the m-th type of electromagnetic interference. EMI,m (t) represents the m-th type of electromagnetic interference base signal, i EMI,m (t)=i NBI (t)+i WBI (t), i NBI (t) represents the narrowband interference signal at time t, i WBI (t) represents the broadband interference signal at time t; w th (t) represents the thermal noise at time t.

[0218] Physical derivation: Thermal noise originates from the thermal motion of electronic devices, while electromagnetic interference originates from external radiation sources. These interferences and target echoes are linearly superimposed at the receiver front end, conforming to the principle of signal superposition. Employing an additive superposition operator is a necessary condition for ensuring the physical authenticity of electromagnetic noise.

[0219] 4. Synchronously record the true value of interference and physical parameters.

[0220] The following interference components' true values ​​and physical parameters are recorded simultaneously to form the scene's physical fingerprint:

[0221] Multipath Channel Impulse Response Carries information about the room's geometry and material properties;

[0222] Displacement disturbance Δd motion (t): Carries biomechanical information about human movement;

[0223] Physiological displacement perturbation d physio (t): carries physiological information about the heart and gastrointestinal tract;

[0224] Electromagnetic interference signal i EMI (t): Carries electromagnetic environment spectrum information;

[0225] Thermal noise w th (t)——Carries receiver noise characteristics information.

[0226] For five core interference / noise components—multipath propagation, human motion, physiological interference, electromagnetic interference, and thermal noise—the ground truth data and key physical parameters of each component are simultaneously collected and recorded to construct a complete "scene physical fingerprint." Among these, the multipath channel impulse response... The underlying characteristics of the room's geometric layout and material properties are solidified, and the motion displacement disturbance Δd motion (t) embodies the biomechanical laws of human movement, and physiological displacement disturbances d physio (t) accurately depicts the rhythms of physiological activities such as heartbeat and intestinal peristalsis, and electromagnetic interference signals i EMI (t) reflects the spectral distribution characteristics of the actual electromagnetic environment, while thermal noise w th (t) represents the inherent noise properties of the receiver front end. This recording mechanism not only achieves accurate physical mechanism-level characterization of all interference sources, but also provides a comprehensive, realistic, and traceable truth benchmark for subsequent controllable data enhancement based on physical eigenvalue equations, thoroughly ensuring the physical consistency and high fidelity between synthetic data and real monitoring scenarios.

[0227] (iv) Constructing a quadruple data structure

[0228] The four types of data in Table 2 are associated and encapsulated to form, as follows: Figure 2 The structured augmented data sample shown.

[0229] Table 2 Quadruple Data

[0230]

[0231] like Figure 2 As shown, the quadruplet data structure diagram clarifies the standardized encapsulation logic and output format of single-sample radar data: Using a "single-sample record" as the basic unit, the core encapsulation consists of a sample pair composed of "clean radar respiratory signal (cleanI / Q)" and "noisy signal (noisyI / Q)," while simultaneously associating two sets of corresponding "interference ground truth labels." The labels cover key physical attributes such as interference type, intensity, time period, and location. The "quadruplet data encapsulation" mechanism achieves a one-to-one correspondence between the four components, ensuring a strong correlation between the clean radar respiratory signal and the noisy signal, and between the real interference components and simulation parameters. Finally, the output is in the form of "sample pair + ground truth + parameter index," providing downstream models with training sample pairs of "noisy signal - clean radar respiratory signal." Furthermore, the complete interference ground truth and parameter index support the model's learning of the physically decoupled representation of pathological signals and environmental interference, laying a data foundation for physically interpretable respiratory monitoring.

[0232] The unique value of this quadruple data structure lies in:

[0233] Supporting decoupled learning: The downstream model can simultaneously learn the mapping relationship between pure radar respiratory signals and interference, realizing a physically decoupled representation of pathology and interference.

[0234] Supports interpretable monitoring: The model can output the type and intensity of interference, enabling end-to-end interpretable inference.

[0235] Supports scene inversion: The physical parameters of the scene can be inverted by perturbing the true value, supporting personalized adaptation.

[0236] (v) Batch generation of augmented datasets

[0237] Repeat steps (i) to (iv) to generate large-scale augmented data samples in batches according to the preset scenario parameter combinations. Finally, a controllable augmented dataset covering all working conditions and multiple interference types is constructed to provide high-quality, high-fidelity data support with physical traceability for subsequent model training, testing and performance verification.

[0238] (vi) Multi-level physical fidelity verification

[0239] As shown in Table 3, a full-link physical fidelity verification system oriented towards physical intrinsic consistency is established. Multi-dimensional quantitative verification is carried out on the multipath channel response, motion phase mapping, physiological micro-motion characteristics, electromagnetic interference spectrum and noise statistical characteristics of the generated data to ensure that the synthesized data strictly follows the physical laws of radar echo generation and has a highly traceable physical consistency with the real scene.

[0240] Table 3 Physical Fidelity Verification System

[0241]

[0242] When any physical fidelity verification indicator fails to reach the preset threshold, the system will automatically trigger a closed-loop parameter tuning process, backtracking the physical parameters (such as multipath channels, motion / physiological models, electromagnetic interference-related parameters) of the corresponding link in the data generation chain for iterative correction until all indicators meet the qualification standards, ensuring the physical consistency and validity of the generated data.

[0243] Based on the data in Table 3, this invention constructs a multi-dimensional physical fidelity verification system covering the entire data generation chain. Through hierarchical quantitative verification, it ensures the consistency of the synthesized data across all dimensions, from the underlying physical mechanism to the upper-level algorithm effectiveness. This system sets nine verification indicators for different interference and noise stages: for the multipath propagation stage, the consistency and geometric matching degree between the multipath distribution and the real room environment are verified through the power delay distribution KS statistic and delay-distance mapping error, respectively, using D... KS <0.15, E ego <5% is the acceptable threshold; Regarding the human motion and physiological interference aspects, the physical consistency, biomechanical constraint compliance, and physiological signal rhythm and coupling laws of motion phase modulation are verified through the correlation between phase change rate and motion speed, amplitude-duration product, heart rate estimation error, and HRV-heart rate correlation, with R... vel >0.95, A·T is within the physiological range, E HR <3%, R HR >0.9 is used as the judgment criterion; for the electromagnetic noise link, the conformity between the electromagnetic interference spectrum matching degree and the physical laws of thermal noise is verified by the spectrum structure similarity and thermal noise power-bandwidth consistency, respectively, using SSIM. spec The acceptable thresholds are >0.85 and error <5%. Finally, the actual utility of the generated data for model training is verified by the F1 score of the downstream respiratory monitoring task, with F1 > 0.85 used as the usability criterion. This system achieves dual quantitative assurance of physical authenticity and engineering usability, providing a traceable and verifiable verification basis for high-fidelity data augmentation.

[0244] (vii) Closed-loop adaptive evolution mechanism

[0245] The uniqueness of this section lies in breaking away from the limitations of static and fixed physical parameter space design in traditional data augmentation methods and constructing a dynamic evolutionary parameter space mechanism based on real data feedback: This parameter space can be continuously iteratively optimized according to the multi-dimensional physical fidelity verification results of the generated data, so that the parameter distribution gradually approaches the real physical characteristics of the target scene, and finally achieves controllable data augmentation scene adaptive adaptation, ensuring the consistency between the generated data and the actual application scenario.

[0246] 1. Extract scene fingerprints

[0247] Using limited real-world monitoring data collected in a real deployment environment D real (Supports unlabeled or weakly labeled formats) Extracts scene feature fingerprints to provide a basis for subsequent adaptive calibration and iterative optimization of physical model parameters.

[0248] (1) Multipath fingerprint

[0249] Multipath fingerprinting includes estimating the power delay spectrum and multipath delay spread. The proportion of the main diameter. Frequency domain correlation or inverse filtering methods are used to determine the proportion of the main diameter from D. real Estimated power delay spectrum :

[0250] ;

[0251] in, Let F be the time delay variable; F{·} is the Fourier transform operator, F -1 {·} is the inverse Fourier transform operator; For complex conjugate operators; s real (t) represents the actual received signal at time t; s ref (t) is the reference signal at time t; Regularization parameters.

[0252] based on Extract the following multipath fingerprint parameters:

[0253] Multipath delay spread :

[0254] ;

[0255] ;

[0256] in, This represents the average latency.

[0257] Main diameter ratio ρ LOS :

[0258] ;

[0259] in, For direct path delay, Power of the LOS path; Let be the power of the i-th path.

[0260] Multipath tail attenuation rate γ tail : Through the The tail portion was obtained by exponential fitting.

[0261] (2) Motion fingerprints

[0262] This section constructs a motion fingerprint extraction mechanism for human rolling over motion artifacts, characterizing real motion features through three dimensions: first, defining the rolling over event frequency p. roll =N roll / T obs N roll T represents the total number of comeback events. obs The observation duration of the actual radar data; and the statistical distribution of the turning amplitude p(A) roll The algorithm maps the peak phase change caused by the rolling over event to the displacement domain and estimates its probability distribution. The third method is an event detection algorithm based on short-time energy detection and adaptive threshold to robustly extract the rolling over duration, providing quantifiable real motion features for subsequent motion artifact modeling and data augmentation.

[0263] (3) Physiological fingerprints

[0264] To address the need for physiological fingerprint extraction, this invention employs an improved autocorrelation method to accurately obtain multi-dimensional physiological parameters from the demodulated phase signal: First, the phase signal is bandpass filtered, dividing the respiratory and heart rate signal frequency bands into 0.1-0.5Hz and 0.8-2.0Hz respectively. The autocorrelation function of the filtered signal is calculated, and the respiratory rate and heart rate f are separated using a peak extraction algorithm. heart Based on continuous heartbeat cycle sequence T card (n) Calculate the standard deviation to obtain the heart rate variability (HRV) (HRV=std(T) card (n)), std(·) is the standard deviation; at the same time, intermittent energy enhancement features are detected in the frequency band of 0.05-0.2Hz, and the intestinal peristalsis frequency λ is extracted. peri This provides reliable real data support for subsequent physiological interference modeling.

[0265] (4) Electromagnetic fingerprint

[0266] This section constructs an electromagnetic fingerprint extraction method, which is based on real monitoring data D. real Extracting pure noise segments from U-bands during periods without target echoes. u (t), using the average power spectrum estimation formula

[0267] (FFT{·} is Fast Fourier Transform), to obtain the background noise spectrum, and simultaneously identify the narrowband interference frequency point f. q This completes the quantitative characterization of electromagnetic environment features.

[0268] (5) Multidimensional vectors

[0269] Based on the four fingerprint types mentioned above, a unified multi-dimensional scene fingerprint vector F is constructed by integrating multipath channel, human motion, physiological activity, and electromagnetic environment feature parameters. scene :

[0270] ;

[0271] This vector provides a complete and traceable real-world benchmark for subsequent adaptive optimization of the parameter space.

[0272] 2. Parameter space mapping

[0273] This step maps the scene fingerprint extracted from real data to specific adjustment amounts for each simulation parameter in the parameter space, constructing a directional mapping relationship from feature differences to parameter correction: scene fingerprint quantization characterizes the feature deviation between the real environment and the current simulation environment, and this mapping determines the adjustment direction and magnitude of each parameter based on the deviation information. For example, when multipath delay spread in the real scene is detected... When the value is greater than the simulation value, the simulation parameters can be improved. The upper limit of the value is used to achieve iterative calibration of the parameter space, so that the distribution of simulation parameters gradually converges to the physical characteristics of the real scene.

[0274] Establish a mapping relationship between a preset parameter space and real-world fingerprints, and define a mapping function. , where Ω (k) Let ΔΩ be the parameter space for the k-th iteration. (k) This represents the parameter space adjustment amount for the k-th iteration.

[0275] Based on this mapping function, a core mapping rule is constructed that follows the laws of physical conservation and the coupling law of constraints (as shown in Table 4, where k represents the number of iterations). The feedback of the scene fingerprint drives the iterative calibration of the parameter space, ensuring that the parameter adjustment fits the logic of the real physical scene throughout the process.

[0276] Table 4 Core Mapping Rules

[0277]

[0278] The core mapping rules in Table 4 construct a parameter space adaptive adjustment strategy based on scene fingerprint deviation. Driven by the characteristic differences between the real scene and the simulation environment, quantifiable parameter update rules are designed for different types of deviations, and the physical basis of each adjustment is clarified to achieve iterative adaptation of the parameter space to the real scene: when the real multipath delay spread is greater than the simulation value, the upper limit of the delay spread is increased by multiplying by an amplification factor to adapt to larger room sizes; when the real turning frequency is higher than the simulation preset value, the sampling probability of motion events is increased by an amplification factor, and physical upper limit constraints are used to avoid exceeding the limit to adapt to scenarios with frequent user activity; when the real heart rate does not match the simulation distribution center, the distribution mean is aligned with the real heart rate and the standard deviation is adjusted to adapt to the user's actual heart rate characteristics; when the real background noise is higher than the simulation value, the lower limit of SNR sampling is reduced by a fixed step size to adapt to high-noise environments; for unmodeled interference types existing in the real scene, a new interference model is introduced to cover scene-specific interference. The entire adjustment process follows physical constraints to ensure the rationality of parameter updates and scene adaptability.

[0279] 3. Parameter Space Evolution

[0280] This section proposes an iterative evolution mechanism for the parameter space. This mechanism does not employ a statically defined fixed parameter space, but rather achieves adaptive convergence to the real-world scenario through closed-loop iteration. Its core evolutionary formula is: Among them, Ω (k+1) Let this be the simulation parameter space for the (k+1)th iteration. (·) is a scene fingerprint extraction operator that encapsulates signal processing, statistical analysis, and feature extraction algorithms. It can extract the real radar echo data corresponding to the k-th iteration. Transform into a structured fingerprint vector F describing the physical properties of the scene scene Then it is fed into the parameter space mapping function F, combined with the old parameter space Ω of the current iteration. (k) The parameter adjustment amount is calculated, and then the new parameter space Ω is updated. (k+1) That is, "new parameter space = old parameter space + scene fingerprint-based adjustment". After each round of evolution, the enhanced data generated by simulation will be closer to the real target environment, realizing dynamic optimization of the parameter space.

[0281] 4. Incremental Iterative Optimization

[0282] Incremental iterative optimization constructs a closed-loop collaborative optimization mechanism oriented towards the target scenario. Its core is to achieve synchronous convergence of the parameter space and downstream model through multiple rounds of iteration: a new round of simulation enhancement dataset is generated based on the evolved parameter space, and the downstream model is retrained and fine-tuned to obtain an updated model; subsequently, a small amount of real data is collected and scene fingerprints are extracted to drive iterative updates of the parameter space, forming a closed-loop process of "data generation - model fine-tuning - feature extraction - parameter optimization" until the parameter space and model performance tend to stabilize. Experimental verification shows that after 2-3 rounds of iteration, the F1 score of the model on the downstream respiratory monitoring task in the target scenario can be further improved by 5%-8%, the robustness to scene-specific interference is significantly enhanced, and the personalized adaptation cycle is greatly compressed from several weeks to several hours.

[0283] II. Supporting Physically Decoupled Representations for Downstream Model Learning

[0284] The downstream monitoring model was trained using a quadruple-enhanced dataset that had been incrementally iteratively optimized. Leveraging the unique advantages of this dataset, multi-dimensional performance breakthroughs were achieved, including:

[0285] ① Decoupled learning: The model is simultaneously input to the noisy signal and the ground truth of the interference. Through a multi-task learning architecture, it learns the ability to decompose the noisy signal into a clean radar breathing signal and various interference components.

[0286] ② Physically interpretable reasoning: After the model is deployed, it can not only output the probability of pathological events, but also simultaneously output the estimated values ​​of interference components such as multipath intensity, motion event type, and background noise level, thus achieving respiratory monitoring with physical interpretability.

[0287] ③ Zero-shot scenario adaptation: The model has learned the physical coupling relationship between interference and signal during the pre-training stage. During the cold start stage, it can distinguish whether the signal change is due to pathology or interference based on physical cognition, thus achieving accurate inference in zero-shot scenarios.

[0288] In summary, by training downstream monitoring models using quadruple-enhanced datasets, and simultaneously endowing the models with three core capabilities—decoupled learning, physically interpretable reasoning, and zero-shot scenario adaptation—we provide crucial support for robust and interpretable end-to-end respiratory monitoring.

[0289] III. Examples

[0290] (I) Example 1: Construction of Parametric Environment Model and Verification of Physical Constraints

[0291] This embodiment illustrates how to construct a parameterized environmental model library around four types of environmental factors: multipath propagation, human motion artifacts, physiological interference, and electromagnetic interference, and verifies that the sampled parameters meet the preset physical constraints.

[0292] 1. Construct a multipath propagation model

[0293] Set the room dimensions to 5m × 4m × 3m, and calculate the main reflection path using ray tracing:

[0294] (1) Direct path delay: Where R0 = 1m;

[0295] (2) First reflection path (ground) delay: , where h is the target height.

[0296] (3) Delay of the first reflection path (wall): , where d w This represents the horizontal distance between the reflector wall and the radar.

[0297] The direct path is given by the shortest Euclidean distance from the radar to the equivalent scattering center of the chest cavity; the ground primary reflection path is calculated using the mirror method; the wall primary reflection path is obtained based on the geometric relationship between the mirror point of the side wall and the target scattering center.

[0298] Amplitude calculation: , where α i Let d be the complex magnitude of the i-th path. i Let Γ be the propagation distance of the i-th path, and Γ be the reflection coefficient (Γ1=0.8 for the ground and Γ2=0.6 for the wall).

[0299] Physical constraint verification: The corresponding propagation distance difference is 0.5m, which is consistent with the actual room geometry; This is consistent with theoretical calculations; where α1 and α2 are the complex amplitudes of the paths to the ground and the wall, respectively, and d1 and d2 are the propagation distances of the paths to the ground and the wall, respectively.

[0300] 2. Constructing a human motion artifact model

[0301] Rolling over motion parameters: Amplitude A roll =0.2m (equivalent radius of torso 0.2m, rotated 90°), duration T roll =1.0s, velocity v max =0.63m / s.

[0302] Physical constraint verification: A roll ·T roll =0.2, within the physiological range of 0.15-0.4; v max =0.63m / s < 2m / s, which meets the physiological safety threshold.

[0303] Coughing movement parameters: Amplitude A cough =0.01m, rise time T rise=0.05s, descent time T fall =0.15s.

[0304] Physical constraint verification: T rise <T fall This conforms to the biomechanical principle of rapid contraction and slow recovery of the thoracic cavity.

[0305] 3. Construct a physiological interference model

[0306] Heart rate parameter: f heart =1.2Hz (72 times / minute), HRV=0.12Hz, A heart =0.5mm.

[0307] Intestinal peristalsis parameters: f peri =0.1 Hz.

[0308] Physical constraint verification: HRV / f heart =0.1, which conforms to the physiological law of heart rate-variability coupling; A heart ∝1 / f heart The amplitude decreases when the heart rate increases.

[0309] 4. Construct an electromagnetic interference model

[0310] Thermal noise parameters: B=200Hz, T=290K .

[0311] Narrowband interference: f q =2.4GHz (WiFi band), duty cycle D=0.3.

[0312] Physical constraint verification: This conforms to Boltzmann's law; f q <F s / 2(F) s =200Hz is not satisfied (the actual bandwidth of the RF front end needs to be considered), and the constraint needs to be met at the RF level.

[0313] 5. Parameter Space Sampling and Physical Constraint Verification

[0314] Sample a set of parameters from the parameter space to verify the physical constraints:

[0315] Multipath: N mp =4, , α1=0.5, which satisfies the time delay-distance constraint.

[0316] Exercise: A roll =0.18m, T roll =1.2s, A roll ·T roll=0.216, within the physiological range, satisfies the amplitude-duration constraint.

[0317] Physiology: f heart =1.1Hz, HRV=0.11Hz, f peri =0.1Hz, ratio 0.1, conforms to coupling, and satisfies the heart rate-variability constraint.

[0318] Electromagnetic: SNR=10dB, B=200Hz It conforms to Boltzmann's law and satisfies the thermal noise power-bandwidth constraint.

[0319] All parameters satisfy the physical constraints, and the sampling is valid. For parameter combinations that pass the above constraint verification, this embodiment records them as a set of valid parameters, and outputs them together with their corresponding scene configuration, event timestamps, and parameter category identifiers for use in the physical mechanism-driven differential synthesis in Embodiment 2. The valid parameter set includes at least one or more of the following: multipath parameter set, motion event parameter set, physiological interference parameter set, and electromagnetic interference parameter set.

[0320] (II) Example 2: Physical Mechanism-Driven Differential Synthesis and Controllable Data Augmentation Verification

[0321] 1. Input

[0322] Pure breathing radar signal sIF clean (t), duration 60 seconds, sampling rate 200Hz, respiratory rate 0.25Hz.

[0323] 2. Differentiated Synthesis Sequence

[0324] First, various interference base signals are generated based on the effective parameter set output in Example 1. Second, time alignment is performed between each interference base signal and the clean breathing radar signal, and spatial alignment is performed if necessary in an array scenario. Then, the gain coefficient of each interference component is calculated based on the target signal-to-noise ratio or a preset injection intensity. Finally, according to the physical mechanism of the interference, convolution, phase modulation, or additive superposition operators are used to combine them to obtain a noisy radar signal. The above process can be completed collaboratively by four sub-modules: interference parameter configuration, interference generation, signal synthesis and alignment, and metadata and tag generation.

[0325] 3. Multipath Synthesis

[0326] The synthesis is achieved using a convolution / time-varying filtering operator. In one implementation, the multipath impulse response obtained in Example 1 is discretized into a finite-length filter kernel, and a convolution operation is performed on the pure complex baseband / intermediate frequency signal to obtain a propagation-type synthesized signal that includes time delay tails and path superposition effects. When considering slow fading, the impulse response can be updated slowly over time. Simultaneously, the multipath impulse response corresponding to the current sample or its equivalent filter kernel is synchronously saved as the true value of the propagation-type interference for that sample, for subsequent physical fidelity verification and supervised training.

[0327] Verification: Based on Figure 3 The data in the table are used to calculate the power delay spectrum and the KS statistic D. KS =0.08<0.15, delay-distance mapping error E geo =3.2% < 5%.

[0328] Figure 3 The comparison results of power delay spectra between real-world and simulated data are shown. The horizontal axis represents signal propagation delay (unit: nanoseconds, reflecting the arrival time of signals along different paths), and the vertical axis represents normalized power (unit: decibels, characterizing the intensity distribution of signals along different paths). The blue curve represents the multipath power distribution collected in the real-world environment, while the red curve represents the power delay spectrum generated by the simulation of this invention. The curve characteristics show a high degree of agreement between the overall attenuation trends, peak positions, and intensities of the main path and each secondary reflection path. This indicates that the multipath propagation model constructed in this invention not only accurately reproduces the intensity and delay of the direct / near-path main signal but also effectively restores the power distribution and delay characteristics of different reflection paths. This verifies the physical consistency of multipath modeling and the high fidelity of the simulation data, proving that this method can generate simulation data that highly matches the multipath characteristics of the real environment, providing reliable multipath scenario support for subsequent model training.

[0329] 4. Motion artifact synthesis

[0330] A phase modulation operator is used, with the turning-over event occurring at the 10th second, an amplitude of 0.2m, and a duration of 1.0s. In this embodiment, the turning-over and coughing events are first generated as equivalent displacement perturbation sequences, then converted into phase modulation quantities according to the radar displacement-phase mapping relationship, and applied to the pure radar breathing signal or the signal after multipath processing. For physiological micro-movements such as heartbeat and intestinal peristalsis, the same method of mapping displacement or equivalent mechanical perturbation to phase perturbation is used for injection. This ensures that motion-related and physiological interferences directly affect the radar echo phase, rather than being incorrectly modeled as simple additive noise. Verification: Correlation between phase change rate and motion velocity R vel =0.97>0.95, which is consistent with physical consistency.

[0331] 5. Electromagnetic interference synthesis:

[0332] An additive superposition operator is used, with an SNR of 10dB. After superimposing thermal noise and narrowband / wideband electromagnetic interference, the actual equivalent signal-to-noise ratio (SNR) of the synthesized signal is further measured. If the deviation between the actual SNR and the target value of 10dB exceeds the preset tolerance, the superposition is re-executed by adjusting the thermal noise amplitude and the electromagnetic interference gain coefficient until the actual SNR falls within the target range. This gain control process ensures that the noise intensity among different samples is controllable and consistent. Verification: The spectral structure is consistent with the preset value, and the thermal noise power P... th =-174+10log 10 (200) = -151dBm, which is consistent with the theoretical value.

[0333] 6. Quadruple Data Output

[0334] The generated data includes clean radar breathing signals, noisy signals, ground truth values ​​for interference, and physical parameter labels, supporting downstream model training. In a preferred implementation, the four-tuple data is encapsulated separately as the core content of the sample, including clean radar breathing signals, noisy radar signals, ground truth values ​​for interference, and physical parameter labels; associated information such as event timestamps, event types, sample numbers, scene identifiers, sampling rates, and target signal-to-noise ratios are additionally stored as metadata. This sample can be saved in a structured format, HDF5, NPZ, or an equivalent structured format to support subsequent training, validation, and backtracking analysis.

[0335] 7. Determination of Sample Validity

[0336] For each enhanced sample generated in this embodiment, a sample-level validity determination must also be performed. If the power delay spectrum, phase continuity, spectral structure, and target signal-to-noise ratio of the sample all meet the preset thresholds, it is written into the enhanced dataset; otherwise, the sample is marked as an invalid sample, and local resampling or local regeneration of the corresponding category parameters is triggered.

[0337] (III) Example 3: Verification of Closed-Loop Adaptive Evolution Mechanism

[0338] 1. Initial Deployment

[0339] Generate the initial augmented dataset D using the universal parameter space. synth (0) The base model M0 is trained. This initial augmentation dataset is preferably generated in batches using the parameterized environment model library constructed in Example 1 and the differentiated synthesis process in Example 2. The base model can be any of a respiratory event detection model, a respiratory waveform reconstruction model, or a risk assessment model, and its training objective is to provide initial parameters and an initial performance baseline for subsequent scene adaptation.

[0340] 2. Real data collection

[0341] Collect 1 hour of real-time monitoring data in the target bedroom. real (1) (Unlabeled). Basic preprocessing is first performed on the collected real monitoring data, including target distance gate selection, phase extraction, spectrum estimation, and coarse event detection. Then, scene fingerprints are extracted: multipath fingerprints are obtained through power delay spectrum, main path ratio, and delay spread estimation; motion fingerprints are obtained through the frequency, duration, and amplitude statistics of events such as turning over / coughing; physiological fingerprints are obtained through estimation of heart rate peak, heart rate variability, and low-frequency visceral activity frequency band energy; and electromagnetic fingerprints are obtained through estimation of background noise spectrum, narrowband interference frequency points, and spectral occupancy range.

[0342] 3. Scene fingerprint extraction

[0343] Multipath fingerprints: (Greater than the simulation default value of 8ns).

[0344] Motion fingerprint: Turning frequency 0.5 times / hour (greater than the simulation default of 0.2 times / hour).

[0345] Physiological fingerprint: f heart =1.0Hz (less than the simulation default value of 1.2Hz).

[0346] Electromagnetic fingerprint: Background noise SNR=5dB (less than the simulation default value of 10dB).

[0347] 4. Parameter space mapping

[0348] The upper limit for sampling has been adjusted from 10ns to 15ns.

[0349] The thermal noise power-bandwidth constraint for turning over was adjusted from 0.2 to 0.5.

[0350] f heart The sampling center was adjusted from 1.2Hz to 1.0Hz.

[0351] The lower limit of SNR sampling has been adjusted from 0dB to -5dB.

[0352] Parameter space mapping is not a simple numerical replacement, but a constrained correction of the original parameter space based on the rule of "real-world scenario deviation - parameter adjustment direction". Specifically, if the real-world scenario's latency spread is higher than the simulation default value, the upper limit of multipath latency spread sampling is increased; if the frequency of real-world turning events is higher than the default value, the sampling probability of motion events is increased; if the real heart rate center deviates from the default value, the sampling center of physiological parameters is adjusted; if the real-world background noise is stronger, the lower limit of signal-to-noise ratio sampling is lowered. After parameter adjustment, the physical constraint coupling conditions defined in Example 1 must still be met to avoid generating parameter combinations that do not conform to the actual scenario.

[0353] 5. Parameter Space Evolution

[0354] Generate a new round of augmented dataset D synth (1) Train model M1. After generating a new round of augmented datasets, retrain or fine-tune the base model using the updated augmented data, and re-evaluate its performance on the target scene test set. If key performance indicators improve compared to the previous round, and validation metrics such as multipath physical fidelity, phase continuity, and spectral consistency do not deteriorate, then accept the current parameter space as the new scene parameter space; otherwise, further adjust the corresponding parameter categories according to the source of the bias. The above process can be repeated for 2-3 rounds or more until the performance improvement tends to saturate or a preset termination condition is reached.

[0355] 6. Effect Verification

[0356] M1 improved its F1 score on the target scenario test set from 0.82 to 0.88, an improvement of 7.3%, and the personalized adaptation time was shortened from 2 weeks to 4 hours.

[0357] like Figure 4 As shown, the curves compare the F1 scores of the two models for downstream respiratory monitoring tasks under different signal-to-noise ratio (SNR) conditions. The horizontal axis represents the SNR (dB), reflecting the change in signal quality from low to high; the vertical axis represents the F1 score, characterizing the detection performance of the model. The solid blue line represents the performance curve of the enhanced model trained using the enhanced data of this invention, while the dashed red line represents the performance curve of the conventional model trained only with real data. It can be seen that the enhanced model of this invention maintains a high and stable F1 score across the entire SNR range. Even in strong noise environments with low SNR (e.g., -5dB, 0dB), its performance remains stable above 0.85, gradually approaching 1 as the SNR increases. In contrast, the conventional model trained only with real data is significantly affected by noise, with its F1 score plummeting to around 0.45 at low SNR (0dB) and only maintaining around 0.78 at high SNR. The results show that the controllable data augmentation method of the present invention effectively improves the robustness of the model to noise interference. Compared with the traditional model that relies solely on real data for training, it significantly enhances the detection accuracy and anti-interference ability of respiratory events in complex low signal-to-noise ratio scenarios.

[0358] The above results demonstrate that the parameter space in this invention is not statically fixed, but can be continuously modified based on feedback from real-world scenarios. By remapping multipath, motion, physiological, and electromagnetic features from real-world scenarios to the parameter space, and then driving a new round of augmented data generation, the distribution of augmented data can gradually approximate the real physical distribution of the target scenario, thereby improving the model's adaptability and stability in that scenario.

[0359] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling, characterized in that, Includes the following steps: Acquire clean radar breathing signals; We constructed a multipath propagation model with physical constraints and coupling, a human motion artifact model, a physiological interference model, and an electromagnetic interference and noise model, and determined the parameter spaces corresponding to the four types of models. Based on the four types of models and the values ​​of their parameters in the parameter space, the true values ​​of each disturbance are calculated. Based on the pure radar breathing signal and the true values ​​of each interference, the noise-added signal is calculated sequentially through convolution / time-varying filtering operator, phase modulation operator, and additive superposition operator according to the physical mechanism. The data is integrated into a four-tuple data structure, which includes the pure radar breathing signal, the noisy signal, the true values ​​of each interference, and the parameter values ​​of each model. Repeat the above steps to generate augmented datasets consisting of quadruple data structures in batches, and perform multi-level physical fidelity verification on the augmented datasets. If the verification fails, trigger an iterative parameter tuning process until the standard is met: extract scene fingerprints from real monitoring data, map scene fingerprints to corresponding parameter space adjustment values, update the parameter space based on the parameter space adjustment values, and realize closed-loop adaptive evolution of the parameter space.

2. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 1, characterized in that, The parameters of the four types of models with physical constraint coupling are: The parameters of the multipath propagation model include path delay, path propagation distance, and path complex amplitude, which satisfy the physical constraint coupling of the multipath propagation model, which is composed of delay-distance constraints, amplitude-distance constraints, and path number-delay spread constraints. The parameters of the human motion artifact model include motion amplitude, motion duration, motion velocity, and motion acceleration, which satisfy the physical constraint coupling of the human motion artifact model, which is composed of amplitude-duration constraints, rise-fall time constraints, and velocity-acceleration constraints. The parameters of the physiological disturbance model include heart rate, heart rate variability, and heart rate amplitude, which satisfy the physical constraint coupling of the physiological disturbance model, which is composed of heart rate-variability constraints and heart rate-amplitude constraints. The parameters of the electromagnetic interference and noise model include thermal noise power, interference frequency, interference bandwidth, and sampling frequency, which satisfy the physical constraint coupling of the electromagnetic interference and noise model, which is composed of thermal noise power-bandwidth constraints and narrowband interference frequency-sampling rate constraints.

3. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 2, characterized in that, The delay-distance constraint is: ； in, Let d be the time delay of the i-th path at time t. i (t) represents the propagation distance of the i-th path at time t; c represents the radar wave speed; The amplitude-distance constraint is: ； Where, α i (t) represents the complex amplitude of the i-th path at time t; ∝ represents the proportionality sign; |Γ| represents the absolute value of the reflection coefficient Γ; Path number-delay spread constraint: Number of effective multipath paths N mp With multipath delay spread There is a positive statistical correlation between them; Amplitude-duration constraint: A during rolling over movement roll ·T roll =constant, A roll For the range of motion of rolling over, T roll The duration of the rolling motion; Ascent-descent time constraint: Coughing movement T rise <T fall T rise T represents the duration of thoracic displacement from resting state to peak value. fall The duration of the movement in the thoracic cavity as it recovers from its peak displacement to its resting state; Velocity-acceleration constraint: the velocity v of the k-th event. k The acceleration a of the motion of the kth event k Meeting the physiological safety threshold, v max ≤2m / s, a max ≤5m / s 2 v max a is the maximum speed of the event. max This represents the maximum acceleration of the event's motion. Heart rate variability constraint: HRV = 0.1·f heart HRV stands for heart rate variability, f heart Heart rate; Heart rate-amplitude constraint: A heart ∝1 / f heart A heart This refers to the amplitude of the heartbeat. Thermal noise power-bandwidth constraint: , Where B is the thermal noise power, and B is the noise bandwidth. Narrowband interference frequency-sampling rate constraint: f q <F s / 2,f q For narrowband interference frequency, F s The sampling frequency.

4. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 1, characterized in that, The calculation process for the noise-added signal is as follows: First, multipath propagation interference is synthesized using a convolution / time-varying filter operator: ； Among them, s multipath (t) represents the convolution / time-varying filter operator at time t, i.e., the multipath propagation interference at time t; The impulse response of a time-varying multipath channel is given by the expression representing the time delay of the multipath channel at time t. The response of the signal components; for The pure radar breathing signal generated only by the target's minute breathing movements. This indicates that time delay has elapsed at time t. The moment after; For time delay The differential; Next, a phase modulation operator is used to synthesize human motion artifact interference and physiological interference: ； Among them, s phase (t) represents the phase modulation operator at time t, i.e., the human motion artifact interference and physiological interference at time t; exp(·) is an exponential function with the natural constant as the base; j is the imaginary unit; π is pi; λ is the carrier wavelength of the radar signal; Δd total (t) represents the total change in the round-trip path caused by the slight movement of the target at time t; Finally, an additive superposition operator is used to synthesize electromagnetic interference and thermal noise: ； Among them, sIF output (t) is the additive superposition operator at time t, i.e., the noisy signal at time t; G m Let i be the gain coefficient for the m-th type of electromagnetic interference. EMI,m (t) represents the m-th type of electromagnetic interference base signal, i EMI,m (t)=i NBI (t)+i WBI (t), i NBI (t) represents the narrowband interference signal at time t, i WBI (t) represents the broadband interference signal at time t; w th (t) represents the thermal noise at time t.

5. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 1, characterized in that, Multi-level physical fidelity verification refers to the verification of data generated by four types of models with physical constraint coupling. When any physical fidelity verification index fails to reach the preset threshold, an iterative parameter tuning process is triggered until the threshold is met. The specific verification content is as follows: Multipath power delay spectrum KS statistic DKS < 0.15, delay-distance mapping error Eego < 5%, phase change rate correlation with motion velocity Rvel > 0.95, motion amplitude A - duration T product A·T ∈ [physiological range], heart rate estimation error EHR < 3%, heart rate variability HRV and heart rate f heart The correlation RHR > 0.9, the spectral structure similarity SSIMspec > 0.85, and the thermal noise power The consistency error with the noise bandwidth B is <5%, and the F1 score of the downstream respiratory monitoring task is >0.

85.

6. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 1, characterized in that, Mapping a scene fingerprint to a corresponding parameter space adjustment amount refers to performing physically constrained directional adjustments within a physically constrained parameter space, based on the deviation between the scene fingerprint and the corresponding parameters in the current parameter space. The specific directional adjustment rules are as follows: When the actual multipath delay spread is greater than the multipath delay spread value in the current parameter space, increase the upper limit of the multipath delay spread sampling in the current parameter space to the target sampling upper limit: ； in, , These are the upper limits for the multipath delay spread in the k-th and k+1th iterations, respectively. This is the amplification factor for multipath delay spread; When the actual rolling event frequency is greater than the rolling event frequency value in the current parameter space, increase the rolling event sampling probability in the current parameter space to the target sampling probability: ； in, , The sampling probabilities of the overturning event in the k-th and k+1th iterations are respectively; δ f p is the amplification factor for the sampling probability of the overturning event. max The physical upper limit of the sampling probability for a comeback event; When the true heart rate deviates from the heart rate value within the current parameter space, adjust the heart rate parameter sampling center in the current parameter space to the target sampling center: ； ； in, , Let be the mean and standard deviation of the heart rate distribution in the (k+1)th iteration, respectively. Let f be the standard deviation of the heart rate distribution in the k-th iteration. heart,real For true heart rate, δ h This is the adjustment factor for the standard deviation of heart rate; When the actual background noise is greater than the background noise value in the current parameter space, reduce the lower limit of the signal-to-noise ratio sampling in the current parameter space to the target sampling lower limit: ； in, , δ represents the sampling lower bound of the signal-to-noise ratio (SNR) for the k-th and k+1-th iterations, respectively. NF This is the adjustment step size for the lower limit of signal-to-noise ratio (SNR) sampling.

7. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 1, characterized in that, Extracting scene fingerprints from real monitoring data involves using noisy signals and quadruples data structures to extract scene fingerprints that correspond one-to-one with four types of models from real monitoring data. Specifically, this includes: Multipath fingerprinting: estimating power delay spectrum, multipath delay spread, and principal path percentage; Motion fingerprints: frequency of rolling over events, distribution of rolling over amplitude, and duration of rolling over; Physiological fingerprints: heart rate, heart rate variability, intestinal motility; Electromagnetic fingerprint: background noise spectrum, narrowband interference frequency points.

8. The controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling according to claim 1, characterized in that, The quadruple data structure is a data structure that supports physically decoupled learning of pathological signals and environmental interference. Specifically, it includes: a pure radar respiratory signal as the carrier of pathological signals, a noisy signal simulating a real monitoring scenario, the ground truth of interference containing all components of multipath, motion, physiological and electromagnetic interference, and parameter value labels for scenario inversion and personalized adaptation.

9. A controllable data augmentation system for millimeter-wave radar respiratory monitoring based on parametric environment modeling, characterized in that, A method for implementing a controllable data augmentation method for millimeter-wave radar respiratory monitoring based on parametric environment modeling as described in any one of claims 1-8 includes: A clean radar breathing signal acquisition module is used to acquire clean radar breathing signals. The parameterized environment model library construction module is used to construct multipath propagation models with physical constraints and coupling, human motion artifact models, physiological interference models, and electromagnetic interference and noise models, and to determine the parameter spaces corresponding to the four types of models. The interference truth value calculation module is used to calculate the true values ​​of each interference based on the four types of models and the values ​​of their parameters in the parameter space. The physical mechanism differentiation synthesis module is used to calculate the noise-added signal based on the pure radar breathing signal and the true values ​​of each interference according to the physical mechanism, through convolution / time-varying filtering operator, phase modulation operator, and additive superposition operator. The quadruple data encapsulation module is used to integrate and form a quadruple data structure containing the pure radar breathing signal, the noise-added signal, the true values ​​of each interference, and the values ​​of each model parameter. The dataset generation and physical fidelity verification module is used to generate augmented datasets in batches and perform multi-level physical fidelity verification on the augmented datasets. If the verification fails, parameter tuning is automatically triggered until the verification is successful. The closed-loop adaptive evolution module is used to extract scene fingerprints from real monitoring data, map scene fingerprints to corresponding parameter space adjustment values, update the parameter space based on the adjustment values, and realize closed-loop adaptive evolution of the parameter space.

10. A millimeter-wave radar for respiratory monitoring, characterized in that, It includes a radar radio frequency transceiver unit, and a millimeter-wave radar respiratory monitoring controllable data enhancement system based on parametric environment modeling as described in claim 9.

Images