Method and device for millimeter wave radar fine-grained respiration monitoring based on diffusion generation model

Abstract

This invention discloses a fine-grained respiratory monitoring method and device based on a diffusion generation model using millimeter-wave radar. The method includes: acquiring echo signals received by the millimeter-wave radar and calculating a one-dimensional range image; locating the range cell containing the human chest cavity and extracting continuous phase information; constructing a diffusion chain from phase observation to the respiratory truth value by using the phase as input to the forward diffusion process and the difference between the respiratory truth value and the phase as progressively added noise; designing a radar diffusion denoising device based on an attention mechanism and a feedforward neural network; training the radar diffusion denoising device according to the reverse diffusion process of the diffusion chain, iteratively optimizing the denoising device parameters by reducing the difference between the respiratory waveform output by the denoising device and the respiratory truth value; inputting the signal to be denoised into the radar diffusion denoising device to complete respiratory waveform denoising and calculate the respiratory frequency. This invention improves the accuracy of respiratory waveform reconstruction and the precision of respiratory frequency estimation.

Description

Technical Field

[0001] This invention relates to respiratory monitoring technology, and more particularly to a fine-grained respiratory monitoring method and device based on a diffusion generation model using millimeter-wave radar. Background Technology

[0002] Continuous, non-invasive monitoring of breathing patterns is crucial for chronic disease management, sleep assessment, and home care. Millimeter-wave radar is a promising sensing method for this purpose due to its non-contact operation and privacy protection. However, achieving fine-grained monitoring in real-world environments remains challenging: subtle human movements introduce non-stationary artifacts, perturbing radar phase measurements and masking subtle respiratory dynamics.

[0003] In radar-based respiratory monitoring research, early methods primarily relied on traditional signal processing techniques (such as filtering and mode decomposition). However, these methods typically rely on assumptions of stationarity or independence, making them highly sensitive to time-varying interference in radar echoes. They often only achieve coarse-grained respiratory rate estimation, struggling to obtain high-fidelity waveform reconstruction. In recent years, deep learning models have attempted to learn direct mappings of clean respiratory waveforms from noisy radar phase or I / Q signals. While these "black-box" regression models perform well, they are prone to overfitting to specific environmental noise, failing to remove interference and potentially outputting results that do not conform to physiological patterns. Some studies have explored generative alternatives, such as variational autoencoders; however, the simplified posterior assumptions (usually Gaussian) used in their variational objectives limit their expressive power, making it difficult to faithfully characterize complex multimodal distributions, leading to inaccurate results. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide a fine-grained respiratory monitoring method and device based on a diffusion generation model for millimeter-wave radar, which can remove micro-motion interference, achieve more accurate respiratory waveform reconstruction results, and more precise respiratory frequency estimation.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] A fine-grained respiratory monitoring method based on a diffusion generation model using millimeter-wave radar includes the following steps:

[0007] (1) Obtain the echo signal received by the millimeter-wave radar and calculate the one-dimensional range profile corresponding to the echo signal;

[0008] (2) Locate the one-dimensional distance unit where the human chest cavity is located in the one-dimensional distance image, and extract the continuous phase information of the one-dimensional distance unit where the human chest cavity is located as the respiratory waveform observation information;

[0009] (3) Using respiratory waveform observation information as the starting point input of the forward diffusion process, the difference between the true respiratory waveform information and the corresponding respiratory waveform observation information is used as noise added step by step to construct a diffusion chain from respiratory waveform observation information to respiratory waveform true information.

[0010] (4) Construct a radar diffusion denoiser, which includes several cascaded blocks, each of which combines a self-attention mechanism, a cross-attention mechanism and a feedforward neural network.

[0011] (5) Train the radar diffusion denoiser. During training, according to the reverse diffusion process of the diffusion chain, take the breathing waveform information of each time step in the diffusion chain as the main input, take the breathing waveform observation information as the explicit condition, take minimizing the difference between the breathing waveform prediction information output by the radar diffusion denoiser and the true breathing waveform information as the loss function objective, and iteratively optimize the radar diffusion denoiser parameters.

[0012] (6) Input the breathing waveform observation information of the echo signal to be reconstructed into the trained radar diffusion denoising device, and calculate the reconstructed fine-grained breathing waveform information.

[0013] (7) Estimate respiratory frequency by spectral peak finding based on fine-grained respiratory waveform information.

[0014] Furthermore, step (1) specifically includes:

[0015] (1.1) The radar transmits millimeter-wave linear frequency modulated signals and receives target echoes. Among them, the transmitted signal and echo signal They are respectively:

[0016]

[0017]

[0018] In the formula, Indicates the amplitude of the transmitted signal. The initial phase of the transmitted signal. B is the frequency modulation slope, and B is the bandwidth. For the frequency sweep period, Indicates the starting frequency of the transmission frequency. These represent the fast time and slow time of a single chirp, respectively. The amplitude of the echo signal. For millimeter waves, the round-trip distance The resulting delay;

[0019] (1.2) By mixing the target echo with the transmitted signal, an intermediate frequency signal is obtained. :

[0020]

[0021] In the formula, For the synthesized amplitude, λ is the carrier wavelength, and c is the speed of light;

[0022] (1.3) The intermediate frequency signal is sampled by an analog-to-digital converter after analog low-pass filtering to form a discrete fast time series:

[0023]

[0024] Among them, 0 Indicates fast time sampling points, 0 Indicates slow time sampling points. The system sampling rate is M; M and N represent the number of fast-time and slow-time sampling points, respectively.

[0025] (1.4) Perform the sampling of the intermediate frequency signal. Point FFT transformation yields a one-dimensional distance image:

[0026]

[0027] Among them, 0 For one-dimensional distance cell index, This represents the value of the nth slow-time sampling point of the kth one-dimensional distance cell on the one-dimensional distance image.

[0028] Furthermore, the characteristic feature is that step (2) specifically includes:

[0029] (2.1) For one-dimensional range images, moving target display technology is used to weaken slow drift and static terms. The specific implementation method is as follows:

[0030]

[0031] In the formula, , This represents the values ​​of the nth and (n-1th)th slow-time sampling points in the kth one-dimensional distance cell. This represents a one-dimensional range image processed by moving target display technology;

[0032] (2.2) Calculate the energy of the signal in each one-dimensional range cell in the one-dimensional range image, and take the one-dimensional range cell with the largest energy. As a one-dimensional distance unit where the human thoracic cavity is located:

[0033]

[0034]

[0035] In the formula, for The energy, where N is the number of slow-time sampling points;

[0036] (2.3) Calculate one-dimensional distance cells using the arctangent function phase :

[0037]

[0038] (2.4) Regarding phase Unwinding is performed to obtain continuous phase, which is used as the respiratory waveform observation sequence. :

[0039] when hour, ,

[0040] when hour, ,in, This represents the phase increment after processing. This represents the phase increment.

[0041] Furthermore, step (3) specifically includes:

[0042] (3.1) Observation sequence of respiratory waveforms Slice the slice to obtain multiple slices of length 1. For each point sample, the samples are normalized to their maximum and minimum values ​​and then tensorized to obtain the respiratory waveform observation tensor. As the starting input for the forward diffusion process, the same slicing and tensorization operations are performed on the respiratory waveform truth sequence x(n) acquired by the respiratory monitoring band to obtain the respiratory waveform truth tensor. And calculate the residuals ;

[0043] (3.2) By gradually injecting residuals , build a The Markov chain, acting as a diffusion chain, progressively maps respiratory waveform observation information to true respiratory waveform information. The edge distribution of the diffusion chain is as follows:

[0044]

[0045] in, Indicates a time step. Indicates from Transferred to The probability, Let w represent the respiratory waveform tensor at time step w. , Let be a set of monotonically increasing noise scheduling coefficients, and satisfy . and , For noise hyperparameters, It is a unit vector. This represents a Gaussian distribution.

[0046] Furthermore, the radar diffusion denoising processor uses a Transformer as its backbone network, comprising several cascaded blocks. Each block's input includes a main string, a conditional string, and a time step embedding. Each block includes:

[0047] The normalization layer is used to normalize the main string and condition string by time step embedding modulation.

[0048] A multi-head self-attention layer is used to extract the main string. and condition strings ;

[0049] A multi-head stripe mask cross-attention layer, where each head of the stripe mask cross-attention layer performs the following calculations:

[0050]

[0051]

[0052]

[0053] in, This represents the output of the striped mask cross-attention layer. For the hidden layer dimension, Indicates matrix transpose. The softmax activation function is used. Represents a strip mask matrix The value of the element in the i-th row and j-th column. These represent the batch index of the main string and the batch index of the condition string, respectively. The radius of the local window;

[0054] A feedforward neural network is used to extract temporal features from the fused outputs of the multi-headed strip mask cross-attention layer, and its output serves as the main string for the next block.

[0055] Furthermore, step (5) specifically includes:

[0056] (5.1) Calculate the observed uniform neighborhood state of the back diffusion process of the diffusion chain according to the following formula:

[0057]

[0058] In the formula, Indicates the observed neighborhood state. This represents a noise vector that follows a standard normal distribution.

[0059] (5.2) Train the radar spread denoiser. During training, the radar spread denoiser observes the consistent neighborhood state. For initialization, the process is run in reverse diffusion along the diffusion chain, and at each time step... Respiratory waveform information at that time step The primary input is the tensor observed using the respiratory waveform. As a condition, the predicted respiratory waveform tensor is obtained. Until time step w=1, the fine-grained respiratory waveform prediction tensor is reconstructed. The loss function used is:

[0060]

[0061] in, Represents the loss function. Represents the mean function, This represents the radar diffusion denoising function.

[0062] Furthermore, during the training of the radar diffusion denoiser, Divided into non-overlapping Each batch, after adding position encoding to obtain the main string, will The same batch segmentation and position encoding are performed to obtain the condition string, and the time step embedding is... Obtained through a linear mapping layer.

[0063] A computer program product includes a computer program that, when executed by a processor, implements the above-described method.

[0064] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0065] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the above-described method.

[0066] Compared with existing technologies, the advantages of this invention are as follows: This invention designs an observation-anchored conditional diffusion framework for fine-grained radar breathing perception. To effectively remove interference caused by micro-motions, the forward process progressively injects perturbations along the direction of the residual between the observed radar phase and the target breathing signal; in the reverse process, the model is explicitly conditioned on the phase observation and the sampling trajectory is initialized within the observation-consistent neighborhood. As the denoising process progresses, micro-motion artifacts are gradually suppressed, ultimately yielding a high-fidelity breathing waveform. This "observation-anchored" strategy avoids the lengthy sampling process required for standard diffusion to start from pure noise, while maintaining consistency with the measurement physics during reconstruction. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating a fine-grained respiration monitoring method for millimeter-wave radar based on a diffusion generation model, provided in an embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of the radar diffusion denoising device provided in an embodiment of the present invention;

[0069] Figure 3 These are respiratory waveforms of a subject under four different movements, reconstructed using the present invention.

[0070] Figure 4 This is a respiratory monitoring performance map of the subject at different distances from the radar, reconstructed using the present invention. Detailed Implementation

[0071] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0072] Example 1

[0073] This invention provides a fine-grained respiratory monitoring method for millimeter-wave radar based on a diffusion generation model. This method can solve the technical problem that subtle respiratory dynamics caused by non-stationary disturbances such as human micro-movements in real-world environments are masked, making it difficult for millimeter-wave radar to accurately monitor fine-grained respiratory waveforms. To verify the effectiveness of this method, this embodiment uses the following scenario as an example to acquire the respiratory waveforms and frequencies of subjects. Using a Texas Instruments IWR6843 AoP frequency-modulated continuous wave radar, a DCA1000 data acquisition card, and a Venier GDX-RB respiratory monitoring chest strap, the radar's slow-time sampling rate and the breathing strap's sampling rate were both 20Hz. The radar bandwidth was set to 2GHz, the ADC sampling rate was 10MHz, and the number of ADC sampling points per chirp was 512. Radar respiratory synchronization data of 12 healthy subjects were collected in a typical office environment, with one hour of data collected from each subject.

[0074] like Figure 1 As shown, the method provided in this embodiment of the invention includes the following steps:

[0075] Step 1: Obtain the echo signal received by the millimeter-wave radar and calculate the one-dimensional range image corresponding to the echo signal.

[0076] Step 1.1: The frequency-modulated continuous wave radar transmits a linear frequency-modulated signal and receives the target echo. Linear frequency modulation refers to frequency modulation within a single chirp sweep cycle. Inside, the transmission frequency starts from the initial frequency. A linear sweep to the termination frequency forms a fixed bandwidth. (FM slope) The frequency scanning process involves the waveform generation module first generating a linear frequency modulated signal, which is then radiated into space by the transmitting antenna. After the electromagnetic wave propagates to the target, it generates an echo signal. The echo signal has a propagation delay relative to the transmitted signal and is accompanied by amplitude attenuation.

[0077] Assume the fast time within chirp is Then transmit signal It can be represented as:

[0078]

[0079] in Indicates the amplitude of the transmitted signal. This represents the initial phase of the transmitted signal. At a time in the slow time dimension... The transmitted signal illuminates the range radar. After the human target is detected, the satellite returns and, after a delay... (where c = 3.0 × 10) 8 (m / s, the propagation speed of millimeter waves), the receiving antenna receives the echo. :

[0080]

[0081] in The amplitude of the received signal. For millimeter waves, the round-trip distance The resulting delay.

[0082] Step 1.2: At the receiving end, the intermediate frequency signal can be obtained by mixing the received signal with the transmitted signal. :

[0083]

[0084] As can be seen from the above formula, the intermediate frequency signal has a single frequency component, and its frequency is determined by... Given, the synthesis amplitude is The constant phase introduced by the carrier frequency can be expressed as:

[0085]

[0086] in Where is the carrier frequency wavelength. This is derived from the squared time delay term. The resulting constant components:

[0087]

[0088] This component is small and is usually ignored. After simplification, the intermediate frequency signal can be expressed as:

[0089]

[0090] The frequency of the intermediate frequency signal can be seen from the above formula. and phase All with Proportional, when the target is stationary Fixed distance between the human body and the radar and chest wall displacement The sum of When the chest wall is generated by respiratory movements When the displacement is , the frequency changes

[0091]

[0092] Due to sampling rate limitations, the frequency resolution of radar signals is [value missing]. Therefore, the distance resolution that can be detected based on frequency changes is:

[0093] Because millimeter-wave radar has a large bandwidth, its range resolution is typically much greater than the chest wall displacement caused by breathing. Therefore, frequency variations can only be used to determine the approximate distance between the target and the radar, and cannot be used to monitor chest wall displacement caused by breathing. The phase change produced by the same displacement in the phase is:

[0094]

[0095] Millimeter-wave radar typically has a high center frequency, corresponding to a transmitted signal wavelength on the order of millimeters. In this case, the phase signal can sensitively detect minute chest wall displacements caused by bodily activity. Therefore, in respiratory monitoring using millimeter-wave radar, the intermediate frequency signal is used to determine the distance between the target and the radar, while the phase signal is used to monitor minute movements caused by breathing.

[0096] Step 1.3: After analog low-pass filtering, the intermediate frequency (IF) signal is sampled by an analog-to-digital converter (ADC) to form a discrete fast time series. The IF signal after discrete sampling can be represented as:

[0097]

[0098] Among them, 0 Indicates fast time sampling points, 0 This represents slow-time sampling points, with data per second. , This represents the system sampling rate.

[0099] Step 1.4: Perform processing on the sampled intermediate frequency signal. Point FFT transformation can separate the beat frequencies corresponding to different distances, resulting in a one-dimensional range image. :

[0100]

[0101] Among them, 0 This is the distance cell index.

[0102] Step 2: Locate the one-dimensional distance unit containing the human chest cavity in the one-dimensional distance image, and extract the continuous phase information of the one-dimensional distance unit containing the human chest cavity as respiratory waveform observation information.

[0103] Step 2.1: To distinguish between static interference targets and slightly moving human targets, it is necessary to use moving target display technology to weaken slow drift and static items. The specific implementation method is as follows:

[0104]

[0105] This technology can not only remove a large number of constants related to static distance, but also enhance the micro-motion component of breathing that changes over time.

[0106] Step 2.2: Calculate the signal energy within each distance cell. :

[0107]

[0108] Take the distance unit with the highest energy As the distance unit where the human thoracic cavity is located,

[0109]

[0110] Step 2.3: Select the correct distance unit Then, the phase corresponding to this distance contains the target's chest wall displacement information. The phase of this distance cell is calculated using the arctangent function:

[0111]

[0112] Step 2.4, because... Internal molding will appear A jump requires phase unwrapping to obtain continuous phase, which serves as the absorption waveform observation sequence. :

[0113] when hour ,

[0114] when hour, ,in, This represents the phase increment after processing. This represents the phase increment.

[0115] Step 3: Using the respiratory waveform observation information as the starting point input of the forward diffusion process, the difference between the respiratory waveform true value information and the corresponding respiratory waveform observation information is used as noise added step by step to construct a diffusion chain from the respiratory waveform observation information to the respiratory waveform true value information.

[0116] Step 3.1, for Slice the slice to obtain multiple slices of length 1. The starting input of the forward diffusion process is obtained by normalizing the samples to their minimum and maximum values ​​and then tensorizing them. True value sequence of respiratory waveforms acquired by respiratory monitoring tape The same slicing and tensor quantization operations are performed to obtain the respiratory truth tensor. To construct a forward diffusion chain from radar phase observations to the true respiratory value, the residual is defined. It encodes waveform distortion information caused by micro-motion.

[0117] Step 3.2: Unlike the standard diffusion model, which injects Gaussian noise unrelated to observations during the forward propagation process, this invention progressively injects residuals that encode disturbances caused by micro-motions. Thus constructing a A step-by-step Markov chain, as a diffusion chain, will Mapped to The process is handled by a set of monotonically increasing noise schedulers. Controlled, satisfying and ,in This indicates the time step. Its one-step transfer kernel is as follows:

[0118]

[0119] Among them, when hour ,and ; For noise hyperparameters, the starting point of the forward diffusion chain is... At any time step The marginal distribution has a closed expression:

[0120]

[0121] The above formula shows that as The increase, The mean will be from Gradually and smoothly transition to This completes the gradual degradation process from the respiratory signal to radar observation. Residuals encoding micro-motions are explicitly injected during the forward propagation process, thus maintaining consistency with radar measurements throughout the entire diffusion chain.

[0122] Step 4: Construct a radar diffusion denoiser, which includes several cascaded blocks, each of which combines a self-attention mechanism, a cross-attention mechanism, and a feedforward neural network.

[0123] like Figure 2 As shown, the radar diffusion Transformer denoiser consists of two cascaded blocks, each with three inputs: the main string (token), the condition string (token), and the time step embedding.

[0124] Each block includes:

[0125] The normalization layer is used to normalize the main string and condition string by time step embedding modulation.

[0126] A multi-head self-attention layer is used to extract the main string. and condition strings ;

[0127] A multi-head stripe mask cross-attention layer, where each head of the stripe mask cross-attention layer performs the following calculations:

[0128]

[0129]

[0130]

[0131] in, This represents the output of the striped mask cross-attention layer. =256 is the hidden layer dimension. Indicates matrix transpose. The softmax activation function is used. Represents a strip mask matrix The value of the element in the i-th row and j-th column is used as a mask to hardcode local priors by allowing only interactions within a local time window, thereby suppressing long-range mismatches and spurious associations. These represent the batch index of the main string and the batch index of the condition string, respectively. =1 represents the local window radius;

[0132] A feedforward neural network is used to extract temporal features from the fused outputs of the multi-headed strip mask cross-attention layer, and its output serves as the main string for the next block.

[0133] Step 5: Train the radar diffusion denoiser.

[0134] During training, following the reverse diffusion process of the diffusion chain, the respiratory waveform information at each time step in the diffusion chain is used as the main input, and the respiratory waveform observation information is used as the explicit condition. The loss function objective is to minimize the difference between the predicted respiratory waveform information output by the radar diffusion denoiser and the true respiratory waveform information. The radar diffusion denoiser parameters are iteratively optimized.

[0135] Step 5.1: To effectively utilize observational information to guide micro-motion removal, the backdiffusion process is initialized at... In the observed consistent neighborhood state:

[0136]

[0137] Instead of initializing from pure Gaussian noise as in standard diffusion.

[0138] Step 5.2: Train the radar diffusion denoiser. During training, the radar diffusion denoiser observes the consistent neighborhood state. For initialization, the process is run in reverse diffusion along the diffusion chain, and at each time step... Respiratory waveform information at that time step The primary input is the tensor observed using the respiratory waveform. As a condition, the output of the last block is used to obtain the breathing waveform prediction tensor through the inverse patch operation. Until time step w=1, the fine-grained respiratory waveform prediction tensor is reconstructed. .in The main token is obtained by dividing the data into 20 non-overlapping patches and adding positional encoding. The condition token is determined by... The same pattern partitioning and positional encoding are performed, which helps extract fine-grained temporal representations and enhances positional alignment. Time-step embedding is achieved by... Obtained through a linear mapping layer.

[0139] Reverse transfer nuclei in the reverse diffusion process By parameters The radar diffusion Transformer denoiser is learned by minimizing the difference between the actual and true posterior values. The Kullback–Leibler (KL) divergence implementation. The true posterior has the following closed form:

[0140]

[0141] Therefore, The parameter is categorized into a Gaussian distribution: its variance is fixed to be consistent with the true posterior, i.e. Its mean is

[0142]

[0143] Among them, noise reduction In Under certain conditions, from the noise state Predict clean breathing signals and gradually remove micro-motion interference.

[0144] The denoiser model parameters are iteratively optimized by reducing the difference between the breathing waveform output by the denoiser and the true breathing value. The training objective of the model is the mean squared error (MSE) loss.

[0145]

[0146] This loss enables the radar diffusion Transformer denoiser to denoise at different noise levels and reconstruct fine-grained breathing waveforms.

[0147] Step 6: Input the respiratory waveform observation information of the echo signal to be reconstructed into the trained radar diffusion denoising device to calculate the reconstructed fine-grained respiratory waveform information. .

[0148] Step 7: Estimate the respiratory rate by performing spectrum peak finding based on fine-grained respiratory waveform information.

[0149] right conduct The respiration spectrum was obtained after point Fourier transform, followed by a cutoff frequency of... of A Butterworth filter of order 1 performs peak finding on the bandpass filtered spectrum and identifies the frequency corresponding to the maximum value. Then, it identifies the frequency corresponding to the largest peak value in the filtered spectrum. As respiratory rate.

[0150] Figure 3 The above parameters are the reconstructed respiratory waveforms of the subjects under four different actions (playing on their phones, chatting, shaking their legs, and turning pages). Figure 4This is a graph showing the cosine similarity of the respiratory waveforms of subjects at distances from the radar from 0.6 to 3 meters, and the mean absolute error of respiratory frequency estimation. Table 1 shows the numerical results for cosine similarity, mean square error, mean absolute error, root mean square error, and standard deviation of respiratory waveform estimation under the above parameters. It can be seen that this invention can resist disturbances and achieves more accurate identification results.

[0151] Table 1 Evaluation Results of the Invention

[0152] Evaluation indicators Cosine similarity Mean square error Mean Absolute Error Root mean square error Standard deviation result 0.811 0.079 0.631 1.175 1.035

[0153] Example 2

[0154] This invention also provides a computer program product, such as an app on a mobile phone or tablet, or an installer on a computer. This product includes a computer program / instructions that, when executed by a processor, implement the method described in Embodiment 1. The code for the computer-executable program used to perform the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0155] Example 3

[0156] This invention provides a computer device that provides services for implementing the method described in Embodiment 1. The device may include: a memory storing a computer-executable program; a processor coupled to the memory; and the processor calling the computer-executable program stored in the memory to execute the steps of the method described in Embodiment 1.

[0157] The memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the memory may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). A program / utility having a set (at least one) of program modules may be stored in, for example, memory. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The computer-executable program of the program modules typically performs the functions and / or methods described in the embodiments of the present invention.

[0158] The processor executes various functional applications and data processing by running programs stored in memory, such as the method provided in Embodiment 1 of the present invention.

[0159] The code of a computer executable program can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages.

[0160] Example 4

[0161] This invention provides a storage medium containing a computer-executable program, which, when executed by a computer processor, is used to perform the method of Embodiment 1.

[0162] The storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0163] Of course, the computer-executable program in the storage medium provided in the embodiments of the present invention is not limited to the above-described method operations, but can also perform related operations in the methods provided in any embodiment of the present invention.

[0164] It should be understood that the embodiments and descriptions above are only the principles, main features and advantages of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the invention, and all such changes and modifications fall within the protection scope of the present invention.

Claims

1. A fine-grained respiration monitoring method for millimeter-wave radar based on a diffusion generation model, characterized in that, Includes the following steps: (1) Obtain the echo signal received by the millimeter-wave radar and calculate the one-dimensional range profile corresponding to the echo signal; (2) Locate the one-dimensional distance unit where the human chest cavity is located in the one-dimensional distance image, and extract the continuous phase information of the one-dimensional distance unit where the human chest cavity is located as the respiratory waveform observation information; (3) Using respiratory waveform observation information as the starting point input of the forward diffusion process, the difference between the true respiratory waveform information and the corresponding respiratory waveform observation information is used as noise added step by step to construct a diffusion chain from respiratory waveform observation information to respiratory waveform true information. (4) Construct a radar diffusion denoiser, which includes several cascaded blocks, each of which combines a self-attention mechanism, a cross-attention mechanism and a feedforward neural network. (5) Train the radar diffusion denoiser. During training, according to the reverse diffusion process of the diffusion chain, take the breathing waveform information of each time step in the diffusion chain as the main input, take the breathing waveform observation information as the explicit condition, take minimizing the difference between the breathing waveform prediction information output by the radar diffusion denoiser and the true breathing waveform information as the loss function objective, and iteratively optimize the radar diffusion denoiser parameters. (6) Input the breathing waveform observation information of the echo signal to be reconstructed into the trained radar diffusion denoising device to obtain the reconstructed fine-grained breathing waveform information. (7) Estimate respiratory frequency by spectral peak finding based on fine-grained respiratory waveform information.

2. The millimeter-wave radar fine-grained respiration monitoring method based on a diffusion generation model according to claim 1, characterized in that, Step (1) specifically includes: (1.1) The radar transmits millimeter-wave linear frequency modulated signals and receives target echoes. Among them, the transmitted signal and echo signal They are respectively: , , In the formula, Indicates the amplitude of the transmitted signal. The initial phase of the transmitted signal. B is the frequency modulation slope, and B is the bandwidth. For the frequency sweep period, Indicates the starting frequency of the transmission frequency. These represent the fast time and slow time of a single chirp, respectively. The amplitude of the echo signal. For millimeter waves, the round-trip distance The resulting delay; (1.2) By mixing the target echo with the transmitted signal, an intermediate frequency signal is obtained. : , In the formula, For the synthesized amplitude, λ is the carrier wavelength, and c is the speed of light; (1.3) The intermediate frequency signal is sampled by an analog-to-digital converter after analog low-pass filtering to form a discrete fast time series: , Among them, 0 Indicates fast time sampling points, 0 Indicates slow time sampling points. The system sampling rate is M; M and N represent the number of fast-time and slow-time sampling points, respectively. (1.4) Perform the sampling of the intermediate frequency signal. Point FFT transformation yields a one-dimensional distance image: , Among them, 0 For one-dimensional distance cell index, This represents the value of the nth slow-time sampling point of the kth one-dimensional distance cell on the one-dimensional distance image.

3. The millimeter-wave radar fine-grained respiration monitoring method based on a diffusion generation model according to claim 1, characterized in that, Step (2) specifically includes: (2.1) For one-dimensional range images, moving target display technology is used to weaken slow drift and static terms. The specific implementation method is as follows: , In the formula, , This represents the values ​​of the nth and (n-1th)th slow-time sampling points in the kth one-dimensional distance cell. This represents a one-dimensional range image processed by moving target display technology; (2.2) Calculate the energy of the signal in each one-dimensional range cell in the one-dimensional range image, and take the one-dimensional range cell with the largest energy. As a one-dimensional distance unit where the human thoracic cavity is located: , , In the formula, for The energy, where N is the number of slow-time sampling points; (2.3) Calculate one-dimensional distance cells using the arctangent function phase : , (2.4) Regarding phase Unwinding is performed to obtain continuous phase, which is used as the respiratory waveform observation sequence. : when hour, , when hour, ,in, This represents the phase increment after processing. This represents the phase increment.

4. The millimeter-wave radar fine-grained respiration monitoring method based on a diffusion generation model according to claim 1, characterized in that, Step (3) specifically includes: (3.1) Observation sequence of respiratory waveforms Slice the slice to obtain multiple slices of length 1. For each point sample, the samples are normalized to their maximum and minimum values ​​and then tensorized to obtain the respiratory waveform observation tensor. As the starting input for the forward diffusion process, the same slicing and tensorization operations are performed on the respiratory waveform truth sequence x(n) acquired by the respiratory monitoring band to obtain the respiratory waveform truth tensor. And calculate the residuals ; (3.2) By gradually injecting residuals , build a The Markov chain, acting as a diffusion chain, progressively maps respiratory waveform observation information to true respiratory waveform information. The edge distribution of the diffusion chain is as follows: , in, Indicates a time step. Indicates from Transferred to The probability, Let w represent the respiratory waveform tensor at time step w. , Let be a set of monotonically increasing noise scheduling coefficients, and satisfy . and , For noise hyperparameters, It is a unit vector. This represents a Gaussian distribution.

5. The millimeter-wave radar fine-grained respiration monitoring method based on a diffusion generation model according to claim 1, characterized in that, The radar diffusion denoising unit uses a Transformer as its backbone network and includes several cascaded blocks. Each block's input includes a main string, a condition string, and a time step embedding. Each block includes: The normalization layer is used to normalize the main string and condition string by time step embedding modulation. A multi-head self-attention layer is used to extract the main string. and condition strings ; A multi-head stripe mask cross-attention layer, where each head of the stripe mask cross-attention layer performs the following calculations: , , , in, This represents the output of the striped mask cross-attention layer, where 'a' is the hidden layer dimension. Indicates matrix transpose. The softmax activation function is used. Represents a strip mask matrix The value of the element in the i-th row and j-th column. These represent the batch index of the main string and the batch index of the condition string, respectively, with U representing the batch quantity. The radius of the local window; A feedforward neural network is used to extract temporal features from the fused outputs of the multi-headed strip mask cross-attention layer, and its output serves as the main string for the next block.

6. The millimeter-wave radar fine-grained respiration monitoring method based on a diffusion generation model according to claim 4, characterized in that, Step (5) specifically includes: (5.1) Calculate the observed uniform neighborhood state of the back diffusion process of the diffusion chain according to the following formula: , In the formula, Indicates the observed neighborhood state. This represents a noise vector that follows a standard normal distribution. (5.2) Train the radar spread denoiser. During training, the radar spread denoiser observes the consistent neighborhood state. For initialization, the process is run in reverse diffusion along the diffusion chain, and at each time step... Respiratory waveform information at that time step The primary input is the tensor observed using the respiratory waveform. As a condition, the predicted respiratory waveform tensor is obtained. Until time step w=1, the fine-grained respiratory waveform prediction tensor is reconstructed. The loss function used is: , in, Represents the loss function. Represents the mean function, This represents the radar diffusion denoising function.

7. The millimeter-wave radar fine-grained respiration monitoring method based on a diffusion generation model according to claim 6, characterized in that, During the training of the radar diffusion denoiser, Divided into non-overlapping Each batch, after adding position encoding to obtain the main string, will The same batch segmentation and position encoding are performed to obtain the condition string, and the time step embedding is... Obtained through a linear mapping layer.

8. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the method of any one of claims 1-7.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes the computer program to implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: The computer program / instructions, when executed by a processor, implement the method of any one of claims 1-7.

Images