Remote control tower air traffic controller visual fatigue monitoring method

Abstract

The invention relates to the technical field of asthenopia monitoring and health monitoring, and discloses a remote control tower air traffic controller asthenopia monitoring method. The method comprises the following steps: synchronously collecting eye movement, electrocardio, workload and environment multi-modal data, generating a labeling label in combination with a visual function test, and constructing a high-dimensional feature data set; key features are extracted through principal component analysis dimensionality reduction, and real-time prediction of the asthenopia state is achieved through LightGBM model training; a sliding median filtering and fixation event merging strategy is adopted for the eye movement data, and baseline drift is removed and HRV features are extracted from the electrocardio data through Bior4.4 wavelet decomposition; generating an asthenopia label based on the weighted fusion visual function index; optimizing model hyper-parameters by using TPE; in real-time monitoring, the system outputs a trigger alarm according to the model. According to the invention, non-intrusive and high-precision asthenopia monitoring is realized, the system effectively adapts to the complex environment of a remote control tower, and the health management efficiency of controllers and the aviation operation safety are remarkably improved.

Description

Technical Field

[0001] The invention relates to the technical field of visual fatigue monitoring and health monitoring, and in particular to a method for monitoring visual fatigue of air traffic controllers at a remote tower. Background Art

[0002] In the aviation field, the job safety of air traffic controllers is of vital importance. With the increase in air traffic and the increasing complexity of control tasks, air traffic controllers are faced with long-term and high-intensity workloads, and visual fatigue has become an important factor affecting air traffic controllers' job performance and safety.

[0003] At present, the most commonly used method in visual fatigue detection technology is based on subjective questionnaire surveys, which usually obtain visual fatigue perception through user self-reporting. Recent studies have proposed a method of using a single physiological signal to identify visual fatigue. The steps usually include: first, collecting a single physiological signal through a sensor, such as eye movement or EEG; then, extracting features related to visual fatigue, such as blinking frequency, eyelid opening and closing, EEG signal θ wave, etc.; then, analyzing these feature changes and combining simple classification models (such as support vector machines) to determine the degree of visual fatigue; finally, outputting the visual fatigue evaluation results.

[0004] Traditional visual fatigue detection methods mainly rely on subjective evaluation or single physiological signal recognition, which often cannot accurately reflect the fatigue status of controllers in real time and are greatly affected by factors such as environmental conditions and individual differences. In a remote tower environment, controllers need to concentrate for a long time and continuously monitor multiple flight targets and routes, which can easily lead to visual and physiological fatigue.

[0005] The shortcomings of the existing technology are mainly manifested in: 1) Lack of accuracy: Methods based on subjective questionnaires are easily affected by individual differences and emotional factors, and cannot achieve real-time monitoring, resulting in low accuracy in visual fatigue assessment.

[0006] 2) Single signal dependence: The use of a single physiological signal (such as eye movement or EEG) cannot fully reflect the multi-dimensional characteristics of visual fatigue, and has high requirements for signal extraction and analysis. It is easily affected by noise interference, which affects the stability and accuracy of the evaluation results.

[0007] 3) Poor real-time performance: Most existing visual fatigue detection methods lack a real-time feedback mechanism and are unable to issue timely warnings when controllers experience visual fatigue, affecting safety management.

[0008] 4) Limited applicability: Similar visual fatigue detection technologies are mostly used in driving, flying and other fields, such as real-time monitoring of the fatigue status of drivers or pilots through eye tracking, facial expression analysis and other technologies. However, visual fatigue detection technology for remote tower controllers has not been fully studied and applied. As an emerging air traffic control technology, the workload and visual pressure of remote tower controllers in the virtual tower environment are more complex, and existing technologies have not yet been able to fully adapt to the needs of this specific work scenario. Summary of the invention

[0009] In view of the above problems, the purpose of the present invention is to provide a remote tower air traffic controller visual fatigue monitoring method, by comprehensively utilizing visual behavior and multimodal data, to evaluate the remote tower air traffic controller's visual fatigue state in real time and objectively, improve the accuracy and reliability of monitoring, and ensure the controller's health and safety in a high-intensity working environment. The technical solution is as follows: A method for monitoring visual fatigue of air traffic controllers at a remote tower comprises the following steps: Step 1: Synchronously collect multi-modal data: Use sensor equipment to obtain the controller's eye movement data, electrocardiogram data, workload data, and environmental data in real time, and use visual function test equipment to obtain labeled visual fatigue label data; Step 2: Data preprocessing and feature extraction: De-noising, interpolation repair and standardization are performed on the eye movement data, electrocardiogram data, workload data and environmental data to extract feature parameters related to visual fatigue; Step 3: Construct a training data set: align the preprocessed feature parameters with the visual fatigue label data according to the time window to form a high-dimensional feature data set; Step 4: Model training: principal component analysis is used to reduce the dimension of the high-dimensional feature data set to obtain a low-dimensional feature data set, which is then input into the LightGBM model for training to generate a visual fatigue prediction model; Step 5: Real-time monitoring and early warning: The multimodal data collected in real time is input into the visual fatigue prediction model to determine the current visual fatigue state of the controller. If visual fatigue is detected, the alarm mechanism is triggered.

[0010] The beneficial effects of the present invention are: 1) This invention combines multi-dimensional information such as eye movement, electrocardiogram, workload and environmental data through multi-modal data synchronous acquisition and fusion technology, which can comprehensively and accurately reflect the visual fatigue status of remote tower air traffic controllers. Compared with the traditional single data source monitoring method, this invention can capture early signs of fatigue from more angles, improve the accuracy and sensitivity of monitoring, and effectively solve the problem that traditional methods cannot comprehensively evaluate the visual fatigue status of controllers.

[0011] 2) Combined with the PCA-LightGBM model, the present invention can efficiently reduce and filter features under the conditions of large amounts of data and high-dimensional features, avoiding the negative impact of redundant features on model performance, and further improving the real-time and accuracy of controller visual fatigue monitoring through the real-time prediction function of the machine learning model. This enables the present invention to not only detect the risk of visual fatigue early, but also dynamically adjust the work tasks of the controller according to real-time data, thereby effectively preventing accidents caused by visual fatigue and ensuring the safety of remote tower operations and the health of controllers.

[0012] 3) Through advanced data fusion and machine learning technology, the present invention provides a more accurate and efficient visual fatigue monitoring and prediction solution, which not only optimizes the monitoring process, but also greatly improves the safety level of air traffic control work. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a remote tower air traffic controller visual fatigue monitoring method. DETAILED DESCRIPTION

[0014] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] The present invention proposes a visual fatigue monitoring method based on multimodal data synchronous acquisition and PCA-LightGBM model. The method simultaneously collects multiple physiological and environmental signals such as eye movement, electrocardiogram, workload and environment, combines PCA (Principal Component Analysis) technology for dimensionality reduction processing, and then uses the LightGBM model for efficient analysis, which can comprehensively reflect the visual fatigue state of the controller, not only considering eye fatigue, but also comprehensively analyzing the influence of factors such as heart rate and workload on fatigue. Compared with the prior art, the present invention has strong real-time performance, accuracy and robustness, can operate stably in a complex environment and avoid noise interference, timely feedback the visual fatigue state of the controller and issue a warning, thereby improving the safety and work efficiency of the remote tower. At the same time, the present invention is specially designed for the working environment of the remote tower controller, which can effectively solve the problem that the prior art cannot cope with the visual fatigue monitoring in this specific scenario. Finally, the present invention aims to provide a real-time, comprehensive and non-invasive visual fatigue monitoring method, improve the safety and health management level of the controller's working environment, avoid work errors and safety hazards caused by visual fatigue, and further optimize the work quality of the remote tower.

[0016] First, we collect the controller's eye movement, electrocardiogram, workload, environment, and visual function data, and collect data through specially designed sensors, equipment, and scales. Then, the collected multimodal data will be input into the data preprocessing module for data cleaning, standardization, and feature extraction. In the feature extraction stage, we extract key features related to visual fatigue from the multimodal data. Subsequently, these features will form the final training data set together with the visual fatigue labels calculated from the visual function data for training the visual fatigue detection model. By modeling the training data using the PCA-LightGBM algorithm, the system will be able to identify the visual fatigue state of the controller.

[0017] In the real-time monitoring stage, we input the controller's real-time eye movement data and ECG data into the trained model for prediction to determine whether the controller is currently in a state of visual fatigue. If visual fatigue is detected, the system will trigger the visual fatigue warning mechanism to remind the controller to take a rest or adjust the working state. If visual fatigue is not detected, the next step of monitoring will be continued. Through this complete process, the system can monitor the visual fatigue state of the controller in real time and issue fatigue warnings in time, thereby improving the controller's work safety and health management level.

[0018] 1. Data collection: The present invention collects eye tracking data through a glasses-type eye tracker to monitor the visual behavior of the controller; collects electrocardiogram data through a dual-lead electrocardiogram; obtains workload data by collecting the work log and shift schedule of the control system; and obtains environmental data through an integrated ambient light quality tester. In this way, multiple data required for the data set can be synchronously obtained without affecting the normal control work of the remote tower controller.

[0019] At the same time, in each data group, visual function data including best corrected distance visual acuity (BCDVA), best corrected near visual acuity (BCNVA), near point of accommodation (NPA), near point of convergence (NPC), and tear break-up time (TBUT) were obtained using a comprehensive ophthalmometer and other equipment, and contrast sensitivity (CSF) was obtained using the Pelli-Robson scale.

[0020] 2. Dataset preprocessing: Preprocess and extract features of the collected data, and divide them into 5-minute time windows.

[0021] A. Eye movement data: The collected eye movement data were preprocessed using sliding median filtering to remove noise and smooth the data. Then, the missing parts in the data were repaired using an interpolation repair method with a maximum gap length of 75 milliseconds. Subsequently, the eye movement data were precisely aligned with the timestamps of the task events to ensure the synchronization and consistency of the data. In order to effectively detect gaze events, the minimum gaze duration was set to 60 milliseconds, and adjacent gazes with a duration of less than 75 milliseconds or a maximum gaze angle of less than 0.5° were merged into one gaze event. Furthermore, in order to identify saccades, the minimum speed threshold was set to 2 pixels / millisecond, and the duration of the saccade was specified to range from 10 milliseconds to 200 milliseconds. Finally, the pupil area was linearly interpolated to ensure smooth changes in the pupil diameter, and the minimum value of the pupil diameter was limited to 2 mm. For the recognition of blink events, a time window of 70 to 350 milliseconds was used for threshold judgment, so as to accurately detect the controller's blink events.

[0022] B. ECG data: The power frequency interference of the collected ECG data is removed by adaptive filtering, and the baseline drift is removed by using Bior4.4 wavelet transform. First, the wavelet decomposition decomposes the signal into different scales and frequency bands through low-pass filtering and high-pass filtering: ; in, Indicates the collected original ECG signal; Indicates the signal through discrete wavelet transform (DWT, Discrete Wavelet Transform) Decomposition treatment; It is The detail coefficient of the layer (high frequency part); It is Approximation coefficient of the layer (low frequency part); is the number of wavelet decomposition levels.

[0023] Baseline drift is usually located in the low-frequency part of the signal, so the baseline drift usually appears as an approximation factor The low-frequency components in the approximation coefficients. By removing or weakening the high-frequency components of these low-frequency parts, the baseline drift can be effectively removed. Set the threshold λ, for the low-frequency parts in the approximation coefficients (such as the first layer approximation coefficients ), process it, remove or reduce the The coefficient of: .

[0024] Next, the bad channel data was repaired by averaging and interpolating adjacent channels to ensure the continuity and integrity of the signal. Subsequently, independent component analysis (ICA) was applied to decompose the ECG signal and the artifact components were manually removed to retain the real signal related to cardiac activity.

[0025] To further analyze the ECG signal, heart-related events are detected from the preprocessed ECG signal. The maximum heart rate is set to 120 bpm, and the R wave threshold is set to 70%. The time difference between adjacent R waves is calculated based on the timestamp of the R wave to form the RR interval sequence. This sequence is used to extract the time domain features of heart rate variability (HRV). Finally, the RR interval sequence is spectrally analyzed by Fast Fourier Transform (FFT) to further extract the frequency domain features of HRV. This process can provide accurate time domain and frequency domain features for subsequent heart health monitoring, effectively reflecting the state and dynamic changes of the heart.

[0026] C. Workload data: The control system logs and controller schedule data are converted into time series data, and missing values ​​and outliers are cleaned. Each record contains information such as the number of flights within 5 minutes, task switching frequency, task complexity, task urgency, and scheduling frequency.

[0027] D. Environmental data: Clean up outliers in the collected environmental data and synchronize them with other data to ensure that the environmental data and physiological data are aligned in the same time window for comprehensive analysis.

[0028] Finally, to ensure the uniformity of the dimensions of each modal data, the Z-score standardization method is used to standardize all data: ; in, is the standardized data, is the characteristic data of each mode after preprocessing, and are their mean and standard deviation respectively. The mean of the standardized data is 0 and the variance is 1, which can eliminate the dimensional differences between different features and thus improve the effect of model training.

[0029] All feature parameters included in the dataset are shown in Table 1.

[0030] Table 1 Characteristic parameters .

[0031] E. Visual fatigue status label: First, obtain contrast sensitivity data from the collected Pelli-Robson sensitivity scale results. The core of this scale is that it measures the subject's ability to recognize letters under low contrast by using letter combinations with gradually decreasing contrast. In the Pelli-Robson sensitivity test, given the contrast level of a letter group, the tester needs to identify the letters and give a response. The contrast of each group of letters gradually decreases until the subject can no longer recognize the letters. By measuring the lowest contrast that the subject can recognize, the contrast sensitivity score is calculated. The formula is as follows: ; in, and The maximum and minimum contrast sensitivity values ​​in the Pelli-Robson test are usually 1.8. This formula can be used to normalize the test results to a score between 0 and 1. Lower contrast sensitivity (i.e., contrast sensitivity Smaller) corresponds to higher contrast sensitivity scores The value indicates that the controller’s visual fatigue level is high.

[0032] Secondly, other visual function data are used for auxiliary judgment. Visual function data include but are not limited to near point of accommodation (NPA), near points convergence (NPC), best corrected distance visual acuity (BCDVA) and other indicators. For these visual function indicators, the standardized score of each indicator is set as ,in The value is between 0 and 1. Taking into account Pelli-Robson sensitivity and other visual function data, a comprehensive score function of visual fatigue status is defined , which takes into account the weighted average of contrast sensitivity and visual function data: ; in, , is the weighted coefficient of each indicator, and , the specific distribution of weighted coefficients is shown in Table 2.

[0033] Table 2 Characteristic weighting coefficient allocation table .

[0034] Comprehensive score function The value of is between 0 and 1, and the higher the comprehensive score function A value of 0 indicates that the controller's visual fatigue is more serious.

[0035] Finally, a threshold is set based on the experience of medical experts , which is used to determine whether the controller is in a state of visual fatigue. If the comprehensive score function Greater than threshold T , then the controller is judged to be in a state of visual fatigue Label is the visual fatigue label, 0 means no visual fatigue, 1 means visual fatigue: ; The obtained visual fatigue status is added to the standardized dataset to obtain a high-dimensional feature dataset with labels.

[0036] 3. PCA-LightGBM: Next, we use the acquired dataset to train the PCA-LightGBM model. The specific workflow is as follows: A. PCA processing: First, the standardized high-dimensional feature data set is reduced in dimension using principal component analysis. Principal component analysis (PCA) is a commonly used dimensionality reduction method that aims to map high-dimensional data to a low-dimensional space through linear transformation to retain the maximum variability in the data. The core idea is to construct a set of new orthogonal bases (principal components) to maximize the explanation of data variance while reducing redundant information.

[0037] First, the standardized data set is converted into a feature matrix , where each row is a sample and each column is a feature. Calculate the covariance matrix of the data : ; in, is the sample size, is the feature matrix The transpose of .

[0038] Solving eigenvalues ​​and eigenvectors: Solving the covariance matrix through eigenvalue decomposition The eigenvalues ​​and corresponding eigenvectors of are: ; in, is the eigenvector matrix, is the eigenvalue matrix. The eigenvectors represent the direction of data variation, and the eigenvalues ​​represent the variance of the data in these directions.

[0039] Select the principal components to construct the dimension reduction matrix: sort the eigenvectors from large to small according to the corresponding eigenvalues, and select the principal components that can explain more than 90% of the data variance. eigenvectors as new principal components. The eigenvectors form a matrix , and then the standardized data Projected into this new space, we get the reduced-dimensional data matrix : .

[0040] B. LightGBM processing: Input the low-dimensional feature data set after PCA processing into the LightGBM model for training. LightGBM (Light Gradient Boosting Machine) is an efficient implementation of the Gradient Boosting Decision Tree (GBDT) algorithm, which is widely used in machine learning tasks of large-scale data sets. The model uses an additive model framework to gradually train several base classifiers (i.e., decision trees), and each classifier corrects the errors of the previous classifier as much as possible.

[0041] At initialization, the model's prediction value is the mean of the target values ​​of the low-dimensional feature training set. Let the initial prediction value f 0( x )for: ; in, It is The true value of the training samples, is the number of training samples.

[0042] In each iteration, LightGBM optimizes the model by calculating the residual (i.e., the difference between the predicted value and the true value). Each iteration adds a new base classifier , adjust the model by minimizing the objective loss function. The goal is to minimize the following loss function by adding a new tree each time: ; in, is the overall loss function; It is The loss function of the training sample is The error between the true value and the predicted value of the training sample; It is The predicted value of the model after iterations.

[0043] C. Model tuning: When training the PCA-LightGBM model (i.e., the visual fatigue prediction model of the present invention), the Tree-structured Parzen Estimator (TPE) is introduced to optimize the model parameter selection. TPE divides the sampling points of the objective function into two parts: a parameter set with better performance (lower objective function value) and parameter sets with poor performance (with a higher objective function value), and model the two sets separately.

[0044] First, through the and To perform estimation, TPE uses Bayes’ theorem to transform the optimization problem into the following form: ; in, It is the performance threshold, which is used to classify good and bad parameters; Indicates the objective function value Less than threshold In the case of The probability distribution of Indicates that given parameters When Less than threshold The probability of Representation parameters TPE uses these two probability distributions to build a generative model, where the optimal parameter set Used to guide the sampling of new parameters, a set of parameters with poor performance It provides exploration information, thereby concentrating sampling near the optimal solution and gradually approaching the optimal solution.

[0045] Then, during the optimization process, the performance of each set of hyperparameters was evaluated through five-fold cross validation, and the average root mean square error (RMSE), mean absolute error (MAE), and adjusted coefficient of determination (Adjusted R-Square) were used as objective function values ​​to quantify the prediction error of the model and guide the optimization direction of the TPE algorithm: ; ; ; ; in, is the true value, is the model prediction value; is the coefficient of determination, is the adjusted coefficient of determination; is the number of training samples; is the number of independent variables (features).

[0046] Finally, the present invention performs 100 iterations in the training phase, and each iteration updates the probability model according to the performance of the current hyperparameters to continuously approach the optimal hyperparameters. The PCA-LightGBM model is retrained using the hyperparameters found by TPE optimization to obtain the best performance.

[0047] 4. Real-time visual fatigue detection: After the model training is completed, the workflow of the real-time visual fatigue detection part is operated through the multimodal data collected in real time. In actual applications, the system will collect multimodal data in real time through a variety of devices. These devices are embedded with wireless communication modules, which can transmit the collected data in real time to the back-end server for further analysis and calculation. The multimodal data collected in real time is processed by the preprocessing module, including denoising, interpolation repair and standardization, to ensure the stability of data quality. The preprocessed data will be input into the trained machine learning model, and the model will make real-time predictions on the current visual fatigue status based on the patterns and features learned during training.

[0048] The system determines whether the controller is in a state of visual fatigue based on the output of the model. If the system detects that the controller is in a state of visual fatigue, it will trigger an alarm mechanism to remind the controller to take a break or adjust his working state in time, thereby avoiding safety hazards caused by visual fatigue. The entire process can monitor visual fatigue in real time without interfering with the normal work of the controller, ensuring the safety of the working environment and the health of the controller.

[0049] In summary, the technical points of the present invention are as follows: 1. Customized design adapted to remote tower environment: The present invention specifically designs a customized data collection and processing method based on the working environment and characteristics of air traffic controllers in remote towers to ensure that the technical solution can effectively adapt to the actual application needs of remote towers.

[0050] 2. Visual fatigue prediction based on PCA-LightGBM model: Combining PCA dimensionality reduction technology and LightGBM machine learning model, the present invention can effectively extract key features from multimodal data and realize real-time monitoring and accurate prediction of visual fatigue status.

[0051] 3. Synchronous acquisition and fusion of multimodal data: The present invention adopts the synchronous acquisition of eye movement, electrocardiogram, workload and environmental data, and provides a comprehensive information basis for monitoring the visual fatigue status of remote tower air traffic controllers through the fusion of multimodal data.

[0052] 4. Optimal selection and efficient modeling of cross-modal features: PCA is used to reduce the dimension of the data, enhance the model's ability to recognize visual fatigue features, improve the model's computational efficiency, and reduce the impact of redundant features.

[0053] 5. Real-time monitoring and early warning function: The present invention can non-invasively monitor the visual fatigue status of air traffic controllers in real time during their work, and issue warnings in advance through prediction results, providing controllers with timely rest and adjustment suggestions.

[0054] 6. Model optimization and parameter selection method: This paper optimizes the hyperparameter selection of the PCA-LightGBM model by introducing the tree-structured Parzen estimator (TPE), further improving the prediction accuracy and stability of the model in visual fatigue monitoring.

Claims

1. A remote tower air traffic controller visual fatigue monitoring method, characterized in that: The following steps are involved: Step 1: Synchronously collect multi-modal data: Use sensor equipment to obtain the controller's eye movement data, electrocardiogram data, workload data, and environmental data in real time, and use visual function test equipment to obtain labeled visual fatigue label data; Step 2: Data preprocessing and feature extraction: De-noising, interpolation repair and standardization are performed on the eye movement data, electrocardiogram data, workload data and environmental data to extract feature parameters related to visual fatigue; Step 3: Construct a training data set: align the preprocessed feature parameters with the visual fatigue label data according to the time window to form a high-dimensional feature data set; Step 4: Model training: principal component analysis is used to reduce the dimension of the high-dimensional feature data set to obtain a low-dimensional feature data set, which is then input into the LightGBM model for training to generate a visual fatigue prediction model; Step 5: Real-time monitoring and early warning: The multimodal data collected in real time is input into the visual fatigue prediction model to determine the current visual fatigue state of the controller. If visual fatigue is detected, the alarm mechanism is triggered.

2. A remote tower air traffic controller visual fatigue monitoring method according to claim 1, characterized in that: In step 1: The eye movement data includes pupil diameter range, blink times, eye saccade times, fixation times and duration; The electrocardiogram data includes time domain characteristics and frequency domain characteristics of heart rate variability; The workload data includes the number of flights, task switching frequency, task complexity and scheduling frequency; The environmental data includes luminous flux, color temperature, flicker index, and screen brightness and contrast.

3. A remote tower air traffic controller visual fatigue monitoring method according to claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Preprocessing of eye movement data; Sliding median filtering was used for denoising, and the missing parts in the data were repaired by interpolation repair method, and the eye movement data were aligned with the timestamps of the task events; the minimum fixation duration threshold was set to merge adjacent fixation events; the minimum saccade speed threshold and the duration range of saccades were set to identify saccades; and the pupil diameter was smoothed by linear interpolation; Step 2.2: Preprocessing of ECG data; The power frequency interference of the collected ECG data was removed by an adaptive filter, and the baseline drift was removed by using Bior4.4 wavelet transform; First, wavelet decomposition decomposes the signal into different scales and frequency bands through low-pass filtering and high-pass filtering: ； in, Indicates the collected original ECG signal; It means that the original ECG signal is decomposed by discrete wavelet transform; It is The detail coefficient of the layer belongs to the high-frequency part; It is The approximation coefficient of the layer belongs to the low-frequency part; is the number of levels of wavelet decomposition; Setting Thresholds , process the low-frequency part of the approximation coefficient and remove or reduce the frequency below the threshold The coefficient of ； in, is the approximation coefficient of the first layer; threshold is the threshold function; Then, the bad channel data is repaired by averaging and interpolating adjacent channels; independent component analysis is applied to decompose the ECG signal and the artifact components are manually removed to retain the real signal related to cardiac activity; Detect cardiac-related events from preprocessed ECG signals: set the maximum heart rate threshold and wave threshold, calculate the time difference between adjacent R waves based on the timestamp of the R wave, and then form the RR interval sequence; finally, extract the time domain and frequency domain features of heart rate variability; Step 2.3: Preprocessing of workload data and environmental data: clean up outliers in workload data and environmental data, and synchronize the environmental data with the physiological data.

4. A remote tower air traffic controller visual fatigue monitoring method according to claim 1, characterized in that: In step 3, visual fatigue label data Generated by the following formula: ； in, , The standardized visual function indicators include best corrected distance visual acuity, best corrected near visual acuity, accommodation near point, convergence near point, contrast sensitivity and tear film break-up time; , is the weighted coefficient of each indicator, and ; when When T is the preset threshold.

5. A remote tower air traffic controller visual fatigue monitoring method according to claim 1, characterized in that: The step 4 specifically includes: Step 4.1: Use principal component analysis to reduce the dimension of the standardized high-dimensional feature data set; Compute the covariance matrix: Convert the standardized data set into a feature matrix , where each row is a sample and each column is a feature; first calculate the covariance matrix of the data : ； in, is the number of feature samples, is the feature matrix The transpose of Solve for eigenvalues ​​and eigenvectors: Solve the covariance matrix through eigenvalue decomposition The eigenvalues ​​and corresponding eigenvectors of are: ； in, is the eigenvector matrix, is the eigenvalue matrix; Select the principal components to construct the dimension reduction matrix: The eigenvectors are sorted from large to small according to their corresponding eigenvalues, and the eigenvectors that can explain the data variance are selected. eigenvectors as new principal components; the selected The eigenvectors form a matrix , and then the standardized feature matrix Projected into this new space, we get the reduced-dimensional data matrix : ； Step 4.2: LightGBM model processing; The low-dimensional feature data set processed by principal component analysis is input into the LightGBM model for training; At initialization, the model's prediction value is the mean of the target values ​​of the low-dimensional feature training set; let the initial prediction value f 0( x )for: ； in, It is The true value of the training samples, is the number of training samples; In each iteration, the LightGBM model optimizes the model by calculating the residual; each iteration adds a new base classifier , adjust the model by minimizing the objective loss function; the goal is to minimize the following loss function by adding a new tree each time: ； in, is the overall loss function; It is The loss function of the training sample is The error between the true value and the predicted value of the training sample is It is The predicted value of the model after iterations; Step 4.3: Model tuning; When training the PCA-LightGBM model, a tree-structured Parzen estimator is introduced to optimize the model parameter selection; the tree-structured Parzen estimator algorithm divides the sampling points of the objective function into two parts: the optimal performance parameter set Sum and Difference Performance Parameter Set , and model these two sets separately; First, by using two probability distributions and To make an estimate, use Bayes’ theorem to transform the optimization problem into the following form: ； in, It is the performance threshold, which is used to classify good and bad parameters; Indicates the objective function value Less than threshold In the case of The probability distribution of Indicates that given parameters When Less than threshold The probability distribution of Representation parameters The prior probability distribution of ; During the optimization process, the performance of each set of hyperparameters was evaluated through five-fold cross validation, and the average root mean square error, mean absolute error, and adjusted coefficient of determination were used as objective function values ​​to quantify the prediction error of the model and guide the optimization direction of the tree-structured Parzen estimator algorithm.

6. A remote tower air traffic controller visual fatigue monitoring method according to claim 1, characterized in that: In step 5, the specific process of real-time monitoring includes: Collect eye movement, ECG, workload and environmental data in real time; After preprocessing the data, input it into the trained visual fatigue prediction model; When the model output is in a visual fatigue state, an audible and visual alarm is triggered or task allocation is adjusted.

Images