An artificial intelligence-based biofeedback device data processing method

Abstract

The application discloses a biofeedback device data processing method based on artificial intelligence and relates to the technical field of biofeedback data processing.The application realizes accurate decoupling of respiratory rhythm and autonomic nervous regulation signals in a nonlinear characteristic space through a latent space orthogonal decomposition mechanism, thereby avoiding the problem of HRV low-frequency band information loss caused by traditional frequency domain filtering;respiratory component estimation based on a variational autoencoder does not need to rely on external sensors, thereby reducing hardware cost and user burden;and through a dynamic orthogonal constraint and online fine-tuning mechanism, individual respiratory mode differences and frequency drifts are adaptively adapted, thereby guaranteeing the stability of long-term analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biofeedback data processing, and particularly relates to a biofeedback device data processing method based on artificial intelligence. BACKGROUND

[0002] Heart rate variability (HRV) is a core indicator for evaluating the function of the autonomic nervous system, and is widely used in the fields of stress monitoring, emotion regulation, and cardiovascular disease warning in biofeedback devices. Traditional HRV analysis relies on frequency domain analysis of RR interval signals, and extracts low frequency (LF, 0.04-0.15 Hz) and high frequency (HF, 0.15-0.4 Hz) components through fast Fourier transform (FFT) or wavelet transform to quantify the balance state of sympathetic and parasympathetic nerves.

[0003] However, the complexity of human physiological signals can lead to interference of respiratory rhythm in HRV analysis. Respiratory movement causes periodic fluctuations in RR intervals through mechanical stretch reflex and central nervous coupling, and the frequency range (0.1-0.3 Hz) overlaps with the LF frequency band of HRV. This spectral confusion makes it difficult for traditional methods to distinguish between autonomic regulation signals and respiratory rhythm coupling noise, reducing the clinical interpretation effectiveness of HRV features. To suppress respiratory interference, some early biofeedback devices use two strategies. One is frequency domain band-pass filtering, which filters out low-frequency respiratory components by setting a fixed cutoff frequency. However, this approach also loses key information in the LF frequency band of HRV, leading to distorted evaluation of autonomic nervous function. The other is multi-sensor fusion, which synchronously collects respiratory signals using an independent respiratory sensor and removes respiratory-related components from heart rate signals using linear regression or blind source separation algorithms. However, the former is difficult to set universal filtering parameters due to individual differences in respiratory frequency, and the latter increases hardware costs and user wear burden, and the linear model cannot analyze the nonlinear dynamic coupling relationship between respiration and HRV.

[0004] However, the influence of respiration on HRV is not simply a spectral superposition or linear modulation, but a non-stationary dynamic process formed through a neural-mechanical feedback loop. For example, in the rapid shallow breathing state of patients with anxiety, the respiratory frequency may temporarily invade the LF frequency band, and traditional FFT analysis may mistakenly attribute the respiratory rhythm to sympathetic nervous hyperactivity, misleading treatment decisions. Although some studies have attempted to introduce adaptive filtering or time-frequency analysis (such as Wigner-Ville distribution), the computational complexity is high and sensitive to noise, making it difficult to meet the real-time requirements of biofeedback devices. Therefore, there is an urgent need to construct a precise decoupling model of the respiratory-HRV coupling relationship from the algorithm level, to achieve sensor-free and low-complexity respiratory interference suppression while preserving the physiological significance of HRV, and to provide a technical foundation for high-precision autonomic nervous function evaluation. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a biofeedback device data processing method based on artificial intelligence to solve the problems of traditional HRV analysis, such as interference of respiratory rhythm coupling, signal distortion caused by frequency domain filtering, increased hardware burden of multi-sensor scheme, and inability of linear model to analyze dynamic nonlinear coupling relationship.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The present application provides a biofeedback device data processing method based on artificial intelligence, which includes,

[0009] Step S1, acquiring electrocardio signals and respiratory signals of a user, wherein the respiratory signals are generated by a chest and abdominal movement sensor or estimated based on periodic fluctuation characteristics from the electrocardio signals;

[0010] Step S2, converting the electrocardio signals into an RR interval sequence and inputting the RR interval sequence into a variational autoencoder for latent space decomposition to obtain a latent space representation containing a respiratory rhythm component and a neural regulation component;

[0011] Step S3, performing orthogonalization processing on the latent space representation to make the respiratory rhythm component and the neural regulation component in the latent space in an orthogonal relationship;

[0012] Step S4, reconstructing an RR interval sequence without respiratory coupling based on the orthogonalized latent space representation;

[0013] Step S5, calculating frequency domain feature parameters according to the reconstructed RR interval sequence and outputting the frequency domain feature parameters to a visualization interface of a biofeedback device.

[0014] As a preferred scheme of the biofeedback device data processing method based on artificial intelligence, the estimation and generation step of the respiratory signal further includes:

[0015] Performing adaptive sliding window Fourier transform on the RR interval sequence to extract amplitude variation characteristics in the 0.1-0.3Hz frequency band;

[0016] Generating a simulated respiratory waveform based on the amplitude variation characteristics as a substitute input of the respiratory signal.

[0017] As a preferred scheme of the biofeedback device data processing method based on artificial intelligence, the latent space decomposition step of the variational autoencoder includes:

[0018] Constructing a latent variable subspace of the respiratory rhythm component and a latent variable subspace of the neural regulation component;

[0019] The distribution of the respiratory rhythm component subspace is constrained to match the time correlation characteristics of the respiratory signal through adversarial training.

[0020] As a preferred scheme of the biofeedback device data processing method based on artificial intelligence, in step S2, the RR interval sequence and the latent space decomposition are generated in the following manner:

[0021] The R-wave peak is detected on the original electrocardiogram signal x(t) using the Pan-Tompkins algorithm to obtain the peak time sequence And calculate the RR interval: Wherein, represents the (n+1)th R-wave peak time, represents the nth R-wave peak time, and RR n represents the nth RR interval.

[0022] The {RR n} is standardized to construct a column vector RR = [RR1, RR2, …, RR N ] T , wherein RR represents the input RR interval sequence column vector, and N represents the sequence length; and the same is sent to a variational autoencoder VAE encoder to obtain a latent space representation: Wherein, z ∈ R d represents the latent space overall vector, represents the encoder mapping, and d

[0023] z is divided into a respiratory rhythm component subspace and a neuromodulation component subspace according to the dimension:

[0024]

[0025] Wherein, z r represents the respiratory rhythm subspace vector, and the dimension is d r , a n represents the neuromodulation subspace vector, and the dimension is d n , and d represents the total dimension of the latent space, and R herein represents the real set.

[0026] Adversarial training is introduced, and the adversarial loss is defined as:

[0027]

[0028] Wherein, L adv represents the adversarial training loss, represents the expectation of the real respiratory feature sample x r , and A represents the expectation of the encoder output respiratory vector z r .φ (·) represents the adversarial discriminator mapping, φ is the discriminator parameter, x r represents the representation of the real respiratory signal in the feature domain, z r represents the respiratory subspace vector generated by the encoder;

[0029] The encoder and the discriminator are aligned through:

[0030] the adversarial process of z r , so that the distribution of z r is gradually aligned with the distribution of x

[0031] As a preferred scheme of the biofeedback device data processing method based on artificial intelligence provided by the application, the orthogonalization processing step comprises:

[0032] Calculate the principal component direction of the respiratory rhythm component in the latent space;

[0033] Apply an orthogonal projection constraint to the neuromodulation component, so that the included angle between the neuromodulation component and the principal component direction is not less than 80 degrees.

[0034] As a preferred scheme of the biofeedback device data processing method based on artificial intelligence provided by the application, in step S3, the step of latent space orthogonalization processing is:

[0035] Calculate the covariance matrix of the respiratory rhythm subspace vector set, formula: C r =Cov(z r ), wherein, represents the covariance matrix of the respiratory subspace vector z r , and Cov(·) is a covariance operator;

[0036] On this basis, by solving principal component analysis, the principal direction vector u of the respiratory subspace is obtained:

[0037]

[0038] wherein, represents the principal component direction vector, and |v| represents the Euclidean norm of the vector v;

[0039] Orthogonal projection is performed on the neuromodulation component vector to obtain the orthogonalized vector:

[0040] z' n =z n -(u T z n )u, wherein, represents the orthogonalized neuromodulation vector, u T z n represents the inner product of the vector.

[0041] Define the orthogonal penalty term:

[0042] L orth =max(0,|u T z n |-|z n |cos(80°)),

[0043] Among them, L orth The loss represents the orthogonality penalty, and max(·) is the maximum value operator. |z n | represents the Euclidean norm of the neural regulation vector, and cos(80°) represents the cosine value of 80 degrees;

[0044] This is incorporated into the overall training loss to suppress overlap between subspaces.

[0045] In a preferred embodiment of the data processing method for a biofeedback device based on artificial intelligence as described in this invention, in step S3, applying orthogonal projection constraints to the neural regulatory components specifically involves:

[0046] Define the angle between vectors as:

[0047]

[0048] in, This represents the principal component direction vector of the breathing subspace. Let |·| denote the neural regulatory component vector, |·| denote the Euclidean norm, arccos(·) is the inverse cosine function, and θ is the angle between the vectors;

[0049] The geometric constraint condition, written as θ≥80°, is equivalent to:

[0050] Where cos(80°) represents the cosine value of 80°;

[0051] In actual training, soft constraint penalty terms are used:

[0052]

[0053] Among them, L orth For orthogonality penalty loss, max(·) is the maximum value operator;

[0054] The threshold setting logic includes: In HRV analysis, 80° corresponds to a cosine value of approximately 0.1736, which ensures high separation of subspaces without excessively weakening neural features; too small a cosine threshold will lead to unstable training, while too large a threshold will result in insufficient separation. 80° is the balance point between performance and robustness in experience; 80° can be used as the initial value for different individuals or application scenarios and fine-tuned according to the performance of the validation set.

[0055] In a preferred embodiment of the data processing method for a biofeedback device based on artificial intelligence as described in this invention, the calculation steps of the frequency domain characteristic parameters include:

[0056] Frequency domain correction was performed on the reconstructed RR interval sequence, specifically by adjusting the power density value of the overlapping region between the dynamic suppression and respiration estimation frequency bands within the LF band (0.04-0.15Hz).

[0057] As a preferred embodiment of the data processing method for a biofeedback device based on artificial intelligence described in this invention, the method further includes a dynamic adjustment step:

[0058] The frequency drift of the respiratory signal is monitored, and when the frequency drift exceeds a preset threshold, the parameters of the variational autoencoder are fine-tuned online.

[0059] The online fine-tuning employs an unsupervised loss function, updating model weights solely based on the latent space orthogonality constraints of the current input signal.

[0060] As a preferred embodiment of the data processing method for a biofeedback device based on artificial intelligence described in this invention, the method is deployed on an edge computing device and specifically includes:

[0061] The variational autoencoder and orthogonalization steps are compiled into a lightweight inference engine; the lightweight inference engine performs model compression through structured pruning and 8-bit integer quantization.

[0062] The RR interval sequence is streamed using memory mapping technology; the memory mapping technology is based on the DMA controller of the target edge device for zero-copy data transmission.

[0063] The beneficial effects of this invention are as follows: This invention achieves precise decoupling of respiratory rhythm and autonomic nervous system regulation signals within a nonlinear feature space through a latent space orthogonal decomposition mechanism, avoiding the loss of low-frequency HRV information caused by traditional frequency domain filtering; respiratory component estimation based on variational autoencoders does not rely on external sensors, reducing hardware costs and user burden; simultaneously, through dynamic orthogonal constraints and online fine-tuning mechanisms, it adapts to individual respiratory pattern differences and frequency drift, ensuring the stability of long-term analysis; lightweight compilation and streaming processing technologies in edge deployment enable complex models to run in real time on low-power devices, overcoming the computational redundancy of traditional adaptive algorithms; and while preserving the physiological significance of HRV, it significantly improves the robustness and system applicability of respiratory interference suppression. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a schematic diagram of the biofeedback data processing flow in Example 1. Detailed Implementation

[0066] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0067] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0068] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0069] Example 1, referring to Figure 1 This embodiment provides an artificial intelligence-based biofeedback data processing method, including:

[0070] Step S1: Acquire the user's electrocardiogram (ECG) signal and respiratory signal, wherein the respiratory signal is generated by a chest and abdominal motion sensor or estimated from the ECG signal based on periodic fluctuation characteristics;

[0071] The steps for estimating and generating respiratory signals further include:

[0072] An adaptive sliding window Fourier transform was performed on the RR interval sequence to extract the amplitude variation characteristics in the 0.1-0.3Hz frequency band;

[0073] Simulated respiratory waveforms are generated based on amplitude variation characteristics, serving as alternative inputs for respiratory signals;

[0074] Step S2: Convert the electrocardiogram signal into an RR interval sequence and input it into a variational autoencoder for latent space decomposition to obtain a latent space representation containing respiratory rhythm components and neural regulation components.

[0075] The latent space decomposition steps of a variational autoencoder include:

[0076] Construct the latent variable subspaces of the respiratory rhythm component and the neural regulation component;

[0077] The distribution of respiratory rhythm subspaces is matched with the temporal correlation characteristics of respiratory signals through adversarial training.

[0078] In step S2, the method for generating the RR interval sequence and latent space decomposition is as follows:

[0079] The Pan-Tompkins algorithm is used to detect R-wave peaks in the raw electrocardiogram signal x(t) to obtain the peak time sequence. And calculate the RR interval: in, This represents the time of the (n+1)th R-wave peak. Represents the nth R-wave peak time, RR n This represents the nth RR interval;

[0080] {RR n After standardization, construct a column vector RR = [RR1, RR2, ..., RR2]. N ] T Where RR represents the input RR interval sequence column vector, and N represents the sequence length; it is fed into the variational autoencoder (VAE) to obtain the latent space representation: Where, z∈R d Represents the total vector of the latent space. Indicates encoder mapping, Its parameters;

[0081] z is divided into a respiratory rhythm component subspace and a neural regulation component subspace according to its dimensions:

[0082]

[0083] Among them, z r This represents a vector of the respiratory rhythm subspace, with dimension d. r , z n This represents a neural regulation subspace vector with dimension d. n d represents the total dimension of the latent space, and R here represents the set of real numbers, which is different from RR;

[0084] Introducing adversarial training, the adversarial loss is defined as:

[0085]

[0086] Among them, L adv Indicates the loss during adversarial training. This represents the actual respiratory feature sample x. r Expectations This represents the encoder output breathing vector z. r A's expectations φ (·) denotes the adversarial discriminator mapping, where φ is the discriminator parameter, and x r z represents the representation of the actual respiratory signal in the feature domain. r This represents the breathing subspace vector generated by the encoder;

[0087] The encoder and discriminator pass through:

[0088] The confrontation process makes z r The distribution of x is gradually aligned r Distribution;

[0089] Step S3: Orthogonalize the latent space representation so that the respiratory rhythm component and the neural regulation component are orthogonal in the latent space.

[0090] The orthogonalization process includes:

[0091] Calculate the principal component orientations of the respiratory rhythm components within the latent space;

[0092] Apply orthogonal projection constraints to the neural regulatory component so that the angle between it and the principal component direction is not less than 80 degrees.

[0093] In step S3, the latent space orthogonalization process is as follows:

[0094] The covariance matrix of the respiratory rhythm subspace vector set is calculated using the formula: C r =Cov(z) r ),in, Represents the breathing subspace vector z r The covariance matrix, where Cov(·) is the covariance operator;

[0095] Based on this, the principal direction vector u of the breathing subspace is obtained by solving the principal component analysis:

[0096]

[0097] in, Let |v| denote the direction vector of the principal component, and |v| denote the Euclidean norm of vector v.

[0098] Orthogonally projecting the neural regulatory component vectors yields the orthogonalized vectors:

[0099] z' n =z n -(u T z n u, where, U represents the orthogonalized neural regulation vector. T zn Represents the dot product of vectors;

[0100] Define the orthogonal penalty term:

[0101] L orth =max(0,|u T z n |-|z n |cos(80°)),

[0102] Among them, L orth The loss represents the orthogonality penalty, and max(·) is the maximum value operator. |z n | represents the Euclidean norm of the neural regulation vector, and cos(80°) represents the cosine value of 80 degrees;

[0103] This is incorporated into the overall training loss to suppress overlap between subspaces;

[0104] In step S3, orthogonal projection constraints are applied to the neural regulatory components, specifically as follows:

[0105] Define the angle between vectors as:

[0106]

[0107] in, This represents the principal component direction vector of the breathing subspace. Let |·| denote the neural regulatory component vector, |·| denote the Euclidean norm, arccos(·) is the inverse cosine function, and θ is the angle between the vectors;

[0108] The geometric constraint condition, written as θ≥80°, is equivalent to:

[0109] Where cos(80°) represents the cosine value of 80°;

[0110] In actual training, soft constraint penalty terms are used:

[0111]

[0112] Among them, L orth For orthogonality penalty loss, max(·) is the maximum value operator;

[0113] The threshold setting logic includes: In HRV analysis, 80° corresponds to a cosine value of approximately 0.1736, which ensures high separation of subspaces without excessively weakening neural features; too small a cosine threshold will lead to unstable training, while too large a threshold will result in insufficient separation. 80° is the balance point between performance and robustness in experience; 80° can be used as the initial value for different individuals or application scenarios and fine-tuned according to the performance of the validation set.

[0114] Specifically, this geometric constraint directly limits the minimum angle between the neural regulatory component and the main respiratory direction in the latent space. The upper bound of the cosine value ensures that the two have significant separation in the high-dimensional space. Compared with the method of pure projection followed by penalty, the soft constraint introduces an angle threshold, making the filtering effect between subspaces more interpretable. Moreover, due to the use of a continuous penalty function, the loss is smoothed during training, which improves the convergence stability of the model. The 80° threshold not only meets the decoupling requirement of HRV signal, but also retains the subtle fluctuation information of neural regulation, providing a purer and physiologically meaningful RR interval reconstruction result for subsequent frequency domain feature extraction.

[0115] Step S4: Based on the orthogonalized latent space representation, reconstruct the RR interval sequence without non-respiratory coupling;

[0116] Step S5: Calculate the frequency domain feature parameters based on the reconstructed RR interval sequence and output them to the visualization interface of the biofeedback device;

[0117] The steps for calculating frequency domain characteristic parameters include:

[0118] Frequency domain correction was performed on the reconstructed RR interval sequence, specifically: the power density value of the overlapping region between the dynamic suppression and respiration estimation frequency bands in the LF band (0.04-0.15Hz).

[0119] In step S5, the frequency domain correction of the reconstructed RR interval sequence is performed as follows:

[0120] The power spectrum is estimated using the formula: P(f)=Welch(RR')(f), where RR' represents the reconstructed RR interval column vector, P(f) represents its power spectral density function, Welch(·) represents the Welch spectrum estimation operator using the overlapping window and averaging method, and f represents the frequency;

[0121] Define the low-frequency band and the breathing overlap region [f LF,low ,f LF,high ]:

[0122] [f LF,low ,f LF,high ] = [0.04, 0.15] Hz,

[0123] F ovl =[f r -δ,f r +δ]∩[f LF,low ,f LF,high ],

[0124] Among them, f LF,low and f LF,high f represents the lower and upper limits of the low-frequency band, respectively. rThis represents the current estimated respiratory rate, and δ represents the width of the respiratory rate drift half-window.

[0125] Construct a suppression mask function, expressed as:

[0126] If f∈F ovl If H(f) = 0, then H(f) = 0; otherwise H(f) = 1.

[0127] Where H(f) is the frequency suppression mask function, when the frequency falls within the overlap region F between the breathing and low frequency bands. ovl If the power is completely suppressed, otherwise the original power is retained;

[0128] Power spectrum correction is performed using the following formula: P corr (f) = H(f)P(f), where P corr (f) is the corrected power spectral density function;

[0129] Calculate the low-frequency band correction power P LF :

[0130]

[0131] Among them, P LF This represents the low-frequency power used for biofeedback after calibration;

[0132] Suppression rules and frequency band selection logic include:

[0133] The low-frequency band [0.04, 0.15] Hz can reflect the balance between the sympathetic and subsympathetic inductance, therefore it is used as the target power calculation range.

[0134] The respiratory frequency range is typically between 0.1 and 0.3 Hz, and fluctuates with respiratory depth and rate. This is achieved by estimating the respiratory rate f in real time. r Its drift half-window width δ can be set as the standard deviation of the respiratory rate sliding window. Half of, or an empirical value of 0.01Hz, dynamically constructs the overlap region F. ovl This ensures that the inhibition rules can adapt to individual respiratory changes;

[0135] A hard threshold mask is used to zero out the power, avoiding the influence of residual breathing peaks on LF power. This method preserves non-breathing components while ensuring computational simplicity and real-time performance.

[0136] Specifically, the above mechanism ensures that all power peaks that may be caused by respiratory coupling are dynamically eliminated in the low-frequency range, thereby providing a purer neural regulation indicator for subsequent biofeedback.

[0137] The data processing method for biofeedback devices further includes a dynamic adjustment step:

[0138] Monitor the frequency drift of the respiratory signal, and when the frequency drift exceeds a preset threshold, trigger online fine-tuning of the variational autoencoder parameters;

[0139] Online fine-tuning employs an unsupervised loss function, updating model weights solely based on the latent space orthogonality constraints of the current input signal;

[0140] Biofeedback device data processing methods are deployed on edge computing devices, specifically including:

[0141] The variational autoencoder and orthogonalization steps are compiled into a lightweight inference engine; the lightweight inference engine compresses the model through structured pruning and 8-bit integer quantization.

[0142] The RR interval sequence is streamed using memory mapping technology; the memory mapping technology performs zero-copy data transfer based on the DMA controller of the target edge device.

[0143] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A data processing method for a biofeedback device based on artificial intelligence, characterized in that: include, Step S1: Acquire the user's electrocardiogram (ECG) signal and respiratory signal, wherein the respiratory signal is generated by a chest and abdominal motion sensor or estimated from the ECG signal based on periodic fluctuation characteristics; Step S2: The electrocardiogram signal is converted into an RR interval sequence and input into a variational autoencoder for latent space decomposition to obtain a latent space representation containing respiratory rhythm components and neural regulation components. Step S3: Orthogonalize the latent space representation so that the respiratory rhythm component and the neural regulation component are orthogonal in the latent space; Step S4: Based on the orthogonalized latent space representation, reconstruct the RR interval sequence without non-respiratory coupling; Step S5: Calculate the frequency domain feature parameters based on the reconstructed RR interval sequence and output them to the visualization interface of the biofeedback device; The orthogonalization process includes: Calculate the principal component orientations of the respiratory rhythm components within the latent space; An orthogonal projection constraint is applied to the neural regulatory component so that the angle between it and the direction of the principal component is not less than 80 degrees. In step S3, the latent space orthogonalization process is as follows: The covariance matrix of the respiratory rhythm subspace vector set is calculated using the following formula: ,in, Represents the respiratory subspace vector The covariance matrix, For covariance operators; Based on this, the principal direction vector of the breathing subspace is obtained by solving the principal component analysis. : ， in, Represents the direction vector of the principal component. Representing vectors The Euclidean norm; Orthogonally projecting the neural regulatory component vectors yields the orthogonalized vectors: ,in, This represents the orthogonalized neural regulation vector. Represents the dot product of vectors; Define the orthogonal penalty term: ， in, Indicates orthogonality penalty loss. To find the maximum value operator, The Euclidean norm of the neural regulation vector. It represents the cosine value of 80 degrees.

2. The data processing method for a biofeedback device based on artificial intelligence as described in claim 1, characterized in that: The step of estimating and generating the respiratory signal further includes: An adaptive sliding window Fourier transform was performed on the RR interval sequence to extract the amplitude variation characteristics in the 0.1-0.3Hz frequency band; A simulated breathing waveform is generated based on the amplitude variation characteristics, serving as a substitute input for the breathing signal.

3. The data processing method for a biofeedback device based on artificial intelligence as described in claim 1, characterized in that: The latent space decomposition step of the variational autoencoder includes: Construct the latent variable subspaces of the respiratory rhythm component and the neural regulation component; The distribution of the respiratory rhythm subspace is constrained to match the temporal correlation characteristics of the respiratory signal through adversarial training.

4. The data processing method for a biofeedback device based on artificial intelligence as described in claim 3, characterized in that, In step S2, the method for generating the RR interval sequence and latent space decomposition is as follows: Using the Pan-Tompkins algorithm on raw electrocardiogram signals The peak of the R-wave was detected to obtain the peak time sequence. And calculate the RR interval: ,in, Indicates the first Second R-wave peak Indicates the first Second R-wave peak Indicates the first One RR interval; Will Construct column vectors after standardization ,in, This represents the column vector of the input RR interval series. The sequence length is represented; it is then fed into a variational autoencoder (VAE) to obtain the latent space representation: ,in, Represents the total vector of the latent space. Indicates encoder mapping, Its parameters; Will Based on dimensions, it is divided into a respiratory rhythm subspace and a neural regulation subspace: ， in, Represents the respiratory rhythm subspace vector, with dimension . Represents the neural regulation subspace vector, with dimension . This represents the total dimension of the latent space. Represents the set of real numbers; Introducing adversarial training, the adversarial loss is defined as: ， in, Indicates the loss during adversarial training. This represents a sample of real respiratory features. Expectations This represents the encoder output breathing vector. Expectations This indicates an adversarial discriminator mapping. For discriminator parameters, This represents the representation of the actual respiratory signal in the feature domain. This represents the breathing subspace vector generated by the encoder; The encoder and discriminator pass through: The confrontation process made The distribution gradually aligns The distribution of .

5. The data processing method for a biofeedback device based on artificial intelligence as described in claim 1, characterized in that, In step S3, orthogonal projection constraints are applied to the neural regulatory components, specifically as follows: Define the angle between vectors as: ， in, This represents the principal component direction vector of the breathing subspace. Represents the neural regulatory component vector. Denotes the Euclidean norm. It is an inverse cosine function. The angle between vectors; Geometric constraints are written as , equivalent to: ,in, Represents the cosine value of 80°; In actual training, soft constraint penalty terms are used: ， in, For orthogonality penalty loss, Maximum value operator.

6. The data processing method for a biofeedback device based on artificial intelligence as described in claim 1, characterized in that, The calculation steps for the frequency domain characteristic parameters include: Frequency domain correction was performed on the reconstructed RR interval sequence, specifically by adjusting the power density value of the overlapping region between the dynamic suppression and respiration estimation frequency bands within the LF band (0.04-0.15Hz).

7. The data processing method for a biofeedback device based on artificial intelligence as described in claim 6, characterized in that, Further, dynamic adjustment steps are included: The frequency drift of the respiratory signal is monitored, and when the frequency drift exceeds a preset threshold, the parameters of the variational autoencoder are fine-tuned online. The online fine-tuning employs an unsupervised loss function, updating model weights solely based on the latent space orthogonality constraints of the current input signal.

8. The data processing method for a biofeedback device based on artificial intelligence as described in claim 7, characterized in that, This method is deployed on edge computing devices and specifically includes: The variational autoencoder and orthogonalization steps are compiled into a lightweight inference engine; the lightweight inference engine performs model compression through structured pruning and 8-bit integer quantization. The RR interval sequence is streamed using memory mapping technology; the memory mapping technology is based on the DMA controller of the target edge device for zero-copy data transmission.

Images