Electroencephalogram signal dynamic monitoring method for fatigue risk early warning

Abstract

The invention relates to the technical field of intelligent fatigue detection, in particular to a fatigue risk early warning-oriented electroencephalogram signal dynamic monitoring method, which comprises the following steps of: constructing an electroencephalogram dynamic sequence, acquiring multi-channel electroencephalogram signals in real time, and generating a time-continuous dynamic electroencephalogram sequence by adopting a sliding time window; time sequence feature extraction: performing time-frequency-space three-dimensional feature analysis on the dynamic electroencephalogram sequence, extracting a time-varying complexity feature and a phase synchronization mutation index, and constructing a dynamic feature vector; fatigue trajectory modeling: inputting the dynamic feature vector into a lightweight time sequence convolutional network, and outputting a fatigue risk trajectory curve; performing graded early warning triggering, calculating a slope value of the risk trajectory curve, generating a self-adaptive threshold value based on historical slope distribution, performing multi-grade fatigue early warning judgment, and transmitting a fatigue early warning signal through a wireless module; according to the method, high-resolution modeling and personalized dynamic early warning of the fatigue state can be realized, and the method is suitable for continuous operation scenes such as driving assistance and industrial security and protection.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fatigue detection, and in particular to a method for dynamic monitoring of electroencephalogram (EEG) signals for fatigue risk early warning. Background Art

[0002] As the demand for workers to maintain continuous attention in fields such as transportation and industrial operations continues to increase, timely identification and proactive warning of fatigue states have become important research directions for ensuring the safe operation of human-machine systems. As a key physiological indicator of central nervous system activity, EEG signals are highly sensitive to changes in cognitive state and have been widely used in fatigue detection research. In recent years, the construction of intelligent fatigue monitoring and warning systems has gradually become a research hotspot, combining multi-channel synchronous EEG signal acquisition, spectral analysis, and nonlinear modeling.

[0003] The following technical problems are common in existing technologies: First, most fatigue recognition methods only rely on single-dimensional features in the time domain or frequency domain, and cannot fully integrate the time-frequency-space multi-dimensional dynamic characteristics of EEG signals, resulting in insufficient recognition sensitivity; second, fatigue modeling methods are mostly based on static classification or shallow regression models, which make it difficult to effectively characterize the continuous evolution trend of fatigue status in time series; third, traditional early warning mechanisms mostly use fixed empirical thresholds, lack the ability to adaptively adjust to individual differences and status fluctuations of users, and easily lead to false alarms or missed alarms. Summary of the Invention

[0004] The present invention provides a dynamic EEG signal monitoring method for fatigue risk warning, a systematic technical solution that can realize dynamic EEG signal modeling, multi-dimensional feature fusion and real-time graded warning to meet the high reliability requirements of fatigue risk monitoring in complex task scenarios.

[0005] A dynamic EEG signal monitoring method for fatigue risk warning includes the following steps:

[0006] S1, EEG dynamic sequence construction: collect the user's multi-channel EEG signals in real time, intercept continuous signal segments with a sliding time window, and output a dynamic EEG sequence;

[0007] S2, time series feature extraction: perform time-frequency-space three-dimensional feature analysis on the dynamic EEG sequence, extract time-varying complexity features and phase synchronization mutation index, and output dynamic feature vectors;

[0008] S3, fatigue trajectory modeling: input the dynamic feature vector into a pre-trained lightweight temporal convolutional network to generate a risk trajectory curve representing the evolution process of the fatigue state;

[0009] S4, graded warning triggering: calculating the real-time slope value of the risk trajectory curve, and when the slope value exceeds the adaptive threshold, outputting a fatigue warning signal of the corresponding level.

[0010] Optionally, the S1 includes:

[0011] S11, multi-channel EEG signal acquisition: real-time acquisition of the user's multi-channel EEG signals, with a sampling rate of ≥256Hz, and channels covering FP1, FP2, Fz, Cz, and Oz leads;

[0012] S12, signal preprocessing and framing: performing power frequency filtering (50 Hz notch) and baseline drift correction on the multi-channel EEG signal, and intercepting continuous signal segments with a sliding time window, where the window length is fixed at 500 ms and the window shift step is 50 ms;

[0013] S13, dynamic EEG sequence generation: The continuous signal segments intercepted by each sliding time window are arranged in chronological order to construct a dynamic EEG sequence with continuous time dimension. Its data structure is a three-dimensional tensor of [number of time frames × number of channels × number of sampling points].

[0014] Optionally, the S11 collects the user's multi-channel EEG signals in real time through an embedded EEG helmet, which includes a dry electrode array, an ARM Cortex-M7 microprocessor and a Bluetooth transmission module.

[0015] Optionally, the S2 includes:

[0016] S21, time-frequency-space three-dimensional feature analysis: performing three-dimensional decomposition on the dynamic EEG sequence to generate a time domain feature matrix, a frequency domain feature matrix, and a space domain feature matrix. The three dimensions include the time domain dimension, the frequency domain dimension, and the space domain dimension, wherein;

[0017] Time domain dimension: Calculate the Hjorth parameter of each channel signal in each time window. The Hjorth parameter includes activity, mobility, and complexity, which are used to measure the fluctuation amplitude, frequency characteristics, and change trend of the EEG signal;

[0018] Frequency domain: By applying Morlet wavelet transform to the signal of each channel, the signal energy in four specific frequency ranges, including the δ band, the θ band, the α band, and the β band, is extracted to characterize the spectral structure under different cognitive states.

[0019] Spatial dimension: Based on the instantaneous phase information of all channel signals in each time window, the phase lag index between each channel is calculated.

[0020] S22, time-varying complexity feature extraction: Based on the time domain feature matrix and the frequency domain feature matrix, the multi-scale permutation entropy and the wavelet packet energy entropy are calculated in time frame order, and the calculated multi-scale permutation entropy and the wavelet packet energy entropy are fused to generate time-varying complexity features, providing a detailed description of the brain function fluctuation pattern for modeling.

[0021] Optionally, the S2 further includes:

[0022] S23, phase synchronization mutation index extraction: Based on the spatial domain feature matrix, a phase synchronization mutation index is calculated to reflect the mutation trend of the neural network. The index is used to reflect the degree of sudden changes in the brain's functional connectivity structure and serves as a leading signal indicator of fatigue state mutation.

[0023] S24, Dynamic Feature Vector Construction: The generated time-varying complexity features and phase-synchronized mutation index are concatenated and integrated according to time frames to form a dynamic feature vector with a unified structure and fixed dimensions. Each time frame corresponds to a set of eigenvalues that comprehensively reflect the changing characteristics of the EEG signal in terms of temporal structure, spectral energy, and functional connectivity at that moment. The resulting dynamic feature vector is continuous in the time dimension and diverse in the feature dimension, providing a unified, high-quality data input foundation for subsequent fatigue trajectory modeling and graded early warning.

[0024] Optionally, the S3 includes:

[0025] S31, dynamic feature vector standardization: performing Z-score standardization processing on the dynamic feature vector so that the mean of each feature dimension is 0 and the variance is 1, and outputting the standardized feature vector;

[0026] S32, lightweight temporal convolutional network forward propagation: input the standardized feature vector into the pre-trained lightweight temporal convolutional network, perform forward propagation operation through the lightweight temporal convolutional network, and generate an original trajectory tensor;

[0027] S33, generating a risk trajectory curve: inputting the original trajectory tensor into a Sigmoid activation function for nonlinear normalization processing to generate a risk trajectory curve.

[0028] Optionally, the S4 includes:

[0029] S41, real-time slope value calculation: performing central difference calculation on the risk trajectory curve according to time frames, and outputting the slope value of each time frame;

[0030] S42, adaptive threshold generation: setting an adaptive threshold based on the user's historical slope value in the last N hours.

[0031] Optionally, the S4 further includes:

[0032] S43, graded warning determination: comparing the slope value of the current time frame with the adaptive threshold, and generating a three-level fatigue warning signal based on the comparison result;

[0033] S44, warning signal transmission: the fatigue warning signal is transmitted to the vehicle-mounted human-computer interaction system via the wireless communication module, triggering the corresponding level of sound, light, vibration, or braking response.

[0034] Beneficial effects of the present invention:

[0035] The present invention, through multi-channel EEG signal acquisition, sliding time window framing and dynamic EEG sequence construction in S1, combined with the three-dimensional feature analysis mechanism of time domain, frequency domain and spatial domain in S2, comprehensively extracts high-temporal and spatial resolution EEG features including Hjorth parameters, wavelet band energy, and phase lag index. At the same time, the integration of multi-scale permutation entropy and wavelet packet energy entropy realizes dynamic tracking of the complexity of neural activity and significantly enhances the response ability to weak fatigue trends. This solution solves the technical shortcomings of existing methods that cannot simultaneously capture spectrum changes and network coupling mutations.

[0036] This paper uses a lightweight temporal convolutional network in the S3 stage to construct a fatigue risk trajectory model. This introduces a causal structure and dilated convolution mechanism, effectively expanding the temporal perception range and enhancing the model's ability to model long-term dependency patterns. Furthermore, the total number of parameters after network compression is kept below 80,000, making it suitable for edge computing scenarios such as wearable devices and in-vehicle platforms, ensuring high-frequency continuous operation and real-time inference without sacrificing accuracy.

[0037] The present invention introduces the risk trajectory slope value as a dynamic change indicator in the S4 stage. Through statistical analysis of the historical risk change rate, it generates an adaptive threshold that can be automatically adjusted over time and according to individual differences. Combined with the graded warning rules, it can accurately distinguish between mild, moderate and severe fatigue states, and transmit the first to third level fatigue warning signals to the human-computer interaction system through the wireless module, realizing a hierarchical response strategy from prompts and warnings to braking intervention, greatly improving driving safety and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 This is a schematic diagram of the S1 process of an embodiment of the present invention;

[0040] Figure 2Schematic diagram of the S2 process of an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0042] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0043] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0044] like Figure 1-2 As shown in FIG, the method for dynamic monitoring of EEG signals for fatigue risk warning includes the following steps:

[0045] S1, EEG dynamic sequence construction: collect the user's multi-channel EEG signals in real time, intercept continuous signal segments with a sliding time window, and output a dynamic EEG sequence;

[0046] S2, time series feature extraction: perform time-frequency-space three-dimensional feature analysis on the dynamic EEG sequence, extract time-varying complexity features and phase synchronization mutation index, and output dynamic feature vectors;

[0047] S3, fatigue trajectory modeling: input the dynamic feature vector into a pre-trained lightweight temporal convolutional network to generate a risk trajectory curve representing the evolution process of the fatigue state;

[0048] S4, graded warning triggering: calculating the real-time slope value of the risk trajectory curve, and when the slope value exceeds the adaptive threshold, outputting a fatigue warning signal of the corresponding level.

[0049] S1 includes:

[0050] S11, multi-channel EEG signal acquisition: Real-time monitoring of the user's EEG activities in the forehead, central and occipital brain regions through the embedded EEG helmet.

[0051] The electrode array covers the following international 10-20 system standard leads: FP1, FP2, Fz, Cz, Oz, a total of 5 channels, to ensure that the acquisition covers a wide range of brain areas and is representative. The sampling rate is set to no less than 256Hz to ensure that the signal's time accuracy is sufficient to capture subtle physiological fluctuations. The collected raw EEG signal is represented as:

[0052] X raw (t,c),c∈{1,2,...,C},t∈R + ;

[0053] Wherein, c represents the channel number (C=5 in this embodiment), t is the sampling time point, and X raw is the original signal amplitude.

[0054] S12 signal preprocessing and framing: To remove non-physiological interference components, the original EEG signal needs to be preprocessed. The processing flow includes:

[0055] (1) Power frequency interference suppression: A notch filter is used to filter out the 50 Hz power frequency interference. The transfer function of the IIR notch filter is:

[0056]

[0057] Where, f0=50Hz, f s =256Hz, r∈(0,1) is the damping factor that controls the bandwidth.

[0058] (2) Baseline drift correction: A moving average filter is used to remove low-frequency baseline drift, and the filter window length is generally 2 to 4 seconds.

[0059] (3) Sliding window framing: The processed continuous signal is sliced in a fixed time window manner. The window length is set to 500ms (corresponding to 128 sampling points) and the window shift step is 50ms (corresponding to 12 to 13 sampling points) to achieve time-continuous segmentation of the signal. Each time window is recorded as:

[0060] x (i) =[X(t i ),X(t i +1),…,X(t i +T-1)]∈R C×T ;

[0061] Wherein, T=128 is the number of sampling points per frame, and i is the number of the current sliding window.

[0062] For example, if the total length of the input signal is 10 seconds, the number of frames that can be obtained under the condition of a 50ms window shift is:

[0063]

[0064] That is, a total of 191 consecutive time frames are generated.

[0065] S13, dynamic EEG sequence generation: stack the signal segments after each time window is divided into frames in chronological order to construct a time-continuous dynamic EEG sequence. This sequence can be represented as a three-dimensional tensor:

[0066] X dyn ∈R N×C×T ;

[0067] Here, N is the number of time frames (191 in the example above), C is the number of channels (5 in this example), and T is the number of sampling points per frame (128 in this example). This tensor preserves the temporal evolution information of the EEG signal, the spatial lead structure, and the instantaneous morphological characteristics of each frame, providing the necessary data foundation for the subsequent time-frequency-space feature extraction step.

[0068] In order to enhance data stability and individual adaptability, Z-score normalization can be used to normalize the data of each channel in the final output dynamic EEG sequence:

[0069]

[0070] Among them, μ c and σ c Represent the mean and standard deviation of channel c in all frames respectively.

[0071] S2 includes:

[0072] S21, Time-Frequency-Space 3D Feature Analysis: Analyze dynamic EEG sequence features from a 3D perspective, including:

[0073] Time domain feature extraction: For each channel signal in each time window, the Hjorth parameters are calculated separately: activity, mobility and complexity.

[0074] Among them, activity measures the degree of change in the overall amplitude of the signal, mobility reflects the change in the frequency component of the signal, and complexity is used to characterize the regularity of signal changes. The specific calculation is as follows:

[0075] Activity A:

[0076] Mobility M:

[0077] Complexity Co:

[0078] in, is the channel signal, t is the time window index, and c is the channel number;

[0079] These parameters are calculated for each channel separately, and the output time domain feature matrix with the size of time frames × number of channels × 3 is expressed as: F time ∈R N×(C×3) .

[0080] Frequency domain feature extraction: Morlet wavelet transform is used to perform time-frequency transformation on the signal of each channel in each time window, and the signal energy in the frequency bands of δ wave (1–4 Hz), θ wave (4–8 Hz), α wave (8–13 Hz), and β wave (13–30 Hz) is extracted. For each frequency band, the square integral of the modulus of the wavelet coefficient is used as the energy index, thus forming a frequency domain feature matrix of the number of time frames × number of channels × 4, expressed as: F freq ∈R N×(C×4) .

[0081] Spatial feature extraction: Based on the instantaneous phase of each channel signal within the time window, the Phase Lag Index (PLI) between channel pairs is calculated to measure the degree of phase synchronization between different brain regions and eliminate volume conduction interference. PLI is calculated for all channel pairs to obtain a spatial feature matrix with the size of time frames × channels × channels, expressed as F space ∈R N×C×C , in the form of a functional connectivity map for each frame, the phase lag index is calculated as:

[0082]

[0083] in, is the instantaneous phase of each channel, i,j is channel i and channel j, <·> τ This value represents the average of all sampling points within the time window, reflecting whether the phase difference between channels is stable and biased in a certain direction. A larger value indicates a more stable connection but non-zero phase synchronization.

[0084] For example, if five channels (FP1, FP2, Fz, Cz, Oz) are used, a 5×5 functional connectivity matrix will be output for each frame, where the diagonal values are zero and the off-diagonal elements represent the PLI strength between channel pairs. S22, Time-varying complexity feature extraction: Extracts nonlinear complexity indicators of dynamic EEG sequences, including:

[0085] (1) Multi-scale permutation entropy (MPE) calculation: The time frame sequence of each channel in the aforementioned time domain feature matrix and frequency domain feature matrix is reconstructed according to multiple scale factors. At each scale, a set of time series embedding vectors with an embedding dimension of k is constructed, and the probability distribution of its permutation pattern is statistically analyzed to calculate the permutation entropy:

[0086]

[0087] in, is the probability of occurrence of the i-th mode under the scale factor τ. By setting multiple scales (such as τ = 1, 3, 5), the dynamic order changes under different granularities can be calculated respectively.

[0088] (2) Calculation of wavelet packet energy entropy (WPEE): Decompose each frame of EEG signal into the 7th layer of wavelet packet structure, count the energy proportions of all sub-bands, and calculate the normalized entropy value:

[0089]

[0090] Among them, E j is the energy of the jth wavelet packet component. A higher entropy value indicates a more uniform energy distribution and a more complex system. This value is calculated independently for each channel.

[0091] Finally, MPE and WPEE are aggregated in the channel dimension to form a time-varying complexity feature set with dimension D1 per frame.

[0092] S23, phase synchronization mutation index extraction: To measure the degree of mutation of EEG functional connectivity, the following operations are performed based on the spatial feature matrix:

[0093] (1) PLI extraction: For each frame’s phase lag index matrix, extract the elements above its main diagonal, which represent the phase connection strength between all channel pairs.

[0094] (2) Standard deviation calculation: Calculate the standard deviation σ of all PLI elements in the frame (t) , used to characterize the discrete degree of the current functional network structure.

[0095] (3) Mutation index generation: The absolute difference of the standard deviation between adjacent time frames is calculated and defined as the phase synchronization mutation index:

[0096] Δσ (t) =|σ (t) -σ (t-1) |;

[0097] This indicator reflects the dramatic changes in the degree of phase coordination of the brain network in the time dimension and is one of the sensitive features for capturing the transition of fatigue state.

[0098] S24, dynamic feature vector construction: The generated time-varying complexity feature set and the phase synchronization mutation index of the corresponding time frame are spliced in columns to construct a dynamic feature vector F with a unified structure. dyn :

[0099]

[0100] Among them, F comp is the complexity feature of each frame, Δσ is the phase mutation feature, and after splicing, they constitute a complete timing description of each frame.

[0101] For example, if the total dimension of the combination of MPE and WPEE is 60, the final dimension of the dynamic eigenvector is 61. This characteristic structure has temporal continuity, multi-scale sensitivity, and spatial connection perception capabilities, and can be directly input into the fatigue trajectory modeling network for risk assessment.

[0102] S3 includes:

[0103] S31 Dynamic feature vector normalization: In order to eliminate the dimensional differences between different feature dimensions and improve the convergence efficiency of network modeling, the dynamic feature vector matrix output in step S2 is normalized. Perform Z-score standardization, the specific steps are:

[0104] For each time frame n∈[1,N] and each feature dimension d∈[1,D1+1], the following transformation is used:

[0105]

[0106] Among them, μ d is the mean of the feature dimension d over all time frames, σ d is the standard deviation of the dimension, is the normalized feature matrix.

[0107] This standardization operation ensures that the mean of each column feature is 0 and the standard deviation is 1, preventing a feature from dominating the model training due to its large scale.

[0108] Example: If the fatigue trajectory modeling input contains 60-dimensional complexity features and a 1-dimensional phase mutation index, each frame still maintains 61 dimensions after normalization, and the amplitude range differences are balanced, which helps avoid gradient explosion or vanishing.

[0109] S32, Lightweight Temporal Convolutional Network forward propagation: The normalized dynamic feature vector is used as input to the Lightweight Temporal Convolutional Network (L-TCN). This network is designed to handle long-term dependencies in time series and has the characteristics of small number of parameters and high computational efficiency. The network structure includes:

[0110] (1) Four-layer causal dilated convolution module: Each layer is a one-dimensional convolution, with dilation coefficients d = 1, 2, 4, and 8, respectively, and the convolution kernel size is set to 3. The causal structure ensures that the model only uses information from the current frame and the previous frame, which meets the requirements of online reasoning; the dilation structure allows the convolution receptive field to expand exponentially, effectively capturing long-term dependency features.

[0111] (2) 1×1 convolutional dimension reduction layer: A point-by-point convolution layer with a channel number not exceeding 64 is connected to perform linear compression on the feature channel of each time step to control the total number of parameters.

[0112] (3) Parameter quantity constraint: The total number of parameters of the entire network is controlled within 80,000 to meet the deployment requirements on embedded devices or low-power computing units.

[0113] The input of the entire network is:

[0114] The output is a raw trajectory tensor: T raw ∈R N×1

[0115] Each element in the original trajectory tensor represents the original fatigue strength prediction value in the corresponding time frame and has not been normalized.

[0116] Example description: The input is 200 frames and 61-dimensional normalized features. After channel compression, the convolution layer still retains the output value corresponding to each frame, so the final result is a 200-dimensional original trajectory vector, each value represents the estimated value of model fatigue at a time point.

[0117] S33, risk trajectory curve generation: In order to make the model output value have probabilistic interpretation and normalization characteristics, the original trajectory tensor T raw Mapping to the [0,1] interval, we get the normalized fatigue risk trajectory curve. The mapping function uses the standard Sigmoid activation function, which is defined as follows:

[0118]

[0119] Among them, T risk ∈R N×1 is the risk trajectory curve, each element T risk (n) represents the fatigue risk intensity corresponding to the nth frame. The closer the value is to 1, the higher the fatigue risk.

[0120] This risk trajectory curve reflects the user's fatigue evolution within the current monitoring window. Its horizontal axis represents the time frame index, and the vertical axis represents the fatigue risk score. A steady curve indicates a stable fatigue state, an upward trend indicates fatigue accumulation, and a sharp rise indicates a high-risk phase.

[0121] Example: When a user experiences a decline in attention or increased EEG synchronization during a certain period of time, the risk trajectory output by the model will show a continuous growth trend, such as increasing from 0.3 to 0.85, prompting the system to further trigger an early warning response.

[0122] S4 includes:

[0123] S41, real-time slope value calculation: the fatigue risk trajectory curve T generated in step S3 risk (n) Perform central difference processing according to the time frame and estimate the fatigue strength change rate of each time frame. For any intermediate time frame n∈[2,N-1], calculate its corresponding slope value as follows:

[0124]

[0125] Among them, T risk (n) represents the fatigue risk intensity value of the nth frame, Δt represents the time interval between adjacent time frames, and the slope value n Represents the local rate of change of the fatigue risk trajectory curve in time frame n.

[0126] Output slope value sequence S(n) to characterize the changing trend of fatigue risk.

[0127] S42, adaptive threshold generation: based on the user's recent N h Slope value sequence within hours {S h}, dynamically generate individualized adaptive thresholds. The specific steps are as follows:

[0128] Calculate the mean and standard deviation of historical slope values:

[0129] The mean is calculated as:

[0130] The standard deviation is calculated as:

[0131] Where L is the number of slope values in the historical time window;

[0132] Construct adaptive threshold: adaptive threshold = μ + K·σ;

[0133] Among them, K is the safety factor, and the empirical value is not less than 2.5;

[0134] The adaptive threshold will be dynamically updated as the user status fluctuates to ensure the sensitivity and stability of the warning;

[0135] S43, graded warning determination: Compare the slope value corresponding to the current time frame with the adaptive threshold, determine the fatigue level according to the following segmentation rules, and output a fatigue warning signal of the corresponding level:

[0136] If 0.15≤slope value n <0.3, output first-level fatigue warning signal;

[0137] If 0.3≤slope value n <0.5, output the second-level fatigue warning signal;

[0138] If the slope value n ≥0.5, output the third-level fatigue warning signal.

[0139] This rule is based on the growth rate of the fatigue risk trajectory curve and provides a graded response mechanism, progressing from mild warning to mandatory intervention.

[0140] S44, Fatigue Warning Signal Transmission: The fatigue warning signal is transmitted to the vehicle-mounted human-computer interaction system via the wireless communication module. Depending on the level of the fatigue warning signal, the system may trigger different forms of prompts or intervention mechanisms, including:

[0141] Level 1 fatigue warning signal: triggers visual prompts on the screen and voice broadcast;

[0142] Secondary fatigue warning signal: additional vibration feedback of the steering wheel or seat;

[0143] Level 3 fatigue warning signal: Activate intelligent brake assistance or forced speed reduction measures.

[0144] The dynamic graded warning system driven by the risk trajectory curve can effectively achieve early perception, accurate judgment and safe intervention of fatigue status.

[0145] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0146] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A dynamic EEG signal monitoring method for fatigue risk warning, characterized by: The following steps are involved: S1, EEG dynamic sequence construction: collect the user's multi-channel EEG signals in real time, intercept continuous signal segments with a sliding time window, and output a dynamic EEG sequence; S2, time series feature extraction: perform time-frequency-space three-dimensional feature analysis on the dynamic EEG sequence, extract time-varying complexity features and phase synchronization mutation index, and output dynamic feature vectors; S3, fatigue trajectory modeling: input the dynamic feature vector into a pre-trained lightweight temporal convolutional network to generate a risk trajectory curve representing the evolution process of the fatigue state; S4, graded warning triggering: calculating the real-time slope value of the risk trajectory curve, and when the slope value exceeds the adaptive threshold, outputting a fatigue warning signal of the corresponding level.

2. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 1, characterized in that: Said S1 comprises: S11, multi-channel EEG signal acquisition: real-time acquisition of the user's multi-channel EEG signals, with a sampling rate of ≥256Hz, and channels covering FP1, FP2, Fz, Cz, and Oz leads; S12, signal preprocessing and framing: performing power frequency filtering (50 Hz notch) and baseline drift correction on the multi-channel EEG signal, and intercepting continuous signal segments with a sliding time window, where the window length is fixed at 500 ms and the window shift step is 50 ms; S13, dynamic EEG sequence generation: The continuous signal segments intercepted by each sliding time window are arranged in chronological order to construct a dynamic EEG sequence with continuous time dimension. Its data structure is a three-dimensional tensor of [number of time frames × number of channels × number of sampling points].

3. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 2, characterized in that: The S11 collects the user's multi-channel EEG signals in real time through an embedded EEG helmet, which includes a dry electrode array, an ARM Cortex-M7 microprocessor and a Bluetooth transmission module.

4. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 3, characterized in that: The S2 includes: S21, time-frequency-space three-dimensional feature analysis: performing three-dimensional decomposition on the dynamic EEG sequence to generate a time domain feature matrix, a frequency domain feature matrix, and a space domain feature matrix, where the three dimensions include the time domain dimension, the frequency domain dimension, and the space domain dimension; S22, time-varying complexity feature extraction: based on the time domain feature matrix and the frequency domain feature matrix, multi-scale permutation entropy and wavelet packet energy entropy are calculated in time frame order, and the calculated multi-scale permutation entropy and wavelet packet energy entropy are fused to generate time-varying complexity features.

5. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 4, characterized in that: Said S2 further comprises: S23, phase synchronization mutation index extraction: based on the spatial feature matrix, calculating the phase synchronization mutation index reflecting the mutation trend of the neural network; S24, dynamic feature vector construction: The generated time-varying complexity features and phase-synchronized mutation index are spliced and integrated according to the time frame to form a dynamic feature vector with unified structure and fixed dimension.

6. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 5, characterized in that: The S3 includes: S31, dynamic feature vector normalization: performing Z-score normalization processing on the dynamic feature vector and outputting a normalized feature vector; S32, lightweight temporal convolutional network forward propagation: input the standardized feature vector into the pre-trained lightweight temporal convolutional network, perform forward propagation operation through the lightweight temporal convolutional network, and generate an original trajectory tensor; S33, generating a risk trajectory curve: inputting the original trajectory tensor into a Sigmoid activation function for nonlinear normalization processing to generate a risk trajectory curve.

7. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 6, characterized in that: The S4 includes: S41, real-time slope value calculation: performing central difference calculation on the risk trajectory curve according to time frames, and outputting the slope value of each time frame; S42, adaptive threshold generation: setting an adaptive threshold based on the user's historical slope value in the last N hours.

8. The method for dynamic monitoring of EEG signals for fatigue risk warning according to claim 7, characterized in that: Said S4 further comprises: S43, graded warning determination: comparing the slope value of the current time frame with the adaptive threshold, and generating a three-level fatigue warning signal based on the comparison result; S44, warning signal transmission: the fatigue warning signal is transmitted to the vehicle-mounted human-computer interaction system via the wireless communication module, triggering the corresponding level of sound, light, vibration, or braking response.

Images