A pilot fatigue real-time monitoring system based on electroencephalogram signals and deep learning and a use method thereof

Abstract

This invention provides a method, system, and device for real-time pilot fatigue monitoring based on electroencephalogram (EEG) signals and deep learning, belonging to the field of aviation safety technology. The method includes: constructing multi-load level flight experimental scenarios and simultaneously collecting pilot personal characteristics and raw EEG signals; preprocessing the EEG signals and extracting power spectral features of δ, θ, α, and β rhythms; constructing a deep sparse contractile autoencoder network model, fusing multimodal features for training, and obtaining a fatigue monitoring model; finally, inputting the real-time collected EEG signals into the model and outputting the pilot's current fatigue level. This invention also provides a corresponding monitoring system and device. This invention achieves objective, high-precision, and real-time monitoring of pilot fatigue state through deep learning technology, possessing adaptability to multi-load scenarios and personalized early warning and intervention capabilities, significantly improving the intelligence level and operational reliability of flight safety management.

Description

Technical Field

[0001] This invention relates to the field of monitoring systems, specifically to a real-time pilot fatigue monitoring system and its usage method based on electroencephalogram (EEG) signals and deep learning. Background Technology

[0002] In the field of aviation safety, pilot fatigue is one of the key contributing factors to human-caused accidents, seriously affecting the safety and reliability of flight operations. Currently, fatigue monitoring methods mainly include subjective questionnaires, behavioral image recognition, and physiological signal-based analysis techniques. Among these, electroencephalogram (EEG) signal analysis methods have received widespread attention due to their objectivity and accuracy. However, existing EEG fatigue monitoring technologies still have significant limitations: most methods rely on traditional frequency domain feature extraction, failing to fully explore the local temporal characteristics of the signals; models are mostly trained on small samples or in single scenarios, making it difficult to adapt to individual pilot differences and dynamic changes in workload; existing systems have low integration levels, lacking real-time monitoring and early warning capabilities deeply integrated with the actual flight environment, and have not yet formed a closed-loop management system of "monitoring-early warning-intervention."

[0003] Furthermore, existing technologies rarely incorporate deep integration of pilot personal characteristic data (such as age, piloting experience, and gender) with EEG signals under multi-load flight scenarios, resulting in insufficient personalization and adaptability of fatigue recognition models. On the hardware side, traditional EEG acquisition devices are highly invasive and complex to wear, making it difficult to achieve long-term, low-interference continuous monitoring in real flight missions. Therefore, there is an urgent need for a pilot fatigue monitoring solution that can combine deep learning technology, support multimodal data fusion, and possess high real-time performance and high accuracy to improve the intelligence level and early warning effectiveness of aviation safety management, and meet the practical application needs of civil aviation, military aviation, and flight training scenarios.

[0004] To address the aforementioned issues, the applicant proposes a real-time pilot fatigue monitoring system and its usage method based on electroencephalogram (EEG) signals and deep learning. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time pilot fatigue monitoring system and its usage method based on electroencephalogram (EEG) signals and deep learning, in order to solve the problems in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a real-time pilot fatigue monitoring method based on electroencephalogram (EEG) signals and deep learning, comprising the following steps:

[0007] S1: Construct flight test scenarios with multiple workload levels, including high workload induced scenarios, medium workload induced scenarios, and low workload induced scenarios; conduct simulated flight missions for pilots at different levels, and simultaneously collect pilots' personal characteristic data and raw EEG signals;

[0008] S2: Preprocess the raw EEG signal, extract the four rhythm components δ, θ, α, and β from the EEG signal, calculate the power spectral density of each rhythm component, and couple the curve region of the power spectral density into a fatigue characteristic index.

[0009] S3: Construct a deep sparse contractile autoencoder network model, taking the fatigue feature indicators and personal feature data as inputs, and the pilot's fatigue state level as the output label. Train and optimize the network model using the labeled EEG dataset to obtain the pilot fatigue monitoring model.

[0010] S4: Process the EEG signal of the pilot to be monitored during the real-time flight mission according to step S2, and input it into the fatigue monitoring model obtained in step S3 to output the real-time fatigue status of the pilot to be monitored.

[0011] Optionally, in step S1, the acquisition frequency of the raw EEG signal is not less than 250Hz, the acquisition channel covers the frontal lobe, parietal lobe, and temporal lobe brain regions, and the power frequency interference and electromyographic interference are eliminated through an anti-interference algorithm during the acquisition process.

[0012] Optionally, in step S2, the power spectral density is calculated using the Welch method, with a calculation window length of 2-4 seconds and an overlap rate of 50%-75%; the fatigue characteristic index is the integral value of the power spectral density of each rhythm component within the corresponding frequency range.

[0013] Optionally, in step S3, the deep sparse shrinking autoencoder network model includes an input layer, a hidden layer, and an output layer. The number of hidden layers is 3-5, and the number of neurons in each layer decreases by 1 / 2 to 2 / 3 of the input feature dimension. An L1 regularization term is introduced into the activation function to achieve sparsity constraints, and the model training is completed by minimizing the cross-entropy loss function through the Adam optimizer.

[0014] Optionally, step S4 may also include generating personalized intervention recommendations based on the output fatigue level. The intervention recommendations include workload adjustment schemes and rest duration recommendations, and are associated with and matched with the pilot's personal characteristic data.

[0015] A real-time pilot fatigue monitoring system based on electroencephalogram (EEG) signals and deep learning includes:

[0016] Data acquisition module: used to collect raw EEG signals and personal characteristic data of pilots during simulated or actual flight missions;

[0017] Signal processing module: used to extract the δ, θ, α, and β rhythm components from the EEG signal and couple the power spectral density of each rhythm component into fatigue characteristic indicators;

[0018] Model calculation module: Stores a pre-trained deep sparse contractile autoencoder network model, used to receive the fatigue feature indicators and personal feature data, and output the pilot's real-time fatigue status level;

[0019] Display and warning module: used to visually display the trend of pilot fatigue status changes, and triggers a warning when the fatigue level reaches moderate or above.

[0020] Optionally, the EEG signal acquisition device in the data acquisition module is a head-mounted dry electrode device, weighing no more than 300 grams, and has wireless transmission capability with a transmission delay of less than 100 milliseconds.

[0021] Beneficial effects: 1. Significantly improves the accuracy and real-time performance of fatigue monitoring.

[0022] This invention employs a deep sparse contractile autoencoder network, which can automatically extract deep local features from EEG signals. Combined with data collected in multi-load level flight test scenarios, it significantly improves the accuracy and sensitivity of fatigue state identification. By optimizing the model structure and feature fusion mechanism, the system can achieve millisecond-level response, meeting the real-time monitoring requirements in flight missions and overcoming the latency and error problems of traditional subjective questionnaires and behavior monitoring methods.

[0023] 2. Achieve personalized and scenario-based precise monitoring

[0024] The system integrates pilots' personal characteristic data (such as age, flying experience, and gender) with multimodal EEG features to construct a personalized fatigue assessment model. It can dynamically adjust the recognition threshold according to different pilots' workload, physiological state, and task type, thereby improving the model's adaptability and reliability in actual flight environments.

[0025] 3. Construct a closed-loop fatigue intervention mechanism

[0026] The system not only enables real-time monitoring and early warning of fatigue status, but also generates personalized intervention suggestions (such as workload adjustment and rest duration recommendations) based on the identification results. It combines cockpit and ground terminals to conduct multi-target monitoring and early warning, forming a closed-loop management process of "monitoring-early warning-intervention" to provide decision support for flight safety management. Attached Figure Description

[0027] Figure 1 Flowchart of data acquisition and feature extraction in an embodiment of the present invention;

[0028] Figure 2Flowchart of model training and optimization in this embodiment of the invention;

[0029] Figure 3 Flowchart of the real-time monitoring and early warning system according to an embodiment of the present invention. Detailed Implementation

[0030] The preferred embodiments of the present invention are described below with reference to the accompanying drawings to make the technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0031] This invention discloses a real-time pilot fatigue monitoring method based on electroencephalogram (EEG) signals and deep learning. Its core lies in achieving accurate identification and real-time early warning of pilot fatigue states through multi-level experimental design, refined signal processing, and advanced deep learning modeling. The implementation process mainly includes four key stages: construction of multi-load level flight experimental scenarios and data acquisition, EEG signal preprocessing and feature extraction, construction and training of a deep sparse contractile autoencoder network model, and implementation of a real-time monitoring and intervention system. These stages will be described in detail below.

[0032] In the multi-workload level flight test scenario construction phase, it is necessary to design flight mission scenarios that can induce different levels of fatigue. In this embodiment, flight test scenarios are divided into three levels: high workload induced scenarios, medium workload induced scenarios, and low workload induced scenarios. High workload scenarios can be designed as instrument approach under complex weather conditions, joint handling of multi-system failures, or route flight under high-density air traffic control environments; medium workload scenarios can include standard instrument departure, routine cruise, and handling of simple special situations; low workload scenarios can be set as visual take-off and landing routes under clear sky conditions or basic simulator familiarization training. The experimental subjects should cover pilots at different levels, including junior flight trainees, intermediate flight trainees, and flight instructors, to obtain representative sample data. During the experiment, two types of data need to be collected simultaneously: first, the pilot's personal characteristic data, including basic information such as age, gender, and years of flying experience, as well as subjective evaluation data obtained through standardized workload scales (such as the NASA-TLX scale) and fatigue questionnaires (such as the Samn-Perelli scale); second, raw EEG signals, recorded using low-invasive EEG acquisition equipment that meets medical standards. To ensure data quality, the frequency of EEG signal acquisition should be no less than 250Hz. The electrode placement should cover key brain regions related to cognition and fatigue, including the frontal, parietal, and temporal lobes. Anti-interference algorithms should be applied in real time during the acquisition process to eliminate 50Hz power frequency interference and electromyographic artifacts.

[0033] In the signal preprocessing and feature extraction stage, the acquired raw EEG signals undergo a series of processing steps to extract effective fatigue feature indicators. Preprocessing steps include, but are not limited to: removal of EEG and ECG artifacts, bandpass filtering to retain effective frequency bands, and segmentation and removal of bad segments. Subsequently, an FIR filter with linear phase characteristics is used to decompose the preprocessed EEG signals into four classic rhythmic components: delta waves, theta waves, alpha waves, and beta waves, with frequency ranges defined as 0.5-4Hz, 4-8Hz, 8-13Hz, and 13-30Hz, respectively. For each rhythmic component, the Welch method is used to calculate its power spectral density. The Welch method, as a classic nonparametric power spectrum estimation method, reduces variance and improves estimation stability by segmenting the data, windowing, and averaging. In this implementation, the calculation window length is set to 2 to 4 seconds, and the overlap rate is set to 50% to 75% to achieve a good balance between time resolution and frequency resolution. Finally, the area under the curve (integral value) of the power spectral density of each rhythmic component within its corresponding frequency range was calculated, serving as four core fatigue characteristic indicators reflecting different brain activity states. Changes in these indicators are closely related to brain arousal, cognitive load, and fatigue levels.

[0034] In the model building and training phases, a deep sparse contractile autoencoder network is constructed as the core classifier. This network structure includes an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer is determined by the fatigue feature indicators (4) and individual feature data (several dimensions after encoding and normalization). It is recommended to set 3 to 5 hidden layers, with the number of neurons in each layer decreasing progressively by 1 / 2 to 2 / 3 of the input feature dimension. This structure helps the network extract and compress feature information layer by layer. To enhance the model's generalization ability and learn more robust feature representations, an L1 regularization term is introduced into the activation function of the hidden layers to achieve sparsity constraints, forcing most neurons to be inactive most of the time, thereby learning more essential and independent features from the data. The output layer corresponds to four fatigue state levels: non-fatigue, mild fatigue, moderate fatigue, and severe fatigue, using a Softmax activation function to output the probability of belonging to each category. The model training aims to minimize the cross-entropy loss function between the predicted and true labels, and an adaptive moment estimation optimizer is used for parameter updates. The dataset used for training consisted of EEG datasets collected in previous experiments and labeled by experts or through standardized procedures. These datasets were divided into training, validation, and test sets according to a specific ratio for model training, hyperparameter tuning, and final performance evaluation. Through iterative optimization, a stable and accurate pilot fatigue monitoring model was finally obtained.

[0035] In the implementation phase of the real-time monitoring and intervention system, the trained model is integrated into a complete hardware and software system. This system consists of a data acquisition module, a signal processing module, a model calculation module, and a display and early warning module. The data acquisition module includes a head-mounted dry electrode EEG acquisition device and a personal feature input terminal. The device must meet the requirements of being lightweight (no more than 300g), low-invasiveness, and comfortable to wear, and possess wireless data transmission capabilities, with an end-to-end transmission latency of less than 100ms to ensure real-time monitoring. The signal processing module incorporates the aforementioned FIR filtering and power spectral density calculation algorithms, responsible for rapidly processing the real-time incoming EEG signals and extracting four fatigue characteristic indicators. The model calculation module loads a trained deep sparse contractile autoencoder network model, receives the characteristic indicators and personal feature data, and calculates and outputs the current pilot's fatigue level in real time. The display and early warning module is responsible for visualizing the results, displaying the changing trend of fatigue status through intuitive charts or color codes (such as green, yellow, orange, and red) on the head-up display or multi-function display in the cockpit. When the system determines that the fatigue level has reached "moderate" or "severe," it immediately triggers a tiered early warning mechanism, alerting the pilot through audible prompts, voice alarms, or visual flashing signals. Simultaneously, relevant data and alarm information are transmitted in real-time to the ground monitoring center via data link, enabling ground control personnel to monitor the crew's status and make collaborative decisions. Furthermore, the system can generate personalized intervention suggestions based on the current fatigue level, individual historical data, and mission information, such as suggesting adjustments to the autopilot takeover level or recommending the optimal rest duration for the next flight segment, forming a complete closed loop of monitoring-early warning-intervention.

[0036] To ensure the accuracy and adaptability of the system in the long term, the model calculation module should also be designed with a model update unit. This unit can periodically or on demand receive new data collected in actual operation (which must be securely processed and labeled), and fine-tune and optimize the original model through incremental learning or online learning algorithms, so that the model can adapt to subtle changes in individual differences among pilots and the concept drift problem that may occur in long-term operation.

[0037] Example 1: Basic Implementation of Real-time Pilot Fatigue Monitoring Based on Standard Scenarios

[0038] This embodiment demonstrates the basic application process of the method of the present invention in a standard simulated flight scenario. Twenty pilots of different levels (including 10 junior trainees, 6 intermediate trainees, and 4 instructors) were selected as experimental subjects. The experiment was conducted on a fixed-base flight simulator, and three typical workload flight mission scenarios—high, medium, and low—were designed and implemented. The high-load scenario simulated instrument approach procedures under complex weather conditions; the medium-load scenario was a standard route cruise and navigation mission; and the low-load scenario was a clear-sky visual takeoff and landing route flight. All pilots were required to complete the NASA-TLX workload scale and the Samn-Perelli fatigue questionnaire before and after participating in each type of mission to obtain a baseline for subjective evaluation. During mission execution, a 32-channel wireless dry electrode EEG cap (a modified version of the NeuroSky MindWave Mobile 2 model) conforming to the international 10-20 system was used to simultaneously collect the pilots' EEG signals. This device weighs 295g, has a sampling frequency set to 256Hz, and effectively suppresses electromagnetic interference and power frequency noise in the simulator environment through built-in active shielding circuitry and digital filtering algorithms.

[0039] The collected raw EEG data first underwent preprocessing: independent component analysis (ICA) was used to automatically identify and remove eye movement and electromyography (EMG) artifacts; an FIR bandpass filter of 0.5-35Hz was used to retain effective frequency bands. Subsequently, the signal was segmented into 4-second continuous data segments with an overlap rate of 66%. For each data segment, the Welch method was applied to calculate the power spectral density, and the power spectral integral values ​​of the four rhythmic frequency bands δ (0.5-4Hz), θ (4-8Hz), α (8-13Hz), and β (13-30Hz) were extracted to form four core time-frequency features. Simultaneously, the questionnaire scores completed by the pilots were standardized and, together with personal characteristics such as age and pilot experience, constituted the auxiliary input features of the model.

[0040] A four-layer deep sparse contractile autoencoder network was constructed, with the number of neurons decreasing layer by layer by 2 / 3 of the total input dimensions (4 EEG features + 4 personal and questionnaire features). ReLU activation functions with L1 regularization were used in the hidden layers to facilitate sparse feature learning. The model was trained using the Adam optimizer with an initial learning rate of 0.001 and a loss function of classification cross-entropy. Training data came from an independent pilot fatigue database containing 100 hours of labeled EEG data.

[0041] The trained model was deployed on an embedded edge computing device (NVIDIA Jetson AGXXavier). In actual monitoring, the system receives and processes EEG signals in real time with an analysis cycle of 4 seconds, extracts features, and inputs them into the model for inference. The model outputs confidence scores for four fatigue levels, and the level corresponding to the highest confidence score is taken as the current monitoring result. The system achieved a fatigue state classification accuracy of 91.5% on the test set, with an average processing latency of 125 milliseconds, initially realizing real-time and objective monitoring of pilot fatigue in a simulated environment.

[0042] Example 2: An Enhanced Monitoring System Integrating Multimodal Data and Personalized Intervention

[0043] This embodiment, building upon Embodiment 1, further enhances the system's personalization capabilities and intervention functions. The experimental subjects were expanded to 50 pilots, and richer individual characteristics were introduced, including cognitive style classification based on the Holland Occupational Interests Test, and baseline data on sleep quality and heart rate variability collected over the past 72 hours via wearable devices (such as Garmin fenix 7). EEG acquisition was upgraded to a 64-channel semi-dry electrode system (ANT Neuro eego sports), improving spatial resolution in fatigue-related brain regions such as the frontal lobe and anterior cingulate cortex while maintaining wearing comfort. Furthermore, during simulated flight missions, flight control parameters (such as stick force and rudder fluctuations) and eye-tracking data (TobiiPro Glasses 3) were simultaneously recorded as auxiliary behavioral performance indicators.

[0044] In the feature engineering phase, in addition to extracting standard frequency band power features, the differential entropy, sample entropy, and functional connectivity strength between different brain regions (such as frontoparietal network connectivity based on phase-locked values) of the EEG signals were also calculated. Personal characteristics, behavioral performance data, and these enhanced EEG features together constitute a high-dimensional feature vector. To handle this multimodal data and address the curse of dimensionality, this embodiment improves the deep learning model architecture. A two-branch network is employed: one branch is a deep sparse contractile autoencoder, specifically for processing high-dimensional EEG features; the other branch is a shallow fully connected network, processing structured personal and behavioral data. The two branches perform feature fusion in the final fully connected layer and then jointly make classification decisions. A focus loss function is introduced during model training to mitigate the impact of class imbalance in the data (such as the scarcity of "severe fatigue" samples).

[0045] The system's core innovation lies in its closed-loop "monitoring-assessment-intervention" mechanism. When the model identifies a pilot entering a state of "mild fatigue," the system first conducts a risk assessment by combining the pilot's historical data (such as the average time it takes for mild fatigue to develop into moderate fatigue under similar workloads). Then, through the in-cockpit interface, it provides preliminary intervention suggestions in a non-intrusive manner, such as: "Slight distraction detected; we suggest adjusting your seat posture and performing a deep breathing exercise." If the fatigue state escalates to "moderate," the system initiates more proactive intervention: on the one hand, it provides clear rest prompts and simple cognitive wake-up tasks (such as mental arithmetic) through a voice synthesis system; on the other hand, it sends warning information and suggested simplified follow-up tasks (such as suggesting engaging autopilot for the next flight segment) to the ground control station for instructors or commanders to consider in their decisions. After six months of simulated flight testing, the system's personalized intervention suggestions achieved an acceptance rate of over 85% and reduced the rate of operational errors caused by fatigue during simulated flights by approximately 30%.

[0046] Example 3: Deployment and Application of Integrated Real-Time Monitoring and Early Warning System in Actual Flight Training

[0047] This embodiment demonstrates a complete system solution for integrating the technology of this invention into a real flight training environment. The system is deployed for formation training missions using primary trainer aircraft (such as Cessna 172R) at a flight academy. The hardware system has undergone comprehensive engineering design and airworthiness considerations. The EEG acquisition device is customized as a highly integrated, lightweight headband, retaining only six key electrodes on the forehead and temples. A novel hydrogel dry electrode is used to ensure comfort and signal stability during extended wear, with the overall weight controlled to within 150g. The device communicates with the onboard data integration unit via Bluetooth Low Energy 5.2, with the end-to-end latency of the entire data link consistently below 80ms.

[0048] The airborne data integration unit is a ruggedized and electromagnetic compatibility certified industrial computer. It not only serves as a data aggregation point for EEG data but also acquires real-time flight data such as aircraft attitude, heading, and engine parameters via an aircraft data bus (e.g., ARINC 429), fusing this data with data from a smart wristband worn by the pilot (monitoring heart rate and skin conductance). This unit incorporates the multimodal deep learning model trained in Example 2, enabling local real-time inference and avoiding reliance on cloud networks.

[0049] The system's display and warning module employs a tiered, multi-terminal strategy. Inside the cockpit, critical information is presented via a small augmented reality head-up display: a semi-transparent halo (green / yellow / orange / red) reflects fatigue status in real time at the periphery of the pilot's normal field of vision, with concise text or icon warnings only appearing in the center of the field of vision when the condition worsens. The main monitoring interface is located on a multi-functional touchscreen display at the instructor's seat. This interface uses spatiotemporal visualization technology to simultaneously monitor the pilot status of up to 12 training aircraft. Each pilot is represented by a dynamic "physiological state identifier," whose color, size, and flashing frequency comprehensively reflect real-time fatigue level, heart rate variability, and workload. When a trainee is determined by the system to be "moderately fatigued" and continues to exceed a preset threshold, their identifier will flash brightly, and the system will automatically display the trainee's personal profile, current mission information, and personalized handling suggestions generated by the system (such as "It is recommended that the commander guide them back to base and arrange a 20-minute ground rest").

[0050] Furthermore, the system establishes a complete continuous model learning mechanism. All anonymized data generated during training flights (including flight data, physiological data, and the final interventions and effects taken by instructors) is periodically transmitted back to the ground data center. The data center's model update unit utilizes this new labeled data to conduct safe and collaborative incremental training on the models deployed on each aircraft through a federated learning framework. This allows the fatigue identification model of the entire fleet to continuously evolve and better adapt to the impact of macro-factors such as seasonal changes and different training phases. Within a full training quarter after its actual deployment, this integrated system successfully issued early warnings for multiple potential high-fatigue-risk events, earning high praise from training commanders and instructors, and becoming an important auxiliary tool for improving flight training quality and safety.

[0051] In summary, the embodiments of the present invention, through a combination of rigorous experimental design, advanced signal processing technology, and deep learning models, construct a complete technical solution from data acquisition to intelligent early warning. This system not only features high precision and high real-time performance but also fully considers individual differences and the needs of actual flight application scenarios, providing a reliable technical means for effectively managing pilot fatigue and improving aviation safety. Those skilled in the art can make appropriate adjustments and modifications based on the above specific embodiments and in light of actual circumstances; all such adjustments and modifications should fall within the protection scope of the appended claims.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0053] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A real-time pilot fatigue monitoring method based on electroencephalogram (EEG) signals and deep learning, characterized in that, Includes the following steps: S1: Construct flight test scenarios with multiple workload levels, including high workload induced scenarios, medium workload induced scenarios, and low workload induced scenarios; conduct simulated flight missions for pilots at different levels, and simultaneously collect pilots' personal characteristic data and raw EEG signals; S2: Preprocess the raw EEG signal, extract the four rhythm components δ, θ, α, and β from the EEG signal, calculate the power spectral density of each rhythm component, and couple the curve region of the power spectral density into a fatigue characteristic index. S3: Construct a deep sparse contractile autoencoder network model, taking the fatigue feature indicators and personal feature data as inputs, and the pilot's fatigue state level as the output label. Train and optimize the network model using the labeled EEG dataset to obtain the pilot fatigue monitoring model. S4: Process the EEG signal of the pilot to be monitored during the real-time flight mission according to step S2, and input it into the fatigue monitoring model obtained in step S3 to output the real-time fatigue status of the pilot to be monitored.

2. The method according to claim 1, characterized in that, In step S1, the acquisition frequency of the raw EEG signal is not less than 250Hz, the acquisition channel covers the frontal lobe, parietal lobe and temporal lobe brain regions, and the power frequency interference and electromyographic interference are eliminated through anti-interference algorithm during the acquisition process.

3. The method according to claim 1, characterized in that, In step S2, the power spectral density is calculated using the Welch method, with a calculation window length of 2-4 seconds and an overlap rate of 50%-75%; the fatigue characteristic index is the integral value of the power spectral density of each rhythm component within the corresponding frequency range.

4. The method according to claim 1, characterized in that, In step S3, the deep sparse shrinking autoencoder network model includes an input layer, a hidden layer, and an output layer. The number of hidden layers is 3-5, and the number of neurons in each layer decreases by 1 / 2 to 2 / 3 of the input feature dimension. An L1 regularization term is introduced into the activation function to achieve sparsity constraints, and the model training is completed by minimizing the cross-entropy loss function through the Adam optimizer.

5. The method according to claim 1, characterized in that, Step S4 also includes generating personalized intervention recommendations based on the output fatigue level. These recommendations include workload adjustment schemes and rest duration recommendations, and are associated with and matched with the pilot's personal characteristic data.

6. A real-time pilot fatigue monitoring system based on electroencephalogram (EEG) signals and deep learning, characterized in that, include: Data acquisition module: used to collect raw EEG signals and personal characteristic data of pilots during simulated or actual flight missions; Signal processing module: used to extract the δ, θ, α, and β rhythm components from the EEG signal and couple the power spectral density of each rhythm component into fatigue characteristic indicators; Model calculation module: Stores a pre-trained deep sparse contractile autoencoder network model, used to receive the fatigue feature indicators and personal feature data, and output the pilot's real-time fatigue status level; Display and warning module: used to visually display the trend of pilot fatigue status changes, and triggers a warning when the fatigue level reaches moderate or above.

7. The system according to claim 6, characterized in that, The EEG signal acquisition device in the data acquisition module is a head-mounted dry electrode device, weighing no more than 300 grams, and has wireless transmission capabilities with a transmission delay of less than 100 milliseconds.

Images