Respiration monitoring analysis and evaluation method based on optical fiber sensor

Abstract

The invention discloses a respiration monitoring analysis and evaluation method based on an optical fiber sensor, and relates to the field of respiration monitoring analysis and evaluation.The respiration monitoring analysis and evaluation method comprises the steps that respiration signals and heart rate signals are collected, and through time alignment, motion compensation and reconstruction analysis, a respiration rate space-time coupling tensor is obtained, a phase difference matrix is constructed to generate an atlas; after decomposition of a Gaussian mixture model, calculating mutual information entropy and a coupling strength index, performing analysis to obtain a mutual information entropy matrix and a dynamic coupling strength curve, performing graph attention network processing, state reproduction and phase synchronization analysis to obtain a dynamic multi-domain joint tensor, and performing projection, reconstruction and characteristic decomposition to obtain parameters. The method comprises the steps of obtaining a mode density manifold graph and a dynamic physiological mode graph, constructing a Gaussian mixture-manifold embedding model, obtaining the Gaussian probability and two-dimensional manifold space coordinates of an abnormal prototype, carrying out corresponding early warning and signal emission, describing a complex physiological mode for respiration and heart rate signals, and carrying out accurate respiration state evaluation and abnormal early warning.

Description

Technical Field

[0001] The present invention relates to the field of respiratory monitoring, analysis and evaluation, and in particular to a respiratory monitoring, analysis and evaluation method based on an optical fiber sensor. Background Art

[0002] With the increasing demand for precise physiological status monitoring in the healthcare sector, respiratory monitoring, as a key means of assessing human physiological health, is crucial for its accuracy and real-time performance. Traditional respiratory monitoring methods often face challenges when processing multi-source signals (such as respiratory and heart rate signals), such as insufficient time alignment accuracy and incomplete motion interference elimination, leading to data distortion. Furthermore, existing analysis and evaluation methods have limited capabilities for analyzing the dynamic coupling characteristics between respiratory and heart rate signals and for fusion analysis of multimodal information, making it difficult to fully characterize complex physiological patterns. Furthermore, there is a lack of in-depth exploration of the spatiotemporal evolution and chaotic properties of physiological signals, making it impossible to accurately capture early signs of respiratory abnormalities. Therefore, there is an urgent need for a method that can effectively process multi-source signals, fuse spatiotemporal dynamic characteristics with multimodal information, and achieve accurate respiratory status assessment and abnormality warning to meet the actual needs of clinical and health monitoring.

[0003] In order to solve the above-mentioned defects, a technical solution is now provided. Summary of the Invention

[0004] In order to solve the technical problems raised by the above background technology, the present invention is proposed. An embodiment of the present invention provides a respiratory monitoring, analysis and evaluation method based on an optical fiber sensor.

[0005] The purpose of the present invention can be achieved by the following technical solution: A respiratory monitoring analysis and evaluation method based on an optical fiber sensor comprises the following steps: Step S100: collecting respiratory signals and heart rate signals, performing time alignment, motion compensation, and reconstruction analysis to obtain a spatiotemporal coupling tensor of respiratory rate; Step S200: constructing a phase difference matrix based on the spatiotemporal coupling tensor of the call rate to generate a spectrum, calculating the mutual information entropy and the coupling strength index after decomposition using a Gaussian mixture model, and analyzing to obtain the mutual information entropy matrix and the dynamic coupling strength curve; Step S300: Based on the mutual information entropy matrix and the dynamic coupling strength curve, the graph attention network is processed to perform state reproduction and phase synchronization analysis to obtain a dynamic multi-domain joint tensor; Step S400: Based on the dynamic multi-domain joint tensor expansion decomposition, parameters are obtained through projection, reconstruction, and eigendecomposition to obtain a pattern density manifold map and a dynamic physiological pattern atlas; Step S500: Construct a hybrid Gaussian-manifold embedding model based on the pattern density manifold map and the dynamic physiological pattern map, obtain the Gaussian probability and two-dimensional manifold space coordinates of the abnormal prototype, and perform corresponding warnings and signal issuance.

[0006] Furthermore, the mutual information entropy matrix and dynamic coupling strength curve analysis steps are: The dynamic phase difference matrix is ​​constructed by extracting the instantaneous phase of the respiratory signal H(t) and the heart rate signal R(t) in the spatiotemporal coupling tensor of the breathing rate through Hilbert transform. , divide the phase difference range into N equal-width intervals, count the frequency of phase difference in each interval, generate a histogram, and use Gaussian kernel function to smooth the phase difference distribution to obtain the probability density , with phase difference As the horizontal axis, the probability density As the vertical axis, a transient phase shift mapping spectrum is established; The instantaneous phase difference distribution map is decomposed into K Gaussian components N through the multimodal Gaussian mixture model. The mean μk of each component is used as the center, and the symbol interval is defined by expanding ɑσk. ɑ represents the confidence interval coefficient. The overlapping intervals are merged to obtain the merged non-redundant symbol interval Sm. The merged non-redundant symbol interval Sm is assigned a unique symbol. The continuous phase difference Mapping to phase symbol sequence and ; Based on phase symbol sequence and The mutual information entropy of respiration and heart rate was calculated using symbolic dynamics to obtain the mutual information entropy matrix I(i, j), where i and j represent the row and column indices of the matrix. The coupling strength index was also calculated, and the dynamic coupling strength curve CSI(t) was obtained with time as the horizontal axis and the coupling strength index as the vertical axis.

[0007] Furthermore, the dynamic multi-domain joint tensor analysis step: Obtain the node betweenness centrality values ​​of the physiological state variable coupling network E(t), and obtain the top K nodes by sorting the node betweenness centrality values ​​Cbetw(k) to obtain a list of key hub nodes. Use the edge weight mean of the physiological state variable coupling network and the node betweenness centrality values ​​of the key hub node list to embed the key hub node list into a graph attention network to obtain a set of hub node feature condensed vectors. The hub node feature condensed vector set is used to establish the state recurrence association matrix Kij through a step function. The sum of all elements in the state recurrence association matrix Kij is divided by the square of the number of hub node feature condensed vector sets to obtain the state recurrence stability rate RL. Map each element of the state recurrence association matrix Kij to a color, and use the state recurrence stability rate RL to control the red depth in the matrix through a linear mapping relationship. Use data visualization tools to arrange the matrix in rows and columns to generate a two-dimensional color block diagram. The horizontal and vertical axes represent the node index respectively. The color of the color block intuitively reflects the state similarity and stability between nodes, and the state recurrence heat map W is obtained; Two-dimensional dynamic phase difference matrix A new empty dimension is added to the dimension to obtain the phase synchronization basis core tensor T, and the two-dimensional state current mapping heat map W is element-wise multiplied with the phase synchronization basis core tensor Tfusi to obtain the space-time recursive fusion tensor Tfusi. Then, the one-dimensional chaotic periodic phase change deterministic indicator DET is integrated through the outer product to obtain the dynamic multi-domain joint tensor Tjoi.

[0008] Furthermore, the chaotic periodic phase transition deterministic indicator DET analysis steps are: Define the diagonal direction, main diagonal and other parallel diagonals in the state recurrence association matrix, scan point by point along each diagonal direction, detect the line segments with consecutive values ​​of 1, if L consecutive points starting from position (i, j) are 1, then record the length L, and obtain the detected line segment length set {L1, v2, ..., LN}. Count the number of times each length L appears and divide it by the total number of appearances to obtain the probability distribution P(L) of the diagonal length L, and perform deterministic formula analysis to obtain the deterministic index of chaotic periodic phase transition. , where Lmin is the minimum effective diagonal length and K represents the number of key hub node lists.

[0009] Furthermore, the pattern density manifold diagram and dynamic physiological pattern atlas analysis steps are: Based on the dominant oscillation frequency wk, mode energy ratio Ek and dynamic calibration desynchronization index DI, the triple [wk, Ek, DI] is formed. The dynamic prototype of the neural gas algorithm It is composed of a triplet, and the neural gas algorithm dynamic prototype is updated through the neural gas algorithm iteration, and the prototype is dynamically adjusted location; Data H without respiratory disease records were obtained from an open-source physiological database. The eigenvector matrix and the principal component latent matrix of the healthy data were obtained through covariance matrix decomposition and principal component analysis. Based on the principal component latent matrix of health data, a basic circadian rhythm template is constructed to form the basic pattern layer of the hierarchical map; The abnormal pattern layer marks the pathological prototype based on the dynamic threshold, and the abnormal pattern layer and the basic pattern layer form a dynamic physiological pattern map; The pattern density manifold is plotted by Algorithms will be Neural Gas Algorithm Dynamic Prototype Mapping to two-dimensional space ,in Represents the coordinates after the two-dimensional manifold space; by connecting the manifold coordinates ym(t) of adjacent time windows, the pattern density manifold map is obtained.

[0010] Furthermore, the dominant oscillation frequency wk, mode energy ratio Ek and dynamically calibrated desynchronization index DI analysis steps are: The dynamic multi-domain joint tensor is first expanded into a time dimension expansion matrix according to the four dimensions of time, frequency band, state recurrence and chaos. , frequency band dimension expansion matrix , Status Dimension Expansion Matrix and Chaos Dimension Expansion Matrix , perform singular value decomposition on each matrix and retain the first ri principal components, and retain the left singular matrix obtained by decomposition , marked as the dimension feature parsing matrix And perform projection operation on the dynamic multi-domain joint tensor Tjoi and the decomposed left singular matrix to obtain the compressed interactive core tensor G; The compressed interaction core tensor G and the projection of the dimensional feature analysis matrix time factor and chaos factor are reconstructed respectively to obtain the spatiotemporal coupling state matrix X and the time sequence evolution state matrix , using the spatiotemporal coupling state matrix X and the temporal evolution state matrix Linear dynamic mode extraction and chaotic characteristic injection are performed to obtain the chaotic weighted modal matrix C. The chaotic weighted modal matrix C is subjected to eigenvalue decomposition to obtain the dynamic coupling eigenvalue ηk and the multimodal rhythm eigenvector δk. The dynamic coupling eigenvalue ηk is subjected to logarithmic operation and imaginary part extraction analysis to obtain the dominant oscillation frequency wk. The contribution degree of the multimodal rhythm eigenvector δk in the coupling is analyzed to obtain the mode energy ratio Ek. The mode energy ratio and the chaotic periodic phase change index DET are dynamically calibrated to the desynchronization index calculation formula to obtain the dynamically calibrated desynchronization index DI.

[0011] Furthermore, the corresponding warning and signal issuing analysis steps: The circadian rhythm template in the basic pattern layer of the dynamic physiological pattern atlas divides a day into T periods and generates multiple Gaussian prototypes through incremental clustering. Each prototype consists of a mean vector μk and a covariance matrix Σk. , where |Ck| represents the number of samples in the dataset Ck belonging to prototype k, xi represents the i-th data point in the dataset, and the abnormal Gaussian prototype is generated by spectral clustering based on the pathological prototype in the abnormal pattern layer. , , where k represents the prototype index, and Represents the covariance matrix of the mean and kth abnormal prototype, cov represents the function of calculating covariance, Represents the data set of the kth abnormal prototype, mean represents the calculation of the arithmetic mean of a set of data; The objective function introduces a prototype position constraint term into the pattern density manifold graph to ensure that the normal prototype is centered in the manifold space and the abnormal prototype is distributed at the edge, thus obtaining the manifold parameter θ. Establish and train a mixed Gaussian-manifold embedding model, process the real-time collected respiratory and heart rate signals, and then calculate and input them into the model to obtain the Gaussian probability of belonging to the abnormal prototype. , and send signals 2 and 1 accordingly; use kernel density estimation to fit the manifold coordinate set {ynormal}, calculate the probability density Pnormal(y) of the normal area, define the boundary threshold δ0 of the normal area according to the confidence interval of normal data, process the real-time respiratory signal and heart rate signal, and map them to the two-dimensional manifold space through the t-SNE manifold term in the model to obtain the current coordinate yt, slide the time window T0 time point, and count the number of deviations within the window , and send out signal two and signal one accordingly; give corresponding warnings and signals for signal one and two.

[0012] Furthermore, the mixed Gaussian-manifold embedding model analysis steps are: Build a mixture Gaussian-manifold embedding model, including: , where μk and Σk are directly taken from the prototype parameters of the base layer and the abnormal layer, including the mean vector μk and covariance matrix Σk of the base pattern layer, and the mean vector of the abnormal pattern layer and covariance matrix , πk is the proportion of prototype data initialized, N(x|μk,Σk) represents the Gaussian distribution function, P(x) is the probability distribution representation of the model for the input data x, λ is the weight coefficient, T(y|θ) represents The optimized manifold distribution function, θ is the manifold parameter, and the parameters {πk, μk, Σk, θ} are iteratively updated through the EM algorithm to minimize the negative log-likelihood function: , where the difference between the newly added prototype and the historical distribution is constrained by the KL divergence, L is the negative log-likelihood function, γ is the weight coefficient, Pcurr is the current distribution, Phist is the historical distribution, and xi represents the i-th input data point.

[0013] Furthermore, the physiological state variable coupling network analysis step: Based on the dynamic coupling intensity curve CSI(t) and the mutual information entropy matrix I(i, j), a time-varying weighted connection matrix Bij(t) is generated. The respiratory cycle length and phase characteristics are extracted from the respiratory signal, and the frequency band energy of the heart rate variability is calculated from the heart rate signal. Each node corresponds to a physiological state parameter, and the node V is represented by the physiological state parameter. The edge A(t) is defined by the time-varying weighted connection matrix Bij(t), that is, the mutual information entropy matrix I(i, j) is greater than θ×max(I), then it is an edge. The weight of the edge W(t) is equal to the time-varying weighted connection matrix Bij(t), and the physiological state-varying coupling network E(t)={V, A(t), W(t)} is obtained.

[0014] Furthermore, the dynamic coupling intensity curve CSI (t) analysis steps are: Based on phase symbol sequence and Calculating the Mutual Information Entropy of Respiration and Heart Rate Using Symbolic Dynamics , specific ,in and represents a subset of symbolized sequences of breathing and heart rate, express , The joint probability of the mutual information entropy matrix I(i, j) is obtained, where i and j represent the index of the rows and columns of the matrix, and the coupling strength index is calculated. , where T represents the total number of time points, express For the gradient of time t, with time as the horizontal axis and the coupling strength index as the vertical axis, the dynamic coupling strength curve CSI (t) is obtained.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention collects respiratory signals and heart rate signals and obtains the spatiotemporal coupling tensor of breathing rate through time alignment, motion compensation and reconstruction analysis; constructs a phase difference matrix based on the spatiotemporal coupling tensor of breathing rate to generate a spectrum, calculates the mutual information entropy and coupling strength index after decomposition by Gaussian mixture model, and analyzes to obtain the mutual information entropy matrix and dynamic coupling strength curve; based on the mutual information entropy matrix and the dynamic coupling strength curve, they are processed by graph attention network, and state reproduction and phase synchronization analysis are performed to obtain a dynamic multi-domain joint tensor; based on the dynamic multi-domain joint tensor, the parameters are obtained through projection, reconstruction and feature decomposition to obtain the pattern density manifold diagram and dynamic physiological pattern spectrum, which can analyze the dynamic coupling characteristics and multimodal information fusion between respiratory and heart rate signals and comprehensively characterize complex physiological patterns.

[0016] 2. The present invention constructs a mixed Gaussian-manifold embedding model based on the pattern density manifold diagram and the dynamic physiological pattern map to obtain the Gaussian probability and two-dimensional manifold space coordinates of the abnormal prototype, and performs corresponding warnings and signals. It can deeply explore the spatiotemporal evolution characteristics and chaotic characteristics of physiological signals, accurately capture the signs of respiratory abnormalities, and provide a method that can effectively process multi-source signals, integrate spatiotemporal dynamic characteristics and multimodal information, and realize accurate respiratory status assessment and abnormal warning, meeting the actual needs of clinical and health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. The following drawings are not intentionally scaled to the actual size, and the focus is on illustrating the main purpose of the present invention.

[0018] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a flow chart of step S200 of the present invention; Figure 3 This is a flow chart of step S300 of the present invention. DETAILED DESCRIPTION

[0019] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of the present invention.

[0020] like Figure 1 As shown, a respiratory monitoring analysis and evaluation method based on an optical fiber sensor includes the following steps: Step S100: collecting respiratory signals and heart rate signals, performing time alignment, motion compensation, and reconstruction analysis to obtain a spatiotemporal coupling tensor of respiratory rate; Step S200: constructing a phase difference matrix based on the spatiotemporal coupling tensor of the call rate to generate a spectrum, calculating the mutual information entropy and the coupling strength index after decomposition using a Gaussian mixture model, and analyzing to obtain the mutual information entropy matrix and the dynamic coupling strength curve; Step S300: Based on the mutual information entropy matrix and the dynamic coupling strength curve, the graph attention network is processed to perform state reproduction and phase synchronization analysis to obtain a dynamic multi-domain joint tensor; Step S400: Based on the dynamic multi-domain joint tensor expansion decomposition, parameters are obtained through projection, reconstruction, and eigendecomposition to obtain a pattern density manifold map and a dynamic physiological pattern atlas; Step S500: Construct a hybrid Gaussian-manifold embedding model based on the pattern density manifold map and the dynamic physiological pattern map, obtain the Gaussian probability and two-dimensional manifold space coordinates of the abnormal prototype, and perform corresponding warnings and signal issuance.

[0021] Specifically, the analysis of step S100 is as follows: Fiber optic sensors collect respiratory and heart rate signals, align the timestamps of multi-source signals through a high-precision clock synchronization module, combine with an adaptive motion compensation algorithm to eliminate interference from body position changes, and use non-uniform sampling interpolation technology to reconstruct the super-resolution time series of respiratory-heart rate signals, generating a spatiotemporal coupling tensor of respiratory rate with dimensions of time × spatial position × signal modality.

[0022] Specifically, such as Figure 2 As shown, the analysis of step S200 is as follows: Specifically, the instantaneous phase of the respiratory signal H(t) and the heart rate signal R(t) in the spatiotemporal coupling tensor of the breathing rate is extracted by Hilbert transform. and , specifically, , , where arg(*) represents the complex angle, j represents the imaginary unit, Ω(*) represents the Hilbert transform of the signal, and the dynamic phase difference matrix is ​​constructed , where sin is the sine function, λ is the attenuation factor, τ is the time window center, and exp is the exponential function with the natural constant e as the base. The phase difference range (−π, π] is divided into N equal-width intervals. The frequency of phase difference in each interval is counted to generate a histogram, and the Gaussian kernel function is used to smooth the phase difference distribution. , where h represents the bandwidth parameter, which controls the degree of smoothing, Represents the phase difference, and the probability density is obtained , with phase difference As the horizontal axis, the probability density The vertical axis is used to establish a transient phase shift mapping spectrum.

[0023] The instantaneous phase difference distribution map is decomposed into K Gaussian components through the multimodal Gaussian mixture model , where μk represents the mean of the kth component and σk is the standard deviation. , where N (*) represents Gaussian distribution, πk represents the weight of the kth component, centered on the mean μk of each component, ɑσk is expanded to define the symbol interval, ɑ represents the confidence interval coefficient, Sk=[μk-ɑσk, μk+ɑσk], and the overlapping intervals are merged to obtain the merged non-redundant symbol interval Sm, , where Si and Sj represent the original symbol intervals, ∪ represents the union operator, ∩ represents the intersection symbol, to avoid symbol redundancy, Sm is assigned a unique symbol to the merged non-redundant symbol interval, and the continuous phase difference Mapping to phase symbol sequence and .

[0024] Based on phase symbol sequence and Calculating the Mutual Information Entropy of Respiration and Heart Rate Using Symbolic Dynamics , specific ,in and represents a subset of symbolized sequences of breathing and heart rate, express , The joint probability of the mutual information entropy matrix I(i, j) is obtained, where i and j represent the index of the rows and columns of the matrix, and the coupling strength index is calculated. , where T represents the total number of time points, express For the gradient of time t, with time as the horizontal axis and the coupling strength index as the vertical axis, the dynamic coupling strength curve CSI (t) is obtained.

[0025] Specifically, such as Figure 3 As shown, the analysis of step S300 is as follows: Based on the dynamic coupling strength curve CSI (t) and the mutual information entropy matrix I (i, j), the time-varying weighted connection matrix Bij (t) is generated. , where θ∈[0,1] represents the sparsification threshold, max(I) represents the maximum value in the matrix I, the respiratory cycle length and phase features are extracted from the respiratory signal, and the frequency band energy of the heart rate variability is calculated from the heart rate signal. Each node corresponds to a physiological state parameter. Specifically: the respiratory cycle length is the time difference between adjacent inhalation starting points, and the frequency band energy is the high-frequency HF power; physiological state parameters, for example, the respiratory cycle is equal to 3.5s, and the HF power is equal to 0.14. The node V is represented by the physiological state parameters, and the edge A(t) is defined by the time-varying weighted connection matrix Bij(t), that is, if the mutual information entropy matrix I(i,j) is greater than θ×max(I), it is an edge, and the weight of the edge W(t) is equal to the time-varying weighted connection matrix Bij(t), and the physiological state variable coupling network is obtained. .

[0026] Obtain the node betweenness centrality value of the physiological variable coupling network E(t) , where σij represents the total number of shortest paths from node i to j, σij(k) represents the number of shortest paths passing through node k, and the top K nodes are sorted by the node betweenness centrality value Cbetw(k) to obtain the key hub node list; the edge weight mean of the physiological variable coupling network and the node betweenness centrality value of the key hub node list are combined to form the node feature hi, and the key hub node list is embedded in the graph attention network to obtain the hub node feature condensed vector set {v1, v2, ..., vK}, vi=GAT(hi, {hj|j∈N(i)}), where GAT represents the graph attention network, which is a neural network model for processing graph structured data, N(i) is the neighbor set of node i, that is, the set of nodes directly connected to node i, and the neighbor node j belongs to the neighbor set of i.

[0027] The hub node feature condensed vector set {v1, v2, ..., vK} is used to establish the state recurrence association matrix Kij through the step function. ,in represents a step function, if x≥0, ; otherwise 0, Indicates the radius threshold, which controls the similarity determination. Represents the norm difference, sums all elements in the state recurrence association matrix Kij and divides it by the square of the number of hub node feature condensation vector sets to obtain the state recurrence stability rate RL.

[0028] Define the diagonal direction in the state recurrence association matrix, specifically the main diagonal (i=j direction) and other parallel diagonals (i=j+k direction, k is the offset). Scan point by point along each diagonal direction to detect line segments with consecutive values ​​of 1. If L consecutive points starting from position (i, j) are 1, the length L is recorded, and the detected line segment length set {L1, v2, ..., LN} is obtained. The number of times each length L appears is counted and divided by the total number of appearances to obtain the probability distribution P(L) of the diagonal length L. Then, a deterministic formula analysis is performed to obtain the deterministic index of chaotic periodic phase transition. , where Lmin is the minimum effective diagonal length, and K represents the number of key hub node lists; The state recurrence association matrix Kij is mapped to a color for each element in the matrix. Specifically, 1 in the matrix is ​​mapped to red and 0 to white. The state recurrence stability rate RL is used to control the depth of the red color of 1 in the matrix through a linear mapping relationship. Using data visualization tools such as Python, the matrix is ​​arranged in rows and columns to generate a two-dimensional color block diagram. The horizontal and vertical axes represent the node index respectively. The color of the color block intuitively reflects the state similarity and stability between the nodes, and the state recurrence heat map W is obtained.

[0029] The specific linear mapping relationship is that if the state recurrence stability rate RL is equal to the set boundary threshold tg1, it corresponds to light red #FF9999, and if it is equal to the set boundary threshold tg2, it corresponds to dark red #8B0000. For the values ​​between the boundary threshold tg1 and the boundary threshold tg2, the red channel R = 255 − 232 × (RL - tg1), the green channel G = 153 − 306 × (RL - tg1), and the blue channel B = 153 − 306 × (RL - tg1). Two-dimensional dynamic phase difference matrix The dimension is Add an empty dimension and get the phase synchronization basis kernel tensor T dimension as , where T, K and 1 represent time, frequency band and potential channel respectively. The two-dimensional state current image heat map W is element-wise multiplied with the phase synchronization basis core tensor Tfusi to obtain the spatiotemporal recursive fusion tensor Tfusi with the dimension of , where J represents the state recurrence dimension, reflecting the similarity pattern of node states in the embedded space. Then, the one-dimensional chaotic periodic phase transition deterministic index DET is integrated through the outer product to obtain the dynamic multi-domain joint tensor Tjoi: , where D represents the chaotic domain, and D quantifies the orderliness or chaos of the system, enhancing the pathological early warning capability.

[0030] Specifically, the phase-synchronized basis kernel tensor ,in represents element-wise multiplication, , by taking the two-dimensional dynamic phase difference matrix The newly added spatial dimension is expanded into a three-dimensional phase-synchronized core tensor, providing a structural basis for multimodal data fusion, dynamically binding the frequency domain characteristics of phase synchronization with the time-series recursive pattern, quantifying the spatiotemporal stability of respiratory-heart rate coupling, and finally integrating the one-dimensional chaotic periodic phase change index DET through outer product to generate a dynamic multi-domain joint tensor, realizing the four-dimensional joint representation of time, frequency band, state recurrence and chaotic characteristics, thereby supporting the holographic analysis of complex physiological systems and accurately capturing the dynamic correlation between periodic stable states and chaotic desynchronization events.

[0031] Specifically, the analysis of step S400 is as follows: The dynamic multi-domain joint tensor is first expanded into a time dimension expansion matrix according to the four dimensions of time, frequency band, state recurrence and chaos. , frequency band dimension expansion matrix , Status Dimension Expansion Matrix and Chaos Dimension Expansion Matrix , specific , , , , perform singular value decomposition on each matrix and retain the first ri principal components, , is the left singular matrix obtained by decomposition in the i-th dimension, is the singular value diagonal matrix obtained by decomposition of the i-th dimension, is the right singular matrix obtained by decomposition of the i-th dimension, T represents the transpose operation, and the left singular matrix obtained by decomposition is retained , marked as the dimension feature parsing matrix And perform projection operation on the dynamic multi-domain joint tensor Tjoi and the decomposed left singular matrix to obtain the compressed interactive core tensor G. The specific projection operation is: , Represents an n-mode product on the first dimension of the tensor, and a tensor-matrix multiplication along a specific dimension.

[0032] The compressed interaction core tensor G and the projection of the dimensional feature analysis matrix time factor and chaos factor are reconstructed respectively to obtain the spatiotemporal coupling state matrix X and the time sequence evolution state matrix , using the spatiotemporal coupling state matrix X and the temporal evolution state matrix Linear dynamic mode extraction and chaotic characteristic injection are performed to obtain the chaotic weighted modal matrix C. The chaotic weighted modal matrix C is subjected to eigenvalue decomposition to obtain the dynamic coupling eigenvalue ηk and the multimodal rhythm eigenvector δk. The dynamic coupling eigenvalue ηk is subjected to logarithmic operation and imaginary part extraction analysis to obtain the dominant oscillation frequency wk. The contribution degree of the multimodal rhythm eigenvector δk in the coupling is analyzed to obtain the mode energy ratio Ek. The mode energy ratio and the chaotic periodic phase change index DET are dynamically calibrated to the desynchronization index calculation formula to obtain the dynamically calibrated desynchronization index DI.

[0033] Specifically, the formulas used above are:

[0034] Specifically, is the compressed interaction core tensor at time t+1, is the pseudo-inverse matrix of the spatiotemporal coupling state matrix X, ζ represents the weight control parameter, which controls the chaotic mode factor The weight of , diag (*) represents the constructed diagonal matrix, δk represents the dynamic coupling eigenvalue obtained by eigenvalue decomposition of matrix C, complex number, the real part represents the modal attenuation rate, the imaginary part is related to the oscillation frequency, δk represents the multimodal rhythm eigenvector obtained by eigenvalue decomposition of matrix C, the spatial distribution pattern of the corresponding mode, Im (*) represents the imaginary part, represents the time interval, i is the index variable in the summation symbol, Ek>0.1 means that the values ​​satisfying Ek greater than 0.1Ek are accumulated, and only the modes with relatively large mode energy ratios and more significant contributions to coupling are considered. represents the weight parameter, Represents the square of the norm.

[0035] Based on the dominant oscillation frequency wk, mode energy ratio Ek and dynamic calibration desynchronization index DI, the triple [wk, Ek, DI] is formed. The dynamic prototype of the neural gas algorithm It is composed of a triplet, and the neural gas algorithm dynamic prototype is updated iteratively through the neural gas algorithm, which is used to dynamically adjust the prototype The specific neural gas algorithm iterative update is: ,in represents the multimodal rhythm feature vector of the current time window, e(t) represents the learning rate that decays over time, Represents the prototype The update amount, e(t)=e0 / (1+t / T), where e0 represents the initial learning rate, T represents the decay period, and h(m, t) represents the neighborhood function. , where σ(t) represents the adaptive neighborhood radius, exp represents the exponential function with the natural constant e as the base, σ(t)=σ0×exp(-γt), σ0 represents the initial neighborhood radius, and γ represents the decay rate.

[0036] Data H without respiratory disease records were obtained from an open-source physiological database, including the dominant oscillation frequency wk, the mode energy ratio Ek, and the dynamically calibrated desynchronization index DI. The eigenvector matrix and the principal component latent matrix of the healthy data were obtained through covariance matrix decomposition and principal component analysis. Specifically: Among them, Cov(*) represents the covariance matrix, U represents the eigenvector matrix, and the column vector represents the potential structural direction of the data. Represents a diagonal matrix, the diagonal elements are the variance contributions of each principal component, U[:,:k] represents taking the first k columns of U, Z represents the principal component potential matrix of health data, and N represents sample data.

[0037] Based on the principal component potential matrix of health data, a basic circadian rhythm template is constructed to form the basic pattern layer of the hierarchical atlas. The specific circadian rhythm template is: Pbase=1 / N∑hi+ω×Z, where hi represents the sub-data of data H without respiratory disease records, ω represents the fusion coefficient, and Pbase represents the basic circadian rhythm template, such as the sleep period template Psleep=[0.1Hz, 0.8, 0.1] in the circadian rhythm template, which represents low-frequency, high-energy and stable coupling.

[0038] The abnormal pattern layer marks the pathological prototype based on the dynamic threshold. The abnormal pattern layer and the basic pattern layer form a dynamic physiological pattern map. The specific dynamic threshold is: Wk>μω+2σω or DI>kj1, where μω and σω represent the historical frequency mean and standard deviation, and kj1 represents the set dynamic calibration desynchronization index, specifically equal to 0.6; Pattern density manifold diagram dynamically prototypes the neural gas algorithm through the t-SNE-Δt algorithm Mapping to two-dimensional space ,in represents the coordinates of the two-dimensional manifold space, where Time weight in the algorithm ,in Represents the coefficient that controls the time decay strength, tcurr represents the current time point, and t represents the time corresponding to the data point; Traditional t-SNE is a nonlinear dimensionality reduction technique that uses Gaussian distribution to calculate the conditional probability between data points in high-dimensional space, and uses t distribution to calculate similarity in low-dimensional space, and minimizes the Kullback-Leibler (KL) divergence between the two to achieve the mapping of high-dimensional data to low-dimensional space in order to preserve the local similarity of the data. It is often used for data visualization. The algorithm introduces the time factor on this basis (through the parameter and time weight w(t)), giving it time perception. During dimensionality reduction, it causes recent data to occupy a larger density region in the manifold space, highlighting its importance and better capturing the dynamic changes in data over time. For example, when analyzing physiological patterns, it can more clearly demonstrate the migration trajectory of patterns over time (such as the clustering of high-frequency breathing patterns in winter). The dimensionality reduction results reflect not only the spatial distribution of the data but also its temporal evolution. This makes it suitable for high-dimensional data visualization and dynamic analysis scenarios that require time series information, such as pattern density manifold plots.

[0039] The pattern density manifold map is obtained by connecting the manifold coordinates ym(t) of adjacent time windows.

[0040] Specifically, the analysis of step S500 is as follows: A mixed Gaussian-manifold embedding model is established based on the pattern density manifold graph and the dynamic physiological pattern map. The specific analysis is as follows: The circadian rhythm template in the basic pattern layer of the dynamic physiological pattern atlas divides a day into T periods and generates multiple Gaussian prototypes through incremental clustering. Each prototype consists of a mean vector μk and a covariance matrix Σk. , where |Ck| represents the number of samples in the dataset Ck belonging to prototype k, xi represents the i-th data point in the dataset, which corresponds to the typical breathing patterns in different time periods. Based on the pathological prototype in the abnormal pattern layer, spectral clustering is used to generate abnormal Gaussian prototypes. , , where k represents the prototype index, and Represents the covariance matrix of the mean and kth abnormal prototype, cov represents the function of calculating covariance, represents the data set of the kth abnormal prototype, and mean represents the calculation of the arithmetic mean of a set of data. The objective function introduces a prototype position constraint term to the pattern density manifold graph to ensure that the coordinates of the normal prototype are centered in the manifold space and the abnormal prototypes are distributed at the edge. The manifold parameter θ is obtained, which includes the prototype coordinate mapping relationship y(μk). Specifically: ,in Represents the similarity in high-dimensional space, Represents the similarity in low-dimensional space, Indicates the preset prototype coordinates, represents the weight coefficient, which makes the normal prototype centered and the abnormal prototype marginalized; , Where KL represents the divergence operation, which measures the difference between two Gaussian distributions, i and j are indices used to distinguish different Gaussian prototypes, yi and yj are coordinate points in the two-dimensional manifold space, and l is an index variable similar to the index k.

[0041] Build a mixture Gaussian-manifold embedding model, including: , where μk, Σk are directly taken from the prototype parameters of the base layer and the abnormal layer, including the mean vector μk and covariance matrix Σk of the base pattern layer, and and , πk is the proportion of data initialized as prototype, πk=|Ck| / N, N is the total amount of data, N(x|μk,Σk) represents the Gaussian distribution function, P(x) is the probability distribution representation of the model for the input data x, λ is the weight coefficient, T(y|θ) represents The optimized manifold distribution function quantifies the local density of the manifold coordinate y and reflects the spatial correlation in the pattern density manifold graph. θ is the manifold parameter. The parameters {πk, μk, Σk, θ} are iteratively updated through the EM algorithm to minimize the negative log-likelihood function: , where the difference between the newly added prototype and the historical distribution is constrained by the KL divergence, L is the negative log-likelihood function, is the objective function of the model optimization, γ is the weight coefficient, Pcurr is the current distribution, Phist is the historical distribution, and xi represents the i-th input data point.

[0042] After processing the real-time collected respiratory and heart rate signals, the Gaussian probability of them belonging to the abnormal prototype is calculated. If it is greater than the set threshold lk1, signal 2 is sent, and if it is less than or equal to it, signal 1 is sent; kernel density estimation is used to fit the manifold coordinate set {ynormal}, and the probability density Pnormal(y) of the normal area is calculated. According to the confidence interval of normal data such as the 95% quantile, the boundary threshold δ0 of the normal area is defined. After processing the real-time collected respiratory signal and heart rate signal, the t-SNE manifold term in the model is used to map them to the two-dimensional manifold space to obtain the current coordinate yt. The sliding time window is the latest T0 time point, and the number of deviations within the window is counted. If it is greater than the set threshold ky1, signal 2 is issued; if it is less than or equal to ky1, signal 1 is issued; if two signal 2s are received, the system issues an early warning, indicating severe breathing abnormalities; if there is one signal 2, breathing abnormalities are displayed, and no operation is performed in other cases.

[0043] The above is an illustration of the present invention and should not be considered as limiting thereof. Although several exemplary embodiments of the present invention have been described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Therefore, all such modifications are intended to be included within the scope of the present invention as defined by the claims. It should be understood that the above is an illustration of the present invention and should not be considered as being limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is defined by the claims and their equivalents.

Claims

1. A respiratory monitoring, analysis and evaluation method based on optical fiber sensor, characterized in that: The following steps are involved: Step S100: collecting respiratory signals and heart rate signals, performing time alignment, motion compensation, and reconstruction analysis to obtain a spatiotemporal coupling tensor of respiratory rate; Step S200: constructing a phase difference matrix based on the spatiotemporal coupling tensor of the call rate to generate a spectrum, calculating the mutual information entropy and the coupling strength index after decomposition using a Gaussian mixture model, and analyzing to obtain the mutual information entropy matrix and the dynamic coupling strength curve; Step S300: Based on the mutual information entropy matrix and the dynamic coupling strength curve, the graph attention network is processed to perform state reproduction and phase synchronization analysis to obtain a dynamic multi-domain joint tensor; Step S400: Based on the dynamic multi-domain joint tensor expansion decomposition, parameters are obtained through projection, reconstruction, and eigendecomposition to obtain a pattern density manifold map and a dynamic physiological pattern atlas; Step S500: Construct a hybrid Gaussian-manifold embedding model based on the pattern density manifold map and the dynamic physiological pattern map, obtain the Gaussian probability and two-dimensional manifold space coordinates of the abnormal prototype, and perform corresponding warnings and signal issuance.

2. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 1, characterized in that: The mutual information entropy matrix and dynamic coupling strength curve analysis steps are: The instantaneous phase of the respiratory signal H(t) and the heart rate signal R(t) in the spatiotemporal coupling tensor of the breathing rate is extracted by Hilbert transform, and the dynamic phase difference matrix is ​​constructed. , divide the phase difference range into N equal-width intervals, count the frequency of phase difference in each interval, generate a histogram, and use Gaussian kernel function to smooth the phase difference distribution to obtain the probability density , with phase difference As the horizontal axis, the probability density As the vertical axis, a transient phase shift mapping spectrum is established; The instantaneous phase difference distribution map is decomposed into K Gaussian components N through the multimodal Gaussian mixture model. The mean μk of each component is used as the center, and the symbol interval is defined by expanding ɑσk. ɑ represents the confidence interval coefficient. The overlapping intervals are merged to obtain the merged non-redundant symbol interval Sm. The merged non-redundant symbol interval Sm is assigned a unique symbol. The continuous phase difference Mapping to phase symbol sequence and ; Based on phase symbol sequence and The mutual information entropy of respiration and heart rate was calculated using symbolic dynamics to obtain the mutual information entropy matrix I(i, j), where i and j represent the row and column indices of the matrix. The coupling strength index was also calculated, and the dynamic coupling strength curve CSI(t) was obtained with time as the horizontal axis and the coupling strength index as the vertical axis.

3. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 1, characterized in that: The dynamic multi-domain joint tensor analysis steps are: Obtain the node betweenness centrality value of the physiological variable coupling network E(t), sort the top K nodes by the node betweenness centrality value Cbetw(k) to obtain the key hub node list; The edge weight mean of the physiological state variable coupling network and the node betweenness centrality value of the key hub node list are used to embed the key hub node list into a graph attention network to obtain a hub node feature condensed vector set; The hub node feature condensed vector set is used to establish the state recurrence association matrix Kij through a step function. The sum of all elements in the state recurrence association matrix Kij is divided by the square of the number of hub node feature condensed vector sets to obtain the state recurrence stability rate RL. Map each element of the state recurrence association matrix Kij to a color, and use the state recurrence stability rate RL to control the red depth in the matrix through a linear mapping relationship. Use data visualization tools to arrange the matrix in rows and columns to generate a two-dimensional color block diagram. The horizontal and vertical axes represent the node index respectively. The color of the color block intuitively reflects the state similarity and stability between nodes, and the state recurrence heat map W is obtained; Two-dimensional dynamic phase difference matrix A new empty dimension is added to the dimension to obtain the phase synchronization basis core tensor T, and the two-dimensional state current mapping heat map W is element-wise multiplied with the phase synchronization basis core tensor Tfusi to obtain the space-time recursive fusion tensor Tfusi. Then, the one-dimensional chaotic periodic phase change deterministic indicator DET is integrated through the outer product to obtain the dynamic multi-domain joint tensor Tjoi.

4. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 3, characterized in that: The chaotic periodic phase transition deterministic indicator DET analysis steps: Define the diagonal direction, main diagonal and other parallel diagonals in the state recurrence association matrix, scan point by point along each diagonal direction, detect the line segments with consecutive values ​​of 1, if L consecutive points starting from position (i, j) are 1, then record the length L, and obtain the detected line segment length set {L1, v2, ..., LN}. Count the number of times each length L appears and divide it by the total number of appearances to obtain the probability distribution P(L) of the diagonal length L, and perform deterministic formula analysis to obtain the deterministic index of chaotic periodic phase transition. , where Lmin is the minimum effective diagonal length and K represents the number of key hub node lists.

5. The respiratory monitoring, analysis and evaluation method based on optical fiber sensor according to claim 1, characterized in that: The pattern density manifold diagram and dynamic physiological pattern atlas analysis steps: Based on the dominant oscillation frequency wk, mode energy ratio Ek and dynamic calibration desynchronization index DI, the triple [wk, Ek, DI] is formed. The dynamic prototype of the neural gas algorithm It is composed of a triplet, and the neural gas algorithm dynamic prototype is updated through the neural gas algorithm iteration, and the prototype is dynamically adjusted location; Data H without respiratory disease records were obtained from an open-source physiological database. The eigenvector matrix and the principal component latent matrix of the healthy data were obtained through covariance matrix decomposition and principal component analysis. Based on the principal component latent matrix of health data, a basic circadian rhythm template is constructed to form the basic pattern layer of the hierarchical map; The abnormal pattern layer marks the pathological prototype based on the dynamic threshold, and the abnormal pattern layer and the basic pattern layer form a dynamic physiological pattern map; The pattern density manifold is plotted by Algorithms will be Neural Gas Algorithm Dynamic Prototype Mapping to two-dimensional space ,in Represents the coordinates after the two-dimensional manifold space; by connecting the manifold coordinates ym(t) of adjacent time windows, the pattern density manifold map is obtained.

6. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 5, characterized in that: The dominant oscillation frequency wk, mode energy ratio Ek and dynamic calibration desynchronization index DI analysis steps are: The dynamic multi-domain joint tensor is first expanded into a time dimension expansion matrix according to the four dimensions of time, frequency band, state recurrence and chaos. , frequency band dimension expansion matrix , Status Dimension Expansion Matrix and Chaos Dimension Expansion Matrix , perform singular value decomposition on each matrix and retain the first ri principal components, and retain the left singular matrix obtained by decomposition , marked as the dimension feature parsing matrix And perform projection operation on the dynamic multi-domain joint tensor Tjoi and the decomposed left singular matrix to obtain the compressed interactive core tensor G; The compressed interaction core tensor G and the projection of the dimensional feature analysis matrix time factor and chaos factor are reconstructed respectively to obtain the spatiotemporal coupling state matrix X and the time sequence evolution state matrix , using the spatiotemporal coupling state matrix X and the temporal evolution state matrix Linear dynamic mode extraction and chaotic characteristic injection are performed to obtain the chaotic weighted modal matrix C. The chaotic weighted modal matrix C is subjected to eigenvalue decomposition to obtain the dynamic coupling eigenvalue ηk and the multimodal rhythm eigenvector δk. The dynamic coupling eigenvalue ηk is subjected to logarithmic operation and imaginary part extraction analysis to obtain the dominant oscillation frequency wk. The contribution degree of the multimodal rhythm eigenvector δk in the coupling is analyzed to obtain the mode energy ratio Ek. The mode energy ratio and the chaotic periodic phase change index DET are dynamically calibrated to the desynchronization index calculation formula to obtain the dynamically calibrated desynchronization index DI.

7. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 1, characterized in that: The corresponding warning and signal issuance analysis steps are: The circadian rhythm template in the basic pattern layer of the dynamic physiological pattern atlas divides a day into T periods and generates multiple Gaussian prototypes through incremental clustering. Each prototype consists of a mean vector μk and a covariance matrix Σk. , where |Ck| represents the number of samples in the dataset Ck belonging to prototype k, xi represents the i-th data point in the dataset, and the abnormal Gaussian prototype is generated by spectral clustering based on the pathological prototype in the abnormal pattern layer. , , where k represents the prototype index, and Represents the covariance matrix of the mean and kth abnormal prototype, cov represents the function of calculating covariance, Represents the data set of the kth abnormal prototype, mean represents the calculation of the arithmetic mean of a set of data; The objective function introduces a prototype position constraint term into the pattern density manifold graph to ensure that the normal prototype is centered in the manifold space and the abnormal prototype is distributed at the edge, thus obtaining the manifold parameter θ. Establish and train a mixed Gaussian-manifold embedding model, process the real-time collected respiratory and heart rate signals, and then calculate and input them into the model to obtain the Gaussian probability of belonging to the abnormal prototype. , and send signals 2 and 1 accordingly; use kernel density estimation to fit the manifold coordinate set {ynormal}, calculate the probability density Pnormal(y) of the normal area, define the boundary threshold δ0 of the normal area according to the confidence interval of normal data, process the real-time respiratory signal and heart rate signal, and map them to the two-dimensional manifold space through the t-SNE manifold term in the model to obtain the current coordinate yt, slide the time window T0 time point, and count the number of deviations within the window , and send out signal two and signal one accordingly; give corresponding warnings and signals for signal one and two.

8. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 7, characterized in that: The mixed Gaussian-manifold embedding model analysis steps: Build a mixture Gaussian-manifold embedding model, including: , where μk and Σk are directly taken from the prototype parameters of the base layer and the abnormal layer, including the mean vector μk and covariance matrix Σk of the base pattern layer, and the mean vector of the abnormal pattern layer and covariance matrix , πk is the proportion of prototype data initialized, N(x|μk,Σk) represents the Gaussian distribution function, P(x) is the probability distribution representation of the model for the input data x, λ is the weight coefficient, T(y|θ) represents The optimized manifold distribution function, θ is the manifold parameter, and the parameters {πk, μk, Σk, θ} are iteratively updated through the EM algorithm to minimize the negative log-likelihood function: , where the difference between the newly added prototype and the historical distribution is constrained by the KL divergence, L is the negative log-likelihood function, γ is the weight coefficient, Pcurr is the current distribution, Phist is the historical distribution, and xi represents the i-th input data point.

9. A respiratory monitoring, analysis and evaluation method based on an optical fiber sensor according to claim 3, characterized in that: The physiological state variable coupling network analysis steps: Based on the dynamic coupling intensity curve CSI(t) and the mutual information entropy matrix I(i, j), a time-varying weighted connection matrix Bij(t) is generated. The respiratory cycle length and phase characteristics are extracted from the respiratory signal, and the frequency band energy of the heart rate variability is calculated from the heart rate signal. Each node corresponds to a physiological state parameter, and the node V is represented by the physiological state parameter. The edge A(t) is defined by the time-varying weighted connection matrix Bij(t), that is, the mutual information entropy matrix I(i, j) is greater than θ×max(I), then it is an edge. The weight of the edge W(t) is equal to the time-varying weighted connection matrix Bij(t), and the physiological state-varying coupling network E(t)={V, A(t), W(t)} is obtained.

10. The method for respiratory monitoring, analysis and evaluation based on an optical fiber sensor according to claim 3, characterized in that: The dynamic coupling intensity curve CSI (t) analysis steps are: Based on phase symbol sequence and Calculating the Mutual Information Entropy of Respiration and Heart Rate Using Symbolic Dynamics , specific ,in and represents a subset of symbolized sequences of breathing and heart rate, express , The joint probability of the mutual information entropy matrix I(i, j) is obtained, where i and j represent the index of the rows and columns of the matrix, and the coupling strength index is calculated. , where T represents the total number of time points, express For the gradient of time t, with time as the horizontal axis and the coupling strength index as the vertical axis, the dynamic coupling strength curve CSI (t) is obtained.

Images