Diver state monitoring method based on biofeedback

Abstract

The invention belongs to the technical field of diving state monitoring, and relates to a diver state monitoring method based on biofeedback, and the method comprises the steps: aligning a sign signal with environment data through a timestamp, extracting motion features through acceleration data, constructing a motion noise model, and removing signal artifacts in real time; in the aspect of EEG signal purification, a db4 wavelet basis is adopted for carrying out wavelet packet decomposition, water pressure related frequency band characteristics are extracted, low-frequency drift is removed, meanwhile, heart rate signals are analyzed through nonlinear dynamics, a Lyapunov index is extracted to evaluate HRV changes, breathing and EEG # imgabs0 # wave coupling are analyzed through wavelet transformation, and the risk evaluation precision is further enhanced; in the aspect of environment compensation, through compensation of a depth and SpO2 relational expression and water temperature on respiratory frequency, in combination with SLTM network extraction time sequence features and a deep causal convolutional layer, risk trend prediction is performed, a diver risk level and a prediction curve are output in real time, a comprehensive risk index is finally calculated, the health state and potential risk of the diver are comprehensively evaluated, and safety is improved.

Description

Technical Field

[0001] This application belongs to the technical field of diving state monitoring. More specifically, it relates to a method for monitoring the state of divers based on biofeedback. Background Art

[0002] With the popularization of underwater operations and diving activities, the health and safety issues of divers have received increasing attention. During diving, the physiological state of divers (such as heart rate, respiration, electroencephalogram activity, etc.) will be affected by various factors, such as environmental parameters like water depth, pressure, temperature, as well as the physical load and psychological state of divers. To ensure the safety of divers, real-time monitoring of their physiological state and timely risk assessment have become a key research area.

[0003] Currently, biofeedback technology has been widely used in the medical and sports fields. Especially in health monitoring, some mature technologies have been formed. Through the real-time acquisition and analysis of biological signals (such as electrocardiogram, electroencephalogram, respiration signal, etc.), the physical state of divers can be accurately evaluated, and the potential risks of divers can be predicted. However, in a complex underwater environment, how to effectively remove noise interference, improve the accuracy of signals, and on this basis, conduct efficient and accurate risk assessment is still a technical problem.

[0004] For example, divers are constantly moving in the underwater environment, which will cause biological signals (especially EEG and heart rate signals) to be severely interfered by motion artifacts. Although existing noise suppression methods (such as traditional Kalman filtering and recursive least squares method) can remove some artifacts, they still cannot effectively remove motion artifacts in a complex environment, resulting in unstable signal quality and affecting subsequent risk assessment. Summary of the Invention

[0005] The present invention provides a method for monitoring the state of divers based on biofeedback, which solves the technical problem that the prior art cannot effectively remove noise interference in a complex environment.

[0006] The method for monitoring the state of divers based on biofeedback includes the following steps: Data acquisition: Collect physiological signal data and environmental data, align the collected data with timestamps, and obtain multi-modal signal stream data aligned on the time axis. Motion noise suppression: Extract motion features from the acceleration data in the physiological signal data to construct a motion noise model, and update the filter weights of the motion noise model in real time through the recursive least squares algorithm. Remove artifacts through the motion noise model to obtain blood flow signals after removing artifacts. EEG Signal Enhancement: For the EEG signal data in the vital sign signal data, perform 6-layer wavelet packet decomposition using the db4 wavelet basis, extract features in different frequency bands from the EEG signal, identify the energy distribution in the water pressure-related frequency band, and remove the low-frequency baseline drift through thresholding to obtain the purified EEG signal; Time-domain Feature Generation: Perform non-linear dynamics analysis on the denoised heart rate signal, evaluate the complexity of the signal, construct a phase space reconstruction diagram, extract the maximum Lyapunov exponent to reflect the dynamic changes of the heart rate signal, and obtain the non-linear feature vector of HRV; Frequency-domain Feature Generation: Based on the purified EEG signal and respiratory signal, use wavelet transform to analyze the coherence between the respiratory signal and the EEG wave, and quantify the coupling degree between the respiratory signal and the EEG wave through the phase synchronization index; obtain the coupling feature of respiration and EEG; Environmental Parameter Compensation: Perform depth compensation on SpO2 based on the relationship established by depth, and compensate the respiratory rate according to the water temperature; Short-term Risk Inference: Based on the non-linear feature vector of HRV and the coupling feature of respiration and EEG, perform time-series feature extraction through the SLTM network structure, and then map it to the short-term risk probability based on the fully connected layer to output the real-time risk level; Medium- and Long-term Trend Prediction: Construct a deep causal convolution layer to predict the future risk evolution, and focus on the key time periods through the attention mechanism to output the risk trend prediction curve; Comprehensive Risk Index Output: Based on the real-time risk level and the risk trend prediction curve, perform weighted fusion using the compensated environmental parameters to obtain the comprehensive risk index.

[0007] In the present invention, in the data acquisition stage, by aligning the timestamps of the vital sign signals and environmental data, the consistency of the multi-modal signals is ensured, providing accurate data support for subsequent processing; for the signal artifacts caused by the movement of divers, the acceleration data is used to extract the motion features, construct a motion noise model, and the filter weights are updated in real time through the recursive least squares algorithm to remove the artifacts and ensure the accuracy of the blood flow signals.

[0008] In terms of EEG signal purification, perform 6-layer wavelet packet decomposition using the db4 wavelet basis, extract the features in the water pressure-related frequency band, and remove the low-frequency baseline drift through thresholding to enhance the quality of the EEG signal. Analyze the heart rate signal through non-linear dynamics, extract the maximum Lyapunov exponent to evaluate the dynamic changes of HRV, providing key features for short-term risk inference; use wavelet transform to analyze the coupling feature between the respiratory signal and the EEG wave, further improving the accuracy of risk assessment.

[0009] In terms of environmental compensation, based on the relationship between depth and SpO2 and the compensation of water temperature for respiratory rate, the accuracy of physiological data is ensured. The time series features are extracted through the SLTM network structure, and the medium- and long-term risk trend prediction is carried out by combining with the deep causal convolution layer, and the risk level and risk prediction curve of the diver are output in real time; finally, by synthesizing the real-time risk level and trend prediction curve, and combining with the compensation of environmental parameters, the comprehensive risk index is obtained, comprehensively and accurately evaluating the health status and potential risks of the diver, and effectively improving the safety of the diver.

[0010] Preferably, the motion noise model is as follows: ; where: represents the motion noise; represents the time-varying weight parameter; represents the filtered acceleration signal; The time-varying weight parameter is updated in real time through the recursive least squares algorithm. After obtaining the motion noise, the motion noise is subtracted from the original blood flow signal to obtain the blood flow signal after removing artifacts.

[0011] Preferably, the EEG signal enhancement includes the following steps: Wavelet packet decomposition: Perform 6-layer decomposition using the db4 wavelet basis. The signal is decomposed into 64 sub-bands, and each band represents the energy of the signal in a specific frequency range; based on the decomposed sub-bands, locate the water pressure interference band and the low-frequency drift band; Water pressure interference suppression: Perform energy analysis in the water pressure interference band, calculate the energy proportion of each band node. If the energy proportion of the water pressure interference band is higher than the predetermined threshold, the corresponding wavelet packet coefficients are attenuated through the soft threshold technique, and the effective components of the EEG signal are retained; Low-frequency baseline drift removal: For the baseline drift below 0.5 Hz, by performing additional decomposition on the wavelet packet tree, the components in the range of 0 - 1 Hz are removed from the signal; Wavelet packet reconstruction: Reconstruct each processed frequency point node, and obtain the purified EEG time-domain signal by synthesizing the wavelet packet coefficients layer by layer; Time-frequency feature matrix generation: Extract the energy within the time window of each node from the processed signal and perform normalization processing to obtain the time-frequency feature matrix.

[0012] Preferably, the specific steps of the time-domain feature generation are as follows: Sample entropy calculation: Based on the denoised heart rate signal, the heart rate signal is converted into a high-dimensional vector through phase space reconstruction. The embedding dimension and time delay are used in the reconstruction process, and the time delay is determined by the autocorrelation function method; then calculate the matching degree between high-dimensional vectors to obtain the sample entropy; Maximum Lyapunov exponent extraction: Based on the denoised heart rate signal, construct the phase space trajectory to obtain the trajectory. Then, find the nearest point for each point, calculate the divergence speed between neighboring points, and further calculate the maximum Lyapunov exponent; Correlation dimension calculation: Based on the denoised heart rate signal, calculate the dimensional characteristics of the phase space through the correlation function to obtain the similarity between two points in the phase space. Then, fit the relationship between the two points in the double logarithmic coordinate to obtain the correlation dimension; Sample entropy calculation: Based on the denoised heart rate signal, obtain the sample entropy value through similarity matching; Detrended fluctuation analysis: Based on the denoised heart rate signal, calculate the cumulative deviation sequence to obtain the cumulative deviation of each time point relative to the mean. Divide the sequence into multiple small intervals, calculate the local volatility within each interval, and finally obtain the DFA scaling exponent through double logarithmic fitting; The results obtained from sample entropy calculation, maximum Lyapunov exponent extraction, correlation dimension calculation, sample entropy calculation, and detrended fluctuation analysis form a five-dimensional non-linear feature vector.

[0013] Preferably, the generation of frequency domain features includes the following steps: EEG EEG wave component extraction: Extract the EEG wave component from the EEG time-frequency matrix based on neurophysiological characteristics; wave component; Continuous wavelet transform of respiratory signal: Use the complex Morlet wavelet to perform continuous wavelet transform on the respiratory signal to obtain the wavelet coefficients of the respiratory signal in the time-frequency domain, that is, the time-frequency representation of the respiratory signal; Cross-wavelet spectrum calculation: Calculate the cross-wavelet spectrum of the time-frequency representation of the respiratory signal and the EEG wave. Obtain the cross-wavelet spectrum by multiplying the time-frequency coefficients of the respiratory signal with the complex conjugate wavelet coefficients of the EEG wave, and calculate the normalized wavelet coherence based on the obtained cross-wavelet spectrum; wave; Phase synchronization index extraction: Extract the phase information from the cross-wavelet spectrum, calculate the instantaneous phase difference between the respiratory signal and the EEG wave at each time-frequency point, calculate the calculated instantaneous phase difference, calculate the global phase synchronization index within the frequency band to obtain the phase synchronization index; wave; Feature vector generation: Calculate the time average of the coherence within the frequency band based on the normalized wavelet coherence to obtain the average coherence intensity; identify continuous time periods with coherence exceeding the threshold, calculate the maximum duration; finally, calculate the variance of the phase synchronization index within the sliding window, and obtain the 3D coupling feature of respiration and EEG based on this. Based on the calculation of the normalized wavelet coherence the time average of the coherence within the frequency band to obtain the average coherence intensity; identify continuous time periods with coherence exceeding the threshold, calculate the maximum duration; finally, calculate the variance of the phase synchronization index within the sliding window, and obtain the 3D coupling feature of respiration and EEG based on this.

[0014] Preferably, the environmental parameter compensation includes the following steps: Nonlinear compensation of depth on blood oxygen saturation: A quadratic regression model is used to correct the blood oxygen saturation, where the linear term and quadratic term of the quadratic regression model compensate for the deviation and overcompensation problems caused by depth respectively; Linear compensation of water temperature on respiratory rate: A respiratory rate compensation model is established according to the Q10 effect, where linear compensation is achieved through the temperature sensitivity coefficient; Treatment of multi-parameter coupling effect: The respiratory rate is adjusted by considering the cross-action of depth and water temperature, and then the pressure gauge and sonar data are fused using a Kalman filter, and accurate depth estimation is achieved by dynamically adjusting the Kalman gain.

[0015] Preferably, the comprehensive risk index output includes the following steps: Definition of membership function: The influence of environmental parameters on the comprehensive risk is quantified through the membership function, where the depth weight gradually increases with the increase of depth, and the body temperature weight is dynamically adjusted according to the degree of deviation of the body temperature from the reference value; Weighted fusion: Weighted fusion is performed according to the results of real-time risk and prediction trend, and the final comprehensive risk index is output.

[0016] The beneficial effects of the present invention include: In the present invention, in the data acquisition stage, by aligning the timestamps of the vital sign signals and environmental data, the consistency of multi-modal signals is ensured, providing accurate data support for subsequent processing; for the signal artifacts caused by the movement of divers, the movement features are extracted using acceleration data, a movement noise model is constructed, and the filter weights are updated in real time through the recursive least squares algorithm, removing the artifacts and ensuring the accuracy of the blood flow signals.

[0017] In the purification of EEG signals, db4 wavelet basis is used for 6-layer wavelet packet decomposition, the features of the water pressure-related frequency band are extracted, and the low-frequency baseline drift is removed through thresholding, enhancing the quality of the EEG signals. The dynamic changes of HRV are evaluated by analyzing the heart rate signals through nonlinear dynamics and extracting the maximum Lyapunov exponent, providing key features for short-term risk inference; the coupling features of the respiratory signals and EEG waves are analyzed using wavelet transform, further improving the accuracy of risk assessment.

[0018] In terms of environmental compensation, based on the relationship between depth and SpO2 and the compensation of water temperature for respiratory rate, the accuracy of physiological data is ensured. The time series features are extracted through the SLTM network structure, and the medium- and long-term risk trend prediction is carried out in combination with the deep causal convolution layer, and the risk level and risk prediction curve of the diver are output in real time; finally, by integrating the real-time risk level and trend prediction curve, and combining the compensation of environmental parameters, the comprehensive risk index is obtained, comprehensively and accurately evaluating the health status and potential risks of the diver, and effectively improving the safety of the diver. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0020] Figure 1 It is the overall step block diagram provided by the embodiment of the present invention.

[0021] Figure 2 It is the EEG signal enhancement step block diagram provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer, the following further details the present application in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0023] See Figure 1 As shown, the diver status monitoring method based on biofeedback includes the following steps: Data acquisition: Collect physiological signal data and environmental data, align the collected data with time stamps to obtain multi-modal signal flow data aligned on the time axis; Wherein the physiological signal data includes optical blood flow signal, piezoelectric respiration signal, EEG electroencephalogram signal, EMG electromyogram signal; the environmental parameters include depth, temperature and motion acceleration; The time stamp alignment uses a hardware time stamp to accurately align asynchronous signals, and the time stamp accuracy is ±1ms; for signals with different sampling rates, resampling is performed to make their sampling rates the same, and the time axis alignment is ensured by using an interpolation method, such as linear interpolation or spline interpolation; finally, the multi-modal signal data stream aligned on the time axis is output.

[0024] Motion noise suppression: Extract motion features from the acceleration data in the vital sign signal data to construct a motion noise model, and update the filter weights of the motion noise model in real time through the recursive least squares algorithm. Remove artifacts through the motion noise model to obtain the blood flow signal after removing artifacts; As a possible implementation of this embodiment, assume the original optical blood flow signal contains the true blood flow signal and motion noise ; The triaxial accelerometer signal corresponds to the acceleration signals of the X, Y, and Z axes respectively; First, apply zero-phase Butterworth band-pass filtering to the accelerometer signal to filter out signal components above or below the 2 - 10 Hz frequency band; obtain the filtered acceleration signal; Represent the motion noise as a weighted combination of the triaxial acceleration signals. Therefore, the constructed motion noise model is as follows: ; Where: represents the motion noise; represents the time-varying weight parameter; represents the filtered acceleration signal; Update the time-varying weight parameter in real time through the recursive least squares algorithm. After obtaining the motion noise, subtract the motion noise from the original blood flow signal to obtain the blood flow signal after removing artifacts; When updating the time-varying weight parameter in real time through the recursive least squares algorithm, minimize the estimation error of the motion noise; construct an acceleration vector based on the filtered acceleration signal, then calculate the gain vector based on the constructed acceleration vector, calculate the a priori error based on the original optical blood flow signal and the acceleration vector, and finally calculate the current weight vector based on the calculated a priori error, gain vector, and the weight of the previous moment; Obtain the updated time-varying weight parameter based on this; How to update the time-varying weight parameter through the recursive least squares algorithm based on the determined data and the desired purpose is a conventional technical means in the art. Therefore, the implementation principle of the recursive least squares algorithm will not be elaborated in this invention.

[0025] In this embodiment, through the extraction and analysis of the acceleration data by the motion noise model, the artifacts caused by the diver's motion can be effectively removed, significantly improving the purity of the blood flow signal. Then, use the recursive least squares algorithm to update the filter weights of the motion noise model in real time, enabling the model to dynamically adapt to the changes in the diver's motion state and further improving the denoising effect.

[0026] EEG Signal Enhancement: For the EEG signal data in the vital sign signal data, perform 6-layer wavelet packet decomposition using the db4 wavelet basis, extract features in different frequency bands from the EEG signal, identify the energy distribution in the frequency bands related to water pressure, and remove the low-frequency baseline drift through thresholding to obtain the purified EEG signal; See Figure 2 As shown, as a possible implementation of this embodiment, the EEG signal enhancement includes the following steps: Wavelet Packet Decomposition: Perform 6-layer decomposition using the db4 wavelet basis. The signal is decomposed into 64 sub-bands, and each band represents the energy of the signal in a specific frequency range; Locate the water pressure interference frequency band and the low-frequency drift frequency band based on the decomposed sub-bands; Exemplarily: The original EEG signal contains underwater electromagnetic interference and low-frequency baseline drift (<0.5Hz), and the sampling rate , that is, 256 data points are sampled per second, and the Nyquist frequency ; The decomposition level L = 6, generating frequency band nodes, and the node numbering rule for each layer is the nth node in the lth layer; The frequency range of each node is as follows: ; In the formula: represents the low-frequency boundary of the nth node in the lth layer; n represents the node number; L represents the decomposition level; represents the high-frequency boundary of the nth node in the lth layer; Based on the above formula, we can obtain that the water pressure interference frequency band of 14 - 18Hz corresponds to nodes (6,7), (6,8), (6,9) (i.e., 14 - 16Hz, 16 - 18Hz, 18 - 20Hz); the low-frequency drift frequency band below 0.5Hz corresponds to node (6,0) (i.e., 0 - 2Hz); Water Pressure Interference Suppression: Perform energy analysis in the water pressure interference frequency band, calculate the energy proportion of each frequency band node. If the energy proportion in the water pressure interference frequency band is higher than a predetermined threshold, attenuate the corresponding wavelet packet coefficients through soft threshold technology to retain the effective components of the EEG signal; Exemplarily: Calculate the energy of each node for the target nodes (6,7), (6,8), (6,9): ; In the formula: represents the energy of node n; represents the wavelet packet coefficient of node n in the time domain; k represents the discrete points in the time domain; Sum up the energies calculated for each node to obtain the total energy, and then calculate the energy proportion of each node in the total energy; If the energy proportion is greater than the threshold, it is determined as significant interference; Perform soft threshold processing on the interference node coefficients: ; ; In the formula: represents the adaptive threshold; N represents the signal length; represents the noise standard deviation of node n, estimated by the minimum scale coefficient; represents sign function of; represents threshold attenuation of the coefficient. If the absolute value of the coefficient is less than then set it to 0 to retain the effective components; Removal of low-frequency baseline drift: For baseline drift below 0.5 Hz, by performing additional decomposition on the wavelet packet tree, the components from 0 - 1 Hz are removed from the signal; Exemplarily, node (6,0) contains components from 0 - 2 Hz. It is necessary to further separate the part less than 0.5 Hz. Perform an additional 1 - layer decomposition on node (6,0) to obtain sub - node (7,0) with 0 - 1 Hz and sub - node (7,1) with 1 - 2 Hz; directly discard the coefficients of sub - node (7,0), that is ; In this way, the components below 0.5 Hz are removed; Wavelet packet reconstruction: Reconstruct each processed frequency - point node, and obtain the purified EEG time - domain signal by synthesizing wavelet packet coefficients layer by layer; Exemplarily: Recursively synthesize wavelet packet coefficients from the 6th layer to the 0th layer: ; In the formula: g and h are the low - pass and high - pass filters for wavelet reconstruction (db4 wavelet basis); represent the low - frequency and high - frequency coefficients of the sub - frequency bands respectively; k, m represent discrete points in the time domain; finally, the purified EEG time - domain signal is obtained through wavelet reconstruction; Time - frequency feature matrix generation: Extract the energy within the time window of each node from the processed signal and perform normalization processing to obtain the time - frequency feature matrix.

[0027] In this embodiment, through 6 - layer decomposition using the db4 wavelet basis, the EEG signal is refined into multiple frequency bands, which can more accurately locate and suppress water - pressure interference and low - frequency drift, effectively remove the interference of water pressure on the signal. Especially when the water pressure changes greatly in the diving environment, the quality of the EEG signal can be significantly improved; and by attenuating the wavelet packet coefficients of the water - pressure interference frequency band and retaining the effective components of the EEG signal, it is beneficial to improve the signal - to - noise ratio and enhance the reliability and accuracy of the signal; for the baseline drift signal below 0.5 Hz, by performing additional decomposition on the wavelet packet tree, the low - frequency drift can be effectively removed; in this way, the low - frequency noise that may affect the signal accuracy and stability is removed, and the time - domain accuracy of the EEG signal is improved.

[0028] Time-domain feature generation: The denoised heart rate signal is used for nonlinear dynamics analysis to evaluate the complexity of the signal, and a phase space reconstruction diagram is constructed to extract the maximum Lyapunov exponent, which reflects the dynamic changes of the heart rate signal, and the nonlinear feature vector of HRV is obtained; As a possible implementation of this embodiment, the specific steps of the time-domain feature generation are as follows: Sample entropy calculation: Based on the denoised heart rate signal, the heart rate signal is converted into a high-dimensional vector through phase space reconstruction. The embedding dimension and time delay are used in the reconstruction process, and the time delay is determined by the autocorrelation function method; then the matching degree between high-dimensional vectors is calculated to obtain the sample entropy; Exemplary: Map the RR interval sequence (heart rate signal) to an m-dimensional phase space and construct a state vector: ; where: m represents the embedding dimension; represents the time delay, which is determined by the autocorrelation function method, and the delay at the first zero crossing is selected; Similarity matching: Define the distance threshold as , where represents the standard deviation of the RR interval sequence; calculate the number of template matches that meet the conditions: 1. The number of pairs of ; 2. B: The number of pairs of distances between (m + 1)-dimensional vectors , represents the distance between (m + 1)-dimensional state vectors; where the sample entropy is defined as: ; In the formula: A and B represent the number of matches in the (m - 1)-dimensional and (m + 1)-dimensional respectively. The smaller the sample entropy, the stronger the regularity of the sequence, and the larger the sample entropy, the higher the complexity of the sequence; Maximum Lyapunov exponent extraction: Based on the denoised heart rate signal, a phase space trajectory is constructed to obtain the trajectory, and then for each point, its nearest neighbor point is found, and the divergence speed between the neighboring points is calculated, and then the maximum Lyapunov exponent is calculated; Exemplary, use the reconstructed m-dimensional phase space vector to describe the phase space of the heart rate signal; For each phase point find its nearest neighbor point , satisfying ; where represents the distance between two phase space points; Track the exponential divergence of neighboring points over time: ; In the formula: Denote discrete time steps (k = 1, 2, …, K, with the maximum step K = 50), represent the number of valid tracking points; represent the divergence rate; Fit the linear part of the divergence rate curve: ; In the formula: represent the maximum Lyapunov exponent, representing the sensitivity to initial conditions; represent the function of the divergence rate varying with time; t represents the time step; c represents the fitting constant; Calculation of correlation dimension: Based on the denoised heart rate signal, calculate the dimensional characteristics of the phase space through the correlation function, obtain the similarity between two points in the phase space, and then fit the relationship between the two points in the double-logarithmic coordinate to obtain the correlation dimension; Exemplarily, calculate the correlation integral: ; where, represent the correlation integral; represent the Heaviside step function, used to judge whether the distance between two points is less than the threshold r; and both represent the state vectors in the phase space; N represents the total number of data points; By calculating whether the distance between two points in the phase space is less than the threshold r and accumulating the qualified pairs; In the double-logarithmic coordinate, select the linear region for fitting: ; In the formula: represent the correlation dimension, which is the dimensional characteristic of the phase space; represent taking the natural logarithm of the distance r; represent the constant term; Calculation of sample entropy of order 2: Based on the denoised heart rate signal, obtain the sample entropy value of order 2 through similarity matching; Exemplarily, set the embedding dimension m = 3, and other parameters are the same as the steps of sample calculation; perform phase space reconstruction and similarity matching in the same way as sample entropy calculation to obtain the sample entropy of order 2 , where and represent the number of matches when m = 3 obtained during the calculation process; Detrended fluctuation analysis: Based on the denoised heart rate signal, calculate the cumulative deviation sequence, obtain the cumulative deviation of each time point relative to the mean value, divide the sequence into multiple small intervals, and calculate the local volatility within each interval. Finally, obtain the DFA scaling exponent through double-logarithmic fitting; Exemplarily, it includes the following steps: Accumulate the deviation sequence: ; In the formula: represents the mean value of the RR interval sequence; represents the i-th RR interval, which represents the time interval between two R waves in the vector of the heart rate signal; represents the mean value of the RR interval sequence; N represents the total length of the RR interval sequence; Sub - interval fitting and fluctuation calculation: Divide the sequence into intervals of length s, and calculate the local trend fitting residual: ; In the formula: represents the local trend fitting of the v-th interval; represents the local volatility measure; s represents the interval length; represents the total number of intervals divided in the data sequence; v represents the interval index; represents the value of the cumulative deviation sequence at the k-th time point; represents the weight of each data point within the interval; Scaling exponent extraction: ; In the formula: represents the DFA scaling exponent, which is extracted by fitting the interval to extract this exponent, reflecting long - range correlation; represents the local volatility of the logarithm; represents the logarithm of the interval length s; represents the interval length range; The results obtained based on sample entropy calculation, maximum Lyapunov exponent extraction, correlation dimension calculation, second - order sample entropy calculation, and detrended fluctuation analysis form a five - dimensional non - linear feature vector.

[0029] In this embodiment, by calculating the sample entropy and the maximum Lyapunov exponent, the complexity and chaotic characteristics of the heart rate signal can be effectively evaluated. These non - linear indicators can reflect the dynamic changes of the diver's physiological state, especially the responses under stress or extreme environments, providing more in - depth information for real - time health monitoring; by calculating the correlation dimension of the heart rate signal, the diver's physiological state can be further revealed. A high - dimensional correlation dimension usually indicates a higher complexity of the signal, and this complexity may be related to the diver's physiological states such as fatigue and stress; the DFA scaling exponent can reveal the long - term dependence and local volatility of the heart rate signal, which helps to evaluate the diver's physiological adaptability in the underwater environment, especially in the case of deep diving and long - term diving; it helps to monitor the physiological changes such as the diver's fatigue and stress.

[0030] Frequency - domain feature generation: Based on the purified EEG signal and respiratory signal, use wavelet transform to analyze the respiratory signal and EEG The coherence between waves is quantified by the phase synchronization index to obtain the coupling degree between the respiration signal and the EEG waves; the coupling characteristics between respiration and EEG are obtained; As a possible implementation of this embodiment, the generation of frequency domain features includes the following steps: EEG wave extraction, extracting the wave (4 - 7 Hz) frequency component from the EEG time-frequency matrix. According to neurophysiological characteristics, the frequency range of the wave is 4 - 7 Hz. First, select the frequency components of this frequency band from the EEG time-frequency matrix: where: represents the wave component at time point t; represents the wavelet coefficient of the EEG time-frequency matrix at time point t and frequency k; t represents the time point; represents selecting from the time-frequency matrix of the EEG signal wave frequency band; Perform continuous wavelet transform on the respiration signal to extract the time-frequency characteristics related to the same frequency band as the wave. First, use the complex Morlet wavelet to perform continuous wavelet transform (CWT) on the respiration signal The Morlet wavelet basis function is defined as follows: ; where: f represents the analysis frequency, and the frequency range 4 - 7 Hz aligned with the wave is selected; represents the bandwidth factor of the wavelet, set to 6 to ensure the balance of time-frequency resolution; j represents the imaginary unit, in complex form; represents the Morlet wavelet basis function; represents the time variable, which is the time scale in the wavelet transform; represents the Gaussian envelope of the wavelet; represents the high-frequency component of the wavelet; Use Morlet to perform continuous wavelet transform on the respiration signal to obtain the wavelet coefficients of the respiration signal in the time-frequency domain: ; where: represents the wavelet coefficient of the respiration signal at time point t and frequency f; represents the complex conjugate of the Morlet wavelet; represents the respiration signal, and the value of the input respiration signal R at time point ; Calculate the respiration signal and EEG The cross-wavelet spectrum of the wave is obtained by multiplying the time-frequency coefficients of the respiratory signal by the complex conjugate wavelet coefficients of the wave signal to get the cross-wavelet spectrum: ; In the formula: is the cross-wavelet spectrum, representing the coupling strength between the respiratory signal and the EEG at time point t and frequency f wave; represents the complex conjugate of the wavelet coefficient of the Calculate the normalized wavelet coherence to reflect the synchronization degree between two signals: ; In the formula: represents the time-frequency smoothing operator; represents the coherence measure; represents the modulus value of the cross-wavelet spectrum; represents the power spectrum of the respiratory signal; represents the EEG wave power spectrum; Extract the phase information from the cross-wavelet spectrum and calculate the instantaneous phase difference between the respiratory signal and the EEG wave at each time-frequency point: ; In the formula: represents the phase angle of the respiratory signal at time point t and frequency f; represents the EEG phase angle of the wave at time t and frequency f; represents the instantaneous phase difference; represents taking the phase angle of the complex number; Calculate the global phase synchronization index within the frequency band: ; In the formula: represents the total number of sampling points within the time window; represents the central frequency of the wave; represents the phase synchronization index, and the value closer to 1 indicates stronger synchronization; represents the complex exponential representation method for calculating the complex form of the phase synchronization index; Based on the above-obtained features, construct a respiratory-EEG coupling feature vector. First, calculate the average value of coherence within the frequency band: ; In the formula: represents the average value of coherence within the frequency band; The maximum coherence duration : Identify consecutive periods with coherence exceeding the threshold (0.6) and calculate their maximum duration; Phase synchronization stability: Calculate the variance of PSI within a sliding window. The window length is set to 30 seconds and the step size is 10 seconds: ; Where: represents the number of sliding windows; represents the phase synchronization index within the i-th window; represents the average value of PSI within the sliding window; Based on this, output a three-dimensional respiratory-electroencephalogram coupling feature vector: .

[0031] In this embodiment, by extracting wave components from the EEG signal, the brain activity state of the diver can be accurately captured, especially the changes in psychological states such as stress, concentration, and relaxation, providing key electroencephalogram physiological characteristics for diver state monitoring; and by performing continuous wavelet transform on the respiratory signal and calculating its cross-wavelet spectrum with the EEG wave, the time-frequency feature coupling between the two can be obtained, revealing the interaction between respiration and brain activity, especially the impact of respiratory patterns on the diver's brain activity; secondly, by calculating the phase difference between the EEG wave and the respiratory signal and extracting the phase synchronization index, the synchrony between respiration and brain electrical activity can be revealed; high synchrony is usually related to characteristics such as the diver's concentration state and psychological stability, which is of great significance for monitoring the mental and physiological states of divers.

[0032] Environmental parameter compensation: Based on the depth, establish a relationship to perform depth compensation on SpO2, and compensate the respiratory rate according to the water temperature; As a possible implementation manner of this embodiment, the environmental parameter compensation includes the following steps: Establish a compensation model: ; Where: represents the oxygen saturation after depth compensation; represents the original oxygen saturation measurement value; D represents the underwater depth, in meters; 0.12 represents the linear compensation coefficient; -0.003 represents the quadratic compensation coefficient, used to suppress overcompensation when the depth is greater than 20 meters; Verify through an underwater pressure chamber experiment and calculate the model error: ; Where: represents the measurement value of the arterial blood gas analyzer; N represents the number of measurement samples; it is required that the error after compensation ; Linear compensation of water temperature for respiratory rate: Establish the compensation formula for respiratory rate: ; In the formula: represents the compensated respiratory rate; represents the originally measured respiratory rate; represents the current water temperature; 25 represents the reference water temperature; 0.023 represents the temperature sensitivity coefficient; When the water temperature is lower than 10°C, the compensation coefficient of the respiratory rate is adjusted to 0.015 to avoid overcompensation in the low-temperature zone; Processing of multi-parameter coupling effect: When the diver is in a deep hypothermic environment (D > 30 m, ), the combined influence of depth and water temperature is more complex, so it is necessary to introduce cross-term compensation: ; In the formula: represents the cross-compensation coefficient of depth and water temperature; To improve the depth measurement accuracy, the depth data of the pressure gauge and the sonar sensor are fused using the Kalman filter algorithm: ; In the formula: represents the filtered depth value; represents the depth value measured by the pressure gauge; represents the depth value measured by the sonar; represents the Kalman gain.

[0033] In this embodiment, by using a quadratic regression model to perform depth compensation on blood oxygen saturation, the non-linear influence of depth change on blood oxygen saturation can be effectively eliminated, ensuring the accuracy of blood oxygen monitoring, helping the diver to maintain accurate physiological data at different depths, thereby enhancing the safety monitoring ability of the diver; secondly, a compensation model is established according to the influence of water temperature on respiratory rate, so that the respiratory rate at different water temperatures can be linearly compensated, thus more accurately reflecting the physiological reactions and needs of the diver; by processing the depth and temperature cross-effect with a Kalman filter, the accuracy of depth estimation is improved, ensuring that the physiological state of the diver in a complex environment can be accurately estimated.

[0034] Short-term risk inference: Based on the non-linear eigenvector of HRV and the coupling feature of respiration and EEG, time-series feature extraction is performed through the SLTM network structure, and then it is mapped to a short-term risk probability based on the fully connected layer to output the real-time risk level; Exemplarily, the input data includes the non-linear eigenvector of HRV (5 dimensions); the respiration-EEG coupling feature (3 dimensions); The time series structure is to construct a time window at a sampling rate of 10 Hz, with a total of 100 time points; Perform Z-score normalization on each feature dimension to eliminate the differences in sensor dimensions; Input the normalized data into the LSTM, calculate the hidden state at each time step t, and pass the sequence of hidden states for 100 time steps to the next layer, retaining only the hidden state at the last time step as the time series feature; Based on the time series features output by the LSTM, use a fully connected layer to map to the risk probability; divide the risk levels based on the output probability values; that is, by setting a threshold, it can be determined which risk level the output probability value belongs to.

[0035] Medium- and long-term trend prediction: Construct a deep causal convolutional layer to predict the future risk evolution, and use the attention mechanism to focus on key time periods, outputting the risk trend prediction curve; In medium- and long-term trend prediction, the input data is a historical feature sequence, which contains 30 time steps, and each time step corresponds to the features within 10 seconds; the feature dimensions in the historical feature sequence are the same as those in short-term risk inference; Causal convolutional network construction: ; In the formula: t represents the current time step; represents the feature at the t-th time step of the output of the l-th layer; K represents the size of the convolutional kernel; d represents the dilation factor; represents the weight matrix of the k-th convolutional kernel in the l-th layer; represents the bias term of the l-th layer; Among them, causal convolution ensures that the prediction at each time step only depends on the current and previous time steps, avoiding future information leakage, and the dilation convolution d = 2 to improve the network's long-range dependence modeling ability by expanding the receptive field; Use the output of the last convolutional network to predict the risk trend in the next 3 minutes: ; In the formula: represents the output of the last causal convolutional network; represents the projection matrix, which is used to map the output of the convolutional network to 18 time steps (each time step is 10 seconds); represents the bias term; represents the normalization function; through the function normalizes the prediction value at each time step to a probability value, representing the risk probability every ten seconds within the next 3 minutes; On this basis, use the attention mechanism for optimization: Calculate the attention score: ; In the formula: represents the learnable query vector; denote the feature vector at the t-th time step (from the last convolutional network layer); Weight the context based on the calculated attention scores: ; Correct the prediction output based on the weighted context: ; where: denote the attention weight matrix; The final output is the risk probability every ten seconds within the next 3 minutes , where each denotes the risk probability at the i-th time step.

[0036] Comprehensive risk index output: Based on the real-time risk level and the risk trend prediction curve, use the compensated environmental parameters for weighted fusion to obtain the comprehensive risk index.

[0037] As a possible implementation of this embodiment: Deep weight calculation: ; where: D represents the diving depth; denote the weight of depth, which increases with the increase of depth; Body temperature weight calculation: ; where: T represents the temperature; denote the weight of body temperature, indicating overheating risk when the body temperature is higher than 28°C and hypothermia risk when it is lower than 28°C; Evaluate the reliability of the prediction trend and avoid overfitting: ; where: denote the prediction mean, representing the mean of the risk probabilities at all time steps; denote the i-th predicted value; Weighted fusion: Based on the real-time risk and the prediction trend, fuse and output the final comprehensive risk index: ; where: denote the real-time risk level; denote the maximum predicted risk within the next 3 minutes.

[0038] In the present invention, during the data acquisition stage, by aligning the timestamps of the vital sign signals and the environmental data, the consistency of the multi-modal signals is ensured, providing accurate data support for subsequent processing; for the signal artifacts caused by the diver's movement, the acceleration data is used to extract the motion features, construct a motion noise model, and the filter weights are updated in real time through the recursive least squares algorithm to remove the artifacts and ensure the accuracy of the blood flow signals.

[0039] In the aspect of EEG signal purification, db4 wavelet basis is used for 6-layer wavelet packet decomposition to extract the features of the frequency band related to water pressure, and the low-frequency baseline drift is removed through thresholding processing, enhancing the quality of the EEG signal. The heart rate signal is analyzed by nonlinear dynamics to extract the maximum Lyapunov exponent to evaluate the dynamic changes of HRV, providing key features for short-term risk inference; the wavelet transform is used to analyze the coupling characteristics between the respiratory signal and the EEG wave, further improving the accuracy of risk assessment.

[0040] In the aspect of environmental compensation, based on the relationship between depth and SpO2 and the compensation of water temperature for respiratory frequency, the accuracy of physiological data is ensured. The time series features are extracted through the SLTM network structure, and combined with the deep causal convolution layer for medium- and long-term risk trend prediction, and the risk level and risk prediction curve of the diver are output in real time; finally, by integrating the real-time risk level and trend prediction curve, and combining the compensation of environmental parameters, the comprehensive risk index is obtained, comprehensively and accurately evaluating the health status and potential risks of the diver, and effectively improving the safety of the diver.

[0041] The above are only the preferred embodiments of the present application, and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for monitoring the status of divers based on biofeedback, characterized in that, It includes the following steps: Data acquisition: Collect physiological signal data and environmental data, align the collected data with timestamps to obtain multi-modal signal flow data with time-axis alignment; Motion noise suppression: Extract motion features from the acceleration data in the physiological signal data to construct a motion noise model, and update the filter weights of the motion noise model in real time through the recursive least squares algorithm. Remove artifacts through the motion noise model to obtain blood flow signals after artifact removal; EEG signal enhancement: Perform 6-layer wavelet packet decomposition on the EEG signal data in the physiological signal data using the db4 wavelet basis, extract features in different frequency bands from the EEG signal, and identify the energy distribution in the frequency bands related to water pressure. Remove low-frequency baseline drift through thresholding to obtain purified EEG signals; Time-domain feature generation: Perform non-linear dynamics analysis on the denoised heart rate signal, evaluate the complexity of the signal, construct a phase space reconstruction diagram, and extract the maximum Lyapunov exponent to reflect the dynamic changes of the heart rate signal, obtaining a non-linear feature vector of HRV; Frequency domain feature generation: Based on the purified EEG signal and respiratory signal, wavelet transform is used to analyze the coherence between the respiratory signal and EEG waves, and the coupling degree between the respiratory signal and EEG waves is quantified by the phase synchronization index; the coupling feature between respiration and EEG is obtained; Environmental parameter compensation: Perform depth compensation on SpO2 based on the established relationship with depth, and compensate the respiratory rate according to the water temperature; Short-term risk inference: Based on the non-linear feature vector of HRV and the coupling features of respiration and EEG, extract time-series features through the SLTM network structure, and then map them to short-term risk probabilities based on the fully connected layer to output the real-time risk level; Medium- and long-term trend prediction: Construct a deep causal convolutional layer to predict the future risk evolution, and focus on key time periods through the attention mechanism to output a risk trend prediction curve; Comprehensive risk index output: Based on the real-time risk level and the risk trend prediction curve, perform weighted fusion using the compensated environmental parameters to obtain a comprehensive risk index.

2. The method for monitoring the status of divers based on biofeedback according to claim 1, characterized in that, The motion noise model is as follows: ; In the formula: represents the motion noise; represents the time-varying weight parameter; represents the filtered acceleration signal; The time-varying weight parameters are updated in real time through the recursive least squares algorithm. After obtaining the motion noise, subtract the motion noise from the original blood flow signal to obtain the blood flow signal after artifact removal.

3. The method for monitoring the status of divers based on biofeedback according to claim 1, characterized in that, The EEG signal enhancement includes the following steps: Wavelet packet decomposition: Perform 6-layer decomposition using the db4 wavelet basis. The signal is decomposed into 64 sub-bands, and each band represents the energy of the signal in a specific frequency range; Locate the water pressure interference band and the low-frequency drift band based on the decomposed sub-bands; Water pressure interference suppression: Perform energy analysis in the water pressure interference band, calculate the energy proportion of each band node. If the energy proportion in the water pressure interference band is higher than a predetermined threshold, attenuate the corresponding wavelet packet coefficients through soft threshold technology to retain the effective components of the EEG signal; Low-frequency baseline drift removal: For baseline drift below 0.5 Hz, remove the components in the range of 0 - 1 Hz from the signal by performing additional decomposition on the wavelet packet tree; Wavelet packet reconstruction: Reconstruct each processed frequency point node, and obtain the purified EEG time-domain signal by synthesizing wavelet packet coefficients layer by layer; Time-frequency feature matrix generation: Extract the energy within the time window of each node from the processed signal and perform normalization to obtain a time-frequency feature matrix.

4. The method for monitoring the status of divers based on biofeedback according to claim 1, characterized in that, The specific steps of the time-domain feature generation are as follows: Sample entropy calculation: Based on the denoised heart rate signal, the heart rate signal is converted into a high-dimensional vector through phase space reconstruction. The embedding dimension and time delay are used in the reconstruction process, and the time delay is determined by the autocorrelation function method; then, the matching degree between high-dimensional vectors is calculated to obtain the sample entropy; Maximum Lyapunov exponent extraction: Based on the denoised heart rate signal, a phase space trajectory is constructed to obtain the trajectory. Then, for each point, its nearest neighbor point is found, and the divergence speed between neighboring points is calculated, and then the maximum Lyapunov exponent is calculated; Correlation dimension calculation: Based on the denoised heart rate signal, the dimensional characteristics of the phase space are calculated through the correlation function to obtain the similarity between two points in the phase space. Then, the relationship between the two points is fitted in the double logarithmic coordinate system to obtain the correlation dimension; Second-order sample entropy calculation: Based on the denoised heart rate signal, the second-order sample entropy value is obtained through similarity matching; Detrended fluctuation analysis: Based on the denoised heart rate signal, the cumulative deviation sequence is calculated to obtain the cumulative deviation of each time point relative to the mean value. The sequence is divided into multiple small intervals, and the local volatility within each interval is calculated. Finally, the DFA scaling exponent is obtained through double logarithmic fitting; The results obtained from sample entropy calculation, maximum Lyapunov exponent extraction, correlation dimension calculation, second-order sample entropy calculation, and detrended fluctuation analysis form a five-dimensional non-linear feature vector.

5. The method for monitoring the status of divers based on biofeedback according to claim 1, characterized in that, The generation of the frequency domain features includes the following steps: EEG EEG wave component extraction: Extract EEG wave components from the EEG time-frequency matrix based on neurophysiological characteristics wave components; Continuous wavelet transform of the respiratory signal: The complex Morlet wavelet is used to perform continuous wavelet transform on the respiratory signal to obtain the wavelet coefficients of the respiratory signal in the time-frequency domain, that is, the time-frequency representation of the respiratory signal; Cross-wavelet spectrum calculation: Calculate the time-frequency representation of the respiratory signal and the cross-wavelet spectrum of the EEG wave, and obtain the cross-wavelet spectrum by multiplying the time-frequency coefficients of the respiratory signal with the complex conjugate wavelet coefficients of the EEG wave, and calculate the normalized wavelet coherence based on the obtained cross-wavelet spectrum; Phase synchronization index extraction: Extract phase information from the cross-wavelet spectrum, and calculate the instantaneous phase difference between the respiration signal and the EEG waves at each time-frequency point, and calculate the calculated instantaneous phase difference The global phase synchronization index within the frequency band is obtained to get the phase synchronization index; Feature vector generation: Based on the calculation of normalized wavelet coherence Calculate the time average of the coherence within the frequency band to obtain the average coherence intensity; identify continuous time periods with coherence exceeding the threshold and calculate the maximum duration; finally, calculate the variance of the phase synchronization index within the sliding window, and based on this, obtain the 3D coupling features of respiration and EEG.

6. The method for monitoring the status of a diver based on biofeedback according to claim 1, wherein The environmental parameter compensation includes the following steps: Non-linear compensation of depth on blood oxygen saturation: A quadratic regression model is used to correct the blood oxygen saturation. The linear term and quadratic term of the quadratic regression model compensate for the deviation and over-compensation problems caused by depth respectively; Linear compensation of water temperature on respiratory rate: A respiratory rate compensation model is established according to the Q10 effect, and linear compensation is achieved through the temperature sensitivity coefficient; Processing of multi-parameter coupling effect: The respiratory rate is adjusted by considering the cross effect of depth and water temperature, and then the pressure gauge and sonar data are fused using a Kalman filter, and accurate depth estimation is achieved by dynamically adjusting the Kalman gain.

7. The method for monitoring the status of a diver based on biofeedback according to claim 1, wherein The output of the comprehensive risk index includes the following steps: Definition of membership function: The influence of environmental parameters on the comprehensive risk is quantified through the membership function. The depth weight gradually increases with the increase of depth, and the body temperature weight is dynamically adjusted according to the degree of deviation of the body temperature from the reference value; Weighted fusion: Weighted fusion is performed according to the results of real-time risk and prediction trend, and the final comprehensive risk index is output.

Images