Air traffic controller cognitive load detection method based on neural network

Abstract

The invention discloses an air traffic controller cognitive load detection method based on a neural network. The method belongs to the field of artificial intelligence and aviation safety management, and comprises the following steps: dynamically inducing a cognitive load through a task induction and interaction module based on a real air traffic control worker working scene (such as aircraft take-off and landing command, flight scheduling and the like), and feeding back a task completion state and a cognitive load level in real time; the electroencephalogram signal processing module carries out time-frequency domain feature extraction on the collected signals, global feature modeling is optimized by utilizing an agency attention mechanism, and the modeling capability and time causal consistency of a long-time sequence are improved by combining the optimized time sequence memory module and the causal convolution module; and the classification network outputs the cognitive load level of the air traffic keeper, and visual feedback is carried out through the evaluation module. According to the method, the real-time performance and accuracy of cognitive load detection are improved, and technical support is provided for aviation safety management and task optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, aviation safety management and biological signal processing, and relates to a method for detecting the cognitive load of air traffic controllers based on a neural network; specifically, it relates to a method for detecting the cognitive load of air traffic controllers based on an attention mechanism and a temporal neural network. Background Art

[0002] Air Traffic Controllers (ATCOs) are important participants in aviation operational safety, and their cognitive load status directly affects the efficiency and safety of air traffic management. Cognitive Load (CL) theory refers to the theory used to explain the load problem caused by the limited working memory capacity when people perform tasks. Studies have shown that when the cognitive load is too high, air traffic controllers are prone to misjudgment and operational errors, which lead to serious safety hazards; when the cognitive load is too low, there may be problems of distraction and reduced work efficiency. Therefore, real-time and accurate detection of the cognitive load of air traffic controllers is of great significance for optimizing task allocation, improving work efficiency and ensuring aviation safety.

[0003] At present, the detection methods of cognitive load mainly include subjective evaluation, behavioral indicators and physiological signal analysis. Subjective evaluation methods such as NASA Task Load Index (NASA-TLX) obtain the subjective load perception of air traffic controllers through questionnaires. Although it is convenient to use, it is difficult to achieve dynamic and real-time monitoring of cognitive load due to the influence of subjective factors such as emotions and memory bias. The method based on behavioral indicators indirectly evaluates the load level by recording characteristics such as task completion time, error rate and reaction time, but has limited ability to capture instantaneous load fluctuations in complex tasks.

[0004] With the development of artificial intelligence technology, deep learning methods have gradually been applied to the processing and analysis of EEG signals. Among many models, time series neural networks (such as long short-term memory networks LSTM and gated recurrent units GRU) are widely used in time series analysis tasks due to their advantages in modeling time dependencies. However, these methods still have certain limitations when processing EEG signals in complex task scenarios, such as insufficient modeling capabilities for long sequence dependencies, low computational efficiency, and poor real-time performance. In addition, although traditional attention mechanisms (such as Softmax Attention) can significantly improve the model's ability to capture key information, their high computational complexity limits their application in long time series data. For multi-task load scenarios, there is also room for improvement in the classification robustness and adaptability of existing models at different load levels. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to propose a method for detecting the cognitive load of air traffic controllers based on a neural network, so as to solve the shortcomings of the current cognitive load detection technology in terms of real-time performance, accuracy and adaptability.

[0006] The technical solution of the present invention is: a method for detecting the cognitive load of air traffic controllers based on a neural network according to the present invention, wherein the system for detecting the cognitive load of air traffic controllers based on electroencephalogram (EEG) signals comprises four modules, namely, a task induction and interaction module, an electroencephalogram (EEG) signal real-time acquisition module, an electroencephalogram (EEG) signal processing module and a cognitive load evaluation module.

[0007] The method comprises the following steps:

[0008] (1): Through the task induction and interaction module, a dynamic simulated task environment is provided for the air traffic controller's real work scene, inducing changes in their EEG signals. The task induction interface can present simulation scenarios related to air traffic control tasks, including aircraft takeoff and landing command, flight scheduling, route planning and conflict detection tasks; the interactive feedback area displays the air traffic controller's task completion status and cognitive load assessment results in real time;

[0009] (2): The EEG signals generated by the air traffic controller during the task execution are collected by the EEG signal real-time acquisition module and transmitted to the EEG signal processing module for subsequent analysis; the EEG signals are covered by the frontal lobe, parietal lobe and occipital lobe brain regions through the electrode array, and the sampling rate is 512Hz. The collected signals are processed by high-pass filtering, low-pass filtering and independent component analysis (ICA) to remove artifacts to generate purified signal data;

[0010] (3): Use the EEG signal processing module to extract features and perform dynamic modeling analysis on the collected signals, and pass the processing results to the cognitive load assessment module, including:

[0011] The power spectral density (PSD) of EEG signals in the α wave, β wave, θ wave and other frequency bands was extracted by fast Fourier transform (FFT), and the time-frequency features were generated by short-time Fourier transform (STFT) to reflect the changes in cognitive load of air traffic controllers in different time periods.

[0012] The global context modeling is optimized through the proxy attention mechanism, the signal query matrix is ​​divided into fixed-size windows, the proxy matrix is ​​generated and combined with the key matrix and the value matrix for interaction, which reduces the computational complexity and enhances the modeling ability of global features;

[0013] The optimized temporal memory module is used to model the dependency of long time series, the dynamic gating mechanism is used to select relevant time step information, and the causal convolution module is combined to ensure causal consistency in time series processing;

[0014] The processed high-dimensional features are input into the classification network, and the air traffic controller's cognitive load level is output through the Softmax activation function, including "low load", "medium load" and "high load";

[0015] (4): The classification results are received and visual feedback is provided through the cognitive load assessment module, including the current cognitive load level, load change curve and trend analysis results, which are displayed in graphical or textual form, and optimization suggestions are provided in combination with the task results.

[0016] Furthermore, the task induction and interaction module includes a task induction interface and an interactive feedback area. The task induction interface induces the cognitive load of air traffic controllers by dynamically simulating air traffic control task scenarios (such as aircraft takeoff and landing command, flight scheduling, route planning, conflict detection), and the task complexity is dynamically adjusted by increasing multi-tasking requirements (such as weather interference, airspace conflicts, etc.); the interactive feedback area displays the task completion time, task accuracy and EEG signal analysis results in real time.

[0017] Specifically, the task induction interface in the task induction module is a simulated air traffic controller's workstation, which accurately reproduces the real work situation by simulating the daily work tasks of the air traffic controller in real time, such as aircraft take-off and landing command, flight scheduling, route planning, conflict detection and other operational tasks; for example, in the aircraft take-off and landing command task, the system will display the flight information of the current aircraft arriving or departing, including flight number, aircraft position, taxiway occupancy, runway usage status, etc., requiring the air traffic controller to make decisions quickly according to the task requirements and command the aircraft to take off and land in an orderly manner; in the multi-flight scheduling task, the task induction interface can dynamically simulate the operation trajectories of multiple flights in the airspace, and add sudden task interference (such as conflict risk prompts or weather change alarms) to induce changes in the cognitive load of air traffic controllers in complex working environments;

[0018] The interactive feedback area is used to display the air traffic controller's operation results and evaluation data in real time; for example, after the air traffic controller completes the aircraft takeoff and landing command, the feedback area will display the completion time, flight scheduling efficiency and operation safety indicators, and at the same time combine the real-time EEG signal processing results to provide feedback on the air traffic controller's current cognitive load level in the form of intuitive graphics or text; this task induction and interactive design based on actual task operations ensures a high degree of fit between the system and the real working environment, providing a guarantee for the accuracy and practicality of cognitive load detection.

[0019] Furthermore, the real-time EEG signal acquisition module records the EEG signals of the air traffic controller through a 64-lead EEG electrode cap. The electrode positions follow the international 10-20 system, covering the frontal lobe (FZ), parietal lobe (CZ, PZ) and occipital lobe (O1, O2), and the sampling rate is 512Hz; the collected signals are band-pass filtered (0.5Hz to 50Hz) to remove low-frequency drift and high-frequency noise, and independent component analysis (ICA) is used to remove eye movement artifacts and electromyographic artifacts to ensure signal quality; the processed signals are transmitted to the EEG signal processing module in real time through the TCP / IP protocol, where the server is responsible for collecting signals and the client is responsible for receiving and analyzing signals.

[0020] Furthermore, the EEG signal processing module includes three stages: signal preprocessing, feature extraction and dynamic modeling. In the signal preprocessing stage, the fast Fourier transform (FFT) is used to convert the time domain signal into a frequency domain signal, and the power spectral density (PSD) of the alpha wave (8-13Hz), beta wave (13-30Hz) and theta wave (4-8Hz) is extracted. The power changes in these frequency bands reflect the cognitive load state of the air traffic controller. In the feature extraction stage, the short-time Fourier transform (STFT) is combined to analyze the time-frequency distribution of the EEG signal to capture the dynamic changes of cognitive load. In the feature modeling stage, the time dependency of the signal is captured through the optimized temporal memory module, the memory state is dynamically updated, the causal consistency of time series modeling is ensured by combining the causal convolution module, and the agent attention mechanism is used to enhance the modeling ability of the global context.

[0021] Specifically, the EEG signal processing module performs multi-step processing and feature extraction on the collected EEG signals, mainly using time-frequency domain analysis methods for feature extraction; in the preprocessing stage, the EEG signal first passes through a bandpass filter (0.5Hz to 50Hz) to remove power frequency interference and other high-frequency noise, and the independent component analysis (ICA) method is used to remove eye movement artifacts and electromyographic artifacts; the preprocessed signal is used to extract features related to cognitive load through time-frequency domain analysis technology, in which the fast Fourier transform (FFT) is used to convert the time domain signal into a frequency domain signal, and the power spectral density (PSD) of different frequency bands is calculated; these frequency bands correspond to different cognitive states, for example, the power changes of alpha and beta waves reflect the attention allocation and workload of the air traffic controller, while the enhancement of theta waves may be related to the increase in task difficulty; through the energy distribution of these frequency bands and their dynamic changes over time, the system can accurately characterize the cognitive load state of the air traffic controller.

[0022] Furthermore, the processing module further uses the optimized temporal memory module to model and classify the extracted frequency domain features; the optimized temporal memory module adjusts the memory state and output state through a dynamic gating mechanism, specifically: the memory state of the current time step is dynamically combined by the forget gate, the input gate and the historical memory, and the output state combines the weight of the current memory state and the output gate, which can deeply model the long time series features; at the same time, the module combines the temporal characteristics of multi-band features to improve the adaptability to complex dynamic task scenarios;

[0023] Specifically, the optimized temporal memory module can effectively capture the temporal dependency of frequency band features in EEG signals, dynamically select feature information related to the current task through a gating mechanism, and thus enhance the adaptability to complex time series; the model achieves deep modeling of long time series through a multi-layer stacking structure, while introducing a causal convolution mechanism to ensure the causal consistency of time series modeling; this design not only improves the robustness of the model when processing dynamic feature data, but also significantly improves the accuracy and real-time performance of classification;

[0024] Step 1: Optimize global context modeling through proxy attention mechanism; The present invention introduces the proxy attention mechanism in the global modeling of EEG signals, and significantly reduces the computational complexity of the traditional attention mechanism by generating proxy tokens, while retaining the modeling ability of global context information; The query matrix of the input signal It is divided into windows of fixed size and generates a proxy matrix through the maximum pooling operation Where m<<n; the generation formula of the proxy matrix is:

[0025]

[0026] In this way, each proxy token represents the global feature of the corresponding window, which significantly reduces the complexity of redundant calculations;

[0027] Then, the surrogate matrix A is related to the bond matrix Sum Matrix Interact to generate the global attention distribution weighted result O A , the formula is as follows:

[0028] O A =σ(QA T +B2)·σ(AK T +B1)·V (2)

[0029] In order to enhance the local feature capture capability, the present invention applies convolution operation to the value matrix V, combines the convolution feature with the global attention result, and finally outputs the comprehensive feature:

[0030] O=OA +Conv(V) (3)

[0031] By introducing the proxy token, the computational complexity of the proxy attention mechanism is reduced from O(n 2 ) is reduced to O(nm), while retaining the integrity of key features globally, providing an efficient solution for processing long sequence EEG signals;

[0032] Step 2: Introduce an optimized temporal memory module to enhance the modeling capability of long time series; In view of the complex dynamic characteristics of EEG signals, the present invention designs an optimized temporal memory module to enhance the dependency modeling capability of long time series and improve the accuracy of feature extraction; the module realizes dynamic update of memory state through a gating mechanism, and the update formula is as follows:

[0033]

[0034] Among them, c t is the memory state of the current time step, h t is the output state of the current time step, f t 、i t and t are the activation values ​​of the forget gate, input gate, and output gate, respectively, which control the retention, update, and output weight of memory information. Represents the input information of the current time step;

[0035] The content of the input gate unit is:

[0036]

[0037] The optimized temporal memory module adopts matrix memory units, expanding the traditional scalar memory into matrix representation to improve the ability to capture multi-dimensional and multi-band features. This matrix design combines the multi-frequency domain characteristics of EEG signals, especially when analyzing the energy changes of different frequency bands such as α waves and β waves, it can better adapt to the diversity of spectral features in EEG signals. In addition, by introducing the dynamic update mechanism of Query, Key and Value, the module can flexibly adjust the information flow according to the correlation between time steps and focus on the time step most relevant to the current task.

[0038] Dynamic feature modeling not only enhances the ability to capture long-term dependencies, but also supports parallel computing through matrix structure, making it more efficient when processing large-scale EEG signals. In addition, the module optimizes the control mechanism of information flow and solves the problem of time dependency loss caused by gradient vanishing in traditional models in long sequence tasks through fine-grained modeling of signal dynamic dependencies.

[0039] Step 3: Combining the causal convolution module to ensure temporal causal consistency; In order to ensure causal consistency in time series modeling, the present invention introduces a causal convolution module in feature modeling; Causal convolution limits the receptive field of the convolution kernel and only uses the input features of the current time step and before for modeling, thereby avoiding interference of future information on the modeling results;

[0040] The specific calculation formula is as follows:

[0041]

[0042] Among them, y t is the output of the current time step, k is the convolution kernel size, w i is the convolution kernel weight, x t-i Represents the input signal before the current time step. By limiting the use of future information, the causal convolution module can strictly follow the temporal causal relationship and ensure that the modeling results are consistent with the actual time series characteristics.

[0043] In addition, in order to enhance the multi-level expression capability of features, the causal convolution module and the optimized temporal memory module are deeply integrated through the residual connection mechanism; the residual connection formula is as follows:

[0044] O final =O causal +O residual # (7)

[0045] Among them, O causal is the output of the causal convolution module, O residual It is the residual output of the previous layer. Residual connection can alleviate the gradient vanishing problem in deep networks, ensure the stability of the model when processing long time series tasks, and improve the ability to capture complex time dependencies.

[0046] Through lightweight convolution kernel design and parameter sharing strategy, the causal convolution module has been optimized in terms of computing resource usage, enabling it to efficiently process EEG signals of air traffic control tasks with high real-time requirements; the module design ensures the causal consistency of the time series and significantly improves the overall performance of the model through multi-level feature fusion.

[0047] The processed high-dimensional features are input into the classification network for classification prediction of cognitive load level. The classification network adopts a fully connected neural network structure and converts the prediction results into probability distribution through the Softmax activation function, thereby outputting the current cognitive load level of the air traffic controller. The final classification results are combined with the task induction and interaction modules for feedback, providing air traffic controllers with real-time load status monitoring and optimization suggestions.

[0048] After receiving the classification results, the cognitive load assessment module provides intuitive feedback on the air traffic controller's cognitive load status in a visual form; the output of the assessment module includes real-time cognitive load level, load change curve and historical statistical data; for example, in the aircraft take-off and landing command task, the system will not only display the air traffic controller's current cognitive load level, but also give operational suggestions based on the changing trend of his cognitive load, such as "pay attention to reducing the density of task allocation" or "arrange rest time appropriately"; in addition, the assessment module supports long-term storage and subsequent analysis of data, which is used to track and manage the work status of air traffic controllers and help optimize the task allocation strategy of air traffic control.

[0049] Through the collaborative work of the task induction and interaction module, the real-time EEG signal acquisition module, the EEG signal processing module and the cognitive load assessment module, the present invention can efficiently detect the cognitive load status in the real working environment of the air traffic controller; the system realizes the accurate analysis of the dynamic load changes in the air traffic control tasks, has real-time, reliability and practicality, and provides important technical support for aviation safety assurance and intelligent air traffic control systems.

[0050] The beneficial effects of the present invention are as follows: the present invention proposes an innovative method combining a proxy attention mechanism and a temporal neural network; by stacking a temporal attention memory network (STAM-Net), the present invention can make full use of the temporal characteristics and dynamic dependencies of EEG signals to accurately capture the changes in cognitive load of air traffic controllers under different task load scenarios; the model optimizes global context modeling through a proxy attention mechanism, significantly reducing computational complexity, while introducing an optimized temporal memory module to enhance the modeling capability of long time series, and combining a causal convolution module to ensure temporal causal consistency, thereby improving network stability and computational efficiency; experimental verification shows that the method exhibits excellent classification accuracy and robustness under multi-task load scenarios, especially in binary and quinary classification tasks, which is significantly superior to traditional methods and existing deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a block diagram of the composition of the present invention;

[0052] Figure 2 is a flow chart of the model in the present invention;

[0053] Figure 3 is a distribution diagram of electrode positions in the present invention;

[0054] Figure 4 It is a schematic diagram of each module in the present invention;

[0055] Figure 5 It is a work flow chart of the present invention. DETAILED DESCRIPTION

[0056] The specific technical scheme of the present invention is further described in detail below with reference to specific examples.

[0057] As shown in the figure, the present invention includes four parts: task induction and interaction module, EEG signal real-time acquisition module, EEG signal processing module and cognitive load assessment module.

[0058] The task induction and interaction module provides air traffic controllers with real task scenarios to induce their EEG signal characteristics in actual operations, and displays the task completion status and cognitive load status in real time. The task induction interface is directly integrated into the air traffic controller's workstation system, using actual air traffic control task operation scenarios as induction conditions, such as aircraft take-off and landing command, flight scheduling, route planning and conflict detection. The interface uses a high-resolution LCD display with a white background and black text and symbols for displayed task information, which is simple and easy to operate. The complexity of the task-induced scenario can be dynamically adjusted according to experimental or actual needs, from a single task (such as commanding a single aircraft take-off and landing) to complex multi-tasks (such as multi-flight route planning and conflict resolution), covering low-load, medium-load and high-load task scenarios. The interactive interface also integrates a real-time feedback function to dynamically display the air traffic controller's task completion status, task accuracy, execution time and the cognitive load level evaluated by the system (such as "low load", "medium load" or "high load"). This design enables air traffic controllers to intuitively understand their own working status and adjust their task strategies.

[0059] The EEG signal real-time acquisition module uses a 64-lead wireless EEG acquisition device. The distribution of electrodes is strictly designed according to the international 10-20 system standard to cover key brain areas such as the frontal lobe, parietal lobe and occipital lobe. The EEG activity in these areas is closely related to changes in cognitive load. The acquisition equipment includes an electrode cap, a signal amplifier, a multi-parameter synchronizer and acquisition software. The experimental sampling frequency is set to 512Hz to ensure that the EEG signal characteristics with high time resolution are captured. The acquisition system transmits the EEG signal to the processing module in real time by wireless means. The data transmission is based on the TCP / IP protocol, in which the acquisition module acts as a server and the processing module acts as a client to ensure the efficiency and stability of real-time data transmission. The real-time transmitted data stream X(t) is the acquired multi-lead signal matrix:

[0060] X(t)=[x1(t),x2(t),…,x 64 (t)] T (8)

[0061] Among them, x i (t) represents the signal value of the i-th electrode.

[0062] The EEG signal processing module is responsible for preprocessing, feature extraction and modeling analysis of the EEG signals collected in real time. First, in the signal preprocessing stage, a bandpass filter is used to filter the EEG signal to remove power frequency interference, low-frequency drift and high-frequency noise. At the same time, the independent component analysis method is used to eliminate artifact signals, such as eye movement and electromyography artifacts. The preprocessed signal is segmented into time windows of fixed length, each window corresponds to a time segment during the task execution process, which is used to capture the dynamic changes of the signal.

[0063] In the preprocessing stage, the signal X(t) is first filtered through a bandpass filter to remove power frequency noise and high-frequency interference, and the filter range is set to [0.5,50] Hz:

[0064]

[0065] Independent component analysis (ICA) is then used to separate artifacts (such as eye movements and myoelectric interference) from the EEG signal to obtain the purified signal X ICA (t); ICA decomposition is achieved through the following formula:

[0066] X ICA (t) = W·X f (t)#(10)

[0067] where W is the demixing matrix for independent component analysis.

[0068] The preprocessed signal X ICA (t) is divided into time windows of length T, and frequency domain features are extracted in each window. First, the time domain signal is converted into the frequency domain signal using the fast Fourier transform (FFT):

[0069]

[0070] In the feature extraction stage, the present invention focuses on analyzing the frequency domain characteristics of EEG signals, especially the frequency band changes that are closely related to cognitive load, such as the power changes of α waves, β waves and θ waves; the feature extraction focuses on combining the time-frequency domain analysis method to dynamically evaluate the spectral changes of the signal in different time segments to capture the attention allocation and brain load level of air traffic controllers during task execution.

[0071] In the modeling and analysis stage, the present invention uses the attention mechanism and temporal memory network to perform deep modeling on the extracted features; the attention mechanism is used to model the global context information of the EEG signal and effectively capture the global dependencies in the long time series; the attention mechanism extracts the key features of the signal and weights the information related to the current task, so that the model can focus more on the important time steps in the air traffic control task and ignore irrelevant background information; first, the proxy matrix is ​​calculated, and the generation formula of the proxy matrix is:

[0072]

[0073] In this way, each proxy token represents the global feature of the corresponding window, which significantly reduces the complexity of redundant computation.

[0074] Then, the surrogate matrix A is related to the bond matrix Sum Matrix Interact to generate the global attention distribution weighted result O A , the formula is as follows:

[0075] O A =σ(QA T +B2)·σ(AK T +B1)·V (13)

[0076] In order to enhance the local feature capture capability, the present invention applies convolution operation to the value matrix V, combines the convolution feature with the global attention result, and finally outputs the comprehensive feature:

[0077] O=O A +Conv(V) (14)

[0078] In addition, the optimized temporal memory module is used to capture the long-term dependencies of EEG signals, which enhances the modeling ability of complex dynamic features. The temporal memory module is designed with a dynamic information flow control mechanism, which enables it to flexibly adjust the weights of historical information and new information, thereby effectively balancing the expression of global dependencies and local details in the modeling process. The optimized temporal memory module dynamically adjusts the memory state c through a gating mechanism. t and output state h t :

[0079]

[0080] Among them, f t ,i t ,o t are the activation values ​​of the forget gate, input gate, and output gate, respectively.

[0081] In order to ensure the causal consistency of time series modeling, the present invention further combines the causal convolution module, and strictly limits the processing of time series to the input of the current time step and before, avoiding the problem of future information leakage; the causal convolution module also reduces the computational complexity through lightweight design, and deeply integrates with the temporal memory module through the residual connection mechanism, significantly enhancing the stability and training efficiency of the network; this module ensures temporal causal consistency:

[0082]

[0083] Finally, the high-dimensional features generated by modeling and analysis are input into the classification network, and the classification network completes the prediction of cognitive load level through layer-by-layer feature transformation and activation function.

[0084] L c =Softmax(W·h+b)#(17)

[0085] The classification results include three states: "low load", "medium load" and "high load", which correspond to the different cognitive load levels of air traffic controllers in task execution.

[0086] The cognitive load assessment module will visually feed back the classification results to the interactive interface, through which air traffic controllers can intuitively understand their own load status, such as whether the current task is within the controllable range and whether the task operation strategy needs to be adjusted. In addition, the assessment module also supports the storage and analysis of historical load data, which is used to track the work status of air traffic controllers in the long term, assist in optimizing task allocation strategies, and improve the overall efficiency and safety of air traffic control work.

[0087] The present invention can efficiently monitor the cognitive load status of air traffic controllers through the organic combination of task induction and interaction module, EEG signal real-time acquisition module, EEG signal processing module and cognitive load assessment module; experimental results show that the present invention has high load detection accuracy and real-time performance in high-complexity task scenarios, providing important technical support for the optimal allocation of air traffic control tasks and aviation safety assurance.

[0088] Example

[0089] The air traffic controller cognitive load detection method based on attention mechanism and temporal neural network of the present invention shows significant advantages in performance, real-time and adaptability compared with traditional air traffic controller load management methods; traditional methods mainly include questionnaires based on subjective evaluation and indirect measurements based on behavioral indicators, but these methods have many limitations; subjective evaluations such as NASA task load index rely on the self-perception of air traffic controllers, which are greatly affected by subjective factors such as emotions and experience, and cannot achieve real-time monitoring; although the method based on behavioral indicators can capture external manifestations such as task completion time and error rate, it is not sensitive to instantaneous changes in task load, and is difficult to be effectively applied in complex multi-tasking scenarios; in contrast, the present invention achieves high-precision cognitive load detection and dynamic change tracking by directly analyzing EEG signals, combining the proxy attention mechanism with the optimized temporal memory module.

[0090] Experimental results show that in two-category tasks (low load and high load), the classification accuracy of the present invention reaches 96.8%, while the traditional method is usually less than 85%; in five-category tasks (extremely low load, low load, medium load, high load and overload), the present invention still maintains an accuracy of 87.6%, which is significantly better than traditional EEG analysis methods and conventional deep learning models; especially in complex task scenarios, such as multi-flight conflict detection or dynamic scheduling, the robustness of the present invention has been further verified, and it can effectively adapt to the classification requirements under different load levels; this performance is due to the introduction of the proxy attention mechanism, which greatly reduces the computational complexity, and at the same time ensures the causal consistency of time series modeling through causal convolution, thereby avoiding the interference of future information on current results.

[0091] In terms of real-time performance, the present invention realizes the rapid processing of EEG signals in tasks by optimizing the modeling process and efficient feature extraction strategy, and the computing efficiency is improved by 37.5% compared with the traditional deep learning model. This advantage enables the present invention to perform real-time load monitoring in a high-frequency sampling signal environment, while traditional evaluation methods based on behavioral indicators often require post-analysis and are difficult to meet the needs of real-time task allocation. In addition, by combining short-time Fourier transform with frequency domain feature extraction, the present invention can capture the time-frequency changes of EEG signals in dynamic task scenarios and accurately reflect the attention allocation and cognitive load level of air traffic controllers. This capability is unmatched by traditional methods, especially in scenarios where multi-task loads change frequently, the dynamic adaptability of the present invention is particularly prominent.

[0092] Therefore, the present invention has verified its excellent performance in real-time, accuracy and adaptability by comparing with traditional load management methods; the experimental results fully demonstrate its superiority in complex mission scenarios, and provide an innovative and efficient technical solution for the optimal allocation of air traffic control tasks and aviation safety assurance.

Claims

1. A method for detecting cognitive load of air traffic controllers based on neural network, characterized in that: The operation steps are as follows: (1): Provide air traffic controllers with a dynamic simulated task environment through the task induction and interaction module to induce changes in their EEG signals; The mission-induced interface presents simulation scenarios related to air traffic control missions. The interactive feedback area displays the air traffic controller’s task completion status and cognitive load assessment results in real time; (2): The EEG signals generated by the air traffic controller during the task execution are collected through the EEG signal real-time acquisition module. The collected signals are processed by high-pass filtering, low-pass filtering and independent component analysis to remove artifacts to generate purified signal data; (3): Use the EEG signal processing module to extract features and perform dynamic modeling on the collected signals; (4): Receive classification results and provide visual feedback through the cognitive load assessment module.

2. The method for detecting cognitive load of air traffic controllers based on a neural network according to claim 1, characterized in that: In step (1), the simulation scenario includes tasks such as aircraft take-off and landing command, flight scheduling, route planning and conflict detection.

3. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 1, characterized in that: The step (1) specifically includes: the task induction and interaction module includes a task induction interface and an interactive feedback area, the task induction interface induces the cognitive load of the air traffic controller by dynamically simulating the air traffic control task scenario, and the task complexity is dynamically adjusted by increasing the multi-tasking processing demand; The interactive feedback area displays the task completion time, task accuracy and EEG signal analysis results in real time, and provides optimization suggestions for air traffic controllers.

4. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 1, characterized in that: The step (2) specifically includes: the EEG signal real-time acquisition module records the EEG signal of the air traffic controller through a 64-lead EEG electrode cap, the electrode position follows the international 10-20 system, covers the frontal lobe, parietal lobe and occipital lobe, and the sampling rate is 512 Hz; The collected signals are band-pass filtered to remove low-frequency drift and high-frequency noise, and independent component analysis is used to remove eye movement artifacts and electromyography artifacts; the processed signals are transmitted to the EEG signal processing module in real time via the TCP / IP protocol, where the server is responsible for collecting signals and the client is responsible for receiving and analyzing signals.

5. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 1, characterized in that: In step (3), the EEG signal processing module includes three stages: signal preprocessing, feature extraction and dynamic modeling; In the signal preprocessing stage, the fast Fourier transform is used to convert the time domain signal into the frequency domain signal, and the power spectrum density of the α wave, β wave and θ wave is extracted. The power changes in these frequency bands reflect the cognitive load state of the air traffic controller. In the feature extraction stage, short-time Fourier transform is used to analyze the time-frequency distribution of EEG signals to capture the dynamic changes of cognitive load; In the feature modeling stage, the temporal dependency of the signal is captured through an optimized temporal memory module, the memory state is dynamically updated, the causal consistency of time series modeling is ensured by combining the causal convolution module, and the proxy attention mechanism is used to enhance the modeling ability of the global context.

6. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 1, characterized in that: The step (3) specifically comprises: extracting the power spectral density of the EEG signal in the α wave, β wave, θ wave and other frequency bands by fast Fourier transform, and generating time-frequency features by combining short-time Fourier transform to reflect the changes in the cognitive load of the air traffic controller in different time periods; The global context modeling is optimized through the proxy attention mechanism, the signal query matrix is ​​divided into fixed-size windows, the proxy matrix is ​​generated and combined with the key matrix and the value matrix for interaction, which reduces the computational complexity and enhances the modeling ability of global features; The optimized temporal memory module is used to model the dependency of long time series, the dynamic gating mechanism is used to select relevant time step information, and the causal convolution module is combined to ensure causal consistency in time series processing; The processed high-dimensional features are input into the classification network, and the air traffic controller's cognitive load level is output through the Softmax activation function, including low load, medium load and high load.

7. A method for detecting cognitive load of air traffic controllers based on neural network according to claim 6, characterized in that: The proxy attention mechanism reduces the computational complexity of global attention by introducing a proxy token matrix; the input query matrix is ​​divided into fixed-size windows, and the generated proxy token matrix represents the global feature information within the window; the global attention distribution is generated by weighted calculation of the proxy token matrix with the key matrix and the value matrix, and finally the global attention features are combined with the local features extracted by the convolution operation to output the comprehensive features for subsequent modeling.

8. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 6, characterized in that: The optimized temporal memory module adjusts the memory state and output state through a dynamic gating mechanism, specifically: the memory state of the current time step is dynamically combined by the forget gate, the input gate and the historical memory, and the output state combines the current memory state and the weight of the output gate to deeply model the long time series features; At the same time, the module combines the timing characteristics of multi-band features to improve its adaptability to complex dynamic mission scenarios.

9. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 6, characterized in that: The classification network adopts a fully connected neural network structure, and maps the high-dimensional features of the EEG signal to the cognitive load level of the air traffic controller through the Softmax activation function; The cognitive load assessment module generates a real-time load change curve based on the classification results, and combines historical data for trend analysis and visual feedback to provide support for air traffic control task control and work efficiency optimization.

10. The method for detecting cognitive load of air traffic controllers based on neural network according to claim 1, characterized in that: The step (4) specifically includes: displaying the current cognitive load level, load change curve and trend analysis results in graphical or textual form, and providing optimization suggestions in combination with task results.

Images