Pilot fatigue classification method based on EEG signal

Abstract

The invention discloses a pilot fatigue classification method based on an EEG signal, and the method comprises the following steps: carrying out the frequency band decomposition of the EEG signal, and extracting differential entropy features corresponding to different frequency bands; a linear dynamic system (LDS) is used for filtering features to remove components irrelevant to fatigue, a deep ResTaylor Network (DRN) model is constructed, the problems of overfitting and gradient disappearance in high-degree polynomial processing are effectively reduced through a residual connection mechanism, and adaptive adjustment of model parameters is carried out in combination with an Adam optimization algorithm; and training the DRN model based on the training set, and verifying on the test set to obtain a pilot fatigue classification result. Through the pilot fatigue state classification method and device, accurate classification of the pilot fatigue state can be achieved, and compared with the prior art, the pilot fatigue state classification method and device have higher classification accuracy and robustness.

Description

Technical Field

[0001] The invention belongs to the technical field of biological signal processing and fatigue detection, and in particular relates to a pilot fatigue classification method based on EEG signals. Background Art

[0002] The fatigue state of pilots is one of the key factors affecting aviation safety. According to statistics, the fatigue of pilots is directly related to the performance of flight missions and flight safety. Excessive fatigue may lead to problems such as decreased attention, slow reaction, and misjudgment, which in turn lead to flight accidents. Therefore, how to effectively monitor and evaluate the fatigue state of pilots is one of the important research topics in the aviation field. In recent years, electroencephalogram (EEG) signals, as a biological signal that directly reflects brain activity, have been widely used in the field of fatigue monitoring. By analyzing the characteristic changes in different frequency bands in EEG signals, the fatigue state of pilots can be obtained. However, EEG signals have the characteristics of nonlinearity and non-stationarity. Traditional fatigue monitoring methods have great challenges in processing these signals, and it is difficult to achieve high-precision fatigue classification. In addition, the flight mission environment is complex and changeable, and EEG signals are easily interfered by noise, which further increases the difficulty of accurately judging the fatigue state of pilots. Therefore, there is an urgent need for a method that can process complex EEG signals and improve the accuracy of fatigue classification to ensure the safety of pilots and the smooth execution of missions. Summary of the invention

[0003] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes a pilot fatigue classification method based on EEG signals. The method utilizes the characteristics of EEG signals that can reflect the state of brain activity, extracts differential entropy features through frequency band decomposition, and combines the linear dynamic system (LDS) to filter out components unrelated to fatigue, thereby obtaining high-quality feature values ​​for fatigue classification. Subsequently, DRN is used for classification, and the overfitting and gradient vanishing problems caused by high-order polynomial expansion are effectively alleviated by introducing the residual connection mechanism, and the adaptive adjustment of parameters is realized in combination with the Adam optimization algorithm to improve the accuracy and robustness of classification.

[0004] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0005] The first aspect of the present invention provides a pilot fatigue classification method based on EEG signals, comprising the following steps:

[0006] Step 1: Obtain EEG signal data, pre-process the signal and decompose the frequency band;

[0007] Step 2: Decompose the EEG signal into multiple sub-bands based on a 2 Hz bandwidth, extract the differential entropy features on each sub-band, and filter the extracted features through a linear dynamic system to remove components unrelated to fatigue.

[0008] Step 3: Construct a deep residual Taylor network classification model, which includes an input layer, an intermediate layer, and an output layer. The input layer receives the EEG signal feature sequence x(w) after frequency band decomposition. The intermediate layer approximates the nonlinear relationship between input and output through a polynomial expansion with the highest power of m. The output layer combines the residual connection mechanism to finally output the classification result y(w) of the model.

[0009] Step 4: Adopt the Adam optimization algorithm to adaptively adjust the parameters of the deep residual Taylor network classification model. Combined with the learning rate attenuation coefficient, dynamically adjust the learning rate of each parameter to ensure that the model finds the global optimal or local optimal solution.

[0010] Step 5: Construct training sets and test sets, train the deep residual Taylor network classification model based on the training set, combine the p80 standard of PERCLOS as the fatigue label value, verify the classification performance of the model on the test set, and obtain the classification results of pilot fatigue.

[0011] Furthermore, the specific steps of step 2 are:

[0012] 1) Perform short-time Fourier transform on the EEG signal, decompose the signal into several sub-bands, and extract the differential entropy features of each sub-band;

[0013] 2) Using a linear dynamic system to filter the extracted eigenvalues ​​to remove noise or interference information that is not related to fatigue;

[0014] 3) Based on the extracted differential entropy feature matrix, the format of the input features is determined to be in the form of leads×samples×bands, where leads correspond to the number of channels of the EEG signal, samples correspond to the number of collected samples, and bands correspond to the number of sub-bands.

[0015] Furthermore, the specific steps of step three are:

[0016] 1) Input feature sequence x(w) = {x1(w), x2(w), …, x n (w)}, nonlinear expansion is performed through the polynomial layer of the deep residual Taylor network classification model to approximate the relationship between input and output;

[0017] 2) A residual connection mechanism is introduced between polynomial layers, allowing part of the input information to be directly passed to the output layer, reducing computational complexity and the risk of overfitting.

[0018] Furthermore, the specific steps of step 4 are:

[0019] 1) The Adam optimization algorithm is used to update the parameters of the deep residual Taylor network classification model, and the first-order moment estimate and the second-order moment estimate are combined to dynamically adjust the learning rate to adapt to the changes of different training data;

[0020] 2) The learning rate decay coefficient is introduced so that the learning rate gradually decreases with the increase of the number of training rounds, thereby ensuring that the model parameter update can gradually approach the global optimal point or the local optimal point.

[0021] Furthermore, the specific steps of step 5 are:

[0022] 1) Perform multiple iterations of DRN model training based on the training set and adjust model parameters to improve classification accuracy;

[0023] 2) Verify the classification performance of the model on the test set, and combine the PERCLOS p80 standard to perform fatigue classification evaluation, and ultimately achieve accurate classification of pilot fatigue status.

[0024] A second aspect of the present invention provides a pilot fatigue classification device based on EEG signals, comprising:

[0025] A signal acquisition unit is used to acquire EEG signal data, pre-process the signal and perform frequency band decomposition;

[0026] A feature extraction unit, used for decomposing the EEG signal into multiple sub-bands based on a frequency bandwidth of 2 Hz, extracting differential entropy features on each sub-band, and filtering the extracted features through a linear dynamic system to remove components unrelated to fatigue;

[0027] A model building unit is used to build a deep residual Taylor network classification model, which includes an input layer, an intermediate layer and an output layer. The input layer receives the EEG signal feature sequence x(w) after frequency band decomposition. The intermediate layer approximates the nonlinear relationship between input and output through a polynomial expansion with a highest power of m. The output layer combines the residual connection mechanism to finally output the classification result y(w) of the model.

[0028] The model solving unit is used to adaptively adjust the parameters of the deep residual Taylor network classification model using the Adam optimization algorithm, and dynamically adjust the learning rate of each parameter in combination with the learning rate attenuation coefficient to ensure that the model finds the global optimal or local optimal solution;

[0029] The classification output unit is used to construct the training set and the test set. The deep residual Taylor network classification model is trained based on the training set. The p80 standard of PERCLOS is used as the fatigue label value. The classification performance of the model is verified on the test set to obtain the classification result of pilot fatigue.

[0030] A third aspect of the present invention provides an electronic device, comprising a processor and a memory connected to the processor for storing executable instructions of the processor, wherein the processor is used to execute the above-mentioned method for classifying pilot fatigue based on EEG signals.

[0031] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned pilot fatigue classification method based on EEG signals.

[0032] Compared with the prior art, the pilot fatigue classification method based on EEG signals described in the present invention has the following advantages:

[0033] The present invention can effectively improve the accuracy of pilot fatigue classification and reduce the influence of noise and environmental factors on classification results; by using the DRN model, the overfitting and gradient vanishing problems that are prone to occur in traditional neural networks in high-order polynomial processing are solved, and the model parameters are adaptively adjusted through the Adam optimization algorithm to further improve the classification performance and robustness of the model; the present invention uses a method combining EEG signal feature extraction with LDS filtering to better capture the physiological characteristics related to the pilot's fatigue state and ensure the accuracy and reliability of the classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0035] Figure 1 It is a flow chart of a pilot fatigue classification method based on EEG signals of the present invention;

[0036] Figure 2 It is a schematic diagram of the structure of the deep residual Taylor network of the present invention;

[0037] Figure 3 A schematic diagram of a confusion matrix for driver fatigue classification according to the present invention;

[0038] Figure 4 It is a schematic diagram of the performance comparison between the deep residual Taylor network classifier of the present invention and other classifiers. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0040] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0041] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0042] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0043] Embodiment 1:

[0044] like Figure 1 As shown, the present invention proposes a pilot fatigue classification method based on EEG signals, comprising:

[0045] Step 1: Get EEG signal data, pre-process and decompose the signal, and extract the differential entropy eigenvalues ​​of each frequency band. After studying the four typical rhythmic waves in EEG signals, it was found that: δ (14-30Hz) reflects drowsiness; θ (4-7Hz) reflects frustration or depression; α (8-13Hz) reflects a state of quietness and concentration; β (0.5-4Hz) reflects tension, emotionality or excitement. The changes in these four rhythms are closely related to the level of mental fatigue. Differential entropy (DE) is Shannon's information entropy-∑ x The generalized form of p(x)log(p(x)) on continuous variables, DE is used to describe the complexity of continuous variables, and its calculation formula is:

[0046]

[0047] Where f(x) represents the probability density function of the random variable. In a fixed frequency band [a, b], the distribution of EEG signals can be approximated by the Gauss distribution N(μ, σ 2 ), so the differential entropy calculation formula is:

[0048]

[0049] Studies have shown that the performance of DE in estimating people's alertness is better than traditional power spectral density and other time-frequency features. Therefore, DE is used as the eigenvalue for alertness estimation.

[0050] Step 2: Use the linear dynamic system (LDS) to filter the extracted eigenvalues ​​to remove noise or interference information that is not related to fatigue, and obtain the final eigenvalues ​​as the input of the classification model. The specific steps are as follows:

[0051] 1) Perform short-time Fourier transform (STFT) on the EEG signal, decompose the signal into several sub-bands, and extract the differential entropy (DE) features on each sub-band; specifically, the data set used divides each channel signal equally according to a bandwidth of 2 Hz, that is, for the frequency band of 0-50 Hz, the signal can be decomposed into 25 frequency bands, [0, 2], [2, 4], ..., [48, 50], and then extract DE from each sub-band. The dimension of the obtained feature matrix is ​​N×M, where N represents the number of channels and M represents the number of divided sub-bands. Taking the experimental data set used as an example, the data format of the selected input features is lead×sample×band (17×885×25), where the first dimension 1-6 corresponds to the temporal lobe brain area T, and 7-17 corresponds to the occipital brain area P, indicating that the number of channels of the EEG signal is 17, the number of divided sub-bands is 25, and the number of samples is 885.

[0052] 2) Using the linear dynamic system (LDS) to filter the extracted eigenvalues, filtering out the noise or interference information irrelevant to fatigue, so as to obtain more accurate eigenvalues;

[0053] 3) Based on the extracted differential entropy feature matrix, the format of the input features is determined to be in the form of leads×samples×bands, where leads correspond to the number of channels of the EEG signal, samples correspond to the number of collected samples, and bands correspond to the number of sub-bands.

[0054] Step 3: Build the DRN model. The model consists of input layer, intermediate layer and output layer. Figure 2 As shown in , the input layer receives the preprocessed EEG signal features, the middle layer approximates the nonlinear relationship between input and output through polynomial expansion, and enhances the expression ability of the model through the residual connection mechanism; Figure 2As shown in the figure, x(w) represents the input feature sequence; y(w) represents the sum of various power polynomials expanded by MTN; W1, W2 are weight matrices; tanh and Softmax are the activation function types used by the network; Loss(w) is the loss value between the output obtained by the model and the label value, which is used for the autonomous update of the model parameters; η * represents the decaying learning rate, ensuring that the model finds at least one optimal solution; P(w) represents the final classification result obtained by the model, which is the output of the classification model. Specifically, it includes:

[0055] 1) Input feature sequence x(w) = {x1(w), x2(w), …, x n (w)}, nonlinearly expanded through the polynomial layer of DRN to approximate the relationship between input and output;

[0056] 2) A residual connection mechanism is introduced between polynomial layers, allowing part of the input information to be directly passed to the output layer, reducing computational complexity and overfitting risk and improving classification accuracy.

[0057] Step 4: Adopt the Adam optimization algorithm to adaptively adjust the parameters of the model, and dynamically adjust the learning rate of each parameter in combination with the learning rate attenuation coefficient to ensure that the model can gradually approach the global optimal point or the local optimal point during the training process; specifically:

[0058] 1) The Adam optimization algorithm is used to update the parameters of the DRN model, and the first-order moment estimate and the second-order moment estimate are combined to dynamically adjust the learning rate to adapt to the changes in different training data;

[0059] For the setting of model parameters, the ResMTN model combines the Adam algorithm, which uses the first-order moment estimate and the second-order moment estimate of the gradient to dynamically and autonomously adjust the learning rate of each parameter. In order to store the moving average of the squared gradient, Adam maintains the exponentially decaying average of the previous gradient and the current gradient. The following is the formula for the first-order moment estimate and the second-order moment estimate of the gradient in the form of a moving average:

[0060] m t =β1m t-1 +(1-β1)g t (3)

[0061]

[0062] Among them, m t represents the gradient first-order moment estimate (mean), v t represents the gradient second-order moment estimate (variance), g tRepresents the first-order gradient of the tth step. β1 and β2 are the exponential decay rates of the first-order moment estimate and the second-order moment estimate of the gradient set by humans, and β1, β2∈[0,1]. The update formula of the Adam algorithm is:

[0063]

[0064] Among them, η is the step size of the Adam algorithm, W t-1 and W t are the weight values ​​at step t-1 and step t, respectively. According to the Adam update formula, the exponential decay rates β1 and β2 of the first-order and second-order moment estimates of the gradient need to be manually configured to obtain satisfactory performance. The Adam algorithm combines the advantages of the AdaGrad algorithm, which is good at processing sparse gradients, and the Adadelta algorithm, which has a small memory requirement. The Adam algorithm can assign different adaptive learning rates to different weight parameters and is suitable for optimization in most non-convex environments. At the same time, as the most popular algorithm, Adam is suitable for high-dimensional spaces and relatively large data sets.

[0065] 2) In order to prevent the gradient descent from being unable to converge to the global optimal point, or even diverging, the learning rate decay coefficient is introduced into the Adam algorithm, so that the learning step size in the early stage of training is as large as possible within an acceptable range. At the same time, as the number of training rounds increases, the learning step size, i.e., the learning rate, becomes smaller and smaller, so that the learning result at least oscillates back and forth within an optimal range, or even gradually approaches the optimal point. Therefore, the update formula of the Adam algorithm combined with the learning rate decay is:

[0066]

[0067] Among them, η * represents the learning rate that decreases as the number of training steps t increases.

[0068] The sample data feature dimension and sample data number of the model are determined by the sample data. In addition, the highest power of the MTN layer polynomial expansion, the initial learning rate, the initial weight matrix W1, W2 and the maximum number of training iterations for each fold experiment need to be set manually according to the sample data. In a fold experiment, calculate the weight matrix W of the t-1th t-1 The first-order gradient of t According to formulas (3) and (4), we can obtain the first-order moment estimate (mean) m of the gradient: t and the second moment estimate (variance) v t , and then the updated weight matrix W is obtained according to formula (6) t , thereby achieving adaptive adjustment of the classification model parameters, obtaining better classification accuracy through repeated iterations, and completing the learning of model parameters through repeated training.

[0069] Step 5: Train the DRN model on the training set and use the PERCLOS p80 standard as the label value to classify fatigue. Finally, verify the classification performance of the model on the test set and obtain the classification results of the pilot's fatigue status. Specifically:

[0070] 1) Perform multiple iterations of DRN model training based on the training set and adjust model parameters to improve classification accuracy;

[0071] 2) Verify the classification performance of the model on the test set, and combine the PERCLOS p80 standard to perform fatigue classification evaluation, and ultimately achieve accurate classification of pilot fatigue status.

[0072] like Figure 3 As shown in the confusion matrix, DRN can achieve a classification accuracy of 97.65%. Figure 4 As shown, the classification performance of DRN is better than that of SVM, KNN, DT and LSTM. Compared with the prior art, the present invention can effectively improve the accuracy of pilot fatigue classification and reduce the impact of noise and environmental factors on the classification results. By using the DRN model, the overfitting and gradient vanishing problems that are prone to occur in traditional neural networks in high-order polynomial processing are solved, and the model parameters are adaptively adjusted through the Adam optimization algorithm to further improve the classification performance and robustness of the model. In addition, the present invention uses a method that combines EEG signal feature extraction with LDS filtering to better capture the physiological characteristics related to the pilot's fatigue state and ensure the accuracy and reliability of the classification results.

[0073] Embodiment 2:

[0074] A pilot fatigue classification device based on EEG signals, comprising:

[0075] A signal acquisition unit is used to acquire EEG signal data, pre-process the signal and perform frequency band decomposition;

[0076] A feature extraction unit, used for decomposing the EEG signal into multiple sub-bands based on a frequency bandwidth of 2 Hz, extracting differential entropy features on each sub-band, and filtering the extracted features through a linear dynamic system to remove components unrelated to fatigue;

[0077] A model building unit is used to build a deep residual Taylor network classification model, which includes an input layer, an intermediate layer and an output layer. The input layer receives the EEG signal feature sequence x(w) after frequency band decomposition. The intermediate layer approximates the nonlinear relationship between input and output through a polynomial expansion with a highest power of m. The output layer combines the residual connection mechanism to finally output the classification result y(w) of the model.

[0078] The model solving unit is used to adaptively adjust the parameters of the deep residual Taylor network classification model using the Adam optimization algorithm, and dynamically adjust the learning rate of each parameter in combination with the learning rate attenuation coefficient to ensure that the model finds the global optimal or local optimal solution;

[0079] The classification output unit is used to train the deep residual Taylor network classification model based on the training set, combine the p80 standard of PERCLOS as the fatigue label value, verify the classification performance of the model on the test set, and obtain the classification result of pilot fatigue.

[0080] Embodiment three:

[0081] An electronic device comprises a processor and a memory which is in communication with the processor and is used to store executable instructions of the processor, wherein the processor is used to execute the above-mentioned pilot fatigue classification method based on EEG signals.

[0082] Embodiment 4:

[0083] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for classifying pilot fatigue based on EEG signals is implemented.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A pilot fatigue classification method based on EEG signals, characterized by: The steps include: Step 1: Obtain EEG signal data, pre-process the signal and decompose the frequency band; Step 2: Decompose the EEG signal into multiple sub-bands based on a 2 Hz bandwidth, extract the differential entropy features on each sub-band, and filter the extracted features through a linear dynamic system to remove components unrelated to fatigue. Step 3: Construct a deep residual Taylor network classification model, which includes an input layer, an intermediate layer, and an output layer. The input layer receives the EEG signal feature sequence x(w) after frequency band decomposition. The intermediate layer approximates the nonlinear relationship between input and output through a polynomial expansion with the highest power of m. The output layer combines the residual connection mechanism to finally output the classification result y(w) of the model. Step 4: Adopt the Adam optimization algorithm to adaptively adjust the parameters of the deep residual Taylor network classification model. Combined with the learning rate attenuation coefficient, dynamically adjust the learning rate of each parameter to ensure that the model finds the global optimal or local optimal solution. Step 5: Construct training sets and test sets, train the deep residual Taylor network classification model based on the training set, combine the p80 standard of PERCLOS as the fatigue label value, verify the classification performance of the model on the test set, and obtain the classification results of pilot fatigue.

2. The pilot fatigue classification method based on EEG signals according to claim 1, characterized in that: The specific steps of step 2 are: 1) Perform short-time Fourier transform on the EEG signal, decompose the signal into several sub-bands, and extract the differential entropy features of each sub-band; 2) Using a linear dynamic system to filter the extracted eigenvalues ​​to remove noise or interference information that is not related to fatigue; 3) Based on the extracted differential entropy feature matrix, the format of the input features is determined to be in the form of leads×samples×bands, where leads correspond to the number of channels of the EEG signal, samples correspond to the number of collected samples, and bands correspond to the number of sub-bands.

3. The pilot fatigue classification method based on EEG signals according to claim 1, characterized in that: The specific steps of step three are: 1) Input feature sequence x(w) = {x1(w), x2(w), …, x n (w)}, nonlinear expansion is performed through the polynomial layer of the deep residual Taylor network classification model to approximate the relationship between input and output; 2) A residual connection mechanism is introduced between polynomial layers, allowing part of the input information to be directly passed to the output layer, reducing computational complexity and the risk of overfitting.

4. The pilot fatigue classification method based on EEG signals according to claim 1, characterized in that: The specific steps of step 4 are: 1) The Adam optimization algorithm is used to update the parameters of the deep residual Taylor network classification model, and the first-order moment estimate and the second-order moment estimate are combined to dynamically adjust the learning rate to adapt to the changes of different training data; 2) The learning rate decay coefficient is introduced so that the learning rate gradually decreases with the increase of the number of training rounds, thereby ensuring that the model parameter update can gradually approach the global optimal point or the local optimal point.

5. The pilot fatigue classification method based on EEG signals according to claim 1, characterized in that: The specific steps of step five are: 1) Perform multiple iterations of DRN model training based on the training set and adjust model parameters to improve classification accuracy; 2) Verify the classification performance of the model on the test set, and combine the PERCLOS p80 standard to perform fatigue classification evaluation, and ultimately achieve accurate classification of pilot fatigue status.

6. A pilot fatigue classification device based on EEG signals, characterized in that: include: A signal acquisition unit is used to acquire EEG signal data, pre-process the signal and perform frequency band decomposition; A feature extraction unit, used for decomposing the EEG signal into multiple sub-bands based on a frequency bandwidth of 2 Hz, extracting differential entropy features on each sub-band, and filtering the extracted features through a linear dynamic system to remove components unrelated to fatigue; A model building unit is used to build a deep residual Taylor network classification model, which includes an input layer, an intermediate layer and an output layer. The input layer receives the EEG signal feature sequence x(w) after frequency band decomposition. The intermediate layer approximates the nonlinear relationship between input and output through a polynomial expansion with a highest power of m. The output layer combines the residual connection mechanism to finally output the classification result y(w) of the model. The model solving unit is used to adaptively adjust the parameters of the deep residual Taylor network classification model using the Adam optimization algorithm, and dynamically adjust the learning rate of each parameter in combination with the learning rate attenuation coefficient to ensure that the model finds the global optimal or local optimal solution; The classification output unit is used to construct the training set and the test set. The deep residual Taylor network classification model is trained based on the training set. The p80 standard of PERCLOS is used as the fatigue label value. The classification performance of the model is verified on the test set to obtain the classification result of pilot fatigue.

7. An electronic device, comprising a processor and a memory connected to the processor for storing instructions executable by the processor, characterized in that: The processor is used to execute the pilot fatigue classification method based on EEG signals as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for classifying pilot fatigue based on EEG signals as described in any one of claims 1 to 5 is implemented.

Images