Wireless-based bedside ECG and blood oxygen sync monitoring system

Abstract

This invention discloses a bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission, belonging to the field of ECG and blood oxygen synchronous monitoring. It includes an acquisition module, a signal processing module, a fusion module, and a judgment module. The system acquires patient signals and converts them into digital signals. A synchronous chaotic sequence is generated using a Lorentz chaotic model combined with a fourth-order Runge-Kutta method, driving a star-shaped topology master-slave architecture. Graph nodes are defined by matrix feature dimensions, edge weights are calculated, and a normalized derived dynamic graph structure is obtained. A high-order spatiotemporal correlation feature matrix is ​​obtained. The system uses gradient-weighted class activation mapping to locate and extract single-modal features from temporal-precision net spectral features of ECG, blood oxygen, and respiratory signals, and calculates the contribution weight of the prediction results. A cross-modal spatiotemporal correlation fusion feature matrix is ​​obtained. Health category probabilities are generated through linear transformation and a softmax function. The system analyzes the patient's health category probability and displays corresponding colors and alarms, making the health status discrimination model more accurate.

Description

Technical Field

[0001] This invention relates to the field of synchronous monitoring of electrocardiogram and blood oxygenation, specifically a bedside synchronous monitoring system for electrocardiogram and blood oxygenation based on wireless transmission. Background Technology

[0002] Real-time monitoring of multimodal physiological signals such as ECG and blood oxygenation at the clinical bedside is crucial for early disease diagnosis and intervention. However, existing monitoring systems face the problem of timestamp misalignment caused by differences in sampling clocks for multimodal signals. Traditional synchronization methods rely on fixed-frequency clocks or simple hardware calibration, which are difficult to adapt to dynamic physiological fluctuations and environmental interference, easily leading to signal analysis errors. At the same time, physiological signals are often affected by electromagnetic noise, motion artifacts, etc., and traditional filtering techniques have limited ability to preserve signal features in complex noise scenarios, resulting in insufficient accuracy of subsequent feature analysis. Therefore, a bedside ECG and blood oxygenation synchronous monitoring system based on wireless transmission has been developed.

[0003] Existing bedside ECG and blood oxygen synchronous monitoring systems based on wireless transmission have technical bottlenecks in mining the spatiotemporal correlation information contained in multimodal signals. Existing fusion methods are unable to effectively capture the deep interaction relationships between ECG, blood oxygen, and respiratory signals in terms of time, frequency, and mode. Furthermore, they lack quantitative evaluation of the contribution of key regions and features of the signals, resulting in insufficient robustness and interpretability of the health status discrimination model.

[0004] To address the aforementioned issues, this system employs techniques such as chaos-driven temporal alignment, multimodal feature enhancement, and dynamic graph convolutional networks to achieve high-precision synchronization, noise suppression, and cross-temporal correlation analysis of multimodal signals, providing an efficient solution for continuous bedside monitoring. Summary of the Invention

[0005] To address the technical problems raised in the background section, this invention is proposed. Embodiments of this invention provide a bedside ECG and blood oxygenation synchronization monitoring system based on wireless transmission.

[0006] The objective of this invention can be achieved through the following technical solution: a bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission, comprising an acquisition module, a signal processing module, a fusion module, and a judgment module.

[0007] The acquisition module is used to acquire the patient's electrocardiogram, blood oxygenation, and respiratory signals, and convert them into digital signals;

[0008] The signal processing module is used to generate synchronous chaotic sequences from the acquired digital signals by combining the Lorentz chaotic model with the fourth-order Runge-Kutta method, drive the star topology master-slave architecture, and then input the signals into the Kalman filter. After prediction and update steps, white noise is added, and the signals are reconstructed by empirical mode decomposition, adjacent mode difference criteria, energy entropy discrimination and soft thresholding to obtain the time-quasi-net spectrum features of ECG, blood oxygen and respiratory signals.

[0009] The fusion module processes the temporal net spectral features of ECG, blood oxygen, and respiratory signals, constructs a fourth-order tensor, and combines it with a sparse projection matrix to reduce the dimensionality of the multimodal spatiotemporal feature matrix. Graph nodes are defined by the matrix feature dimension, edge weights are calculated, and the normalized derivation dynamic graph structure is analyzed. The neighborhood information of the normalized derivation dynamic graph structure is analyzed by constructing a multimodal graph convolutional network model to obtain a high-order spatiotemporal correlation feature matrix. The single-modal features of the temporal net spectral features of ECG, blood oxygen, and respiratory signals are located and extracted by gradient weighted class activation mapping, and the contribution weight of the prediction results is calculated. The cross-modal spatiotemporal correlation fusion feature matrix is ​​then analyzed.

[0010] The determination module is used to input the cross-modal spatiotemporal correlation fusion feature matrix into the logistic regression model, generate the health category probability through linear transformation and softmax function, analyze the patient's health category probability, and display and alarm accordingly with different colors.

[0011] Furthermore, the time-precision net spectral characteristic ECG, blood oxygen, and respiratory signal analysis steps are as follows:

[0012] By adding m sets of white noise to time-accurate denoising ECG, blood oxygen, and respiration signals to satisfy the intelligent signal-to-noise ratio adaptation rule, empirical mode decomposition is performed on the three types of signals after incorporating the rule-incorporated white noise, and the intrinsic mode function sets of each signal are obtained. k represents the order. The average value of each intrinsic mode function in the m decomposition results is calculated. The intrinsic mode function sets of ECG, blood oxygenation and respiration are processed respectively. The decomposition process is terminated by using the neighboring mode difference stopping criterion. High-frequency noise modes are screened by energy entropy value discrimination and dynamically removed by soft thresholding. The retained effective modes are reconstructed to obtain the time-precise net spectral characteristics of ECG, blood oxygenation and respiration signals.

[0013] Furthermore, a Lorentz chaotic model is constructed using the third-order autonomous ordinary differential equations of dynamics, and a fourth-order Runge-Kutta method numerical solution equation is introduced to generate a synchronous chaotic sequence G. sy ;

[0014] A chaotic sequence-driven ECG and blood oxygenation timing alignment mechanism is constructed, and a star topology master-slave architecture is built. The ECG sensor is used as the master node and as the core transmitter, transmitting a synchronous chaotic sequence G to the blood oxygenation sensor, which is used as the slave node. sy After obtaining the sequence from the node, the local clock phase is dynamically calibrated using a phase-locked loop. A time deviation Δt is defined as the objective function, and iterative optimization using the least squares method is employed until the time deviation Δt is less than a set threshold, resulting in the time-aligned ECG signal X′. ECG [n] and blood oxygen signal X′ SpO[n], where n represents the sampling time point number. Then, by repeating the steps, a chaotic sequence-driven ECG-respiratory timing alignment mechanism is constructed, and a star topology master-slave architecture is built. The ECG sensor is used as the master node and as the core transmitter, transmitting a synchronous chaotic sequence G to the respiratory sensor, which is used as the slave node. sy The time-aligned electrocardiogram signal X′ is obtained by the same processing described above. ECG [n] and respiratory signal X′ brea [n], and then the electrocardiogram signal X′ ECG [n], Blood oxygen signal X′ SpO [n] and respiratory signal X′ brea [n] Alignment, input the synchronized ECG signal, blood oxygen signal, and respiration signal into a Kalman filter, and through prediction and update steps, obtain the time-accurate denoising ECG signal X″. ECG [n]、Blood oxygen signal X″ SpO [n] and respiratory signal X″ brea [n].

[0015] Furthermore, the steps of the dimensionality-reduced multimodal spatiotemporal feature matrix analysis are as follows:

[0016] The time-frequency matrix W is generated from the time-frequency net spectrum characteristic electrocardiogram signal through continuous wavelet transform. ECG ∈R T×F The spectral matrix S of the time-precision net spectral characteristic blood oxygen signal was extracted by short-time Fourier transform. SpO ∈R T×F The empirical wavelet transform of the time-precision net spectral characteristic respiratory signal is decomposed into multimodal components. Extract the top three components with the largest energy percentage and construct a fourth-order tensor. Where T represents the time step, F represents the frequency band, A represents the number of respiratory rate components, and C represents the number of channels, and is filled using a fourth-order tensor filling rule, initializing the factor matrices R, L, U, and B and the core tensor G, where R ∈ R T×H1 , L∈R F×H2 ,U∈R A×H3 , B∈R C×H4 , G∈R H1×H2×H3×H4 Where T, L, A, and C represent the original dimensions of time step, frequency band, number of respiratory rate components, and number of channels, respectively, and H1, H2, H3, and H4 represent the ranks of each dimension after dimensionality reduction. The objective function is minimized through iterative optimization using the alternating direction multiplier method to obtain the optimized factor matrix and the optimized four-dimensional core tensor G′. The optimized four-dimensional core tensor G′ is flattened into a two-dimensional matrix in dimensional order. The expanded matrix is ​​multiplied by the sparse projection matrix to obtain the dimensionality-reduced multimodal spatiotemporal feature matrix Mfus.

[0017] Furthermore, the cross-modal spatiotemporal correlation fusion feature matrix analysis steps are as follows:

[0018] A multimodal graph convolutional network model is constructed. The neighborhood information in the derived dynamic graph structure S′ is aggregated and normalized using the graph convolutional layer feature update formula in the model. After three layers of graph convolutional operations, the model is trained by jointly optimizing the multi-task objective function of the model, and outputting a high-order spatiotemporal correlation feature matrix Hfinal.

[0019] Key regions are identified by locating temporal net spectral features of ECG, blood oxygen, and respiration signals using gradient-weighted class activation mapping. Single-modal features extracted from these signals are used to calculate contribution weights for prediction results. Based on key region location and feature contribution weights, the edge weights between nodes in the dynamic graph structure are dynamically adjusted, and the normalized derived dynamic graph structure S′ is updated. Graph convolution, feature optimization, and multi-task joint training are iteratively performed, and the model finally outputs a cross-modal spatiotemporal correlation fusion feature matrix.

[0020] Furthermore, the naturalization-derived dynamic graph structure analysis steps are as follows:

[0021] Based on the dimension-reduced multimodal spatiotemporal feature matrix Mfus, each column has d feature dimensions, defining graph nodes. Each node corresponds to a feature vector, and the edge weights between nodes are determined by the formula. Where ni and nj are the eigenvectors of the i-th and j-th columns of the eigenma matrix Mfus, σ0 is the Gaussian kernel parameter, θij is the phase difference between the two eigenvectors extracted by the Hilbert transform, and all edge weights wij constitute the dynamic graph structure S. The dynamic graph structure S is normalized by adding self-loops to obtain the normalized derived dynamic graph structure S′, S′=S+I, where I is the identity matrix.

[0022] Furthermore, the key region analysis steps are as follows:

[0023] The time-defined net spectrum characteristics of electrocardiogram, blood oxygen, and respiration signals were segmented into segments U∈R according to time windows. T×C Calculate the gradient of segment U at each time point. Where y c It is the target category probability, U i，j Given the signal value at time point i and channel j, global average pooling is performed on the gradients of each channel to obtain the channel importance weights. Where L is the length of the time series, a time-dimensional heatmap is generated by weighted summation, and the importance value H of the classification decision at the i-th time point is used. i It means that H i =∑ j (α j ×U i，j If the importance value H for classification decision iIf the threshold is greater than or equal to the set threshold, then the signal segment at time point i is the critical region K.

[0024] Furthermore, the feature contribution weight dynamically adjusts the edge weight analysis steps between nodes in the dynamic graph structure:

[0025] A multimodal feature set F is established from the quasi-net spectrum characteristic ECG signal, the time-based quasi-net spectrum characteristic blood oxygenation signal, and the time-based quasi-net spectrum characteristic respiration signal, and then divided into all possible subset combinations. Where F = (f1, f2, ..., fn), and fi represents a single-modal feature. Using the formula φ(fi) = ∑Q|Q|! (|F|-|Q|-1)! / |F|! ×[v(Q∪{fi}-v(Q)), we obtain the contribution weight φ(fi) of feature fi to the prediction result, where ! represents the factorial symbol and v(Q) represents the predicted output of the model;

[0026] For feature dimensions located within the critical region K, their feature vectors are adjusted using contribution weights: n′i=ni×(1+∑Kφ(fi)×δ(ni∈K)), where δ(*) represents an indicator function, which takes a value of 1 when the feature dimension ni belongs to the critical region and 0 otherwise. ni represents the feature vector, and n′i represents the adjusted feature vector. The edge weights between nodes are then recalculated using the adjusted feature vector.

[0027] n′j represents the adjusted feature vector, and w′ij represents the adjusted weights of the edges between nodes.

[0028] Furthermore, the different color displays and alarm analysis steps:

[0029] The cross-modal spatiotemporal correlation fusion feature matrix is ​​input into the logistic regression model, and each feature dimension is standardized. in This represents the normalized value of the eigenvector in the cross-modal spatiotemporal correlation fusion feature matrix, where μ and σ are the mean and standard deviation, respectively. The model is obtained through linear transformation. The score zi is calculated, where w is the weight vector, b is the bias term, and T represents the transpose of the matrix. The health category probability is generated using the softmax function, with categories being healthy, sub-healthy, and unhealthy. The parameters are optimized by minimizing the cross-entropy loss function and incorporating a regularization term. The Adam optimizer iteratively updates w and b, ultimately outputting the patient's health category probability. Different colors are used to display and alert patients based on their different health category probabilities, as shown in the following formula:

[0030] P(yi=c=softmax(z) i，c ) = exp(z i，c ) / ∑ k (exp(zi，k ))

[0031]

[0032] Where c represents the category of physical health, and exp is an exponential function with the natural constant e as its base. Let λ0 represent the square of the L2 norm, λ0 represent the regularization coefficient, k represent the category index, traversing all health categories, P represent the probability, N represent the sample size, and M represent the number of samples. i This represents the factor matrix of the decomposition.

[0033] Furthermore, the optimizer iteratively updates w and b in the analysis step:

[0034] Initialize the weight vector w∈R d And bias b∈R, where w is initially the parameter W of the third layer of the dynamic graph convolution. (3) ∈R d×d The main diagonal elements, b initialized to zero, are used. The Adam optimizer iteratively updates the parameters, updating w and b through the weight gradients of logistic regression, until the weight and bias gradients are less than a set threshold. The weight and bias gradients of logistic regression are:

[0035]

[0036] in Let Hfinal be the i-th eigenvector of the high-order spatiotemporal correlation feature matrix Hfinal, and L be the loss function.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. This invention utilizes a bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission. First, the system acquires physiological signals through a data acquisition module. ECG, blood oxygen, and respiratory signals are collected using lead electrodes, photoelectric sensors, and a respiration sensor, respectively. These signals are converted into digital signals by a 24-bit ADC. The signal processing module generates a synchronous chaotic sequence using a Lorentz chaotic model, driving a star-shaped master-slave architecture to achieve multimodal signal time alignment. After removing dynamic noise using Kalman filtering, white noise is added according to the intelligent signal-to-noise ratio. Through empirical mode decomposition, adjacent mode difference criteria, energy entropy discrimination, and soft thresholding, the time-precise net spectral characteristic signal is reconstructed, providing high-quality data for subsequent analysis.

[0039] 2. This invention performs multimodal transformation on the temporal-linear net spectral feature signal through a fusion module, constructs a fourth-order tensor and optimizes its dimensionality reduction. Graph nodes are defined by feature dimensions and edge weights are calculated to form a normalized derivation dynamic graph structure. High-order spatiotemporal correlation features are extracted through a multimodal graph convolutional network. Gradient-weighted class activation mapping is used to locate key regions, and Shapley values ​​are used to calculate feature contribution weights. The graph structure is dynamically adjusted and iteratively optimized to output a cross-modal spatiotemporal correlation fusion feature matrix. The judgment module inputs the fused features into a logistic regression model, which generates health category probabilities through standardization, linear transformation, and softmax. Parameters are updated using an Adam optimizer combined with regularization. Finally, health status is displayed in different colors according to probability thresholds, triggering alarms and enabling real-time assessment and early warning of patient health. It effectively captures the deep interaction relationships between ECG, blood oxygen, and respiratory signals in time, frequency, and mode through the mining of spatiotemporal correlation information contained in multimodal signals. It can quantitatively assess key signal regions and feature contributions, making the health status discrimination model more accurate. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to show the main idea of ​​the present invention.

[0041] Figure 1 This is a system block diagram of the present invention;

[0042] Figure 2 This is a flowchart of the signal processing module of the present invention;

[0043] Figure 3 This is a flowchart of the fusion module of the present invention. Detailed Implementation

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0045] like Figure 1 As shown, the bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission includes an acquisition module, a signal processing module, a fusion module, and a judgment module.

[0046] The acquisition module is used to acquire the patient's electrocardiogram (ECG), blood oxygen saturation, and respiratory signals, and convert them into digital signals;

[0047] Specifically, the acquisition module includes lead electrodes, a photoelectric sensor, and a respiration sensor. The lead electrodes are attached to specific areas of the patient to acquire electrocardiogram (ECG) signals. The photoelectric sensor is clipped to the patient's finger to obtain blood oxygenation signals, and the respiration sensor is worn on the patient's chest to obtain respiratory signals. The ECG, blood oxygenation, and respiration signals are then converted into digital signals via a high-precision analog-to-digital converter (e.g., a 24-bit ADC) using dual-wavelength red light measurement for subsequent analysis. Specifically, X... ECG [n] = ADC(V ECG (t n )), X SpO [n] = ADC(V SpO (t n )), X brea [n] = ADC(V brea (t n Where n represents the discrete-time index, t n This represents the nth sampling time point, ADC(*) represents the analog-to-digital conversion function, and V ECG (t n ) represents the electrocardiogram signal, V SpO (t n ) represents the blood oxygen signal, V brea (t n () indicates a breathing signal.

[0048] like Figure 2 As shown, the signal processing module generates a synchronous chaotic sequence from the acquired digital signal using the Lorentz chaotic model combined with the fourth-order Runge-Kutta method. This sequence drives a star-shaped master-slave architecture. The signal is then input into a Kalman filter and undergoes prediction and update steps. White noise is added, and through empirical mode decomposition, adjacent mode difference criteria, energy entropy discrimination, and soft thresholding, the temporal net spectral characteristics of ECG, blood oxygen, and respiratory signals are reconstructed. The specific analysis is as follows:

[0049] A Lorentz chaotic model is constructed using the third-order autonomous ordinary differential equation of dynamics, and a fourth-order Runge-Kutta method is introduced to solve the equations, generating a synchronous chaotic sequence G. sy This is used to drive the sampling clocks of the ECG, blood oxygen, and respiration sensors, ensuring that the timestamps of the ECG, blood oxygen, and respiration signals are aligned.

[0050] The specific third-order autonomous ordinary differential equation of the dynamics is as follows: Where x, y, and z are state variables used to generate a synchronous chaotic sequence, α, β, and χ are classical chaotic parameters, x′, y′, and z′ are the first derivatives of the state variables, and ξ(t) is a Gaussian white noise term used to simulate real-world environmental disturbances. x(t i ) represents the chaotic sequence value generated at discrete time points, and I is the length of the chaotic sequence;

[0051] A chaotic sequence-driven ECG and blood oxygenation timing alignment mechanism is constructed, and a star topology master-slave architecture is built. The ECG sensor is used as the master node and as the core transmitter, transmitting a synchronous chaotic sequence G to the blood oxygenation sensor, which is used as the slave node. sy After obtaining the sequence from the node, the local clock phase is dynamically calibrated using a phase-locked loop. A time deviation Δt is defined as the objective function, and iterative optimization using the least squares method is employed until the time deviation Δt is less than a set threshold, resulting in the time-aligned ECG signal X′. ECG [n] and blood oxygen signal X′ SpO [n], and then by repeating the steps, a chaotic sequence-driven ECG and respiratory timing alignment mechanism is constructed, and a star topology master-slave architecture is built, with the ECG sensor as the master node and as the core transmitter, transmitting a synchronous chaotic sequence G to the respiratory sensor as the slave node. sy The time-aligned electrocardiogram signal X′ is obtained by the same processing described above. ECG [n] and respiratory signal X′ brea [n], and then the electrocardiogram signal X′ ECG [n], Blood oxygen signal X′ SpO [n] and respiratory signal X′ brea [n] Alignment, input the synchronized ECG signal, blood oxygen signal, and respiration signal into a Kalman filter, and through prediction and update steps, obtain the time-accurate denoising ECG signal X″. ECG [n]、Blood oxygen signal X″ SpO [n] and respiratory signal X″ brea [n];

[0052] Specifically, the process of obtaining a time-accurate noise-cleared ECG signal can be expressed by the following formula:

[0053]

[0054] x - h =Q1×x h-1 +Q2×v h-1

[0055] P - h =Q1×P h-1 ×Q1 T +O

[0056] K h =P - h ×O1 T ×(O1×P - h ×O1 T +R) -1

[0057] x h =X - h +K h ×(d h -O1×X - h )

[0058] P h =(O2-K) h ×O1)×P - h

[0059] Where argmin γ This indicates the search for the value of γ that minimizes the objective function. This represents a chaotic sequence of electrocardiogram (ECG) signals from the master node. This represents the chaotic sequence of blood oxygen signals from a node at time (t+γ), where γ represents the time delay parameter, ||*| represents the norm, Q1 represents the state transition matrix, Q2 represents the control matrix, and v h-1 Indicates control input, x - h Let x represent the predicted value of the prior state. h P represents the posterior estimate. - h Let P represent the prior covariance matrix. h-1 Let K represent the posterior covariance matrix at time k-1. h O1 represents the Kalman gain, O2 represents the observation matrix, and O2 represents the identity matrix. h The observed values ​​represent the time-aligned ECG and blood oxygen inputs, R represents the observation noise covariance matrix, and O represents the process noise covariance matrix.

[0060] By targeting the timed noise-cleared ECG signal X″ ECG [n]、Blood oxygen signal X″ SpO [n] and respiratory signal X″ brea [n] Add m sets of white noise to satisfy the intelligent signal-to-noise ratio adaptation rule. Perform empirical mode decomposition on the three types of signals after incorporating the rule-incorporated white noise to obtain the eigenmode function set of each signal. k represents the order. The average value of each intrinsic mode function in the M-group decomposition results is calculated. The intrinsic mode function sets of ECG, blood oxygen, and respiration are processed separately. The decomposition process is terminated by using the neighboring mode difference stopping criterion. High-frequency noise modes are screened by energy entropy value discrimination and dynamically removed by soft thresholding. The remaining effective modes are reconstructed to obtain the time-aligned net spectrum characteristic ECG signal, time-aligned net spectrum characteristic blood oxygen signal, and time-aligned net spectrum characteristic respiration signal.

[0061] Specifically, the intelligent signal-to-noise ratio adaptation rule is: STNR = 20log(‖X″) ECG ||2 / ||n m (t)‖2)=∈0, where‖*‖2 represents the 2-norm, the square root of the sum of squares of the signal's sample values, and n m (t) represents the m-th group of adaptive white noise, STNR represents the signal-to-noise ratio, and ∈0 represents the set target value of the signal-to-noise ratio. The specific operation of blood oxygen signal and respiratory signal is the same and will not be described again.

[0062] The stopping criterion for adjacent modal differences is STD = ∑ t |MF k (t)-MF k+1 (t)| 2 / MF k 2 (t)<δ0, where MF k (t) represents the mean of the k-th eigenmode function, STD represents the standard deviation, t represents the time point, and δ0 represents the preset threshold.

[0063] The energy entropy value discrimination method is W k =-∑|MF k (t i )| 2 / ∑MF k 2 (t)×log(|MF k (t i )| 2 / ∑MF k 2 (t))>W th If , it is determined to be a high-frequency noise mode, where i represents the index of the time point, W th Represents the energy entropy threshold, ∑MF k 2 (t) represents the total energy of the k-th MF at all time points;

[0064] Soft thresholding is handled as MF / k (t)=sign(MF k (t))×max(|MF k (t)|-β, 0), where β=ɑ0×(2logN) 1 / 2 , of which MF / k (t) represents the value of the k-th MF after denoising at time t, sign(*) represents the sign function, and max(|MF) k (t)|-β,0) represents MF kThe absolute value of (t) is thresholded. If it is equal to or less than the threshold β, it is zero. If it is greater than the threshold β, it is subtracted from the threshold β. α0 represents the standard deviation of the noise, and N represents the length of the signal.

[0065] like Figure 3 As shown, the fusion module processes the temporal net spectral features of ECG, blood oxygen, and respiratory signals, constructs a fourth-order tensor, and combines it with a sparse projection matrix to reduce the dimensionality of the multimodal spatiotemporal feature matrix. Graph nodes are defined by the matrix feature dimension, edge weights are calculated, and the normalized derived dynamic graph structure is analyzed. By constructing a multimodal graph convolutional network model, the neighborhood information of the normalized derived dynamic graph structure is analyzed to obtain a high-order spatiotemporal correlation feature matrix. The single-modal features extracted from the temporal net spectral features of ECG, blood oxygen, and respiratory signals are located and the contribution weights of the prediction results are calculated. The cross-modal spatiotemporal correlation fusion feature matrix is ​​then obtained. The specific analysis is as follows:

[0066] The time-frequency matrix W is generated from the time-frequency net spectrum characteristic electrocardiogram signal through continuous wavelet transform. ECG ∈R T×F The spectral matrix S of the time-precision net spectral characteristic blood oxygen signal was extracted by short-time Fourier transform. SpO ∈R T×F The empirical wavelet transform of the time-precision net spectral characteristic respiratory signal is decomposed into multimodal components. Extract the top three components with the largest energy percentage and construct a fourth-order tensor. Where T represents the time step, F represents the frequency band, A represents the number of respiratory rate components, and C represents the number of channels, and is filled using a fourth-order tensor filling rule, initializing the factor matrices R, L, U, and B and the core tensor G, where R ∈ R T×H1 , L∈R F×H2 ,U∈R A×H3 , B∈R C×H4 , G∈R H1×H2×H3×H4 Where T, L, A, and C represent the original dimensions of time step, frequency band, number of respiratory rate components, and number of channels, respectively, and H1, H2, H3, and H4 represent the ranks of each dimension after dimensionality reduction. The objective function is minimized through iterative optimization using the alternating direction multiplier method to obtain the optimized factor matrix and the optimized four-dimensional core tensor G′. The optimized four-dimensional core tensor G′ is flattened into a two-dimensional matrix in dimensional order. The expanded matrix is ​​multiplied by the sparse projection matrix to obtain the dimensionality-reduced multimodal spatiotemporal feature matrix Mfus.

[0067] Specifically, the fourth-order tensor filling rule is as follows: Where ζ, ψ and χ are modal weights, which are obtained through optimization. Я(t, f, a, c) represents the element value of the fourth-order tensor Я at a specific coordinate position (t, f, a, c), where t, f, a, and c represent the time step index, frequency band index, respiratory rate component index and channel index, respectively.

[0068] More specifically, the modal weights ζ, ψ, and χ are classified as vectors in a vector space. Values ​​are randomly selected from a uniform distribution in the interval [0,1] and labeled as vectors ζ0, ψ0, and χ0. ζ0 is determined as the first orthogonal vector ζ1, specifically ζ1 = ζ0. The projection of ψ0 onto the orthogonal vector ζ1 is calculated as projζ1(ψ0) = <ψ0, ζ0> / <ζ1, ζ1> × ζ1, where <, *> represent the vector dot product. The projection is subtracted from the vector ψ0 to obtain the vector ψ1 orthogonal to ζ1. The projections of χ0 onto vectors ζ1 and ψ1 are calculated using the above projection formula. The two projections are subtracted from the vector χ0 to obtain the vector χ1 orthogonal to ζ1 and ψ1. Each orthogonal vector is divided by its magnitude to obtain the optimized modal weights.

[0069] Specifically, the objective function to be optimized and minimized is: Where λ0 represents the sparse regularization coefficient, i represents the summation index, and M... i Represents the factor matrix, Let represent the square of the Frobenius norm, and ‖*‖1 represent the L1 norm. This yields the optimized four-dimensional core tensor G′ with different scale information corresponding to each dimension in terms of time scale, frequency scale, respiratory component, and channel interaction.

[0070] The specific steps for solving the dimensionality-reduced multimodal spatiotemporal feature matrix are as follows: Mfus = flatten(G′) × Ppr, where flatten(*) represents the dimensionality-order flattening operation, Ppr represents the sparse projection matrix, and Ppr ∈ R. (H1×H2×H3×H4)×d Principal component analysis is used to optimize the mapping of high-dimensional flattened vectors to d-dimensional feature space;

[0071] Based on the dimension-reduced multimodal spatiotemporal feature matrix Mfus, each column has d feature dimensions, defining graph nodes. Each node corresponds to a feature vector, and the edge weights between nodes are determined by the formula. Where ni and nj are the eigenvectors of the i-th and j-th columns of the eigenma matrix Mfus, σ0 is the Gaussian kernel parameter, which controls the weight decay rate, θij is the phase difference between the two eigenvectors extracted by Hilbert transform, which reflects the temporal correlation, and all edge weights wij constitute the dynamic graph structure S. The dynamic graph structure S is normalized by adding self-loops to obtain the normalized derived dynamic graph structure S′, S′=S+I, where I is the identity matrix, indicating the addition of self-loops;

[0072] Specifically, the dynamism of the normalized derived dynamic graph structure S′ is reflected in the training process. Each forward propagation recalculates the edge weights based on the updated reduced-dimensional multimodal spatiotemporal feature matrix Mfus, enabling the graph structure to adapt to feature changes.

[0073] A multimodal graph convolutional network model is constructed. The neighborhood information in the derived dynamic graph structure S′ is aggregated and normalized using the graph convolutional layer feature update formula in the model. After three layers of graph convolutional operations, the model is trained by jointly optimizing the multi-task objective function of the model, and outputting a high-order spatiotemporal correlation feature matrix Hfinal.

[0074] The formula for updating features in a graph convolutional layer is: H (l+1) =σ(D′) -1 / 2 S′D′ -1 / 2 ×H (l) ×K (l) ), where H (l+1) K represents the feature representation of the nodes in the (l+1)th layer. (l) Let represent the trainable parameter matrix of the l-th layer graph convolution, D′ represent the degree matrix of the aggregated normalized derived dynamic graph structure S′, and σ represent the activation function;

[0075] Joint optimization multi-task objective function: L=δ1×Lcla+γ1×Lre+σ1×Lsp, where Lcla is the cross-entropy loss function. It is the reconstruction loss. It is a sparse regularization term, and δ1, γ1 and σ1 are weight coefficients.

[0076] Key regions are identified by locating time-space net spectral features of ECG, blood oxygen, and respiration signals using gradient-weighted class activation mapping. Single-modal features extracted from these signals are used to calculate the contribution weights of the prediction results. Based on the key region location and feature contribution weights, the edge weights between nodes in the dynamic graph structure are dynamically adjusted, and the normalized derived dynamic graph structure S′ is updated. Graph convolution, feature optimization, and multi-task joint training are iteratively performed, and the model finally outputs a cross-modal spatiotemporal correlation fusion feature matrix.

[0077] Specifically, the analytical steps for determining key areas are as follows:

[0078] The time-defined net spectrum characteristics of electrocardiogram, blood oxygen, and respiration signals were segmented into segments U∈R according to time windows. T×C Calculate the gradient of segment U at each time point. Where y c It is the target category probability, U i，jGiven the signal value at time point i and channel j, global average pooling is performed on the gradients of each channel to obtain the channel importance weights. Where L is the length of the time series, a time-dimensional heatmap is generated by weighted summation, and the importance value H of the classification decision at the i-th time point is used. i It means that H i =∑ j (α j ×U i，j If the importance value H for classification decision i If the signal segment at time point i is greater than or equal to the set threshold, then the signal segment at time point i is the critical region K.

[0079] Specifically, the analytical steps for extracting single-modal features and calculating the contribution weights of the prediction results are as follows:

[0080] A multimodal feature set F is established based on the quasi-net spectrum characteristic ECG signal, time-based quasi-net spectrum characteristic blood oxygenation signal, and time-based quasi-net spectrum characteristic respiration signal, and then divided into all possible subset combinations. Where F = (f1, f2, ..., fn), and fi represents a single-modal feature, specifically the RR interval of the electrocardiogram, the standard deviation of the difference between adjacent RR intervals, the baseline fluctuation of blood oxygen, the maximum range of SpO2 value change per unit time, respiratory rate, the slope of the rising branch of the respiratory waveform, and the ratio of inspiration to expiration duration. Using the formula φ(fi) = ∑Q|Q|! (|F|-|Q|-1)! / |F|! ×[v(Q∪{fi}-v(Q)), the contribution weight φ(fi) of feature fi to the prediction result is obtained, where ! represents the factorial symbol, v(Q) represents the predicted output of the model, and the summation is performed from S = empty set, which represents the predicted output of the model when there are no feature inputs;

[0081] Specifically, the analysis steps for adjusting the edge weights between nodes in a dynamic graph structure are as follows:

[0082] For feature dimensions located within the critical region K, their feature vectors are adjusted using contribution weights, specifically: n′i=ni×(1+∑Kφ(fi)×δ(ni∈K)), where δ(*) represents an indicator function, taking a value of 1 when the feature dimension ni belongs to the critical region, and 0 otherwise. n′i represents the adjusted feature vector, and the edge weights between nodes are recalculated using the adjusted feature vector. n′j represents the adjusted eigenvector.

[0083] The judgment module is used to input the cross-modal spatiotemporal correlation fusion feature matrix into the logistic regression model. After linear transformation and softmax function, it generates the health category probability, analyzes the patient's health category probability, and displays and alarms accordingly with different colors. The specific analysis is as follows:

[0084] The cross-modal spatiotemporal correlation fusion feature matrix is ​​input into the logistic regression model, and each feature dimension is standardized. in This represents the normalized value of the eigenvector in the cross-modal spatiotemporal correlation fusion feature matrix, where μ and σ are the mean and standard deviation, respectively. The model is obtained through linear transformation. The score zi is calculated, where w is the weight vector, b is the bias term, and T represents the transpose of the matrix. The health category probability is generated by the softmax function, and the specific categories are healthy, sub-healthy, and unhealthy. The parameters are optimized by minimizing the cross-entropy loss function and combining it with the regularization term. The Adam optimizer iteratively updates w and b, and finally outputs the patient's health category probability. If the patient's health category probability is greater than 90%, it is displayed in green. If the patient's health category probability is greater than 30%, it is displayed in red and an alarm is issued. Other cases are displayed in orange.

[0085] The formula used in the above analysis is:

[0086]

[0087] Where c represents the category of physical health, and exp is an exponential function with the natural constant e as its base. Let λ0 represent the square of the L2 norm, λ0 represent the regularization coefficient, k represent the category index, traversing all health categories, P represent the probability, N represent the sample size, and M represent the number of samples. i The factor matrix representing the decomposition;

[0088] Specifically, the analysis steps for the Adam optimizer to iteratively update w and b are as follows:

[0089] Initialize the weight vector w∈R d And bias b∈R, where w is initially the parameter W of the third layer of the dynamic graph convolution. (3) ∈R d×d The main diagonal elements are used, and b is initialized to zero. The Adam optimizer is used to iteratively update the parameters. w and b are updated through the weight gradient of logistic regression until the loss converges. Finally, each dimension of w is calculated. i Representing standardized features The weight of contribution to the classification results;

[0090] Logistic regression weight gradient:

[0091]

[0092] in Let Hfinal be the i-th eigenvector of the high-order spatiotemporal correlation feature matrix Hfinal, and L be the loss function.

[0093] Specifically, it achieves high-precision synchronous acquisition and processing of bedside ECG, blood oxygen, and respiratory signals, solving the timestamp alignment problem caused by asynchronous sampling clocks in multimodal physiological signals. Through a chaotic-driven temporal alignment mechanism and techniques such as Kalman filtering and empirical mode decomposition, it effectively suppresses noise and enhances signal features. Utilizing fourth-order tensor decomposition, dynamic graph convolutional networks, and multi-task joint optimization, it overcomes the technical bottleneck of cross-temporal and spatial correlation fusion of multimodal features, accurately locating key signal regions and quantifying the contribution weight of single-modal features, achieving adaptive updating of the dynamic graph structure. Finally, through a logistic regression model, it outputs the probability of health categories in real time and triggers corresponding alarms, addressing the practical needs of real-time analysis of multimodal physiological signals, accurate assessment of health status, and abnormal early warning in clinical scenarios. It provides efficient and reliable technical support for continuous bedside monitoring and early intervention. Through iterative optimization, it forms a complete monitoring process with a closed-loop flow of signal acquisition, processing, modeling, optimization, and classification, ensuring accurate analysis of multimodal physiological signals and real-time assessment of health status.

[0094] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it 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 invention is defined by the claims and their equivalents.

Claims

1. A bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission, comprising an acquisition module, a signal processing module, a fusion module, and a judgment module, characterized in that: The acquisition module is used to acquire the patient's electrocardiogram, blood oxygenation, and respiratory signals, and convert them into digital signals; The signal processing module is used to generate synchronous chaotic sequences from the acquired digital signals by combining the Lorentz chaotic model with the fourth-order Runge-Kutta method, drive the star topology master-slave architecture, and then input the signals into the Kalman filter. After prediction and update steps, white noise is added, and the signals are reconstructed by empirical mode decomposition, adjacent mode difference criteria, energy entropy discrimination and soft thresholding to obtain the time-quasi-net spectrum features of ECG, blood oxygen and respiratory signals. The fusion module processes the temporal net spectral features of ECG, blood oxygen, and respiratory signals, constructs a fourth-order tensor, and combines it with a sparse projection matrix to reduce the dimensionality of the multimodal spatiotemporal feature matrix. Graph nodes are defined by the matrix feature dimension, edge weights are calculated, and the normalized derivation dynamic graph structure is analyzed. The neighborhood information of the normalized derivation dynamic graph structure is analyzed by constructing a multimodal graph convolutional network model to obtain a high-order spatiotemporal correlation feature matrix. The single-modal features of the temporal net spectral features of ECG, blood oxygen, and respiratory signals are located and extracted by gradient weighted class activation mapping, and the contribution weight of the prediction results is calculated. The cross-modal spatiotemporal correlation fusion feature matrix is ​​then analyzed. The judgment module is used to input the cross-modal spatiotemporal correlation fusion feature matrix into the logistic regression model, generate health category probabilities through linear transformation and softmax function, analyze the health category probabilities of patients, and display and alarm accordingly with different colors; The time-precision net spectral characteristic ECG, blood oxygen, and respiratory signal analysis steps are as follows: A Lorentz chaotic model is constructed by using a third-order autonomous ordinary differential equation of dynamics and a fourth-order numerical solution equation is introduced using the Runge-Kutta method to generate a synchronous chaotic sequence Gsy. A chaotic sequence-driven ECG and blood oxygenation timing alignment mechanism was constructed. A star-topology master-slave architecture was built, with the ECG sensor as the master node and the core transmitter, transmitting a synchronous chaotic sequence Gsy to the blood oxygen sensor as the slave node. After the slave node acquires the sequence, it dynamically calibrates the local clock phase through a phase-locked loop. The time deviation Δt is defined by the objective function, and iterative optimization is performed using the least squares method until the time deviation Δt is less than a set threshold, thus obtaining the timing-aligned ECG signal. ECG [n] and blood oxygen signal SpO [n], where n represents the sampling time point number. Then, a chaotic sequence-driven ECG-respiratory time-series alignment mechanism is constructed by repeating the steps. A star-topology master-slave architecture is built, with the ECG sensor as the master node and core transmitter, transmitting a synchronous chaotic sequence Gsy to the respiratory sensor as the slave node. Following the same processing described above, the time-aligned ECG signal is obtained. ECG [n] and respiratory signals brea [n], and then the electrocardiogram signal ECG [n], blood oxygen signal SpO [n] and respiratory signals brea [n] Alignment is performed, and the synchronized ECG, blood oxygen, and respiration signals are input into a Kalman filter. Through prediction and update steps, a time-accurate denoising ECG signal is obtained. ECG [n], blood oxygen signal SpO [n] and respiratory signals brea [n]; By adding m sets of white noise to time-accurate denoising ECG, blood oxygen, and respiration signals to satisfy the intelligent signal-to-noise ratio adaptation rule, empirical mode decomposition is performed on the three types of signals after incorporating the rule-incorporated white noise, and the intrinsic mode function set {MF} of each signal is obtained. }, k represents the order, calculate the average value of each intrinsic mode function in the m decomposition results, process the intrinsic mode function sets of ECG, blood oxygen and respiration respectively, use the neighboring mode difference stopping criterion to determine the end of the decomposition process, screen high frequency noise modes by energy entropy value discrimination method and dynamically remove them by soft thresholding, reconstruct the retained effective modes to obtain the time-precise net spectral features of ECG, blood oxygen and respiration signals.

2. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 1, characterized in that, The steps of the dimensionality-reduced multimodal spatiotemporal feature matrix analysis are as follows: The time-frequency matrix W is generated from the time-frequency net spectrum characteristic electrocardiogram signal through continuous wavelet transform. ECG ∈R T×F The spectral matrix S of the time-precision net spectral characteristic blood oxygen signal was extracted by short-time Fourier transform. SpO ∈R T×F The empirical wavelet transform of the time-precision net spectral characteristic respiratory signal is decomposed into multimodal components {Ea}. ∈R T×A Extract the top three components with the largest energy percentage and construct a fourth-order tensor Я∈R. T ×F×A×C Where T represents the time step, F represents the frequency band, A represents the number of respiratory rate components, and C represents the number of channels, and is filled using a fourth-order tensor filling rule, initializing the factor matrices R, L, U, and B and the core tensor G, where R ∈ R T×H1 , L∈R F×H2 ,U∈R A×H3 , B∈R C×H4 , G∈R H1×H2×H3×H4 Where T, L, A, and C represent the original dimensions of time step, frequency band, number of respiratory rate components, and number of channels, respectively, and H1, H2, H3, and H4 represent the ranks of each dimension after dimensionality reduction. The objective function is minimized through iterative optimization using the alternating direction multiplier method, resulting in the optimized factor matrix and the optimized four-dimensional core tensor. The optimized four-dimensional core tensor Flattened into a two-dimensional matrix in dimensional order, the expanded matrix is ​​multiplied by the sparse projection matrix to obtain the dimensionality-reduced multimodal spatiotemporal feature matrix Mfus.

3. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 1, characterized in that, The cross-modal spatiotemporal correlation fusion feature matrix analysis steps are as follows: Construct a multimodal graph convolutional network model, and aggregate and normalize the derived dynamic graph structure using the feature update formula of the graph convolutional layer in the model. The neighborhood information in the model is processed through a three-layer graph convolution operation, and the model is trained by jointly optimizing the multi-task objective function to output a high-order spatiotemporal correlation feature matrix Hfinal. Key regions are identified by locating time-net-spectrum ECG, blood oxygenation, and respiration signals using gradient-weighted class activation mapping. Single-modal features extracted from these signals are then used to calculate their contribution weights. Based on key region location and feature contribution weights, the edge weights between nodes in the dynamic graph structure are dynamically adjusted, and the normalized derived dynamic graph structure is updated. The model iteratively performs graph convolution, feature optimization, and multi-task joint training, and finally outputs a cross-modal spatiotemporal correlation fusion feature matrix.

4. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 3, characterized in that, The following steps are involved in the analysis of the structure of the naturalized derivation dynamic graph: Graph nodes are defined based on each column of the dimensionality-reduced multimodal spatiotemporal feature matrix Mfus, which has d feature dimensions. Each node corresponds to a feature vector, and the edge weights between nodes are expressed by the formula wij=exp(-‖ni-nj‖). / (2σ0)×cos(θij), where ni and nj are the eigenvectors of the i-th and j-th columns of the eigenma matrix Mfus, σ0 is the Gaussian kernel parameter, and θij is the phase difference between the two eigenvectors extracted by the Hilbert transform. All edge weights wij constitute the dynamic graph structure S. The dynamic graph structure S is normalized and a normalized derived dynamic graph structure is obtained by adding self-loops. , =S+I, where I is the identity matrix.

5. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 3, characterized in that, The key area analysis steps are as follows: The time-defined net spectrum characteristics of electrocardiogram, blood oxygen, and respiration signals were segmented into segments U∈R according to time windows. T×C Calculate the gradient of segment U at each time point. y c / U i，j , where y c It is the target category probability, U i，j Given the signal value at time point i and channel j, global average pooling is performed on the gradients of each channel to obtain the channel importance weight α. j =1 / L×∑ i ( y c / U i，j (), where L is the length of the time series, and a time-dimensional heatmap is generated by weighted summation, using the importance value H of the i-th time point for the classification decision. i It means that H i =∑ j (α) j ×U i，j If the importance value H of the classification decision i If the threshold is greater than or equal to the set threshold, then the signal segment at time point i is the critical region K.

6. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 3, characterized in that, The steps for dynamically adjusting the edge weights between nodes in the dynamic graph structure based on the feature contribution weights are as follows: A multimodal feature set F is established from the quasi-net spectrum characteristic ECG signal, the time-quasi-net spectrum characteristic blood oxygenation signal, and the time-quasi-net spectrum characteristic respiration signal, and then divided into all possible subset combinations Q. F, where F = (f1, f2, ..., fn), and fi represents the single-modal feature, using the formula (fi) = ∑ Q |Q|!(|F|-|Q|-1)! / |F|!×[v(Q∪{fi}-v(Q),to obtain the contribution weight of feature fi to the prediction result. (fi), where ! denotes the factorial symbol, and v(Q) represents the model's predicted output; For feature dimensions located within the critical region K, their feature vectors are adjusted using contribution weights: n′i=ni×(1+∑ K (fi)×δ(ni∈K)), where δ( The function () represents the indicator function, which takes a value of 1 when the feature dimension ni belongs to the key region, and 0 otherwise. ni represents the feature vector, and n′i represents the adjusted feature vector. The edge weights between nodes are recalculated using the adjusted feature vector. w′ij=exp(-‖n′in′j‖ / (2σ0)×cos(θij), where n′j represents the adjusted feature vector and w′ij represents the adjusted weight of the edges between nodes.

7. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 1, characterized in that, The different color displays and alarm analysis steps: The cross-modal spatiotemporal correlation fusion feature matrix is ​​input into the logistic regression model, and each feature dimension is standardized, n′ =(n′i-μ) / σ, where n′ This represents the normalized value of the eigenvector in the cross-modal spatiotemporal correlation fusion feature matrix, where μ and σ are the mean and standard deviation, respectively. The model is transformed by linear transformation zi=w T ×n′ The score zi is calculated using +b, where w is the weight vector, b is the bias term, and T represents the transpose of the matrix. The health category probability is generated using the softmax function, with categories being healthy, sub-healthy, and unhealthy. The parameters are optimized by minimizing the cross-entropy loss function and incorporating a regularization term. The Adam optimizer iteratively updates w and b, ultimately outputting the patient's health category probability. Different colors are used to display and alert patients based on their different health category probabilities, as shown in the following formula: P(yi=c)=softmax(z i，c )=exp(z i，c ) / ∑ k (exp(z i，k )) L=1 / N×∑ i ∑ c (y i，c ×log(P i，c ))+λ0×(‖w‖ +∑ ‖M i ‖1) Where c represents the category of physical health, and exp is an exponential function with the natural constant e as its base. || Let λ0 represent the square of the L2 norm, λ0 represent the regularization coefficient, k represent the category index, traversing all health categories, P represent the probability, N represent the sample size, and M represent the number of samples. i This represents the factor matrix of the decomposition.

8. The bedside ECG and blood oxygen synchronous monitoring system based on wireless transmission according to claim 7, characterized in that, The optimizer iteratively updates w and b in the analysis steps: Initialize the weight vector w∈R d And bias b∈R, where w is initially the parameter W of the third layer of the dynamic graph convolution. (3) ∈R d×d The main diagonal elements, b is initialized to zero, the Adam optimizer is used to iteratively update the parameters, and w and b are updated through the weight gradient of logistic regression until the weight and bias gradients are less than the set threshold. Weights and bias gradients for logistic regression: L / w=1 / N×∑ i (P i -y i )×h +2λ0×w; L / b=1 / N×∑ i (P i -y i ); Where h Let Hfinal be the i-th eigenvector of the high-order spatiotemporal correlation feature matrix Hfinal, and L be the loss function.

Images