Millimeter wave radar respiration monitoring method and device based on gated attention wavelet network

Abstract

The invention discloses a millimeter-wave radar respiration monitoring method and device based on a gated attention wavelet network, and the method comprises the steps: a millimeter-wave radar sends a detection signal to a human body target, receives an echo signal returned by the human body target, converts the echo signal into two paths of orthogonal lacing signals, and carries out the sampling and combination of the two paths of orthogonal lacing signals into a complex baseband signal; fourier transform is carried out on the fast time dimension of the complex baseband signal to obtain a one-dimensional distance image, a distance unit with the maximum average energy in the one-dimensional distance image is selected as a distance unit where the human chest is located, and a continuous phase of the distance unit where the human chest is located is extracted; inputting the continuous phase of the distance unit where the human body is located into a time gating attention encoder to obtain a phase after primary interference suppression; inputting the phase after the primary interference suppression into a frequency gating attention decoder, and reconstructing a pure respiration waveform; fourier transform and band-pass filtering are conducted on the pure breathing waveform, and the frequency where a peak value is located serves as the estimated breathing frequency. The method is higher in accuracy.

Description

Technical Field

[0001] This invention relates to respiratory monitoring technology, and more particularly to a millimeter-wave radar respiratory monitoring method and device based on a gated attention wavelet network. Background Technology

[0002] Millimeter-wave radar, as a non-contact and privacy-preserving sensing method, has become a promising solution for long-term respiratory monitoring, meeting a key need in modern home healthcare. Despite these advantages, accurately extracting fine-grained respiratory waveforms from highly non-stationary radar echoes remains challenging due to various interferences in real-world scenarios, particularly non-periodic body micro-movements. The spectral components of these interferences often significantly overlap with the respiratory frequency band; furthermore, the amplitude of body micro-movement interferences is typically much larger than the amplitude of the respiratory signal, which originates from chest wall movements typically less than 12 millimeters.

[0003] Traditional radar-based respiration monitoring methods primarily rely on signal processing techniques such as mode decomposition and blind source separation. However, these methods heavily depend on prior assumptions, such as predefined frequency bands or signal independence, which often fail due to the spectral aliasing and highly non-stationary nature of real-world signals. To overcome these limitations, recent advances in deep learning have achieved significant improvements through end-to-end modeling. Techniques such as adversarial learning and self-attention have enhanced the accuracy and robustness of waveform reconstruction. However, existing deep learning methods suffer from drawbacks; for example, some methods ignore micro-motion interference located in the same distance cell as the respiration signal, while others directly use the raw I / Q signal as model input, neglecting the direct correlation between phase and target distance. These shortcomings hinder the model's generalization ability and physical interpretability in complex dynamic scenarios, leading to inaccurate reconstruction results. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide a more accurate millimeter-wave radar respiratory monitoring method and device based on a gated attention wavelet network.

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

[0006] A millimeter-wave radar respiratory monitoring method based on a gated attention wavelet network includes the following steps:

[0007] Step 1: The millimeter-wave radar sends a detection signal to the human target and receives the echo signal returned by the human target. After converting the echo signal into two orthogonal band signals, it samples and combines them into a complex baseband signal.

[0008] Step 2: Perform Fourier transform on the fast time dimension of the complex baseband signal to obtain a one-dimensional range image. Select the range cell with the largest average energy in the one-dimensional range image as the range cell where the human chest cavity is located, and extract the continuous phase of the range cell where the human chest cavity is located.

[0009] Step 3: Input the continuous phase of the distance unit where the human body is located into the time-gated attention encoder to obtain the phase after the initial interference suppression. The time-gated attention encoder is used to suppress the mid-to-high frequency interference and non-periodic body micro-motion interference in the continuous phase using the time-gated attention mechanism.

[0010] Step 4: Input the phase after the initial interference suppression into the frequency-gated attention decoder to reconstruct the pure breathing waveform. The frequency-gated attention decoder is used to perform secondary suppression on the phase after the initial interference suppression and reconstruct the wavelet coefficients of the breathing waveform. Then, the pure breathing waveform is reconstructed through inverse wavelet transform. The time-gated attention encoder and the frequency-gated attention decoder constitute a gated attention wavelet network.

[0011] Step 5: Perform Fourier transform and bandpass filtering on the reconstructed pure respiratory waveform, and use the frequency of the peak as the estimated respiratory frequency.

[0012] Furthermore, step 1 specifically includes the following steps:

[0013] Step 1.1: The detection signal transmitted by the millimeter-wave radar's transmitting antenna towards the human target is as follows:

[0014]

[0015] Among them, s tx (t k A represents the detected signal transmitted within a single linear frequency sweep cycle. t f is the signal amplitude. c The starting carrier frequency is 0 ≤ t k ≤T c Fast time refers to the sampling time within a single linear sweep cycle, B is the modulation bandwidth, and T is the sampling time. c For linear frequency sweep period, Let Re{·} denote the modulation slope, where Re is the real part, and j is the imaginary unit. 2 =-1;

[0016] Step 1.2: The millimeter-wave radar's receiving antenna receives the echo signal s returned by the human target. rx (t m ,t k ), t m =nT PRF For slow time, The pulse repetition period is denoted by PRF, the pulse repetition frequency is denoted by PRF, one linear sweep frequency period signal is one Chirp signal, n = 1, 2, ..., N represents the index of the transmitted probe signal Chirp, and N is the total number of probe signal Chirps;

[0017] Step 1.3: After passing the echo signal through the mixer and low-pass filter inside the millimeter-wave radar, the intermediate frequency signal s is obtained. IF (t m ,t k );

[0018] Step 1.4: Convert the intermediate frequency signal s IF (t m ,t k The baseband signal is multiplied in phase and quadrature with the local oscillator signal, respectively, and then low-pass filtered before being passed through a 90° phase shifter to obtain two orthogonal baseband signals I(t). m ,t k ) and Q(t m ,t k );

[0019] Step 1.5: Convert the baseband signal I(t) m ,t k ) and Q(t m ,t k The signals are fed into an analog-to-digital converter, and each Chirp signal is uniformly sampled to obtain a discrete sequence. n = 1, 2, ..., N, m = 0, 1, ..., M-1, where M is the number of samples for each Chirp signal;

[0020] Step 1.6: Combine the two data streams into a complex baseband signal.

[0021] Furthermore, step 2 includes the following steps:

[0022] Step 2.1: Perform an M-point Fourier transform on the fast time dimension of the complex baseband signal z(n,m) to obtain a one-dimensional range profile:

[0023]

[0024] Where Z(n,k) represents the distance cell k in the one-dimensional distance image of the nth Chirp;

[0025] Step 2.2: Enhance the one-dimensional range image using moving target display technology to obtain the enhanced one-dimensional range image:

[0026]

[0027] in, Z(n-1,k) represents the enhanced one-dimensional range image, and Z(n-1,k) represents the one-dimensional range image of the (n-1)th Chirp.

[0028] Step 2.3: Calculate the average energy E(k) of range cell k in the enhanced one-dimensional range image of all Chirps according to the following formula:

[0029]

[0030] Where N is the total number of Chirp probes;

[0031] Step 2.4: Select the distance unit with the highest average energy. As a unit of distance to the human thoracic cavity:

[0032]

[0033] Step 2.5: Calculate the phase θ(n) of the distance unit containing the human thoracic cavity:

[0034] θ(n)=atan2(Q z (n),I z (n))

[0035] In the formula, I z (n) represents Real part, Q z (n) represents The virtual part;

[0036] Step 2.6: Unwrap the phase θ(n) to obtain the continuous phase x(n).

[0037] Furthermore, the untangling in step 2.6 specifically involves:

[0038]

[0039] Δ′=θ(n)-θ(n-1) represents the phase difference before unwinding, and Δ represents the continuous phase difference.

[0040] Furthermore, the continuous phases in step 3 are processed as follows before being used as input to the time-gated attention encoder:

[0041] The continuous phase x(n) is sliced ​​using a window with a length of l and a step size of s. Each slice is considered as a sample, and the total number of samples after slicing is... Each sample has a length of l points, where This indicates rounding down, where N is the total number of Chirp probes. Each b samples in the total sample is grouped into a batch and concatenated into a continuous phase tensor X, which serves as the input to the time-gated attention encoder.

[0042] Furthermore, the time-gated attention encoder is composed of several cascaded time-gated attention blocks, each time-gated attention block including:

[0043] The input layer is used to calculate the query vector Q based on the input. E Key vector K E and value vector V E In this case, the input to the first time-gated attention block is a continuous phase X, and Q... E =K E =V E =X;

[0044] Convolutional gated networks are used to process query vector Q. E and bond vector K E Convolutional gating is performed using the following formula:

[0045]

[0046] Where LP(·) represents a linear layer, Conv1D(·) represents a one-dimensional convolutional layer, and Sigmoid(·) represents the Sigmoid activation function. and They are Q E With K E The query vector and key vector have undergone convolutional gating;

[0047] The output layer is used to calculate the output of the current time-gated attention block according to the following formula:

[0048]

[0049] Among them, K' E express The transpose of , SoftMax(·) denotes the SoftMax activation function, and l is the continuous phase length; The output of the current time-gated attention block is used as the input to the next time-gated attention block, and the output of the last time-gated attention block is used as the phase after the initial interference suppression.

[0050] Furthermore, the frequency-gated attention decoder includes I parallel sub-decoders and an inverse wavelet transform module connected to all sub-decoders. The input to each sub-decoder is the phase after initial interference suppression. Each sub-decoder includes several cascaded frequency-gated attention blocks and a one-dimensional convolutional layer at the last layer. Each frequency-gated attention block includes:

[0051] The input layer is used to calculate the query vector Q based on the input. D Key vector KD and value vector V D The input to the first frequency-gated attention block is the phase after initial interference suppression. and

[0052] Frequency-gated networks are used to query vector Q. D With key vector K D Perform frequency gating:

[0053]

[0054] Where FFN(·) represents a feedforward neural network, FFT(·) represents a fast Fourier transform, LP(·) represents a linear layer, and Sigmoid(·) represents the Sigmoid activation function. and These are the query vectors Q D With key vector K D The query vector and key vector have been frequency-gated;

[0055] The output layer is used to calculate the output of the current frequency-gated attention block according to the following formula:

[0056]

[0057] Where IFFT(·) denotes the inverse fast Fourier transform, SoftMax(·) denotes the SoftMax activation function, l is the continuous phase length, and Conj(·) denotes the conjugate operation. The output of the current frequency-gated attention block is used as the input to the next frequency-gated attention block. The output of the last frequency-gated attention block is then fed into a one-dimensional convolutional layer to obtain the wavelet coefficients C. i ;

[0058] The inverse wavelet transform module is used to convert the wavelet coefficients C output by all sub-decoders. i The pure respiratory waveform Y is reconstructed by inverse wavelet transform of order I-1.

[0059] Furthermore, step 5 includes the following steps:

[0060] Step 5.1: Perform a Fourier transform on the reconstructed pure respiratory waveform Y to obtain the respiratory spectrum ψ;

[0061] Step 5.2: Let ψ pass through the cutoff frequency [f]. l ,f h A J-order Butterworth bandpass filter is used to obtain the filtered breathing spectrum, where f l ,f h These are the lowest and highest frequencies of the filter, respectively.

[0062] Step 5.3: Take the frequency F corresponding to the largest peak on the filtered respiratory spectrum as the estimated respiratory rate.

[0063] Furthermore, the gated attention wavelet network is trained using a preset loss function, which is specifically:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] Where Loss is the total loss, and w1 to w4 are the weights used to balance the losses of each item. The loss represents the respiratory waveform value, b is the number of consecutive phase samples in each training batch, l is the length of the consecutive phase, and Y, Let ||·|| represent the clean breathing waveform and the true value of the breathing waveform output by the gated attention wavelet network, respectively. F Denotes the Frobenius norm. The symbol represents the shape loss of the respiratory waveform, <, ·> represents the element-wise inner product, and I represents the number of sub-decoders in the frequency-gated attention decoder. C represents the wavelet coefficient loss. i , Let i and n represent the i-th wavelet coefficients of the breathing waveform obtained from the gated attention wavelet network and the true value of the breathing waveform, respectively. The true value of the i-th wavelet coefficient, d i Indicate C i Dimensions Indicates respiratory spectrum loss, ψ, Represent the respiratory spectrum obtained by the gated attention wavelet network and the true value of the respiratory waveform, respectively. The true value of the obtained respiratory spectrum.

[0070] A millimeter-wave radar respiratory monitoring device based on a gated attention wavelet network includes:

[0071] Millimeter-wave radar is used to send detection signals to human targets and receive echo signals returned by human targets.

[0072] The signal preprocessing module is used to convert the echo signal into two orthogonal baseband signals, and then sample and combine them into a complex baseband signal.

[0073] The phase extraction module is used to perform Fourier transform on the fast time dimension of the complex baseband signal to obtain a one-dimensional range image, select the range cell with the largest average energy in the one-dimensional range image as the range cell where the human chest cavity is located, and extract the continuous phase of the range cell where the human chest cavity is located.

[0074] The interference suppression module is used to input the continuous phase of the distance unit where the human body is located into the time-gated attention encoder to obtain the phase after initial interference suppression. The time-gated attention encoder is used to use the time-gated attention mechanism to initially suppress the mid-to-high frequency interference and non-periodic body micro-motion interference in the continuous phase.

[0075] The breathing waveform reconstruction module is used to input the phase after the initial interference suppression into the frequency-gated attention decoder to reconstruct the pure breathing waveform. The frequency-gated attention decoder is used to perform secondary suppression on the phase after the initial interference suppression and reconstruct the wavelet coefficients of the breathing waveform. Then, the pure breathing waveform is reconstructed through inverse wavelet transform. The time-gated attention encoder and the frequency-gated attention decoder constitute a gated attention wavelet network.

[0076] The respiratory rate estimation module is used to perform Fourier transform and bandpass filtering on the reconstructed clean respiratory waveform, and use the frequency of the peak as the estimated respiratory rate.

[0077] Compared with existing technologies, the advantages of this invention are as follows: This invention utilizes a designed gated attention wavelet network to combine the nonlinear modeling advantages of deep neural networks with the multi-scale analysis capabilities of wavelet transform. It operates directly based on radar phase information and employs a two-stage suppression mechanism: a time-gated attention encoder enhances respiratory features while initially suppressing various interferences, especially those from non-periodic body micro-movements; subsequently, a frequency-gated attention decoder adaptively separates and refines the wavelet coefficients related to breathing, thereby providing more precise secondary suppression of residual interference. Finally, the purified signal is reconstructed through inverse discrete wavelet transform, resulting in a high-fidelity respiratory waveform with higher generalization ability and stronger interpretability. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the millimeter-wave radar respiratory monitoring method based on a gated attention wavelet network provided by the present invention.

[0079] Figure 2 This is the structure of the time-gated attention encoder provided by the present invention;

[0080] Figure 3 This is the structure of the frequency-gated attention decoder provided by the present invention;

[0081] Figure 4This is the structure of the sub-decoder provided by the present invention;

[0082] Figure 5 This is a diagram showing the respiratory waveform of a subject in a quasi-static state, restored using the gated attention wavelet network of this invention.

[0083] Figure 6 This is a diagram showing the respiratory waveform of a subject under motion conditions, restored using the gated attention wavelet network of this invention. Detailed Implementation

[0084] 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.

[0085] Example 1

[0086] This invention provides a millimeter-wave radar respiratory monitoring method based on a gated attention wavelet network. This method can solve the technical problem of accurately extracting fine-grained respiratory waveforms from highly non-stationary radar echoes, which are subject to various interferences in real-world scenarios, especially non-periodic body micro-movements. This embodiment uses a 60GHz IWR6843FMCW radar and a DCA1000 acquisition board to collect radar data from 12 subjects in an office environment. True-value respiratory data were simultaneously acquired using a GDX-RB respiratory belt. Each subject was approximately 1.2 meters away from the radar, and four 15-minute recordings were performed.

[0087] like Figure 1 As shown, the embodiments of the present invention include the following steps:

[0088] Step 1: The millimeter-wave radar sends a detection signal to the human target and receives the echo signal returned by the human target. After converting the echo signal into two orthogonal tethered signals, it samples and combines them into a complex baseband signal.

[0089] Step 1.1: The detection signal transmitted by the millimeter-wave radar's transmitting antenna towards the human target is as follows:

[0090]

[0091] Among them, s tx (t k A represents the detected signal transmitted within a single linear frequency sweep cycle. t f is the signal amplitude. c The starting carrier frequency is 0 ≤ t k ≤T c Fast time refers to the sampling time within a single linear sweep cycle, B is the modulation bandwidth, and T is the sampling time. c For linear frequency sweep period, Let Re{·} denote the modulation slope, where Re is the real part, and j is the imaginary unit. 2= -1. In this embodiment, the setting parameters of the millimeter-wave radar are as shown in Table 1 below:

[0092] Table 1 Millimeter-wave Radar Parameter Settings

[0093]

[0094] Step 1.2: The receiving antenna of the millimeter-wave radar receives the echo signal s rx (t m ,t k ) returned by the human target. s rx (t m ,t k ) can be expressed as:

[0095]

[0096] where t m = nT PRF is the slow time, is the pulse repetition period, PRF is the pulse repetition frequency, a linear frequency sweep period signal is a Chirp signal, n = 1, 2,..., N represents the index of the transmitted detection signal Chirp, N is the total number of detection signal Chirps. In this embodiment, the total number of Chirps in 15 minutes of data is 18000. is the propagation delay of the signal traveling to and from the target, and the electromagnetic wave propagation speed c is 3.0×10 8 m / s.

[0097] R(t m ) is the distance from the human chest wall to the radar and can be expressed as:

[0098] R(t m ) = R0 + x b (t m )

[0099]

[0100] where the static distance R0 of the human target is 1.2 m, x b (t m ) is the chest wall micro-motion caused by breathing, A b is the chest wall displacement amplitude caused by breathing, and usually satisfies A b << R0, f is the breathing frequency, is the initial phase.

[0101] Step 1.3: After passing the echo signal through the mixer and low-pass filter inside the millimeter-wave radar, the intermediate frequency signal s IF (t m ,t k ) is obtained. s IF(t m ,t k This can be represented as:

[0102]

[0103] Where A is the amplitude. For beat frequency, For breathing phase modulation, carrier wavelength The value is 5mm. The formula uses "≈" because, in this invention, the quadratic term in the mixing result has been included. Treating these as fixed errors that can be canceled out or calibrated, only the first-order beat frequency term 2πf, which is most sensitive to distance and respiratory micro-movements, is retained. b t k and phase shift

[0104] Step 1.4: Convert the intermediate frequency signal s IF (t m ,t k The baseband signal is multiplied in phase and quadrature with the local oscillator signal, respectively, and then low-pass filtered before being passed through a 90° phase shifter to obtain two orthogonal baseband signals I(t). m ,t k ) and Q(t m ,t k );

[0105]

[0106] Step 1.5: Convert the baseband signal I(t) m ,t k ) and Q(t m ,t k The signals are fed into an analog-to-digital converter separately, and for each Chirp signal, the sampling rate is... For 10 Msps, uniform sampling is performed on M = 512 points, T s The sampling period is used to obtain a discrete sequence. n=1,2,...,N, m=0,1,...,511.

[0107] Step 1.6: Combine the two data streams into a complex baseband signal.

[0108] Step 2: Perform Fourier transform on the fast time dimension of the complex baseband signal to obtain a one-dimensional range image. Select the range cell with the largest average energy in the one-dimensional range image as the range cell where the human chest cavity is located, and extract the continuous phase of the range cell where the human chest cavity is located.

[0109] Step 2 includes the following steps:

[0110] Step 2.1: Perform a 512-point Fourier transform on the fast time dimension of the complex baseband signal z(n,m) to obtain a one-dimensional range profile:

[0111]

[0112] Where Z(n,k) represents the distance unit k in the one-dimensional distance image of the nth Chirp.

[0113] Step 2.2: Enhance the one-dimensional range image using moving target display technology to obtain the enhanced one-dimensional range image:

[0114]

[0115] in, Z(n-1,k) represents the enhanced one-dimensional range image, and Z(n-1,k) represents the one-dimensional range image of the (n-1)th Chirp.

[0116] Step 2.3: Calculate the average energy E(k) of range cell k in the enhanced one-dimensional range image of all Chirps according to the following formula:

[0117]

[0118] Where N is the total number of chirps detected; for 15 minutes of data, N = 18000.

[0119] Step 2.4: Select the distance unit with the highest average energy. As a unit of distance to the human thoracic cavity:

[0120]

[0121] The complex sequence is:

[0122]

[0123] Where Re{·} represents taking the real part of the signal, I z (n) represents the real part of the distance cell, Im{·} denotes taking the imaginary part of the signal, and Q z (n) is the imaginary part of the distance cell.

[0124] Step 2.5: Calculate the phase θ(n) of the distance unit containing the human chest cavity using the arctangent function atan2:

[0125] θ(n)=atan2(Q z (n),I z (n))

[0126] In the formula, I z (n) represents Real part, Qz (n) represents The virtual part;

[0127] Step 2.6: Unwrap the phase θ(n) to obtain the continuous phase x(n):

[0128]

[0129] Δ′=θ(n)-θ(n-1) represents the phase difference before unwinding, and Δ represents the continuous phase difference.

[0130] Step 3: Input the continuous phase of the distance unit where the human body is located into the time-gated attention encoder to obtain the phase after the initial interference suppression.

[0131] The continuous phases are processed as follows before being used as input to the time-gated attention encoder:

[0132] The continuous phase x(n) is sliced ​​using a window with a length of l = 400 points and a step size of s = 20 points. Each slice is considered as a sample, and the total number of samples after slicing is: in Indicates rounding down, N is the total number of chirps in the probe signal, and the total number of samples after 12 hours of data slicing is 42288, with each sample being 400 points long. Every b = 1024 samples from the total sample are grouped into a batch and concatenated into a continuous phase tensor. As input to the time-gated attention encoder.

[0133] The time-gated attention encoder is used to initially suppress mid-to-high frequency interference and aperiodic body micro-motion interference in continuous phase using a time-gated attention mechanism. Figure 2 As shown, the time-gated attention encoder consists of several cascaded time-gated attention blocks; in this embodiment, there are specifically two. Each time-gated attention block includes:

[0134] The input layer is used to calculate the query vector based on the input. key vector and value vector The input to the first time-gated attention block is a continuous phase X, and Q... E =K E =V E =X;

[0135] Convolutional gated networks are used to process query vector Q. E and bond vector K E Convolutional gating is performed using the following formula:

[0136]

[0137] Where LP(·) represents a linear layer, Conv1D(·) represents a one-dimensional convolutional layer, and Sigmoid(·) represents the Sigmoid activation function. and They are Q E With K E The query vector and key vector have undergone convolutional gating;

[0138] The output layer is used to calculate the output of the current time-gated attention block according to the following formula:

[0139]

[0140] Among them, K' E express The transpose of , SoftMax(·) denotes the SoftMax activation function, and l is the continuous phase length; The output of the current time-gated attention block is used as the input to the next time-gated attention block, and the output of the last time-gated attention block is used as the phase after the initial interference suppression.

[0141] Step 4: Use the phase input frequency-gated attention decoder after the initial interference suppression to reconstruct the pure breathing waveform.

[0142] The frequency-gated attention decoder is used to perform secondary suppression on the phase after the initial interference suppression and reconstruct the wavelet coefficients of the breathing waveform. Then, the pure breathing waveform is reconstructed through inverse wavelet transform. The time-gated attention encoder and the frequency-gated attention decoder constitute a gated attention wavelet network.

[0143] like Figure 3 As shown, the frequency-gated attention decoder includes I = 5 parallel sub-decoders and an inverse wavelet transform module connected to all sub-decoders. Figure 4 As shown, the input to each sub-decoder is the phase after initial interference suppression. Each sub-decoder comprises two cascaded frequency-gated attention blocks and a one-dimensional convolutional layer at the last layer. Each frequency-gated attention block includes:

[0144] The input layer is used to calculate the query vector based on the input. key vector and value vector The input to the first frequency-gated attention block is the phase after initial interference suppression. and

[0145] Frequency-gated networks are used to query vector Q. DWith key vector K D Perform frequency gating:

[0146]

[0147] Where FFN(·) represents a feedforward neural network, FFT(·) represents a fast Fourier transform, LP(·) represents a linear layer, and Sigmoid(·) represents the Sigmoid activation function. and These are the query vectors Q D With key vector K D The query vector and key vector have been frequency-gated;

[0148] The output layer is used to calculate the output of the current frequency-gated attention block according to the following formula:

[0149]

[0150] Where IFFT(·) denotes the inverse fast Fourier transform, SoftMax(·) denotes the SoftMax activation function, l is the continuous phase length, and Conj(·) denotes the conjugate operation. This indicates the output of the current frequency-gated attention block.

[0151] The output of the last frequency-gated attention block is fed into a one-dimensional convolutional layer as the input to the next frequency-gated attention block, yielding wavelet coefficients C. i .

[0152] The one-dimensional convolutional layers of the five sub-decoders output respectively

[0153] The inverse wavelet transform module is used to convert... The pure respiratory waveform is reconstructed by inverse wavelet transform of order I-1=4. Figure 5 and Figure 6 The images show the results of respiratory waveform recovery under quasi-static and dynamic conditions, respectively. From top to bottom, the images represent the phase, the true value of the respiratory waveform, and the waveform recovered by the proposed gated attention wavelet network.

[0154] Step 5: Perform Fourier transform and bandpass filtering on the reconstructed pure respiratory waveform, and use the frequency of the peak as the estimated respiratory frequency.

[0155] Step 5 includes the following steps:

[0156] Step 5.1: Perform a 4096-point Fourier transform on the reconstructed pure respiratory waveform Y to obtain the respiratory spectrum.

[0157] Step 5.2: Pass ψ through a 4th-order Butterworth bandpass filter with a cutoff frequency of [0.1, 0.6] to obtain the filtered breathing spectrum;

[0158] Step 5.3: Find the frequency corresponding to the largest peak in the filtered respiratory spectrum. As an estimated respiratory rate.

[0159] Furthermore, the gated attention wavelet network is trained using a preset loss function, with the ground truth values ​​used during training being: the respiratory waveform and frequency ground truth values ​​acquired by the breathing chest strap, respectively. The wavelet coefficients obtained by discrete wavelet transform of the waveform true value are: The spectrum obtained by performing a 4096-point Fast Fourier Transform on the actual respiratory waveform is as follows: The process of establishing the preset loss function is as follows:

[0160] To reduce the difference between the reconstructed respiratory waveform and the true value, mean squared error loss is used.

[0161]

[0162] Among them ||·|| F Let f denote the Frobenius norm, b = 1024, l = 400.

[0163] To enhance the shape similarity between the reconstructed respiratory waveform and the ground truth, cosine similarity loss is used.

[0164] Where <·,·> represents the element-wise inner product.

[0165] The accuracy of the reconstructed waveform can be further improved by reducing the difference between the reconstructed wavelet coefficients and the true values, using mean square error loss.

[0166]

[0167] in Let d1 = 31, d2 = 31, d3 = 56, d4 = 105, d5 = 206 represent the mean square error loss among the i-th wavelet coefficients.

[0168] To enhance the accuracy of frequency estimation, mean square error loss is used. To reduce the spectrum ψ of the reconstructed respiratory waveform from the true value Differences between:

[0169]

[0170] Where p = 4096.

[0171] The total loss function (Loss) consists of the following parts:

[0172]

[0173] Where w1 = 0.1, w2 = 1, w3 = 1, and w4 = 0.01 are the weights used to balance the various losses. The results of the cosine similarity and mean error between the reconstructed respiratory waveform and the true respiratory waveform, as well as the mean absolute error, root mean square error, and standard deviation between the estimated respiratory rate and the true respiratory rate, are shown in Table 2 below:

[0174] Table 2 Error Results

[0175] Evaluation indicators Cosine similarity Mean square error Mean Absolute Error Root mean square error Standard deviation result 0.808 0.076 0.667 0.905 0.857

[0176] Example 2

[0177] This invention also provides a millimeter-wave radar respiratory monitoring device based on a gated attention wavelet network, comprising:

[0178] Millimeter-wave radar is used to send detection signals to human targets and receive echo signals returned by human targets.

[0179] The signal preprocessing module is used to convert the echo signal into two orthogonal baseband signals, and then sample and combine them into a complex baseband signal.

[0180] The phase extraction module is used to perform Fourier transform on the fast time dimension of the complex baseband signal to obtain a one-dimensional range image, select the range cell with the largest average energy in the one-dimensional range image as the range cell where the human chest cavity is located, and extract the continuous phase of the range cell where the human chest cavity is located.

[0181] The interference suppression module is used to input the continuous phase of the distance unit where the human body is located into the time-gated attention encoder to obtain the phase after initial interference suppression. The time-gated attention encoder is used to use the time-gated attention mechanism to initially suppress the mid-to-high frequency interference and non-periodic body micro-motion interference in the continuous phase.

[0182] The breathing waveform reconstruction module is used to input the phase after the initial interference suppression into the frequency-gated attention decoder to reconstruct the pure breathing waveform. The frequency-gated attention decoder is used to perform secondary suppression on the phase after the initial interference suppression and reconstruct the wavelet coefficients of the breathing waveform. Then, the pure breathing waveform is reconstructed through inverse wavelet transform. The time-gated attention encoder and the frequency-gated attention decoder constitute a gated attention wavelet network.

[0183] The respiratory rate estimation module is used to perform Fourier transform and bandpass filtering on the reconstructed clean respiratory waveform, and use the frequency of the peak as the estimated respiratory rate.

[0184] The signal preprocessing module includes:

[0185] The baseband signal acquisition unit is used to acquire the intermediate frequency signal S. IF (t m ,t k The baseband signal is multiplied in phase and quadrature with the local oscillator signal, respectively, and then low-pass filtered before being passed through a 90° phase shifter to obtain two orthogonal baseband signals I(t). m ,t k ) and Q(t m ,t k );

[0186]

[0187] Analog-to-digital converter, used to convert baseband signal I(t) m ,t k ) and Q(t m ,t k Each Chirp signal, at a sampling rate For 10 Msps, uniform sampling is performed on M = 512 points, T s The sampling period is used to obtain a discrete sequence.

[0188] n=1,2,...,N, m=0,1,...,511.

[0189] A synthesizer is used to combine two data streams into a complex baseband signal.

[0190] The phase extraction module includes:

[0191] The Fourier transform unit is used to perform a 512-point Fourier transform on the fast time dimension of the complex baseband signal z(n,m) to obtain a one-dimensional range image:

[0192]

[0193] Where Z(n,k) represents the distance unit k in the one-dimensional distance image of the nth Chirp.

[0194] The enhancement unit is used to enhance a one-dimensional range image using moving target display technology, resulting in an enhanced one-dimensional range image:

[0195]

[0196] in, Z(n-1,k) represents the enhanced one-dimensional range image, and Z(n-1,k) represents the one-dimensional range image of the (n-1)th Chirp.

[0197] An energy calculation unit is used to calculate the average energy E(k) of range cell k in the enhanced one-dimensional range image of all Chirps according to the following formula:

[0198]

[0199] Where N is the total number of chirps detected; for 15 minutes of data, N = 18000.

[0200] Selecting a cell is used to select the distance cell with the highest average energy. As a unit of distance to the human thoracic cavity:

[0201]

[0202] The complex sequence is:

[0203]

[0204]

[0205]

[0206] Where Re{·} represents taking the real part of the signal, I z (n) represents the real part of the distance cell, Im{·} denotes taking the imaginary part of the signal, and Q z (n) is the imaginary part of the distance cell.

[0207] The phase calculation unit is used to calculate the phase θ(n) of the distance unit where the human chest cavity is located using the arctangent function atan2:

[0208] θ(n)=atan2(Q z (n),I z (n))

[0209] In the formula, I z (n) represents Real part, Q z (n) represents The virtual part;

[0210] Unwinding unit, used to unwind phase θ(n) to obtain continuous phase x(n):

[0211]

[0212] Δ′=θ(n)-θ(n-1) represents the phase difference before unwinding, and Δ represents the continuous phase difference.

[0213] The continuous phases are processed as follows before being used as input to the time-gated attention encoder:

[0214] The continuous phase x(n) is sliced ​​using a window with a length of l = 400 points and a step size of s = 20 points. Each slice is considered as a sample, and the total number of samples after slicing is: in Indicates rounding down, N is the total number of chirps in the probe signal, and the total number of samples after 12 hours of data slicing is 42288, with each sample being 400 points long. Every b = 1024 samples from the total sample are grouped into a batch and concatenated into a continuous phase tensor. As input to the time-gated attention encoder.

[0215] The time-gated attention encoder is used to initially suppress mid-to-high frequency interference and aperiodic body micro-motion interference in continuous phase using a time-gated attention mechanism. Figure 2 As shown, the time-gated attention encoder consists of several cascaded time-gated attention blocks; in this embodiment, there are specifically two. Each time-gated attention block includes:

[0216] The input layer is used to calculate the query vector based on the input. key vector and value vector The input to the first time-gated attention block is a continuous phase X, and Q... E =K E =V E =X;

[0217] Convolutional gated networks are used to process query vector Q. E and bond vector K E Convolutional gating is performed using the following formula:

[0218]

[0219]

[0220] Where LP(·) represents a linear layer, Conv1D(·) represents a one-dimensional convolutional layer, and Sigmoid(·) represents the Sigmoid activation function. and They are Q E With K E The query vector and key vector have undergone convolutional gating;

[0221] The output layer is used to calculate the output of the current time-gated attention block according to the following formula:

[0222]

[0223] Among them, K' E express The transpose of , SoftMax(·) denotes the SoftMax activation function, and l is the continuous phase length; The output of the current time-gated attention block is used as the input to the next time-gated attention block, and the output of the last time-gated attention block is used as the phase after the initial interference suppression.

[0224] The frequency-gated attention decoder comprises I=5 parallel sub-decoders and an inverse wavelet transform module connected to all sub-decoders. The input to each sub-decoder is the phase after initial interference suppression. Each sub-decoder comprises two cascaded frequency-gated attention blocks and a one-dimensional convolutional layer at the last layer. Each frequency-gated attention block includes:

[0225] The input layer is used to calculate the query vector based on the input. key vector and value vector The input to the first frequency-gated attention block is the phase after initial interference suppression. and

[0226] Frequency-gated networks are used to query vector Q. D With key vector K D Perform frequency gating:

[0227]

[0228] Where FFN(·) represents a feedforward neural network, FFT(·) represents a fast Fourier transform, LP(·) represents a linear layer, and Sigmoid(·) represents the Sigmoid activation function. and These are the query vectors Q D With key vector K D The query vector and key vector have been frequency-gated;

[0229] The output layer is used to calculate the output of the current frequency-gated attention block according to the following formula:

[0230]

[0231] Where IFFT(·) denotes the inverse fast Fourier transform, SoftMax(·) denotes the SoftMax activation function, l is the continuous phase length, and Conj(·) denotes the conjugate operation. This indicates the output of the current frequency-gated attention block.

[0232] The output of the last frequency-gated attention block is fed into a one-dimensional convolutional layer as the input to the next frequency-gated attention block, yielding wavelet coefficients C. i .

[0233] The one-dimensional convolutional layers of the five sub-decoders output respectively

[0234] The inverse wavelet transform module is used to convert... The pure respiratory waveform is reconstructed by inverse wavelet transform of order I-1=4. Figure 5 and Figure 6 The images show the results of respiratory waveform recovery under quasi-static and dynamic conditions, respectively. From top to bottom, the images represent the phase, the true value of the respiratory waveform, and the waveform recovered by the proposed gated attention wavelet network.

[0235] The respiratory rate estimation module includes:

[0236] The respiratory spectrum calculation unit is used to perform a 4096-point Fourier transform on the reconstructed pure respiratory waveform Y to obtain the respiratory spectrum.

[0237] A bandpass filter is used to pass ψ through a 4th-order Butterworth bandpass filter with a cutoff frequency of [0.1, 0.6] to obtain the filtered breathing spectrum;

[0238] The peak-finding unit is used to find the frequency corresponding to the largest peak in the filtered respiratory spectrum. As an estimated respiratory rate.

[0239] In addition, the device also includes a network training module for training using a preset loss function, which is specifically:

[0240]

[0241]

[0242]

[0243]

[0244]

[0245] Where Loss is the total loss, and w1 to w4 are the weights used to balance the losses of each item. The loss represents the respiratory waveform value, b is the number of consecutive phase samples in each training batch, l is the length of the consecutive phase, and Y, Let ||·|| represent the clean breathing waveform and the true value of the breathing waveform output by the gated attention wavelet network, respectively. F Denotes the Frobenius norm. The symbol represents the shape loss of the respiratory waveform, <, ·> represents the element-wise inner product, and I represents the number of sub-decoders in the frequency-gated attention decoder. C represents the wavelet coefficient loss. i , Let i and n represent the i-th wavelet coefficients of the breathing waveform obtained from the gated attention wavelet network and the true value of the breathing waveform, respectively. The true value of the i-th wavelet coefficient, d i Indicate C i Dimensions Indicates respiratory spectrum loss, ψ, Represent the respiratory spectrum obtained by the gated attention wavelet network and the true value of the respiratory waveform, respectively. The true value of the obtained respiratory spectrum.

[0246] The apparatus provided in this embodiment of the invention can be used to execute the method provided in Embodiment 1 of the invention, and has the corresponding functions and beneficial effects of executing the method.

[0247] It is worth noting that in the embodiments of the above-mentioned device, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of the present invention.

[0248] The embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art will clearly understand that each implementation can be achieved using software plus necessary general-purpose hardware platforms, or it can be implemented solely through hardware, as long as the function or purpose can be achieved.

[0249] 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 millimeter-wave radar respiratory monitoring method based on a gated attention wavelet network, characterized in that, Includes the following steps: Step 1: The millimeter-wave radar sends a detection signal to the human target and receives the echo signal returned by the human target. After converting the echo signal into two orthogonal band signals, it samples and combines them into a complex baseband signal. Step 2: Perform Fourier transform on the fast time dimension of the complex baseband signal to obtain a one-dimensional range image. Select the range cell with the largest average energy in the one-dimensional range image as the range cell where the human chest cavity is located, and extract the continuous phase of the range cell where the human chest cavity is located. Step 3: Input the continuous phase of the distance unit where the human body is located into the time-gated attention encoder to obtain the phase after the initial interference suppression. The time-gated attention encoder is used to suppress the mid-to-high frequency interference and non-periodic body micro-motion interference in the continuous phase using the time-gated attention mechanism. Step 4: Input the phase after the initial interference suppression into the frequency-gated attention decoder to reconstruct the pure breathing waveform. The frequency-gated attention decoder is used to perform secondary suppression on the phase after the initial interference suppression and reconstruct the wavelet coefficients of the breathing waveform. Then, the pure breathing waveform is reconstructed through inverse wavelet transform. The time-gated attention encoder and the frequency-gated attention decoder constitute a gated attention wavelet network. Step 5: Perform Fourier transform and bandpass filtering on the reconstructed pure respiratory waveform, and use the frequency of the peak as the estimated respiratory frequency.

2. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1: The detection signal transmitted by the millimeter-wave radar's transmitting antenna towards the human target is as follows: Among them, s tx (t k A represents the detected signal transmitted within a single linear frequency sweep cycle. t f is the signal amplitude. c The starting carrier frequency is 0 ≤ t k ≤T c Fast time refers to the sampling time within a single linear sweep cycle, B is the modulation bandwidth, and T is the sampling time. c For linear frequency sweep period, Let Re{·} denote the modulation slope, where Re is the real part, and j is the imaginary unit. 2 =-1; Step 1.2: The millimeter-wave radar's receiving antenna receives the echo signal s returned by the human target. rx (t m ,t k ), t m =nT PRF For slow time, The pulse repetition period is denoted by PRF, the pulse repetition frequency is denoted by PRF, one linear sweep frequency period signal is one Chirp signal, n = 1, 2, ..., N represents the index of the transmitted probe signal Chirp, and N is the total number of probe signal Chirps; Step 1.3: After passing the echo signal through the mixer and low-pass filter inside the millimeter-wave radar, the intermediate frequency signal s is obtained. IF (t m ,t k ); Step 1.4: Convert the intermediate frequency signal s IF (t m ,t k The baseband signal is multiplied in phase and quadrature with the local oscillator signal, respectively, and then low-pass filtered before being passed through a 90° phase shifter to obtain two orthogonal baseband signals I(t). m ,t k ) and Q(t m ,t k ); Step 1.5: Convert the baseband signal I(t) m ,t k ) and Q(t m ,t k The signals are fed into an analog-to-digital converter, and each Chirp signal is uniformly sampled to obtain a discrete sequence. n = 1, 2, ..., N, m = 0, 1, ..., M-1, where M is the number of samples for each Chirp signal; Step 1.6: Combine the two data streams into a complex baseband signal.

3. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Perform an M-point Fourier transform on the fast time dimension of the complex baseband signal z(n,m) to obtain a one-dimensional range profile: Where Z(n,k) represents the distance cell k in the one-dimensional distance image of the nth Chirp; Step 2.2: Enhance the one-dimensional range image using moving target display technology to obtain the enhanced one-dimensional range image: in, Z(n-1,k) represents the enhanced one-dimensional range image, and Z(n-1,k) represents the one-dimensional range image of the (n-1)th Chirp. Step 2.3: Calculate the average energy E(k) of range cell k in the enhanced one-dimensional range image of all Chirps according to the following formula: Where N is the total number of Chirp probes; Step 2.4: Select the distance unit with the highest average energy. As a unit of distance to the human thoracic cavity: Step 2.5: Calculate the phase θ(n) of the distance unit containing the human thoracic cavity: θ(n)=attan2(Q z (n),I z (n)) In the formula, I z (n) represents Real part, Q z (n) represents The virtual part; Step 2.6: Unwrap the phase θ(n) to obtain the continuous phase x(n).

4. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 3, characterized in that, The untangling process in step 2.6 specifically involves: Δ′=θ(n)-θ(n-1) represents the phase difference before unwinding, and Δ represents the continuous phase difference.

5. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, The continuous phases in step 3 are processed as follows and then used as the input to the time-gated attention encoder: The continuous phase x(n) is sliced ​​using a window with a length of l and a step size of s. Each slice is considered as a sample, and the total number of samples after slicing is... Each sample has a length of l points, where This indicates rounding down, where N is the total number of Chirp probes. Each b samples in the total sample is grouped into a batch and concatenated into a continuous phase tensor X, which serves as the input to the time-gated attention encoder.

6. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, The time-gated attention encoder is composed of several cascaded time-gated attention blocks, each of which includes: The input layer is used to calculate the query vector Q based on the input. E Key vector K E and value vector V E In this case, the input to the first time-gated attention block is a continuous phase X, and Q... E =K E =V E =X; Convolutional gated networks are used to process query vector Q. E and bond vector K E Convolutional gating is performed using the following formula: Where LP(·) represents a linear layer, Conv1D(·) represents a one-dimensional convolutional layer, and Sigmoid(·) represents the Sigmoid activation function. and They are Q E With K E The query vector and key vector have undergone convolutional gating; The output layer is used to calculate the output of the current time-gated attention block according to the following formula: Among them, K' E express The transpose of , SoftMax(·) denotes the SoftMax activation function, and l is the continuous phase length; The output of the current time-gated attention block is used as the input to the next time-gated attention block, and the output of the last time-gated attention block is used as the phase after the initial interference suppression.

7. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, The frequency-gated attention decoder includes I parallel sub-decoders and an inverse wavelet transform module connected to all sub-decoders. The input to each sub-decoder is the phase after initial interference suppression. Each sub-decoder includes several cascaded frequency-gated attention blocks and a one-dimensional convolutional layer at the last layer. Each frequency-gated attention block includes: The input layer is used to calculate the query vector Q based on the input. D Key vector K D and value vector V D The input to the first frequency-gated attention block is the phase after initial interference suppression. and Frequency-gated networks are used to query vector Q. D With key vector K D Perform frequency gating: Where FFN(·) represents a feedforward neural network, FFT(·) represents a fast Fourier transform, LP(·) represents a linear layer, and Sigmoid(·) represents the Sigmoid activation function. and These are the query vectors Q D With key vector K D The query vector and key vector have been frequency-gated; The output layer is used to calculate the output of the current frequency-gated attention block according to the following formula: Where IFFT(·) denotes the inverse fast Fourier transform, SoftMax(·) denotes the SoftMax activation function, l is the continuous phase length, and Conj(·) denotes the conjugate operation. The output of the current frequency-gated attention block is used as the input to the next frequency-gated attention block. The output of the last frequency-gated attention block is then fed into a one-dimensional convolutional layer to obtain the wavelet coefficients C. i ; The inverse wavelet transform module is used to convert the wavelet coefficients C output by all sub-decoders. i The pure respiratory waveform Y is reconstructed by inverse wavelet transform of order I-1.

8. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, Step 5 includes the following steps: Step 5.1: Perform a Fourier transform on the reconstructed pure respiratory waveform Y to obtain the respiratory spectrum ψ; Step 5.2: Let ψ pass through the cutoff frequency [f l ,f h A J-order Butterworth bandpass filter is used to obtain the filtered breathing spectrum, where f l ,f h These are the lowest and highest frequencies of the filter, respectively. Step 5.3: Take the frequency F corresponding to the largest peak on the filtered respiratory spectrum as the estimated respiratory rate.

9. The millimeter-wave radar respiration monitoring method based on gated attention wavelet networks according to claim 1, characterized in that, The gated attention wavelet network is trained using a preset loss function, which is specifically: Where Loss is the total loss, and w1 to w4 are the weights used to balance the losses of each item. The loss represents the respiratory waveform value, b is the number of consecutive phase samples in each training batch, l is the length of the consecutive phase, and Y, Let ||·|| represent the clean breathing waveform and the true value of the breathing waveform output by the gated attention wavelet network, respectively. F Describing the Frobenius norm, The symbol represents the shape loss of the respiratory waveform, <, ·> represents the element-wise inner product, and I represents the number of sub-decoders in the frequency-gated attention decoder. C represents the wavelet coefficient loss. i , Let i and n represent the i-th wavelet coefficients of the breathing waveform obtained from the gated attention wavelet network and the true value of the breathing waveform, respectively. The true value of the i-th wavelet coefficient, d i C represents i Dimensions Indicates respiratory spectrum loss, ψ, Represent the respiratory spectrum obtained by the gated attention wavelet network and the true value of the respiratory waveform, respectively. The true value of the obtained respiratory spectrum.

10. A millimeter-wave radar respiratory monitoring device based on a gated attention wavelet network, characterized in that, include: Millimeter-wave radar is used to send detection signals to human targets and receive echo signals returned by human targets. The signal preprocessing module is used to convert the echo signal into two orthogonal baseband signals, and then sample and combine them into a complex baseband signal. The phase extraction module is used to perform Fourier transform on the fast time dimension of the complex baseband signal to obtain a one-dimensional range image, select the range cell with the largest average energy in the one-dimensional range image as the range cell where the human chest cavity is located, and extract the continuous phase of the range cell where the human chest cavity is located. The interference suppression module is used to input the continuous phase of the distance unit where the human body is located into the time-gated attention encoder to obtain the phase after initial interference suppression. The time-gated attention encoder is used to use the time-gated attention mechanism to initially suppress the mid-to-high frequency interference and non-periodic body micro-motion interference in the continuous phase. The breathing waveform reconstruction module is used to input the phase after the initial interference suppression into the frequency-gated attention decoder to reconstruct the pure breathing waveform. The frequency-gated attention decoder is used to perform secondary suppression on the phase after the initial interference suppression and reconstruct the wavelet coefficients of the breathing waveform. Then, the pure breathing waveform is reconstructed through inverse wavelet transform. The time-gated attention encoder and the frequency-gated attention decoder constitute a gated attention wavelet network. The respiratory rate estimation module is used to perform Fourier transform and bandpass filtering on the reconstructed clean respiratory waveform, and use the frequency of the peak as the estimated respiratory rate.

Images