Real-time fatigue monitoring system based on EEG signals

Abstract

The application discloses a real-time fatigue monitoring system based on an EEG signal and relates to the technical field of neural regulation, and comprises a chaotic reconstruction module, an original EEG signal is collected, phase space reconstruction is performed on the original EEG signal by a chaotic attractor processor, a chaotic trajectory tensor is generated, and slow-band oscillation components and fast-band oscillation components are output when the Lyapunov index of the chaotic trajectory tensor exceeds a preset index threshold; and a closed-loop termination module, which terminates physical light field projection when the updated vigilance drop probability value, the cognitive overload probability value and the physiological fatigue probability value are all lower than a preset safety threshold in a continuous detection period.

Description

Technical Field

[0001] The present invention relates to the field of neural regulation technology, and in particular to a real-time fatigue monitoring system based on EEG signals. Background Art

[0002] In recent years, fatigue monitoring technology based on EEG signals has rapidly developed in areas such as driving safety and high-risk operations. Existing technologies primarily use frequency domain analysis to extract fatigue-related EEG features and combine them with eye tracking to identify fatigue states. Some advanced researchers have implemented machine learning models to fuse and analyze multimodal physiological signals, using visual stimulation interventions to alleviate fatigue and blue light pulses to modulate alertness neural pathways, demonstrating the feasibility of neurofeedback interventions.

[0003] Existing technologies lack robustness against artifacts in non-steady-state EEG signals. Myoelectric and eye movement artifacts lead to excessive frequency aliasing, distorting fatigue feature separation. Traditional frequency-domain filters exhibit significant spectral leakage at the theta / alpha band boundary, making it difficult to separate fatigue-related theta waves from artifact-laden alpha waves. Summary of the Invention

[0004] In order to solve the problem of aliasing distortion of slow wave and fast wave frequency bands caused by electromyographic artifact interference in non-steady-state EEG signals, the present invention provides a real-time fatigue monitoring system based on EEG signals.

[0005] The present invention provides a real-time fatigue monitoring system based on EEG signals, comprising:

[0006] The chaos reconstruction module is used to collect the original EEG signal, perform phase space reconstruction on the original EEG signal through the chaotic attractor processor, generate a chaotic trajectory tensor, and output a slow frequency band oscillation component and a fast frequency band oscillation component when the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold;

[0007] an adversarial discrimination module, configured to input the slow-frequency oscillation component and the fast-frequency oscillation component into a generator unit of a generative adversarial network to generate a synthetic feature vector, and simultaneously transmit the synthetic feature vector in parallel to an alertness discrimination unit, a cognitive load discrimination unit, and a physiological fatigue discrimination unit to generate a probability value of decreased alertness, a probability value of cognitive overload, and a probability value of physiological fatigue;

[0008] A visual tracking module is used to capture the corneal reflection point and generate the visual focus plane coordinates when the maximum value of any probability value exceeds a preset probability value threshold;

[0009] The light field construction module is used to calculate the field dynamic parameters of the dynamic optical flow field based on the visual focus plane coordinates and the maximum value of any probability value, and generate the optical flow field equation;

[0010] The optical neural control module is used to convert the optical flow field equation into a physical light field and project it onto the user's retinal fovea, and update the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue;

[0011] The closed-loop termination module is used to terminate the physical light field projection when the updated probability value of alertness decrease, cognitive overload probability value and physiological fatigue probability value are all lower than the preset safety threshold in the continuous detection cycle.

[0012] As a preferred solution of the present invention, the steps of generating the chaotic trajectory tensor are as follows:

[0013] The collected original EEG signal is decomposed into multiple scales by a recursive Gaussian kernel integrator to generate multiple frequency band time domain sub-signals;

[0014] According to the instantaneous energy gradient ratio of the time domain sub-signals in each frequency band, the optimal embedding dimension is calculated within the dynamic time window;

[0015] The time-domain sub-signals of each frequency band and the corresponding optimal embedding dimensions are input into the recursive tensor coupler to generate a chaotic trajectory tensor.

[0016] As a preferred solution of the present invention, the outputting of the slow frequency band oscillation component and the fast frequency band oscillation component comprises the following specific steps:

[0017] The instantaneous exponential spectrum calculation based on the time window is performed on the chaotic trajectory tensor, and the maximum component of the exponential value of each dimension is extracted as the Lyapunov index;

[0018] When the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold, the frequency band components are decoupled through an oscillation coupling separator, and a slow frequency band oscillation component and a fast frequency band oscillation component are output;

[0019] When the Lyapunov exponent of the chaotic trajectory tensor does not exceed the preset exponential threshold, the embedding dimension attenuation coefficient of the phase space reconstruction is updated and the chaotic trajectory tensor is regenerated.

[0020] As a preferred solution of the present invention, the steps of generating a synthetic feature vector are as follows:

[0021] Input the slow frequency band oscillation component and the fast frequency band oscillation component into the feature tensor folder to generate a frequency band fusion tensor;

[0022] The frequency band fusion tensor is input into the quantum probability encoder and transformed into a quantum state vector through the quantum entanglement gate;

[0023] The quantum state vector is input into the Hamiltonian driven generator, a unitary matrix transformation is performed, a multi-basis Pauli operator measurement is performed on the quantum state vector after the unitary matrix transformation, and a synthetic eigenvector is output.

[0024] As a preferred embodiment of the present invention, the steps of generating the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue are as follows:

[0025] The synthesized feature vector is spatiotemporally encoded to generate a spatiotemporal feature tensor, which is then fed into the attention mask controller to generate a weighted feature tensor.

[0026] The weighted feature tensor is simultaneously fed into three differential decision tree engines to independently calculate the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue;

[0027] When the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue meet the conservation constraints, they are output; otherwise, the weighted feature tensor is regenerated.

[0028] As a preferred solution of the present invention, the steps of capturing corneal reflection points to generate visual focus plane coordinates are as follows:

[0029] When the maximum value of any probability value exceeds the preset probability value threshold, the annular light source array is activated and the light source intensity and spatial distribution are dynamically adjusted to output a light intensity modulation signal;

[0030] Based on the light intensity modulation signal, a three-wavelength infrared camera is used to synchronously collect corneal reflection points and extract multispectral reflection features. The multispectral reflection feature matrix is ​​output and a reflection feature map is generated through weighted fusion.

[0031] According to the reflection feature map, the Lucas-Kanade optical flow method is used to calculate the eyelid contour motion velocity vector and output the eyelid motion parameters.

[0032] The eyelid motion parameters and the reflection feature map are vector dot producted to generate the compensated pupil center offset vector, and spherical integral mapping is performed to generate the visual focus plane coordinates.

[0033] As a preferred solution of the present invention, the field dynamics parameters include intensity attenuation coefficient, spatial wave number and retinal projection distance parameters.

[0034] As a preferred solution of the present invention, the field dynamics parameters of the dynamic optical flow field are calculated to generate the optical flow field equation, and the specific steps are as follows:

[0035] Calculate the intensity attenuation coefficient according to the maximum value of the depth component of the visual focus plane coordinates and any probability value;

[0036] According to the spatiotemporal gradient of the visual focal plane coordinates and the pre-calibrated retinal foveal coordinates, the spatial wave number and retinal projection distance parameters are calculated;

[0037] The intensity attenuation coefficient, spatial wave number and retinal projection distance parameters are input into the viscoelastic field equation to generate the optical flow field equation.

[0038] As a preferred solution of the present invention, the outputting of the updated probability value of alertness decrease, the probability value of cognitive overload, and the probability value of physiological fatigue may be performed in the following steps:

[0039] The optical flow field equation is input into the quantum state amplitude converter to generate the quantum state amplitude distribution, and the phase hologram adapted to the fovea of ​​the retina is generated through the curvature compensation phase modulator;

[0040] The phase hologram is synthesized into a physical light field and projected onto the fovea of ​​the user's retina. After the projection is completed, the EEG signal is collected to obtain the differential energy ratio of the oscillation component of the frequency band before and after the stimulation;

[0041] Based on the differential energy ratio and the preset weight vector, the updated probability value of alertness decrease, the probability value of cognitive overload and the probability value of physiological fatigue are output.

[0042] As a preferred solution of the present invention, the specific steps of terminating the physical light field projection are as follows:

[0043] According to the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue updated during the continuous detection cycle, the changing trends of each probability value are analyzed and a multidimensional data structure is constructed;

[0044] Perform time series pattern recognition and related entropy evaluation on multidimensional data structures and output decision indicators;

[0045] When the continuous detection cycles of the decision indicators are all lower than the preset safety threshold, a quantum state conversion command signal is generated to control the physical light field to terminate the projection.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] This method maps the raw EEG signal into a three-dimensional chaotic trajectory tensor space through multi-scale recursive phase space reconstruction. Lyapunov exponent thresholds are used to trigger dynamic frequency band decoupling. A recursive Gaussian kernel integrator is used to perform nonlinear separation of the δ / θ and β / γ frequency bands, suppressing transient artifacts. Dynamic Lyapunov exponent analysis also improves the success rate of frequency band separation at low levels of chaos, ensuring stable output of valid frequency band components. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Fig. 1 The module diagram of the real-time fatigue monitoring system based on EEG signals.

[0050] Fig. 2 Flowchart for chaotic trajectory tensor generation.

[0051] Fig. 3 Flowchart for quantum state feature synthesis.

[0052] Fig. 4 This is a flow chart of the closed-loop control of the physical light field. DETAILED DESCRIPTION

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0054] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0056] Reference Figs. 1-4 , is an embodiment of the present invention, which provides a real-time fatigue monitoring system based on EEG signals, comprising the following steps:

[0057] The chaos reconstruction module is used to collect the original EEG signal, perform phase space reconstruction on the original EEG signal through the chaotic attractor processor, generate a chaotic trajectory tensor, and output a slow frequency band oscillation component and a fast frequency band oscillation component when the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold.

[0058] Furthermore, the collected original EEG signal is decomposed into multiple scales by a recursive Gaussian kernel integrator to generate multiple frequency band time domain sub-signals;

[0059] Specifically, a wide-window Gaussian kernel is used to low-pass filter the original EEG signal to extract the slow frequency band oscillation component. The medium frequency band oscillation component and the fast frequency band oscillation component are separated in sequence by gradually reducing the Gaussian kernel window size. The Gaussian kernel convolution result of each scale is processed by zero-phase filtering to eliminate time shift and output the δ frequency band time domain sub-signal, θ frequency band time domain sub-signal, α frequency band time domain sub-signal, β frequency band time domain sub-signal and γ frequency band time domain sub-signal.

[0060] Specifically, according to the instantaneous energy gradient ratio of the time domain sub-signals in each frequency band, the optimal embedding dimension is calculated within the dynamic time window, and the expression is:

[0061] ;

[0062] Where, Indicates the The optimal embedding dimension of the time domain sub-signal of each frequency band, Indicates the frequency band index, represents the time domain sub-signal, Indicates taking the minimum value, Indicates the history The gain coefficient of the time domain sub-signal in each frequency band, Represents the ceil function, represents the Gaussian error function, represents the normalization constant, represents the dynamic time window integral, Indicates the current time, represents the integration time window, Indicates the starting point of the time window, represents the spatial gradient operator, Indicates the The spatial gradient of the time domain sub-signal in each frequency band, Indicates the moment, Indicates the The time domain sub-signals of the frequency band are The modulus of the space vector at time, Indicates the The frequency band noise threshold of the time domain sub-signal of each frequency band, Indicates the The time domain sub-signals of the frequency band are The cube root of the moment amplitude, represents the moment differential, represents the hyperbolic tangent function, represents the gradient amplification factor, Indicates the The maximum gradient modulus within the time window of the spatial gradient of the time domain sub-signal of the frequency band, Represents the dynamic gain factor.

[0063] It should be noted that the first The gain coefficient of the time-domain sub-signal of each frequency band is based on the spectrum attenuation characteristics of the neural signal, and the nonlinear energy loss of high-frequency oscillation is compensated by the power law function. The example value is when hour, The gradient amplification factor is set according to the saturation threshold of the maximum gradient modulus of the cortical signal, and the saturation threshold is an empirical parameter predetermined by the physical properties of the signal and the experimental characteristics. The example value is 0.3; the dynamic gain factor is dynamically generated by the instantaneous gradient energy integral and the noise floor, reflecting the time-varying characteristics of the excitability of the neural cluster. The example value range is [0.3, 1.8]; The frequency band noise threshold of the time domain sub-signal of each frequency band is set based on the fixed ratio of the root mean square of the time domain sub-signal of each frequency band. The example value is low frequency band .

[0064] The time-domain sub-signals of each frequency band and the corresponding optimal embedding dimensions are input into the recursive tensor coupler to generate a chaotic trajectory tensor.

[0065] Specifically, the recursive tensor coupler inputs the δ-band time domain sub-signal, θ-band time domain sub-signal, α-band time domain sub-signal, β-band time domain sub-signal and γ-band time domain sub-signal with the corresponding optimal embedding dimensions into the multi-dimensional phase space reconstruction algorithm, and maps each frequency band time domain sub-signal into a high-dimensional phase space trajectory through the delayed coordinate embedding method. The phase space trajectories of the time domain sub-signals in each frequency band are coupled across frequency bands through the Kronecker product operation to generate a chaotic trajectory tensor.

[0066] Specifically, the instantaneous exponential spectrum calculation based on the time window is performed on the chaotic trajectory tensor, and the maximum component of the exponential value of each dimension is extracted as the Lyapunov index, which is expressed as:

[0067] ;

[0068] Where, Indicates the current time Time dimension The instantaneous Lyapunov exponent of Represents the dimension index, represents the instantaneous Lyapunov exponential calculation window length, represents the normalized time factor, represents the natural logarithm function, Representation Dimension The weight coefficient of represents the instantaneous time window integral, Indicates the end time of integration, Indicates at time When the chaotic trajectory tensor is The component of the dimension, represents the time decay kernel, represents the decay rate constant, Indicates time With the current time The absolute value of the difference, represents the minimum divergence basis, represents the sum of the divergences of the first three main dimensions, Indicates the current time Time dimension and interactive dimensions The trajectory divergence of represents the interaction dimension index, represents the normalized Sigmoid function, Indicates the cutoff frequency of the effective component of the EEG signal, Indicates the EEG signal sampling rate.

[0069] It should be noted that the normalized time factor is based on the finite time Lyapunov exponent theory of phase space trajectory, and the window length is calculated by the instantaneous Lyapunov exponent The reciprocal of normalizes the cumulative exponential growth to the rate of change per unit time, and the example value is 2; the dimension The weight coefficient is determined by principal component analysis of phase space reconstruction. The example value is when the variance of the first three dimensions of the EEG signal accounts for [0.6, 0.25, 0.15]. , , .

[0070] When the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold, the frequency band components are decoupled through an oscillation coupling separator, and a slow frequency band oscillation component and a fast frequency band oscillation component are output;

[0071] It should be noted that the preset index threshold is set based on the percentile of the Lyapunov exponent distribution of resting-state EEG signals of healthy people, and the example value is 0.372 bits per second.

[0072] Specifically, when the oscillation coupling separator detects that the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold, it decomposes the chaotic trajectory tensor into independent modal components including the δ-band time domain sub-signal, the θ-band time domain sub-signal, the α-band time domain sub-signal, the β-band time domain sub-signal and the γ-band time domain sub-signal. The slow frequency band oscillation component is generated by weighted fusion of the δ-band time domain sub-signal and the θ-band time domain sub-signal, and the fast frequency band oscillation component is generated by principal component analysis fusion of the β-band time domain sub-signal and the γ-band time domain sub-signal.

[0073] When the Lyapunov exponent of the chaotic trajectory tensor does not exceed the preset exponential threshold, the embedding dimension attenuation coefficient of the phase space reconstruction is updated and the chaotic trajectory tensor is regenerated.

[0074] Specifically, when the Lyapunov exponent of the chaotic trajectory tensor does not exceed the preset exponential threshold, the gradient descent optimization algorithm is used to adjust the embedding dimension attenuation coefficient of the phase space reconstruction, and the updated embedding dimension attenuation coefficient is input into the recursive tensor coupler to recalculate the chaotic trajectory tensor.

[0075] The adversarial discrimination module is used to input the slow frequency band oscillation component and the fast frequency band oscillation component into the generator unit of the adversarial generative network to generate a synthetic feature vector, and at the same time transmit the synthetic feature vector in parallel to the alertness discrimination unit, the cognitive load discrimination unit and the physiological fatigue discrimination unit to generate the probability value of alertness decrease, the probability value of cognitive overload and the probability value of physiological fatigue.

[0076] Furthermore, the slow frequency band oscillation component and the fast frequency band oscillation component are input into the feature tensor folder to generate a frequency band fusion tensor;

[0077] Specifically, after the slow frequency band oscillation component and the fast frequency band oscillation component are time-aligned preprocessed, the input feature tensor folder performs a third-order tensor outer product operation, where the slow frequency band oscillation component is expanded into a two-dimensional matrix along the time axis, and the fast frequency band oscillation component is expanded into a two-dimensional matrix along the frequency axis. Through cross-scale coupling, a frequency band fusion tensor is generated.

[0078] The frequency band fusion tensor is input into the quantum probability encoder and transformed into a quantum state vector through the quantum entanglement gate;

[0079] Specifically, the frequency band fusion tensor is mapped to the high-dimensional Hilbert space through the tensor direct product operation to form the initial quantum state. The initial quantum state is input into the quantum entanglement gate network composed of a cascade of controlled NOT gates and Hadamard gates for unitary transformation. The superposition state after the unitary transformation is the quantum state vector.

[0080] The quantum state vector is input into the Hamiltonian driven generator, a unitary matrix transformation is performed, a multi-basis Pauli operator measurement is performed on the quantum state vector after the unitary matrix transformation, and a synthetic eigenvector is output.

[0081] Specifically, the quantum state vector is input into the Hamiltonian-driven generator to perform a unitary matrix transformation implemented by a parameterized quantum circuit. The quantum state vector after the unitary matrix transformation sequentially undergoes projection operations using the Pauli X operator basis measurement, the Pauli Y operator basis measurement, and the Pauli Z operator basis measurement. The projection results of the three basis measurements are linearly combined according to a fixed weight ratio, and the combined scalar sequence constitutes a synthetic eigenvector.

[0082] It should be noted that the fixed weight ratio refers to the constant contribution coefficient of the Pauli X operator basis measurement expectation value, the Pauli Y operator basis measurement expectation value, and the Pauli Z operator basis measurement expectation value in the synthetic eigenvector.

[0083] The synthesized feature vector is spatiotemporally encoded to generate a spatiotemporal feature tensor, which is then fed into the attention mask controller to generate a weighted feature tensor.

[0084] Specifically, the synthesized feature vector is spatiotemporally encoded through a combination of segmented position encoding and a time-domain convolutional network. The position encoding generates a timestamp tag matrix, and the time-domain convolutional network extracts the frequency-time joint features. The timestamp tag matrix and the frequency-time joint features are then concatenated into a four-dimensional spatiotemporal feature tensor. The four-dimensional spatiotemporal feature tensor is input into the attention mask controller to implement a dual-channel gating mechanism. Channel one uses dot-product attention based on inner products to calculate the temporal relevance score, while channel two uses a one-dimensional convolution kernel to weight the spatial features. The outputs of the two channels are fused into an attention mask matrix through an affine transformation. The attention mask matrix is ​​then element-wise multiplied with the four-dimensional spatiotemporal feature tensor to generate a weighted feature tensor.

[0085] Specifically, the weighted feature tensor is simultaneously input into three differential decision tree engines to independently calculate the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue. The expression is:

[0086] ;

[0087] Where, Indicates the probability value of decreased alertness, represents the probability value of cognitive overload, represents the probability value of physiological fatigue, represents the differential decision tree integration function, represents the adaptive time window integration operation, represents the quantum decoherence time window, represents the quantum bubble niche projection operator, represents the weighted feature tensor, represents the Pauli projection quantization operation, represents the Hadamard product, represents the exponential decay kernel function, represents the exponential decay rate constant, Indicates the absolute time offset, represents the starting time of quantum observation, represents the Hadamard element except, represents a numerical stability constant, represents the real part of the complex function, represents the quantum entangled unitary matrix, represents the tensor convolution operation, represents the Dirac function sampler, represents the Dirac characteristic sampling function.

[0088] When the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue meet the conservation constraints, they are output; otherwise, the weighted feature tensor is regenerated.

[0089] Specifically, when the absolute difference between the arithmetic sum of the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue and the fixed deviation threshold is less than the preset tolerance, it is determined that the conservation constraint is met. At this time, the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue are output. Otherwise, a feedback signal is sent to the attention mask controller to trigger the regeneration of the weighted feature tensor.

[0090] It should be noted that the fixed deviation threshold is set based on the arithmetic sum and mean of the probability values ​​of decreased alertness, cognitive overload and physiological fatigue of healthy people in a resting state with eyes closed, and the example value is 0.95; the preset tolerance is set based on the measurement accuracy of the EEG signal acquisition equipment and the individual physiological variation range, and the example value is ±0.05; the conservation constraint condition means that the absolute difference between the arithmetic sum of the probability values ​​of decreased alertness, cognitive overload and physiological fatigue and the fixed deviation threshold is less than or equal to the preset tolerance.

[0091] The visual tracking module is used to capture the corneal reflection point and generate the visual focus plane coordinates when the maximum value of any probability value exceeds a preset probability value threshold.

[0092] Furthermore, when the maximum value of any probability value exceeds a preset probability value threshold, the annular light source array is activated and the light source intensity and spatial distribution are dynamically adjusted to output a light intensity modulation signal;

[0093] It should be noted that the preset probability value threshold is set based on the maximum duration threshold when healthy people perform standard cognitive tasks, and the example value is 0.6.

[0094] Specifically, when any probability value of the probability value of decreased alertness, the probability value of cognitive overload or the probability value of physiological fatigue exceeds the preset probability value threshold, the ring light source array activation instruction is triggered. The degree of excess of any probability value is proportional to the light source intensity control factor of the ring light source array. The light source intensity control factor drives the PWM modulation circuit to generate a pulse width modulation signal to control the driving current intensity of each sub-light source array. The spatial distribution is allocated to the corresponding sub-light source array according to the probability value type. When the probability value of decreased alertness exceeds the standard, the inner ring near-infrared LED light source group turns on the radial scanning mode. When the probability value of cognitive overload exceeds the standard, the middle ring red light LED light source group executes the circular pulsation mode. When the probability value of physiological fatigue exceeds the standard, the outer ring infrared LED light source group is activated for spiral gradient enhancement. The dynamic control result outputs a light intensity modulation signal.

[0095] Based on the light intensity modulation signal, a three-wavelength infrared camera is used to synchronously collect corneal reflection points and extract multispectral reflection features. The multispectral reflection feature matrix is ​​output and a reflection feature map is generated through weighted fusion.

[0096] Specifically, the light intensity modulation signal triggers the three-wavelength infrared camera to expose synchronously in the near-infrared, red and far-infrared bands. The three-wavelength infrared camera captures the dynamic sequence of corneal reflection points at a high frame rate. The reflection intensity matrix of each band is decomposed into amplitude component and phase component through the time-frequency analysis algorithm. The amplitude component of the near-infrared band and the phase component of the red light band are mixed in the complex domain to generate a fusion feature layer. The amplitude component of the far-infrared band is band-pass filtered to extract high-frequency detail features. The fusion feature layer is weighted linearly superimposed according to the optical transmittance characteristics of the corneal tissue. The superposition result is processed by noise reduction and then output as a reflection feature map.

[0097] Specifically, according to the reflection feature map, the Lucas-Kanade optical flow method is used to calculate the eyelid contour motion velocity vector and output the eyelid motion parameters, which are expressed as follows:

[0098] ;

[0099] Where, represents the eyelid movement parameters, Indicates the integration of the eyelid contour area. Indicates the eyelid contour area, represents the reflection feature map, represents the horizontal coordinate in space, represents the vertical coordinate in space, represents the natural exponential function, represents the square of the dynamic characteristic displacement vector, represents the Gaussian kernel scaling factor, represents the spatial transverse coordinate differential element, represents the spatial longitudinal coordinate differential element, represents the sum of multi-spectral channels, represents the spectral channel index, Represents spectral channels The weight of Represents spectral channels The spectral reflectance characteristic matrix, represents the Hilbert–Schmidt norm.

[0100] It should be noted that the Gaussian kernel scaling factor is based on the standard deviation statistics of eyelid micro-movements in healthy people and is determined by fitting the blink displacement distribution through maximum likelihood estimation. The example value is 1.7; spectral channel The weight is distributed according to the difference in corneal reflectivity of infrared light of different wavelengths and the proportion of spectral energy. For example, if it is the infrared spectral channel , if it is the red light spectrum channel , if it is the near-infrared spectral channel .

[0101] The eyelid motion parameters and the reflection feature map are vector dot producted to generate the compensated pupil center offset vector, and spherical integral mapping is performed to generate the visual focus plane coordinates.

[0102] Specifically, the eyelid motion parameters and the reflection feature map are used to generate a joint feature matrix through the Kronecker product operation. Each column vector of the joint feature matrix is ​​subjected to a dot product operation with the pre-calibrated eye rotation basis vector to output the pupil center offset vector. The pupil center offset vector is input into the spherical integral mapper, and the offset vector is surface integrated in the unit spherical coordinate system. The spherical integral result is mapped to a two-dimensional plane through orthogonal projection to generate the visual focus plane coordinates.

[0103] The light field construction module is used to calculate the field dynamic parameters of the dynamic optical flow field based on the visual focus plane coordinates and the maximum value of any probability value, and generate the optical flow field equation.

[0104] Furthermore, the field dynamics parameters include intensity attenuation coefficient, spatial wave number and retinal projection distance parameters.

[0105] It should be noted that the intensity attenuation coefficient represents the optical flow energy attenuation rate under the combined effect of the depth component of the visual focus plane coordinates and the maximum brain state probability;

[0106] Spatial wave number refers to the frequency of optic flow phase oscillation induced by the difference between the coordinate gradient of the quantified visual focal plane and the coordinate of the retinal fovea;

[0107] The retinal projection distance parameter describes the effective optical projection path length from the visual focal plane to the retinal surface.

[0108] Specifically, the intensity attenuation coefficient is calculated based on the maximum value of the depth component of the visual focus plane coordinates and any probability value. The expression is:

[0109] ;

[0110] Where, represents the intensity attenuation coefficient, represents the depth absorption constant, Represents the depth component of the visual focus plane coordinates, represents the depth component, represents the Gaussian kernel normalization term, represents the standard deviation of visual depth distribution, represents the probability cumulative integral, represents the maximum value of any probability value, represents the probability density function, represents the probability integral dummy variable, The health baseline value representing the probability of brain state, represents the squared measure of probability deviation, represents the standard deviation of the probability distribution, represents the probability element, represents the depth modulation factor, represents the depth normalization term.

[0111] It should be noted that the intensity attenuation coefficient is based on the optical-neural joint attenuation model of the visual focus depth component and the maximum brain state probability. The example value is: =1.0, =0.6, The depth absorption constant is the depth-dependent attenuation rate constant fitted from the absorption spectra of infrared light by the cornea and aqueous humor, and the example value is 0.48; the depth modulation factor is the depth response saturation parameter determined by the refractive accommodation limit of the eye lens, and the example value is 0.33.

[0112] Specifically, according to the spatiotemporal gradient of the visual focus plane coordinates and the pre-calibrated retinal fovea coordinates, the spatial wave number and retinal projection distance parameters are calculated, and the expression is:

[0113] ;

[0114] Where, represents the spatial wave number, represents the retinal projection distance parameter, represents the visual focus plane coordinate vector, represents the visual focus plane coordinate vector, Represents the visual focus plane coordinate gradient vector, represents the vector cross product, represents the pre-calibrated foveal coordinates, represents the spatial scale factor, represents the wave field coupling coefficient, represents the retinal surface, Integrate the retinal surface, represents the plane wave phase term, represents the imaginary unit, represents the wave vector, represents the position vector, Represents the retinal surface differential element.

[0115] It should be noted that the spatial scale factor is derived from the spatial density of cone cells in the macula of the retina. For example, the value of the fovea of ​​a standard adult eye is 0.17, and it rises to 0.45 in the peripheral area of ​​the retina. The wave field coupling coefficient is calculated by the phase locking value of the EEG gamma wave and the retinal ganglion discharge. For example, the value range of the alert state is [0.15, 0.22], and it drops to [0.02, 0.05] in deep sleep.

[0116] Specifically, the intensity attenuation coefficient, spatial wave number, and retinal projection distance parameters are input into the viscoelastic field equation to generate the optical flow field equation, which is expressed as follows:

[0117] ;

[0118] Where, represents the optical flow field equation, represents the three-dimensional space coordinates, Represents the partial derivative of the optical flow velocity field with respect to the current time, represents the Laplace operator, represents the optical flow velocity field, represents the divergence operator, represents the pressure difference phase velocity term, represents the viscoelastic coefficient, represents the convection term, represents the neural relaxation time constant, Represents the normalized time factor.

[0119] It should be noted that the viscoelastic coefficient is calibrated by the dynamic resistance of the vitreous body and the delay characteristics of the retinal ganglion cells, and the example value is 0.62; the normalized time factor is the neural relaxation time constant, measured by the synaptic transmission delay in the thalamocortical visual pathway, for example in the alert state .

[0120] The optical neural control module is used to convert the optical flow field equation into a physical light field and project it onto the user's retinal fovea, and update the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue.

[0121] Furthermore, the optical flow field equation is input into a quantum state amplitude converter to generate a quantum state amplitude distribution, and a phase hologram adapted to the fovea of ​​the retina is generated through a curvature compensation phase modulator;

[0122] Specifically, the optical flow field equation is mapped to the quantum Hilbert space through the Schrödinger operator to form the initial wave function. The initial wave function is input into the quantum state amplitude converter to perform amplitude modulation based on the Pauli Z gate and quantum Fourier transform to generate a quantum state amplitude distribution. The quantum state amplitude distribution is input into the curvature compensation phase modulator for three-step processing: the first step is to use the Fresnel diffraction integral to calculate the wavefront curvature of the retinal surface, the second step is to solve the phase compensation matrix through the conjugate gradient method, and the third step is to perform tensor contraction operation on the compensation matrix and the quantum state amplitude distribution. The operation result is optimized by Gauss-Seidel iterative optimization to output a phase hologram adapted to the fovea of ​​the retina.

[0123] The phase hologram is synthesized into a physical light field and projected onto the fovea of ​​the user's retina. After the projection is completed, the EEG signal is collected to obtain the differential energy ratio of the oscillation component of the frequency band before and after the stimulation;

[0124] Specifically, the phase hologram is converted into a coherent light field through the phase grayscale encoding of the spatial light modulator. After the spatial light modulator is loaded with the phase hologram, it is projected to the fovea area of ​​the retina through the collimating lens group and the eyeball optics. The projection process synchronously triggers the three-wavelength infrared camera to record the trajectory of the corneal reflection point to ensure the accuracy of light field positioning. After the light field projection is completed, EEG signal acquisition is immediately started. The EEG signal is decomposed into time domain sub-signals of each frequency band through a recursive Gaussian kernel integrator. According to the energy difference of the time domain sub-signals of each frequency band before and after stimulation, the differential energy ratio of the frequency band oscillation components before and after stimulation is obtained through the differential operator.

[0125] Based on the differential energy ratio and the preset weight vector, the updated probability value of alertness decrease, the probability value of cognitive overload and the probability value of physiological fatigue are output.

[0126] It should be noted that the preset weight vector is set based on the statistical distribution of the energy contribution of each EEG frequency band in healthy people under standard cognitive tasks. The example values ​​are δ band weight 0.2, θ band weight 0.3, α band weight 0.25, β band weight 0.15, and γ band weight 0.1.

[0127] Specifically, the differential energy ratio and the preset weight vector are used to generate a joint feature tensor through the Kronecker product operation. The joint feature tensor is input into three independent differential decision tree engines for probability value update calculation. The update process of the three probability values ​​is executed synchronously. The calculation result of each probability value must be verified by the conservation constraint condition. If the conditions are met, the updated probability value of decreased alertness, cognitive overload probability value and physiological fatigue probability value are output. Otherwise, the frequency band weight coefficient in the preset weight vector is readjusted and iterative calculation is performed.

[0128] The closed-loop termination module is used to terminate the physical light field projection when the updated probability value of alertness decrease, cognitive overload probability value and physiological fatigue probability value are all lower than the preset safety threshold value during the continuous detection cycle.

[0129] Furthermore, based on the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue updated during the continuous detection period, the changing trends of the probability values ​​are analyzed and a multidimensional data structure is constructed;

[0130] Specifically, the updated probability values ​​of decreased alertness, cognitive overload probability values, and physiological fatigue probability values ​​within the continuous detection cycle are used to extract first-order difference features through time series autoregression. The first-order difference features and the updated probability values ​​of decreased alertness, cognitive overload probability values, and physiological fatigue probability values ​​constitute a six-dimensional time series feature vector. The six-dimensional time series feature vector is input into the Gram angular field converter to generate a Gram angular field matrix. The Gram angular field matrix extracts spatial correlation features through a deep convolutional network. The spatial correlation features and the updated probability values ​​of decreased alertness, cognitive overload probability values, and physiological fatigue probability values ​​are formed into a nine-dimensional mixed feature tensor through tensor splicing. The nine-dimensional mixed feature tensor is output as a multi-dimensional data structure through the cross-cycle dependency relationship of the multi-head self-attention mechanism.

[0131] Perform time series pattern recognition and related entropy evaluation on multidimensional data structures and output decision indicators;

[0132] Specifically, the multidimensional data structure extracts the periodic and trend characteristics in the time dimension through the long short-term memory network. The hidden state output of the long short-term memory network is fused element-by-element with the multidimensional data structure through the Hadamard product to generate the temporal pattern recognition result. The related entropy evaluation uses the Kullback-Leibler divergence to obtain the difference in probability distribution in adjacent time windows. The difference is fused with the spectral energy of the Gram angular field matrix through the Hadamard product into an entropy change feature vector. The temporal pattern recognition result and the entropy change feature vector are input into the gated recursive unit for joint encoding. The hidden state output of the gated recursive unit is mapped to a decision indicator through the softmax function.

[0133] When the continuous detection cycles of the decision indicators are all lower than the preset safety threshold, a quantum state conversion command signal is generated to control the physical light field to terminate the projection.

[0134] It should be noted that the preset safety thresholds are set based on the percentiles of the energy distribution of the EEG frequency bands of healthy people in a resting state with their eyes closed. Example values ​​are the alertness maintenance index of 0.35, the cognitive load balance index of 0.4, and the physiological recovery index of 0.3.

[0135] Specifically, when the continuous detection cycle of the decision indicator is lower than the preset safety threshold, the quantum state transition command signal is triggered. The quantum state transition command signal is transmitted to the physical light field controller through optical fiber. After the physical light field controller parses the command, it performs three termination operations: the first step is to stop the phase hologram refresh, the second step is to turn off the radio frequency drive of the spatial light modulator, and the third step is to disconnect the trigger synchronization signal of the three-wavelength infrared camera to complete the termination projection of the physical light field.

[0136] In summary, the present invention maps the raw EEG signal into a three-dimensional chaotic trajectory tensor space through multi-scale recursive phase space reconstruction and uses Lyapunov exponent thresholds to trigger dynamic frequency band decoupling. A recursive Gaussian kernel integrator is used to perform nonlinear separation of the δ / θ and β / γ frequency bands, suppressing transient artifacts. Simultaneously, dynamic Lyapunov exponent analysis improves the success rate of frequency band separation at low levels of chaos, ensuring stable output of valid frequency band components.

[0137] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A real-time fatigue monitoring system based on EEG signals, characterized by: include, The chaos reconstruction module is used to collect the original EEG signal, perform phase space reconstruction on the original EEG signal through the chaotic attractor processor, generate a chaotic trajectory tensor, and output a slow frequency band oscillation component and a fast frequency band oscillation component when the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold; an adversarial discrimination module, configured to input the slow-frequency oscillation component and the fast-frequency oscillation component into a generator unit of a generative adversarial network to generate a synthetic feature vector, and simultaneously transmit the synthetic feature vector in parallel to an alertness discrimination unit, a cognitive load discrimination unit, and a physiological fatigue discrimination unit to generate a probability value of decreased alertness, a probability value of cognitive overload, and a probability value of physiological fatigue; a visual tracking module configured to capture a corneal reflection point and generate visual focus plane coordinates when the maximum value of any of the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue exceeds a preset probability value threshold; The light field construction module is used to calculate the field dynamic parameters of the dynamic optical flow field based on the visual focus plane coordinates and the maximum value of any probability value, and generate the optical flow field equation; The optical neural control module is used to convert the optical flow field equation into a physical light field and project it onto the user's retinal fovea, and update the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue; The closed-loop termination module is used to terminate the physical light field projection when the updated probability value of alertness decrease, cognitive overload probability value and physiological fatigue probability value are all lower than the preset safety threshold value during the continuous detection cycle.

2. The real-time fatigue monitoring system based on EEG signals according to claim 1, characterized in that: The specific steps of generating the chaotic trajectory tensor are as follows: The collected original EEG signal is decomposed into multiple scales by a recursive Gaussian kernel integrator to generate multiple frequency band time domain sub-signals; According to the instantaneous energy gradient ratio of the time domain sub-signals in each frequency band, the optimal embedding dimension is calculated within the dynamic time window; The time-domain sub-signals of each frequency band and the corresponding optimal embedding dimensions are input into the recursive tensor coupler to generate a chaotic trajectory tensor.

3. The real-time fatigue monitoring system based on EEG signals according to claim 1, characterized in that: The specific steps of outputting the slow frequency band oscillation component and the fast frequency band oscillation component are as follows: The instantaneous exponential spectrum calculation based on the time window is performed on the chaotic trajectory tensor, and the maximum component of the exponential value of each dimension is extracted as the Lyapunov index; When the Lyapunov exponent of the chaotic trajectory tensor exceeds a preset exponential threshold, the frequency band components are decoupled through an oscillation coupling separator, and a slow frequency band oscillation component and a fast frequency band oscillation component are output; When the Lyapunov exponent of the chaotic trajectory tensor does not exceed the preset exponential threshold, the embedding dimension attenuation coefficient of the phase space reconstruction is updated and the chaotic trajectory tensor is regenerated.

4. The real-time fatigue monitoring system based on EEG signals according to claim 1, wherein: The specific steps of generating the synthetic feature vector are as follows: Input the slow frequency band oscillation component and the fast frequency band oscillation component into the feature tensor folder to generate a frequency band fusion tensor; The frequency band fusion tensor is input into the quantum probability encoder and transformed into a quantum state vector through the quantum entanglement gate; The quantum state vector is input into the Hamiltonian driven generator, a unitary matrix transformation is performed, a multi-basis Pauli operator measurement is performed on the quantum state vector after the unitary matrix transformation, and a synthetic eigenvector is output.

5. The real-time fatigue monitoring system based on EEG signals according to claim 1, characterized in that: The specific steps of generating the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue are as follows: The synthesized feature vector is spatiotemporally encoded to generate a spatiotemporal feature tensor, which is then fed into the attention mask controller to generate a weighted feature tensor. The weighted feature tensor is simultaneously fed into three differential decision tree engines to independently calculate the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue; When the probability value of decreased alertness, the probability value of cognitive overload, and the probability value of physiological fatigue meet the conservation constraints, they are output; otherwise, the weighted feature tensor is regenerated.

6. The real-time fatigue monitoring system based on EEG signals according to claim 1, characterized in that: The steps of capturing the corneal reflection point and generating the visual focus plane coordinates are as follows: When the maximum value of any probability value exceeds the preset probability value threshold, the annular light source array is activated and the light source intensity and spatial distribution are dynamically adjusted to output a light intensity modulation signal; Based on the light intensity modulation signal, a three-wavelength infrared camera is used to synchronously collect corneal reflection points and extract multispectral reflection features. The multispectral reflection feature matrix is ​​output and a reflection feature map is generated through weighted fusion. According to the reflection feature map, the Lucas-Kanade optical flow method is used to calculate the eyelid contour motion velocity vector and output the eyelid motion parameters. The eyelid motion parameters and the reflection feature map are vector dot producted to generate the compensated pupil center offset vector, and spherical integral mapping is performed to generate the visual focus plane coordinates.

7. The real-time fatigue monitoring system based on EEG signals according to claim 1, characterized in that: The field dynamics parameters include intensity attenuation coefficient, spatial wave number and retinal projection distance parameters.

8. The real-time fatigue monitoring system based on EEG signals according to claim 7, characterized in that: The field dynamics parameters of the dynamic optical flow field are calculated to generate the optical flow field equation. The specific steps are as follows: Calculate the intensity attenuation coefficient according to the maximum value of the depth component of the visual focus plane coordinates and any probability value; According to the spatiotemporal gradient of the visual focal plane coordinates and the pre-calibrated retinal foveal coordinates, the spatial wave number and retinal projection distance parameters are calculated; The intensity attenuation coefficient, spatial wave number and retinal projection distance parameters are input into the viscoelastic field equation to generate the optical flow field equation.

9. The real-time fatigue monitoring system based on EEG signals according to claim 1, characterized in that: The steps of outputting the updated probability value of decreased alertness, probability value of cognitive overload, and probability value of physiological fatigue are as follows: The optical flow field equation is input into the quantum state amplitude converter to generate the quantum state amplitude distribution, and the phase hologram adapted to the fovea of ​​the retina is generated through the curvature compensation phase modulator; The phase hologram is synthesized into a physical light field and projected onto the fovea of ​​the user's retina. After the projection is completed, the EEG signal is collected to obtain the differential energy ratio of the oscillation component of the frequency band before and after the stimulation; Based on the differential energy ratio and the preset weight vector, the updated probability value of alertness decrease, the probability value of cognitive overload and the probability value of physiological fatigue are output.

10. The real-time fatigue monitoring system based on EEG signals according to claim 9, characterized in that: The specific steps of terminating the physical light field projection are as follows: According to the probability values ​​of decreased alertness, cognitive overload, and physiological fatigue updated during the continuous detection cycle, the changing trends of each probability value are analyzed and a multidimensional data structure is constructed; Perform time series pattern recognition and related entropy evaluation on multidimensional data structures and output decision indicators; When the continuous detection cycles of the decision indicators are all lower than the preset safety threshold, a quantum state conversion command signal is generated to control the physical light field to terminate the projection.

Images