Flight state monitoring device and method based on electroencephalogram signal data classification

Abstract

The invention provides a flight state monitoring device and method based on electroencephalogram data classification, and relates to the technical field of data classification mode recognition, and the device specifically comprises an electroencephalogram data acquisition module, a data preprocessing and feature extraction module, a graph sequence attention network module and a state monitoring and sharing control module. According to the invention, by integrating the space domain and time sequence domain features and combining the graph convolution network, the long and short term memory network and the attention mechanism, the accuracy and effectiveness of flight state recognition are improved, the flight state change is described more accurately, sharing control is realized, and the driving safety is improved; the method has wide application value in the fields of intelligent driving, aerospace, fatigue monitoring and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of data classification pattern recognition, and in particular to a flight status monitoring device and method based on electroencephalogram signal data classification. Background Art

[0002] As human-machine interaction technology becomes increasingly prominent in aviation, Aircraft Shared Control (ASC), or Human-Autonomy Teaming (HAT), has become a key technology for flight safety and mission execution. This shared control system allows the autopilot to control the aircraft, dynamically adjusting control weights to enhance operational flexibility while ensuring safety. For example, during long-distance flights or high-load missions, the system proactively intervenes to share operational tasks, reducing the risk of human error.

[0003] However, the effectiveness of shared control relies on accurate perception and adaptation of flight conditions. Traditionally, flight conditions are assessed through expert interviews and questionnaires, such as the NASA-TLX Task Load Scale and the Fatigue Symptom Questionnaire. However, these indirect analysis methods have significant shortcomings. Questionnaires rely on individual subjective feedback, making the results susceptible to subjective factors and individual differences and lacking stability. Questionnaire assessments often require the subject to interrupt their task, making them difficult to implement in real-world flight environments. Furthermore, subjective assessments can only be conducted at specific time points and fail to provide continuous state monitoring data that closely correlates with dynamic mission changes. These issues limit the effectiveness and accuracy of assessment results. With the continuous advancement of ergonomic sensor and chip technologies, as well as the continued development of machine learning and deep learning algorithms, physiological computing has gradually become an application model. By leveraging multimodal physiological data and building classification models using machine learning or deep learning techniques, an effective mapping relationship between physiological data and flight conditions can be established. This allows for the development of real-time, highly accurate, and robust flight condition monitoring devices, providing reliable support for dynamic state perception.

[0004] In existing research, the application of EEG (Electroencephalogram) signals has mainly focused on analyzing and classifying them using their frequency domain characteristics. However, this approach has certain limitations in practice, namely, it ignores the multi-channel spatial characteristics and temporal information characteristics exhibited by EEG data. Based on this, the present invention proposes an improved data classification method that improves the performance of EEG signal classification by utilizing the spatial and temporal domain information of EEG signals. Summary of the Invention

[0005] In order to address the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a flight status monitoring device and method based on EEG signal data classification. By integrating spatial domain and temporal domain features, combining graph convolutional networks, long short-term memory networks and attention mechanisms, the accuracy of state recognition is improved and flight status changes are more accurately portrayed; based on data pattern recognition technology, a shared control scheme is implemented to improve driving safety; and it has broad application value in intelligent driving, aerospace, fatigue monitoring and other fields.

[0006] Specifically, on one hand, the present invention provides a flight status monitoring device based on EEG signal data classification, which includes: an EEG signal data acquisition module, a data preprocessing and feature extraction module, a graph sequence attention network module, and a status monitoring and sharing control module;

[0007] The EEG signal data acquisition module acquires EEG signal data corresponding to the flight status and annotates the EEG signal data through fatigue monitoring;

[0008] The data preprocessing and feature extraction module is used to perform bandpass filtering and normalization to filter out environmental noise and physiological artifacts, and use differential entropy to preprocess the EEG signal data, perform multi-band feature extraction to extract key information of the frequency band, and obtain differential evolution features;

[0009] The graph sequence attention network module is used to learn and extract the composite features of EEG signal data in spatial topology and time series. Its structure includes: graph convolutional network layer, bidirectional long short-term memory network layer, attention layer, fully connected layer and softmax output layer; the graph convolutional network layer forms the spatial correlation between electrodes by constructing the adjacency matrix between EEG channels, and updates the features of each node through graph convolution operation; the bidirectional long short-term memory network layer captures the temporal evolution of EEG signals from the forward and backward time series paths respectively, and fuses them to obtain temporal context information; the attention layer weights each time parameter of the output to improve classification performance and model interpretability; the fully connected layer and softmax output layer map the results to flight status classification results;

[0010] The state monitoring and shared control module identifies and classifies the flight state, and uses the flight state classification results as the adjustment factor for the dynamic weight of the shared control system. It includes a monitoring submodule and a shared control submodule; the monitoring submodule inputs data into the graph sequence attention network module, outputs the flight state label and continuously updates it; the shared control submodule adjusts the control weight of the autopilot device according to the flight state level output by the flight state.

[0011] Preferably, the flight status classification results specifically include: normal state Alert, reminder state Mild Fatigue and alarm state Severe Fatigue; in the normal state Alert, the autopilot device only provides basic flight status monitoring and assistance; in the reminder state Mild Fatigue, the autopilot device moderately intervenes in flight control, and ensures safety through attitude correction, voice and visual reminders, etc.; in the alarm state Severe Fatigue, the autopilot device performs flight control and simultaneously starts the multi-modal wake-up and alarm mechanism.

[0012] Preferably, the frequency ranges corresponding to the differential evolution features are: first feature Delta: 1-4 Hz, second feature Theta: 4-8 Hz, third feature Alpha: 8-14 Hz, fourth feature Beta: 14-31 Hz and fifth feature Gamma: 31-50 Hz.

[0013] Preferably, the graph convolutional network layer is constructed based on the spatial distribution relationship of the EEG signal data electrodes, specifically:

[0014]

[0015] Among them, A is the adjacency matrix; A ij is the connection strength between electrode i and electrode j; d i,j is the Euclidean distance between electrodes i and j; δ is a normalization factor, which is set to 9 to ensure the sparsity of the matrix; i is the first electrode number of the adjacency matrix; j is the second electrode number of the adjacency matrix.

[0016] Preferably, the graph convolutional network layer updates the features of each node through graph convolution operations, and the input feature matrix of the lth layer is:

[0017]

[0018] Among them, H( l ) is the input feature matrix of the lth layer; D is the degree matrix; W( l ) is the training weight matrix; σ1 is the nonlinear activation function.

[0019] Preferably, the bidirectional long short-term memory network layer learns the temporal characteristics of the EEG signal data signal, specifically:

[0020]

[0021] in, is the hidden state of the forward LSTM; is the hidden state of the backward LSTM; is the output of the bidirectional long short-term memory network layer, which is composed of the output of the forward and backward LSTM; t is the time parameter; LSTM forward Forward long short-term memory network; LSTM backward It is a backward long short-term memory network.

[0022] In another aspect, the present invention provides a flight status monitoring method based on EEG signal data classification, comprising:

[0023] S1: Obtain EEG signal data corresponding to the flight state, annotate the EEG signal data through fatigue monitoring, preprocess the EEG signal data using differential entropy, extract differential evolution features, and perform bandpass filtering and normalization;

[0024] S2: Construct a flight status monitoring model based on a neural network structure of a graph convolutional network-long short-term memory network and an attention layer to analyze EEG signal data and complete flight status classification. The model includes a graph convolutional network layer, a long short-term memory network layer, an attention layer, and a fully connected layer. The EEG signal data after normalization in step S1 is input into the graph convolutional network layer to extract spatial topological features. The long short-term memory network layer learns the time evolution pattern of the EEG signal data signal through a gating mechanism. The attention layer performs weighted processing on the output of the long short-term memory network layer to obtain information on key time parameters. The flight status classification result is output through the fully connected layer.

[0025] S3: Using the EEG signal dataset processed in step S1 to train the flight status monitoring model constructed in step S2, completing the flight status monitoring model training, and using the cross entropy loss function to improve the accuracy of the flight status classification results;

[0026] S4: Using the flight status monitoring model obtained in step S3 to identify and classify the flight status, and using the flight status classification result as a regulating factor for the dynamic weight of the shared control system.

[0027] Preferably, step S1 specifically includes:

[0028] S11: Acquire multiple types of EEG signal data in flight status and annotate the EEG signal data through fatigue monitoring;

[0029] S12: Use differential entropy to preprocess EEG signal data and perform bandpass filtering to reduce artifacts and noise;

[0030] S13: Extract differential evolution features, calculated using short-term Fourier transform of non-overlapping windows;

[0031] S14: Use standardization methods to normalize EEG signal data to eliminate individual differences, improve training stability, and enhance generalization ability.

[0032] Preferably, step S2 specifically includes:

[0033] S21: Using graph convolutional networks to construct the topological structure of EEG signal data to extract features using spatial correlation between electrodes;

[0034] S22: Use the long short-term memory network layer to learn the time series characteristics of EEG data signals, and selectively retain or discard historical information through a gating mechanism;

[0035] S23: Introduce the attention layer to calculate the importance weight of the time parameter, and perform weighted summation on the output of the long short-term memory network layer, input it into the fully connected layer, and calculate the probability distribution of the flight state.

[0036] Preferably, step S3 is specifically:

[0037] S31: partitioning EEG signal dataset;

[0038] S32: independent partitioning combined with 5-fold cross-validation and leave-one-subject-out cross-validation;

[0039] S33: The cross entropy loss function is used in the training phase to measure the classification error of the flight status monitoring model.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) Compared with traditional EEG monitoring methods that usually only focus on a single feature dimension, such as frequency domain or time domain information, the present invention combines the characteristics of multiple dimensions such as space, time, and frequency, fully utilizes the complex information structure of EEG signals, and improves the accuracy of flight status recognition.

[0042] (2) Compared with the existing EEG classification methods, most of them are based on traditional CNN or RNN structures, which cannot effectively capture the spatial topological structure of EEG signals. GCN is used to construct the spatial dependency relationship between EEG electrodes, making the information transmission between different brain regions more reasonable and improving the ability to characterize complex brain activities.

[0043] (3) Based on the strong time dependence of EEG signals, the present invention uses a long short-term memory network (LSTM) to process time series data, effectively modeling the trend of flight status evolution over time and improving the ability to detect progressive fatigue states.

[0044] (4) Compared with the traditional long short-term memory network (LSTM) method, which is easily affected by the "information decay" problem of long sequence information, the present invention introduces the attention layer, which can dynamically allocate weights in the entire EEG signal time series, highlighting the characteristics of key time parameters, thereby improving the accuracy of state recognition.

[0045] (5) The feature information of the five frequency bands of Delta, Theta, Alpha, Beta, and Gamma selected by the differential evolution DE feature in the present invention can effectively reduce the dimension of the original EEG data while retaining the key information of the EEG signal; the Z-score standardization is used to reduce the influence of physiological differences between individuals, so that the model can adapt to different individuals and improve the generalization ability.

[0046] (6) The present invention dynamically adjusts the driving control mode by monitoring the flight status as the input of the shared control system; this shared control mode based on EEG data can dynamically optimize the allocation of driving rights, ensure that the automatic driving device can intervene in time, and improve driving safety and system intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a structural diagram of a flight status monitoring device based on EEG signal data classification according to the present invention;

[0048] Figure 2 This is a flow chart of the flight status monitoring method based on EEG signal data classification of the present invention;

[0049] Figure 3 This is a diagram showing the specific locations of the wearable EEG acquisition electrodes in the present invention;

[0050] Figure 4 Schematic diagram of a three-class confusion matrix in an embodiment of the present invention;

[0051] Figure 5 Schematic diagram showing the performance comparison between the method of the present invention and various comparative methods on different indicators. DETAILED DESCRIPTION

[0052] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0053] This embodiment discloses a flight status monitoring device based on EEG signal data classification, the structure of which is as follows: Figure 1 As shown in the figure, it has been systematically designed at both the hardware and software levels to meet the high reliability and timeliness requirements for data acquisition, processing and state recognition in actual flight scenarios. It includes: EEG signal data acquisition module, data preprocessing and feature extraction module, image sequence attention network module and state monitoring and shared control module.

[0054] The EEG signal data acquisition module acquires EEG signal data corresponding to the flight status and annotates the EEG signal data through fatigue monitoring; the EEG signal data acquisition module is responsible for acquiring EEG signal data during the flight. The core equipment in the embodiment is a multi-channel EEG acquisition helmet or an EEG electrode array; each electrode channel collects signals from different brain regions for subsequent feature extraction and status analysis; the acquisition end and the main control center transmit data via wired or wireless means, and are equipped with necessary front-end filtering and amplification circuits to determine signal quality and transmission stability.

[0055] The data preprocessing and feature extraction module performs bandpass filtering and normalization to remove environmental noise and physiological artifacts. It also uses differential entropy to preprocess the EEG signal data and extract differential evolution features. These features correspond to the following frequency ranges: Delta (1-4 Hz), Theta (4-8 Hz), Alpha (8-14 Hz), Beta (14-31 Hz), and Gamma (31-50 Hz). By applying multi-band feature extraction to the preprocessed EEG data, key frequency band information is extracted. This effectively preserves the spatiotemporal patterns of the EEG signals while significantly reducing data dimensionality, improving the computational efficiency of subsequent analysis.

[0056] The Graph-Sequential Attention Network (GSAN) module, also known as the Graph Convolutional Network-Long Short-Term Memory Network and Attention Layer model, is used to learn and extract the composite features of EEG signal data in terms of spatial topology and temporal sequence. Its structure includes a graph convolutional network layer, a bidirectional long short-term memory network layer, an attention layer, a fully connected layer, and a softmax output layer. The graph convolutional network layer forms spatial associations between different electrodes by constructing an adjacency matrix between EEG channels. The graph convolutional network layer is constructed based on the spatial distribution of electrodes in the EEG signal data. Specifically,

[0057]

[0058] Among them, A is the adjacency matrix; A ij is the connection strength between electrode i and electrode j; d i,j is the Euclidean distance between electrodes i and j; δ is a normalization factor, which is set to 9 to ensure the sparsity of the matrix; i is the first electrode number of the adjacency matrix; j is the second electrode number of the adjacency matrix.

[0059] The graph convolutional network layer updates the features of each node through graph convolution operations. The input feature matrix of the lth layer is:

[0060]

[0061] Among them, H( l ) is the input feature matrix of the lth layer; D is the degree matrix; W( l ) is the training weight matrix; σ1 is the nonlinear activation function.

[0062] The bidirectional long short-term memory network layer captures the temporal evolution of EEG signals from the forward and backward time series paths respectively, and fuses them to obtain temporal context information. The bidirectional long short-term memory network layer learns the temporal characteristics of EEG signal data signals, specifically:

[0063]

[0064] in, is the hidden state of the forward LSTM; is the hidden state of the backward LSTM; is the output of the bidirectional long short-term memory network layer, which is composed of the output of the forward and backward LSTM; t is the time parameter; LSTM forward Forward long short-term memory network; LSTM backward It is a backward long short-term memory network.

[0065] The attention layer assigns different weights to each output time parameter to improve classification performance and model interpretability; the fully connected layer and the Softmax output layer map the results into flight status classification results; the flight status classification results specifically include: normal state Alert, reminder state Mild Fatigue and alarm state Severe Fatigue; in the normal state Alert, the autopilot device only provides basic flight status monitoring and assistance; in the reminder state Mild Fatigue, the autopilot device moderately intervenes in flight control and ensures safety through attitude correction, voice and visual reminders; in the alarm state Severe Fatigue, the autopilot device performs flight control and simultaneously activates the multimodal wake-up and alarm mechanism.

[0066] The state monitoring and shared control module identifies and classifies flight states and uses the classification results as a factor in adjusting the dynamic weights of the shared control system. Specifically, it comprises a monitoring submodule and a shared control submodule. After acquiring EEG signals and extracting features, the monitoring submodule feeds the data into a trained GCN-LSTM-attention layer model for inference, outputting a flight state label and continuously updating it. The shared control submodule dynamically adjusts the control weights of the autopilot based on the flight state levels (Alert, Mild Fatigue, and Severe Fatigue).

[0067] By integrating these functional modules, the present invention's flight status monitoring device, based on EEG signal data classification, not only accurately and timely identifies flight status but also dynamically adjusts shared control strategies. Compared to traditional methods, this device demonstrates superior flight safety, generalizability, and effectiveness, providing new technical approaches and theoretical support for ensuring flight safety and mission success in increasingly complex and volatile flight environments.

[0068] The second aspect of the present invention proposes a flight status monitoring method based on EEG signal data classification, such as Figure 2 As shown, EEG signal data corresponding to the flight state is obtained and processed; a flight state monitoring model based on a neural network structure of a graph convolutional network-long short-term memory network and an attention layer is constructed; the flight state monitoring model is trained using the preprocessed EEG signal data set; and the flight state is identified using the flight state monitoring model. The process includes the following steps:

[0069] Step S1: Obtain EEG signal data corresponding to the flight state, annotate the EEG signal data through fatigue monitoring, preprocess the EEG signal data using differential entropy, extract differential evolution features, and perform bandpass filtering and normalization.

[0070] Step S11: Obtain multiple types of EEG signal data in flight status through the public EEG signal dataset SEED-VIG, and annotate the EEG signal data using the fatigue monitoring PERCLOS method.

[0071] This paper uses the SEED-VIG dataset, a publicly available multimodal dataset on driver fatigue and alertness. This dataset primarily includes EEG (electroencephalogram) data, EOG (electrooculogram) data, and eye tracking data. This dataset, collected by the Bio-Intelligence Laboratory of Shanghai Jiao Tong University, is specifically designed to study flight status during extended driving. The SEED-VIG dataset, based on a large number of subjects' EEG data collected in driving simulation experiments, provides high-quality EEG data for flight status classification.

[0072] During the experiment, a long-term driving task was carried out in a VR driving simulation environment. The experiment lasted about 2 hours. During this period, the driving environment was controlled, such as monotonous scenes, long-term driving, etc.; EEG signal data corresponding to different flight states were obtained, and the EEG signal data in the three states of normal state Alert, reminder state Mild Fatigue and alarm state Severe Fatigue were recorded respectively to construct a complete flight state classification data set. The SEED-VIG data set provides the present invention with high-quality EEG signal data under a variety of flight states, and determines the availability and reliability of the data through experimental design. The use of this data set enables the present invention to train a flight state classification model based on the GCN-LSTM-attention layer in a real driving simulation environment, and to determine that the model has good generalization ability and practical application value. Figure 3 The figure shows the specific location of the wearable EEG acquisition electrode points in the present invention, provides a schematic diagram of the electrode distribution of the wearable EEG acquisition device in actual application, and shows the specific layout of multi-channel electrodes at different locations on the scalp. The embodiment of the present invention can also use the EEG signal data acquisition device to collect EEG signals in flight. The number of signal channels of the EEG signal data is set to 18 channels, and the electrode placement is as follows: Figure 3 As shown in the figure, this system, based on a 10-20 electrode placement system, determines the EEG signals collected to accurately identify flight status. Using the method in step S11, the collected data is processed to construct a dataset. This data is then input into the GCN-LSTM-attention layer flight status monitoring model trained in step S2 to obtain the output result vector. The category with the highest probability is selected as the final status classification result.

[0073] This paper uses the raw data from the SEED-VIG dataset as the data source. The data labels are divided based on the PERCLOS method of eye tracking data to calculate the percentage of duration within a specified time interval, specifically:

[0074]

[0075] interval=blink+fixation+saccade+CLOS;

[0076] PERCLOS is the fatigue monitoring assessment index for EOG data, an important indicator of fatigue or distraction, commonly used in driving monitoring, fatigue detection, and other fields. Interval is the total time interval of the EOG data, which can also be understood as the total observation time during the monitoring period. Blink is the total blink time of the EOG data, which is the time spent closing and then quickly opening the eyes during a monitoring period. CLOS is the total length of time spent closed, typically referring to eye closure other than normal blinking, such as that caused by fatigue. Fixation is the total gaze time of the EOG data, which is the time spent looking at a fixed position, usually associated with focused attention. Saccade is the total time spent in EOG data, which is rapid eye movement, the process of rapidly moving the eyes from one fixation point to another.

[0077] Electrooculogram (EOG) data was used to label each time point in the electroencephalogram (EEG) data, using two thresholds of 0.35 and 0.7 to classify flight states. After preprocessing, each sample in the SEED-VIG dataset collected an average of approximately 18 to 20 valid EEG segments, each lasting approximately 60 seconds. Based on the thresholds of 0.35 and 0.7 set by the PERCLOS method, the final label distribution was: Alert (normal state), accounting for approximately 40%; Mild Fatigue (alert state), accounting for approximately 35%; and Severe Fatigue (alarm state), accounting for approximately 25%. These distributions align with the probability of occurrence of different fatigue levels in real-world driving environments, further enhancing the dataset's credibility and representativeness.

[0078] Step S12: Use differential entropy to preprocess the EEG signal data; the original EEG signal data is processed with a bandpass filter between 1 and 75 Hz to reduce artifacts and noise, and downsampled to 200 Hz to reduce computational complexity. The differential entropy expression is:

[0079]

[0080] Where h(X) is the differential entropy of the EEG eigenvalues ​​of the EEG data; σ is the standard deviation of the EEG eigenvalues ​​of the EEG data; X is the original EEG eigenvalues ​​of the EEG data; μ is the mean of the EEG eigenvalues ​​of the EEG data; f(x) is the probability density function PDF of the normal distribution, which is used to describe the probability density of the random variable X at the value x; e is a natural constant, approximately 2.71828; x is a variable with a value, which is an arbitrary real number in the range (-∞, +∞)(-∞, +∞); log is a logarithmic function, generally referring to the natural logarithm; exp is an exponential function; and π is the pi constant.

[0081] Step S13: For each EEG segment, differential evolution DE features are extracted from five frequency bands: Delta: 1-4 Hz, Theta: 4-8 Hz, Alpha: 8-14 Hz, Beta: 14-31 Hz, and Gamma: 31-50 Hz. All differential evolution DE features are calculated using a short-term Fourier transform with a 0.5 s non-overlapping window. The main reason for selecting the feature information of the five frequency bands of Delta, Theta, Alpha, Beta, and Gamma for differential evolution DE features is that it effectively reduces the dimensionality of the original EEG data, improves computational efficiency, and retains the key information of the EEG data, enabling the model to more accurately learn changes in flight status.

[0082] Step S14: Since the numerical range of EEG signal data fluctuates greatly due to individual differences, the present invention uses the Z-score normalization method to normalize the data, eliminate individual differences, improve training stability, and enhance generalization ability. The calculation expression is:

[0083]

[0084] Among them, X z is the normalization result of the original EEG signal data EEG feature value.

[0085] Step S2: Construct a flight status monitoring model based on a neural network structure of graph convolutional network-long short-term memory network and attention layer to analyze EEG signal data to complete flight status classification. The specific structure of the model is as follows: Figure 2 The figure shows a schematic diagram of the flight status monitoring model of the present invention, which includes a graph convolutional network layer, a long short-term memory network layer, an attention layer and a fully connected layer; the EEG signal data after normalization processing in step S1 is input into the graph convolutional network layer to extract spatial topological features, the long short-term memory network layer learns the time evolution pattern of the EEG signal data signal through a gating mechanism, the attention layer performs weighted processing on the output of the long short-term memory network layer to obtain information on key time parameters, and outputs the flight status classification result through the fully connected layer; the overall process of the model is: first, the graph convolutional network GCN layer receives the EEG signal data and extracts spatial topological features through a graph convolution operation; then, the features extracted by the graph convolutional network GCN layer are input into the long short-term memory network LSTM layer, and the long short-term memory network LSTM layer learns the time evolution pattern of the EEG signal data through a gating mechanism; then, the attention layer performs weighted processing on the output of the long short-term memory network LSTM layer to obtain information on key time parameters; finally, the flight status classification is completed through the fully connected layer.

[0086] Step S21: Construct a graph convolutional network (GCN) layer. The GCN layer leverages the spatial structure of EEG electrodes in the EEG data to extract local and global spatial patterns through the feature aggregation mechanism of graph neural networks. The core function of the long short-term memory (LSTM) layer is to model the temporal trends of the EEG data and combine historical information to predict the current flight state. The attention layer is designed to address the limitations of the LSTM by assigning different weights to different time parameters, thereby giving greater attention to the characteristics of key time parameters during the classification process. The fully connected layer is used for final state prediction. Using a softmax activation function, it calculates the probability distribution of the flight state and outputs the classification results for the normal flight state (Alert), the warning state (Mild Fatigue), and the alarm state (Severe Fatigue).

[0087] The GCN layer is designed to fully capture the rich spatial dependency information in EEG signals. Compared to traditional GCNs based on fixed topology, this paper innovatively proposes a two-level spatial feature fusion GCN architecture. This architecture extracts EEG spatial features at both the local and global levels, effectively integrating multi-scale spatial information to improve the accuracy and robustness of flight status monitoring.

[0088] The present invention makes innovative improvements based on the traditional graph convolutional network (GCN) and proposes a refined and dynamically optimized local spatial feature extraction method. Traditional graph convolutional networks (GCNs) are often based on simple binary adjacency relationships, ignoring the impact of the actual physical distance between electrodes on the propagation and spatial correlation of EEG signals. To overcome this problem, the present invention proposes a method for constructing a weighted adjacency matrix of the three-dimensional spatial distance between electrodes: First, the Euclidean distance between any two electrode nodes i and j is calculated based on the three-dimensional spatial coordinates of the electrode layout:

[0089]

[0090] Where: d ij is the three-dimensional Euclidean distance between electrode node i and node j; (x i ,y i ,z i ) is the three-dimensional space coordinate of electrode node i; (x j ,y j ,z j ) is the three-dimensional space coordinate of electrode node j.

[0091] On this basis, the adjacency matrix is ​​proposed to be set as:

[0092]

[0093] Among them, A ij is the element value of the adjacency matrix between node i and node j; δ is the sparse factor, which is used to control the overall sparsity of the adjacency matrix; d ij is the spatial distance between electrode node i and node j.

[0094] Through the above-mentioned refined design, the adjacency matrix more realistically and accurately reflects the actual physiological spatial relationship between local electrode nodes, and improves the feature learning accuracy of the model at the local scale. In order to further improve the flexibility and adaptability of local spatial feature extraction, the present invention proposes to design the non-zero elements in the above-mentioned initial adjacency matrix as parameters learned during the network training process, thereby realizing dynamic optimization of the adjacency relationship. Specifically, a training parameter matrix P with the same shape as the initial adjacency matrix A is introduced, and its elements are mapped to (0,1) through the sigmoid function, and the optimized adjacency matrix is ​​set to:

[0095] A dynamic =A ij ⊙M,M=σ(P);

[0096] Among them: A dynamic is the adjacency matrix after dynamic optimization; M is the training parameter matrix after mapping by the sigmoid function; P is the parameter matrix learned during network training; σ(x) is the Sigmoid activation function, which is set to ⊙ is the element-wise Hadamard product.

[0097] Using the backpropagation algorithm, the network loss function is optimized by continuously adjusting the parameter matrix P, thereby dynamically updating the strength of local adjacency relationships. The dynamically optimized adjacency matrix adapts to the varying spatial dependencies of the dataset during training, making local spatial feature extraction more accurate.

[0098] As the number of layers in traditional graph convolutional networks (GCNs) increases, node features tend to become similar, resulting in a serious over-smoothing problem, which reduces the model's sensitivity to local feature changes. To address this issue, the present invention innovatively designs a cross-layer residual connection mechanism to explicitly retain the rich spatial features of each layer. The mathematical setting of the cross-layer residual connection is:

[0099]

[0100] Where: H( l+1 ) is the node feature output by the l+1 layer; H( l ) is the node feature of the l-th layer input; W( lis the training parameter matrix of the lth graph convolutional layer; σ(·) is the activation function (e.g., ReLU); is the normalized graph Laplacian matrix; is the dynamic adjacency matrix after adding the identity matrix; is the node degree diagonal matrix; I is the identity matrix.

[0101] Through cross-layer residual connections, the output of each layer not only contains the results of the current convolution calculation, but also directly accumulates the feature inputs of the previous layer. This mechanism effectively alleviates the phenomenon of node features gradually converging with increasing network depth, enhances the network's sensitivity to subtle changes in local spatial patterns, and improves the accuracy of flight status monitoring.

[0102] To fully capture these cross-regional, global-scale spatial features, this paper innovatively proposes a global spatial feature extraction framework based on the self-attention mechanism. Compared to traditional graph convolutional networks (GCNs), which focus only on fixed local topological structures, this method automatically discovers and learns the dynamic spatial connections between EEG nodes at a global scale.

[0103] In order to fully preserve the local spatial structure information and the global information of the original EEG data, the present invention innovatively proposes to extract the global spatial features using the original EEG feature X i征put and the spatial feature H extracted by local GCN local After splicing and fusion, the input features of the global feature extraction module are as follows:

[0104] X global =concat(X i征put ,H locHl );

[0105] Where: X global is the input feature for graph convolution; It is the original feature matrix of EEG node, which contains complete information without convolution processing. is the node feature matrix after local spatial feature extraction; concat(·) is the feature concatenation operation, which determines that the model combines both the original signal and local structural information, thus laying a richer and more discriminative foundation for global scale feature learning.

[0106] To accurately capture the dynamically changing functional connections between EEG nodes, this paper innovatively introduces a multi-head self-attention mechanism to dynamically learn and model the global connections between nodes. The specific mathematical expression is:

[0107] First, set the query, key, and value vectors to:

[0108] Q=X global W q ;

[0109] K=X global W k ;

[0110] V=X global W v ;

[0111] Where: Q is the query vector; K is the key vector; V is the value vector; X global is the training parameter influence coefficient; is the first training parameter; is the second training parameter; is the third training parameter; D k is the feature dimension in the attention head.

[0112] Calculate the multi-head attention weight as:

[0113]

[0114] Among them, Attention(Q,K,V) is the output feature of a single attention head; softmax(·) is a normalization function that ensures that the sum of attention weights is 1.

[0115] The results of using multi-head attention to further enhance the network to capture complex global spatial associations are:

[0116] MultiHead(X global )=concat(head1,...,head h )W o

[0117] Among them, MultiHead(X global ) is the output of the multi-head attention mechanism; head i is the attention of the th head; concat is the connection function; h is the total number of attention heads; W o is the weight matrix for output integration.

[0118] The calculation result for each head is:

[0119]

[0120] The multi-head attention mechanism enables the model to focus on the interactions between nodes from multiple different spatial subspaces at the same time, thereby mining the complex spatial feature change patterns in a more detailed and comprehensive manner.

[0121] To avoid the problem of over-smoothing caused by the adjacency relationship calculated by the global attention mechanism being too dense, this paper innovatively designs a dynamic sparsification strategy based on attention weights. Set the sparse global adjacency matrix:

[0122]

[0123] in: is the global scale adjacency matrix after dynamic sparsification; Attention ij is the attention weight from node i to node j; Top-k (Attention i,: ) is to retain the top k% of connections with the largest attention weights for each node i to ensure that the global spatial relationship is sparse and accurate; this sparsification strategy avoids the problem of excessive smoothing of the network at the global scale, while also reducing the computational redundancy caused by invalid connections and improving the generalization performance of the network.

[0124] Using the dynamic adjacency matrix after the above sparse processing The global feature information is further propagated and aggregated through the graph convolution network. The graph convolution is set as:

[0125]

[0126] Among them: O( l+1 ) is the output feature after the l+1th layer global scale graph convolution; is the training weight matrix of the global graph convolutional network at layer l; σ(·) is the activation function; is the normalized global graph Laplacian matrix, used to aggregate global node features; X global is the input feature for graph convolution.

[0127] Normalized global graph Laplacian matrix Set to:

[0128]

[0129] in, Add the identity matrix to the dynamically sparsified adjacency matrix; is the global node degree diagonal matrix; I is the identity matrix.

[0130] This innovative design makes the propagation of global spatial information more dynamic and precise, effectively capturing the complex, sparse, and time-varying functional connectivity patterns between regions. Both the local and global GCN models of this invention use a two-layer GCN structure, with output dimensions of 10×18×10 and 10×18×15, respectively; the input and output dimensions of the graph convolutional network GCN layer are 10×18×5 and 10×18×25, respectively. Dimensions are defined as: time parameter length × electrode channel × feature.

[0131] Step S22: Construct a long short-term memory network (LSTM) layer. The LSTM layer is used to learn the time series characteristics of the EEG signal and selectively retain or discard historical information through a gating mechanism. The core calculation process of the LSTM includes a forget gate, an input gate, and an output gate.

[0132] The forget gate is used to determine whether the state information of the parameters at the previous time should be retained. The calculation expression is:

[0133] f t =σ1(W f ·[h t-1 ,x t ]+b f );

[0134] Among them, f t is the output of the forget gate; x t is the input feature of the current time parameter; h t-1 is the hidden state of the parameter at the previous time; W f is the first training parameter of the forget gate; b f is the second training parameter of the forget gate.

[0135] The input gate is used to determine whether the new information of the current time parameter should be stored in the cell state and perform nonlinear transformation on it. The calculation expression is:

[0136] i t =σ1(W i ·[h t-1 ,x t ]+b i );

[0137]

[0138] Among them, i t is the activation value of the input gate, between 0 and 1, indicating how much information of the input of the current time parameter t is written into the memory unit. is the candidate memory value, which is the “original content” of the current input’s impact on the memory unit; W i is the weight matrix of the input gate; b i is the bias term of the input gate; ta is a hyperbolic tangent function with an output range between (-1, 1), which is used to provide nonlinear transformation for candidate memory; W C is the weight matrix of the candidate memory unit; b C is the bias term of the candidate memory unit.

[0139] The calculation expression of the output gate is:

[0140] o t =σ1(W o ·[h t-1 ,x t ]+b o );

[0141] h t =o t tanh(C t );

[0142] Among them, C t is the unit state of the current time parameter; h t is the hidden state of the current time parameter; W o is the weight matrix of the output gate; b o is the bias term of the output gate; o t Activation value for the output gate.

[0143] The use of a bidirectional long short-term memory network layer Bilstm can better learn the temporal characteristics of EEG signals. It consists of two long short-term memory network LSTM layers, a forward LSTM: from t = 1 to t = T, and a reverse LSTM: from t = T to t = 1.

[0144] Step S23: Since the LSTM layer only uses the last hidden state for classification, some information about the time parameters is lost. To solve this problem, the present invention introduces an attention layer to calculate the importance weights of the time parameters and perform a weighted summation of the LSTM outputs. The attention layer calculation process is as follows:

[0145] e t =V T tanh(W a h t +b a );

[0146]

[0147] Among them, e tis the attention energy at the current time parameter t, also known as the "alignment score", which measures the current hidden state h t Relevance to the context; b a is the bias term of the attention layer; h att is the final attention context vector, which is the sum of all hidden states h t The weighted sum of a is the first training parameter of the attention layer; V is the second training parameter of the attention layer; a t is the weight of the time parameter t; T is the total amount of time.

[0148] The features output by the attention layer are input into the fully connected layer, and the probability distribution of the flight state is calculated as:

[0149] y=W f h att +b f ;

[0150] Where y is the model output vector, which represents the original distribution or unnormalized probability distribution of each flight state; b f is the output layer bias term, which is used to adjust the baseline value of each category score.

[0151] Finally, the Softmax function is used to calculate the probability distribution of the three states: normal state Alert, reminder state Mild Fatigue, and alarm state Severe Fatigue, and output the flight state classification result.

[0152] Step S3: Use the EEG signal dataset processed in step S1 to train the flight status monitoring model constructed in step S2 to complete the flight status monitoring model training, and use the cross entropy loss function to improve the accuracy of the flight status classification results.

[0153] Step S31: Divide the EEG signal dataset. In the experimental design of this invention, reasonable data partitioning is crucial to the generalization ability of the model. To determine the sufficiency of the training data and simultaneously evaluate the practical application of the model, this invention adopts an 80% training set + 20% test set partitioning method to maximize data utilization while leaving sufficient data for performance evaluation.

[0154] Step S32: Building on the data partitioning strategy of step S31, the present invention utilizes a subject-independent partitioning approach for EEG signal data sources, i.e., the data sources of the training and test sets are different. This method can better simulate actual flight status monitoring scenarios and enhance the model's generalization capabilities. Furthermore, to further investigate the impact of internal data on model training, the present invention designed a set of control experiments using a subject-dependent partitioning approach, i.e., the same data set was used for both training and testing. By comparing the classification performance of these two data partitioning approaches, the model's adaptability to individual differences was evaluated and the optimal data processing approach was determined. To reduce the impact of random data partitioning on model performance, the present invention employed 5-fold cross-validation, which involves equally partitioning the training dataset into five subsets. Four subsets are used for training at a time, with the remaining subset serving as the validation set, and this cycle repeated five times to improve model stability. Furthermore, within the independent partitioning mode, the present invention further employs leave-one-subject-out (LOSO) cross-validation, i.e., each set of data is selected as the test set, with the remaining data from all subjects serving as the training set. This method can more rigorously evaluate the model's generalization ability across individuals, making it applicable to a variety of aircraft.

[0155] Step S33: During the training phase, the flight status monitoring model constructed by the present invention, based on a neural network structure consisting of a graph convolutional network (GCN)-long short-term memory (LSM) network and an attention layer, accurately classifies flight status through multi-layer feature extraction and time series modeling. The present invention uses the cross-entropy loss function to measure the classification error of the model. This loss function is widely used in multi-class classification tasks and its mathematical expression is:

[0156]

[0157] Among them, y i is the true label of the i-th category predicted by flight status monitoring, is the category probability output by the flight status monitoring model; N is the number of flight status categories; L is the final calculated loss value, which represents the difference between the current prediction of the flight status monitoring model and the true label; The probability of the i-th category predicted for flight status monitoring is usually the output of the softmax function; log is the logarithmic function.

[0158] The cross-entropy loss function can effectively measure the gap between the model's predicted distribution and the true distribution, and determine the accuracy of the classification results. In order to improve the stability and convergence speed of model training, the present invention uses the Adam optimizer Adaptive Moment Estimation, Adam, for parameter optimization. The Adam optimizer combines the advantages of momentum gradient descent Momentum and adaptive learning rate adjustment Adaptive Learning Rate, can converge quickly and is applicable to complex non-convex optimization problems. The initial learning rate is set to 0.001, and the cosine annealing scheduling Cosine Annealing Learning Rate Decay strategy is used for dynamic adjustment to ensure that the model gradually converges to the optimal solution during training and avoids falling into local optimality. The present invention optimizes the hyperparameters in the training process. The training batch size Batch Size is set to 150 to ensure that sufficient sample information can be obtained in each iteration, thereby improving the stability of gradient updates. The number of training epochs is set to 50, and the early stopping strategy Early Stopping is adopted, that is, training is terminated early when there is no significant improvement in the performance of the validation set to prevent model overfitting. In terms of network regularization, the present invention adopts the Dropout mechanism and sets the Dropout ratio to 0.2, that is, 20% of the neurons are randomly discarded during the training process to reduce the model's dependence on specific features and improve generalization ability.

[0159] In addition, in order to further optimize computing performance and reduce video memory usage, the present invention introduces a Mixed Precision Training (MPT) strategy, which enables the model to combine 16-bit floating point numbers FP16 and 32-bit floating point numbers FP32 during the calculation process to accelerate calculations and improve video memory utilization. In the flight status monitoring model training process of the present invention, although the traditional full 32-bit floating point FP32 calculation method can provide higher numerical accuracy, it has problems such as high video memory usage, slow calculation speed, and limited gradient updates when training large-scale neural networks. In particular, when processing high-dimensional time series data, such as electroencephalogram (EEG) signal data, the demand for computing resources is extremely high. To improve computing efficiency and optimize video memory utilization, the present invention introduces a Mixed Precision Training (MPT) strategy, which effectively accelerates the training process while ensuring computing stability.

[0160] In the model training process of the present invention, the calculation process of mixed precision training mainly includes the following three stages. First, in the forward propagation Forward Propagation stage, the weight parameters of the model are calculated using FP16 to reduce the amount of calculation and improve the calculation throughput efficiency. At the same time, key normalization operations, such as Batch Normalization, still use FP32 to avoid model instability caused by decreased numerical precision. Secondly, in the loss calculation and gradient scaling Loss Scaling stage, due to the low calculation accuracy of FP16, the gradient value will be reduced to near zero, thereby affecting the gradient propagation effect. Therefore, the present invention adopts a loss scaling strategy to amplify the loss when calculating it so that the gradient value remains within a reasonable range, and then scale it back to the original value when updating the gradient to ensure the calculation stability. Finally, in the Backward Propagation & Parameter Update stage, the gradient calculation still uses FP16 to reduce storage overhead and improve calculation efficiency, while when the model parameters are updated, FP32 storage is still maintained to ensure the accuracy of the gradient update and the stability of the training convergence. After adopting mixed-precision training, the model's memory usage during both forward and backward propagation phases was reduced by an average of approximately 35%. Furthermore, with the same batch size of 150, the single iteration speed was approximately 1.5 to 1.8 times faster than pure FP32 training. In actual testing, most subject data achieved over 85% accuracy within 25 to 30 epochs, while pure FP32 training required 35 to 40 epochs to achieve a similar level.

[0161] The test results of the flight status monitoring model are as follows: Figure 4 Figure 2 shows a schematic diagram of a three-class confusion matrix for an embodiment of the present invention, illustrating the prediction accuracy for three different states. When using the proposed method to identify three flight states on the SEED-VIG dataset, 138 correct predictions were made out of 150 test samples, for an accuracy of approximately 92%. The diagonal elements 47, 45, and 46 in the confusion matrix represent the number of correctly classified classes, demonstrating the high accuracy and robustness of the proposed method for classifying different fatigue levels.

[0162] like Figure 5The figure shows a schematic diagram comparing the performance of the present invention's method with various comparison methods on various metrics, comparing the accuracy, precision, and recall of the present invention's method with those of traditional CNN, LSTM alone, and GCN-LSTM without an attention layer. The present method significantly outperforms the other comparison methods in terms of accuracy, precision, and recall, achieving an accuracy of 0.92, a precision of 0.90, and a recall of 0.91. This demonstrates that the present invention, while comprehensively utilizing the spatial topological and temporal characteristics of EEG signal data and introducing an attention layer, significantly improves classification performance and enables dynamic monitoring, providing a more reliable state perception foundation for subsequent shared control.

[0163] Step S4: Using the flight state monitoring model obtained in step S3, the flight state is identified and classified, and the flight state classification results are used as a factor for adjusting the dynamic weight of the shared control system. Based on state identification, the present invention further implements a shared control strategy for flight state monitoring based on EEG signals. The flight state is classified using the flight state monitoring model trained in step S3, and the flight state classification results are used as an input basis for adjusting the dynamic weight of the shared control system. The specific implementation method is as follows:

[0164] When the model identifies that the flight status is normal Alert, the autopilot device only provides necessary flight parameter monitoring, environmental information prompts and basic auxiliary functions to ensure that subjective initiative and decision-making flexibility are maximized under normal conditions.

[0165] When the flight status is identified as Mild Fatigue, the autopilot proactively and appropriately intervenes in flight control tasks to reduce cognitive and operational burdens. At this point, the system can provide auxiliary measures such as automatic heading stabilization, attitude correction, voice reminders, and visual warnings to prevent further deterioration of the flight status. It also ensures a gradual and smooth transition of control authority to avoid operational risks caused by sudden changes in status.

[0166] When the flight status is identified as the Severe Fatigue alarm, the autopilot proactively takes over control of the aircraft, temporarily exiting the flight control loop. Simultaneously, the system activates a multimodal warning and wake-up mechanism, including audible and visual alarms, voice prompts, and seat vibration, to assist in recovery. Once the model detects that the flight status has returned to a safe level, the system gradually reduces the autopilot's control weight until control is smoothly and safely fully returned to the aircraft, achieving closed-loop dynamic control adjustment. To further verify the system's feasibility in a flight environment, a shared control experiment scenario was constructed on a virtual flight simulator platform. The experiment involved a flight of approximately 90 minutes along a pre-planned route, with random environmental disturbances such as mild turbulence and monotonous terrain triggered during flight to simulate fatigue. Upon detecting the "Severe Fatigue" state, the model rapidly transitioned to full autopilot control, waking up and returning to the Mild Fatigue or Alert state after an average of 10 to 15 seconds. In the normal Alert state, control authority is maintained at approximately 90% to 100%, with autonomous driving occupying only 0% to 10%, primarily performing parameter monitoring and information prompts. When the model identifies a Mild Fatigue alert state, control authority gradually decreases to 60% to 70%, while autonomous driving authority increases to 30% to 40%, reducing cognitive and operational burdens. Once the Severe Fatigue alert state is reached, autonomous driving authority immediately increases to 90% to 100% and operates, briefly decreasing to 0.5% to 10%. After the multimodal warning is activated, the system gradually returns control authority in a 10% to 30% to 60% gradient, with autonomous driving authority smoothly decreasing from 90% to 40% until the state returns to the Alert level. The system gradually returns control authority through a smooth transition mechanism, and no dangerous maneuvers or sudden crashes were simulated throughout the entire process.

[0167] The beneficial effects of the present invention are as follows: by combining the features of multiple dimensions such as space, time, and frequency, the complex information structure of EEG signals is fully utilized to improve the accuracy of flight state recognition; GCN is used to construct the spatial dependency relationship between EEG electrodes of EEG data, making the information transmission between different brain regions more reasonable; according to the strong time dependency of EEG signals, the long short-term memory network LSTM is used to process time series data, effectively modeling the trend of flight state evolution over time, and improving the detection ability of progressive fatigue state; the introduction of the attention layer dynamically allocates weights in the entire EEG signal time series, so that the features of key time parameters are highlighted, thereby improving the accuracy of state recognition; the selection of differential evolution features effectively reduces the dimension of the original EEG data while retaining the key information of the EEG signal; the use of Z-score standardization to reduce the influence of physiological differences between individuals, so that the model can adapt to different individuals and improve generalization ability. The shared control mode based on EEG data of the present invention can dynamically optimize the allocation of driving rights, ensure that the automatic driving device can intervene in time, and improve driving safety and the intelligence level of the device.

[0168] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A flight status monitoring device based on EEG signal data classification, characterized in that: It includes: EEG signal data acquisition module, data preprocessing and feature extraction module, image sequence attention network module and state monitoring and sharing control module; The EEG signal data acquisition module acquires EEG signal data corresponding to the flight status and annotates the EEG signal data through fatigue monitoring; The data preprocessing and feature extraction module is used to perform bandpass filtering and normalization to filter out environmental noise and physiological artifacts, and use differential entropy to preprocess the EEG signal data, perform multi-band feature extraction to extract key information of the frequency band, and obtain differential evolution features; The graph sequence attention network module is used to learn and extract the composite features of EEG signal data in spatial topology and time series. Its structure includes: graph convolutional network layer, bidirectional long short-term memory network layer, attention layer, fully connected layer and softmax output layer; the graph convolutional network layer forms the spatial correlation between electrodes by constructing the adjacency matrix between EEG channels, and updates the features of each node through graph convolution operation; the bidirectional long short-term memory network layer captures the temporal evolution of EEG signals from the forward and backward time series paths respectively, and fuses them to obtain temporal context information; the attention layer weights each time parameter of the output; the fully connected layer and softmax output layer map the results to flight state classification results; The state monitoring and shared control module identifies and classifies the flight state, and uses the flight state classification results as the adjustment factor for the dynamic weight of the shared control system. It includes a monitoring submodule and a shared control submodule; the monitoring submodule inputs the EEG signal data into the graph sequence attention network module, outputs the flight state label and continuously updates it; the shared control submodule adjusts the control weight of the autopilot device according to the flight state level output by the flight state.

2. The flight status monitoring device based on EEG signal data classification according to claim 1, characterized in that: The flight status classification results specifically include: normal state Alert, reminder state Mild Fatigue and alarm state Severe Fatigue; in the normal state Alert, the autopilot device provides flight status monitoring and assistance; in the reminder state Mild Fatigue, the autopilot device intervenes in flight control, which includes attitude correction, voice and visual reminders; in the alarm state Severe Fatigue, the autopilot device performs flight control and simultaneously starts the multi-modal wake-up and alarm mechanism.

3. The flight status monitoring device based on EEG signal data classification according to claim 1, characterized in that: The frequency range corresponding to the differential evolution feature is as follows: The first feature Delta: 1-4Hz, second characteristic Theta: 4-8Hz, third characteristic Alpha: 8-14Hz, fourth characteristic Beta: 14-31Hz and fifth characteristic Gamma: 31-50Hz.

4. The flight status monitoring device based on EEG signal data classification according to claim 1, characterized in that: The graph convolutional network layer is constructed based on the spatial distribution relationship of the electrodes of the EEG signal data, specifically: Among them, A is the adjacency matrix; A ij is the connection strength between electrode i and electrode j; d i,j is the Euclidean distance between electrodes i and j; δ is the normalization factor; i is the first electrode number in the adjacency matrix; j is the second electrode number in the adjacency matrix.

5. The flight status monitoring device based on EEG signal data classification according to claim 1, characterized in that: The graph convolutional network layer updates the features of each node through graph convolution operations. The input feature matrix of the lth layer is: Among them, H( l ) is the input feature matrix of the lth layer; D is the degree matrix; W( l ) is the training weight matrix; σ1 is the nonlinear activation function.

6. The flight status monitoring device based on EEG signal data classification according to claim 1, characterized in that: The bidirectional long short-term memory network layer learns the temporal characteristics of the EEG signal data signal, specifically: in, is the hidden state of the forward LSTM; is the hidden state of the backward LSTM; is the output of the bidirectional long short-term memory network layer, which is composed of the output of the forward and backward LSTM; t is the time parameter; LSTM forward Forward long short-term memory network; LSTM backward It is a backward long short-term memory network.

7. A flight status monitoring method based on EEG signal data classification, characterized in that: It includes: S1: Obtain EEG signal data corresponding to the flight state, annotate the EEG signal data through fatigue monitoring, preprocess the EEG signal data using differential entropy, extract differential evolution features, and perform bandpass filtering and normalization; S2: Construct a flight status monitoring model based on a graph sequence attention network to analyze EEG signal data and complete flight status classification. The graph sequence attention network includes a graph convolutional network layer, a long short-term memory network layer, an attention layer, and a fully connected layer. The EEG signal data after normalization in step S1 is input into the graph convolutional network layer to extract spatial topological features. The long short-term memory network layer learns the time evolution pattern of the EEG signal data signal through a gating mechanism. The attention layer performs weighted processing on the output of the long short-term memory network layer to obtain information on key time parameters, and outputs the flight status classification result through the fully connected layer. S3: Using the EEG signal dataset processed in step S1 to train the flight status monitoring model constructed in step S2, completing the flight status monitoring model training, and using the cross entropy loss function to improve the accuracy of the flight status classification results; S4: Using the flight status monitoring model obtained in step S3 to identify and classify the flight status, and using the identification and classification results as adjustment factors for the dynamic weight of the shared control system.

8. The flight status monitoring method based on EEG signal data classification according to claim 7, characterized in that: Step S1 specifically includes: S11: Acquire multiple types of EEG signal data in flight status and annotate the EEG signal data through fatigue monitoring; S12: Use differential entropy to preprocess EEG signal data and perform bandpass filtering to reduce artifacts and noise; S13: Extract differential evolution features, calculated using short-term Fourier transform of non-overlapping windows; S14: Normalize the EEG signal data.

9. The flight status monitoring method based on EEG signal data classification according to claim 7, characterized in that: Step S2 specifically includes: S21: Using graph convolutional networks to construct the topological structure of EEG signal data to extract features using spatial correlation between electrodes; S22: Use the long short-term memory network layer to learn the time series characteristics of EEG data signals, and selectively retain or discard historical information through the gating mechanism; S23: Introduce the attention layer to calculate the importance weight of the time parameter, and perform weighted summation on the output of the long short-term memory network layer, input it into the fully connected layer, and calculate the probability distribution of the flight state.

10. The flight status monitoring method based on EEG signal data classification according to claim 7, characterized in that: Step S3 is specifically as follows: S31: partitioning EEG signal dataset; S32: independent partitioning combined with 5-fold cross-validation and leave-one-subject-out cross-validation; S33: The cross entropy loss function is used in the training phase to measure the classification error of the flight status monitoring model.

Images