Air traffic controller fatigue determination method based on random forest classification

Abstract

The application discloses an air controller fatigue determination method based on random forest classification and relates to the technical field of data processing. The method comprises the following steps: collecting screen parameters; constructing a DAG graph, inputting the screen parameters, and calculating causal Markov property and real causal effect; based on the real causal effect, performing Bayesian optimization on a random forest classification model; and determining the fatigue degree of the controller by using the random forest classification model. The method takes multi-dimensional screen parameters as intervention variables, can maintain high interpretability and anti-interference ability under the interference of confusing factors, has strong scene adaptability and robustness, and can realize real-time determination of the fatigue degree of the controller.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for determining air traffic controller fatigue based on random forest classification. Background Technology

[0002] Traditional methods for assessing air traffic controller fatigue mainly include subjective and objective methods. Subjective methods include questionnaires and the completion of subjective scales such as the Karolinska Sleepiness Scale (KSS) and the Stanford Sleepiness Scale (SSS). While simple and quick, these methods are highly subjective. Objective methods include the analysis of physiological signals such as electroencephalograms (EEG) and electrocardiograms (ECG), as well as the recognition of behavioral features such as facial expressions. While unaffected by subjective will, these methods have low error tolerance and are prone to errors. Commonly used decision-making algorithms include decision trees, support vector machines (SVMs), neural networks, and random forests. Decision tree models are simple in structure but prone to overfitting; SVMs perform well with small sample data but are not adaptable to high-dimensional features; neural networks require a large amount of labeled data, resulting in high training costs; random forests improve stability by integrating multiple decision trees and are suitable for multi-feature fusion scenarios, but their parameter selection significantly affects the decision-making performance.

[0003] The publicly known technologies involved include:

[0004] Random Forest model: An ensemble learning method that improves decision performance by constructing multiple independent decision trees and combining the decision results of all trees (majority voting for classification problems and mean for regression problems).

[0005] Fatigue level assessment: Fatigue level assessment is a state monitoring technology that evaluates an individual's fatigue level by analyzing and processing their physiological, behavioral, and environmental characteristics. Each person's fatigue state has unique physiological characteristics, and fatigue level assessment utilizes these characteristics to infer the state. This technology is widely used in air traffic control, transportation, and industrial production. The accuracy of fatigue level assessment is typically affected by factors such as environmental interference, individual differences, and data quality.

[0006] Decision models are tools that use mathematical or algorithmic frameworks to estimate future events, states, or values ​​by analyzing patterns and regularities in historical data. Their core principle is to infer unknown outcomes from known information, and they are widely used in fields such as financial forecasting, medical diagnosis, and environmental monitoring. Common decision models include statistical methods such as linear regression, logistic regression, and time series models, such as ARIMA (Auto-Regressive Moving Average Model); machine learning models such as decision trees, random forests, and support vector machines; and deep learning-based neural networks such as LSTM (Long Short Term Memory) and CNN (Convolutional Neural Network).

[0007] Existing methods for assessing fatigue among air traffic controllers largely rely on physiological and behavioral characteristic analysis techniques, which are widely used to ensure the safety of air traffic control work and monitor the health of personnel. This approach typically employs the following steps for fatigue assessment:

[0008] 1. Multi-source data acquisition and preprocessing: First, physiological or behavioral data of controllers during their work are collected, including eye movement features, facial expression features, and physiological signals. The collected data undergoes noise reduction, standardization, and time alignment to remove outliers and redundant information, ensuring that the data quality meets the requirements of subsequent analysis.

[0009] 2. Fatigue Feature Extraction: Extract features related to fatigue state from preprocessed multi-source data, integrate these features to form a multi-dimensional feature set that can reflect the fatigue level of controllers.

[0010] 3. Fatigue State Classification and Model Training: The extracted features are input into the classification model for training to achieve fatigue state classification and assessment. Commonly used classification algorithms include Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Random Forest. During training, sample data is divided into categories such as "normal state," "mild fatigue," "moderate fatigue," and "severe fatigue." The model is trained and its parameters optimized using a large number of labeled samples to improve the model's accuracy in recognizing different fatigue states.

[0011] 4. Fatigue status assessment and early warning: Input the real-time collected and processed feature data into the trained model to obtain the current fatigue status assessment results of the controller.

[0012] The existing technology has the following drawbacks:

[0013] 1. Insufficient scenario adaptability: Existing random forest applications have significant limitations in determining controller fatigue levels. Existing models do not fully incorporate unique influencing factors such as remote tower screen parameters, resulting in incomplete feature capture of controller fatigue states and poor adaptability in determination.

[0014] 2. Limited Feature Dimensions: Existing technologies for driver fatigue detection rely heavily on limited-dimensional data such as subjective feelings or facial expressions. However, controller fatigue is the result of multiple factors, including physiological and environmental factors. Existing models lack the ability to fuse multi-source heterogeneous data and fail to deeply explore the correlation between environmental factors and fatigue characteristics, thus limiting the accuracy of judgments.

[0015] 3. Lack of real-time response: Controller fatigue assessment needs to meet the real-time monitoring requirements of remote control towers, but existing random forest and improved methods suffer from computational efficiency bottlenecks. The extensive iterative computations of heuristic optimization methods prolong model training and assessment time, making it difficult to handle the frequently updated dynamic data on controller physiological states and screen interactions, resulting in delayed real-time warnings and failure to trigger timely intervention measures. Summary of the Invention

[0016] To address the aforementioned technical problems in the existing technology, this invention provides a method for determining air traffic controller fatigue based on random forest classification.

[0017] Specifically, the technical solution includes the following steps:

[0018] Step S1: Collect screen parameters;

[0019] Step S2: Construct a DAG graph, input the screen parameters, and calculate the causal Markov property and the true causal effect;

[0020] Step S3: Based on the true causal effect, perform Bayesian optimization on the random forest classification model;

[0021] Step S4: Use a random forest classification model to determine the fatigue level of the controller.

[0022] Compared with existing technologies, the technical solution provided by this invention uses multi-dimensional screen parameters as intervention variables, which can maintain high interpretability and anti-interference ability under the interference of confusing factors, has strong scene adaptability and robustness, and can realize real-time determination of controller fatigue level. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the process of determining the fatigue level of air traffic controllers in one embodiment of the present invention.

[0024] Figure 2 This is a flowchart of causal inference in one embodiment of the present invention.

[0025] Figure 3 This is a DAG graph structure in one embodiment of the present invention.

[0026] Figure 4 This is a flowchart illustrating the decision-making process for random forest classification in one embodiment of the present invention.

[0027] Figure 5 This is a flowchart of Bayesian optimization in one embodiment of the present invention.

[0028] Figure 6 This is the confusion matrix before Bayesian optimization in one embodiment of the present invention.

[0029] Figure 7 This is a confusion matrix optimized by Bayes in one embodiment of the present invention.

[0030] Figure 8 This is a flowchart of a graded early warning simulation in one embodiment of the present invention. Detailed Implementation

[0031] The technical solutions provided by the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0032] Example 1:

[0033] Accurate and real-time fatigue assessment is crucial for air traffic controllers working in remote control towers, ensuring air traffic safety. However, traditional assessment methods struggle to maintain stable accuracy in practice due to insufficient adaptability to complex interactive scenarios, highly subjective feature selection, and dynamic fluctuations in fatigue levels. This increases the risk of safety accidents caused by delayed or misjudged fatigue warnings. This embodiment addresses the limitations of traditional methods by proposing a random forest classification-based fatigue assessment method for air traffic controllers. It introduces causal inference to filter directly relevant features, utilizes Bayesian methods for efficient parameter optimization, and incorporates a tiered warning mechanism. This enhances the model's adaptability to complex screen interaction scenarios in remote control towers, thereby improving the accuracy, real-time performance, and reliability of assessments in complex situations.

[0034] The specific process of this embodiment is as follows: Figure 1 As shown, this is to address the challenge of real-time fatigue assessment of controllers in remote control scenarios under the influence of screen parameters, and to improve the accuracy, real-time performance, and scenario adaptability of fatigue assessment.

[0035] Step 1: Real-time collection of controllers' physiological data and screen parameter data, followed by cleaning and integration through a data preprocessing module to effectively reduce the impact of data noise and outliers on the quality of judgment. Step 2: Addressing the issue of strong subjectivity in traditional random forest feature selection, causal inference technology is introduced to uncover core fatigue-related features in the data. During fatigue level determination, an optimized random forest model is used for feature fusion and fatigue level assessment. Causal inference technology, by constructing a causal graph model, identifies the core factors affecting controller fatigue, filters directly relevant features, and improves the objectivity of feature selection. Step 3: Bayesian methods are used to efficiently optimize random forest parameters through prior knowledge modeling and posterior probability updates, reducing ineffective iterations. Simultaneously, its dynamic update capability allows the model to adaptively adjust parameters to match the dynamic changes in fatigue state. Finally, the determined fatigue level is compared with a set warning threshold, triggering corresponding tiered warnings based on the comparison results, thereby accurately obtaining the controller's current fatigue level.

[0036] Specifically, it includes the following steps:

[0037] Step 1: Screen parameters and controller eye movement data acquisition and preprocessing.

[0038] By simulating a real air traffic control environment, considering only the mapping relationship between screen parameters and controller fatigue state while strictly controlling other environmental variables, screen parameter data was collected. Eye trackers were used to collect controller pupil diameter and single-point annotation time data. The raw data underwent preliminary processing to remove obvious outliers and abnormal values, forming the original dataset. Then, the filtered screen light features were used as input features, and the corresponding fatigue level was used as the output label, forming a structured feature dataset.

[0039] Research by the Narcolepsy Research Center has found that pupils constrict when certain occupational groups are fatigued, while pupil diameter remains relatively stable when awake. As fatigue intensifies, the average time spent focusing on a single point also increases, sometimes leading to prolonged periods of staring blankly at a single point, indicating decreased attention span. However, due to physiological and mental factors, the physiological responses and characteristics of fatigue vary among individual controllers. Therefore, this embodiment, based on existing research on human physiological monitoring and experience summarized from the PubMed database, classifies fatigue levels according to pupil diameter and single-point annotation time in eye-tracking data.

[0040] Let the pupil diameter be The time spent staring at a single point is The estimated level of fatigue is :

[0041] ;

[0042] ;

[0043] ;

[0044] In the formula, For sampling point sequence The serial number in The number of sampling points. , For the first Pupil diameter at each sampling point For the first Single-point annotation time for each sampling point; For the first The estimated fatigue level at each sampling point is categorized into mild fatigue, moderate fatigue, and severe fatigue.

[0045] like ,and Then determine the first Predicted fatigue level at each sampling point Mild fatigue;

[0046] like ,and Then determine the first Predicted fatigue level at each sampling point Moderate fatigue;

[0047] like ,or Then determine the first Predicted fatigue level at each sampling point It is severe fatigue.

[0048] Step 2: Input causal features and construct a causal graph model.

[0049] Remote control towers are a typical environment with complex human-computer interaction. Traditional random forests learn the association between features and labels through correlations in the data, but they can be interfered with by spurious associations introduced by confounding variables, such as controller age and working hours, leading to incorrect judgments by the model. Therefore, this embodiment introduces causal inference. By identifying causal paths between variables and adjusting the formula through a backdoor to remove confounding biases, Bayesian optimization focuses more on the direct causal effects of features, dynamically adjusts feature weights, avoids the model learning spurious associations, and thus improves the "causal authenticity" of the judgments.

[0050] Specifically, the causal inference process is as follows: Figure 2 As shown:

[0051] 1. Structure and joint probability distribution;

[0052] Let the set of variables be The screen parameters include brightness. Blue light ratio and refresh rate Other variable parameters include the controller's age. and cumulative continuous working hours , It indicates the level of fatigue of the controller, and also includes three levels: mild fatigue, moderate fatigue and severe fatigue.

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] ;

[0059] In the formula, For the first The brightness of each sampling point For the first The proportion of blue light at each sampling point For the first Refresh rate of each sampling point For the first The age of controllers at each sampling point is generally... , For the first The cumulative continuous working time of each sampling point For the first The fatigue level of each sampling point.

[0060] Construct a DAG (Directed Acyclic Graph), the structure of which is as follows: Figure 3 As shown. In the DAG diagram structure, the brightness is... Blue light ratio and refresh rate As an intervention variable, the controller's age and cumulative continuous working hours As a confounding variable, fatigue level As an outcome variable, the intervention variable directly affects the outcome variable, forming a true causal path; controller age. and cumulative continuous working hours It directly affects fatigue, and is both a natural influencing factor of the outcome variable and may have a potential association with the intervention variable, thus jointly affecting the outcome variable; therefore, it is a confounding variable.

[0061] The causal Markov property representing the joint probability of the intervention variable, confounding variable, and outcome variable is calculated using the following formula:

[0062] ;

[0063] In the formula, It represents the prior probability, indicating the likelihood of an event occurring given the corresponding variable; When the intervention variable is known, the degree of fatigue The conditional probability of occurrence reflects the causal logic that "the probability of fatigue level is directly driven by brightness, blue light ratio, and refresh rate," that is, given the brightness... Blue light ratio and refresh rate fatigue level The conditional probability of occurrence.

[0064] 2. Causal intervention;

[0065] Of the variables mentioned above, controller age is the most important. and cumulative continuous working hours In addition to the screen parameters considered in this embodiment, other confounding variables also contribute to the controller's fatigue level. It is necessary to ensure the performance of multiple factors (brightness, etc.). Blue light ratio and refresh rate Coordinated intervention is needed, but it also requires intervention to address confounding variables (controller age). and cumulative continuous working hours The distribution of () is analyzed, and non-causal relationships are removed from the observation data to accurately calculate the true causal effect of "screen parameters → fatigue level". The formula is as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] In the formula, This indicates the true causal effect of the intervention variables combined. This indicates expectations, i.e., the level of labor. The average value under a specific intervention, Indicates causal intervention, For the first The brightness of each sampling point For the first The proportion of blue light at each sampling point For the first Refresh rate of each sampling point; For the first The brightness of each sampling point For the first The proportion of blue light at each sampling point For the first Refresh rate of each sampling point; express , , , , At that time, the fatigue level was The conditional probability, This represents the pre-defined probability distribution of the confounding variables, which is used as a weight to isolate the interference of age and working hours on the results.

[0070] above, no. The variable at the nth sampling point represents the variable after disturbance, and the nth sampling point represents the variable after disturbance. The variables at each sampling point represent the variables before the intervention, and are used to compare the difference in effect "before intervention - after intervention". For a single sampling point, there are multiple true causal effects, the specific number of which is related to its sequence number. For example, for the second sampling point, there is only one corresponding true causal effect, that is, in Take 2, The result is obtained when the value is 1. It is easy to deduce that for the ... The number of sampling points corresponds to the following true causal effects: One. Clearly, when calculating the true causal effect, The value cannot be 1.

[0071] It is evident that the calculation is expected to be adjusted through the backdoor path. Represents the degree of fatigue Controller age and cumulative continuous working hours The summation of all possible values ​​covers all scenario combinations. The formula for true causal effects ensures that the results are independent of the controller's age. and cumulative continuous working hours To minimize interference, ensure that only brightness is calculated. Blue light ratio and refresh rate Regarding the degree of fatigue The net causal effect.

[0072] While eliminating the interference of confounding variables, the interaction effect is used. Further verification shows that the intervention variables jointly affect the outcome variables, as shown in the following formula:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] In the formula, To calculate the net causal effect of luminance L, The net causal effect of the blue light ratio B. The net causal effect of refresh rate R; express , , At that time, the fatigue level was The conditional probability, express , , At that time, the fatigue level was The conditional probability, express , , At that time, the fatigue level was The conditional probability;

[0081] like This indicates brightness. Blue light ratio and refresh rate A synergistic effect exists, demonstrating the complex relationship between screen parameters and their interaction influencing controller fatigue levels. Otherwise, no synergistic effect exists.

[0082] This embodiment introduces causal inference, which can eliminate spurious correlation features and avoid being misled by spurious correlation features when using Bayesian methods to select hyperparameters. This reduces ineffective optimization of spurious correlation features and significantly improves the targeting and efficiency of the optimization. The backdoor path intervention also provides counterfactual judgment capabilities for Bayesian optimization.

[0083] Step 3: Construct the initial random forest classification model.

[0084] The feature vector formed by concatenating the intervention variable and confounding variables. As input, fatigue level As the output, the classification result is used. The preprocessed dataset is divided into training and test sets in a 7:3 ratio to ensure that the two sets of data follow the same distribution. The parameter range of the random forest is set, and an initial random forest classification model is built.

[0085] The specific process is as follows: Figure 4 As shown:

[0086] When splitting at each node of the decision tree, a subset of features is randomly selected from all features to find the optimal split threshold, reducing the dependence of a single tree on specific features. While a single decision tree achieves classification by recursively splitting nodes, a random forest performs classification by combining multiple decision trees. Each decision tree is built based on a different subset of samples and features. The final classification result is determined by a vote from all decision trees; that is, the classification results of all decision trees are statistically analyzed, and the category that appears most frequently is taken as the final decision of the random forest.

[0087] pass The decision trees determine the classification results, thus obtaining the final decision result of the random forest model, i.e., the first decision tree. fatigue level at each sampling point The formula is as follows:

[0088] ;

[0089] In the formula, An index for a single decision tree. Fatigue levels are categorized into three types: "mild fatigue," "moderate fatigue," and "severe fatigue." For the first The classification labels determined by each decision tree; For indicator functions, if ,but ,otherwise, .

[0090] Step 4: Bayesian optimization.

[0091] To improve the accuracy of controller fatigue assessment, a Bayesian optimization algorithm was used to optimize the random forest model. Utilizing the quantified results after causal inference, feature weights were dynamically adjusted during hyperparameter search. For features with high causal importance, a more refined splitting strategy was applied; for features with low causal importance, their role in the tree was simplified. This collaborative strategy allows the optimized random forest to focus on key causal features while maximizing model performance through hyperparameter adaptation, thereby improving the accuracy of controller fatigue assessment.

[0092] The specific process is as follows: Figure 5 As shown:

[0093] (1) Constructing the proxy model:

[0094] Using a Gaussian process as a surrogate model for Bayesian optimization, the goal of the Gaussian process is to obtain the posterior distribution based on the observed sample points through Bayesian updates, thereby making a probability determination on the function values ​​of unobserved points and providing a basis for selecting the next evaluation point for subsequent function acquisition.

[0095] For the objective function A Gaussian process can be represented as:

[0096] ;

[0097] ;

[0098] ;

[0099] In the formula, Indicates including the first The combination of intervention and confounding variables for each sampling point. Let be the objective function. Represents a Gaussian process. For kernel function, The mean function is used. The mean is initialized using the causal analysis results from step 2, which improves the rationality of the prior judgment.

[0100] (2) Construct the acquisition function:

[0101] The acquisition function is used to select the next point to be evaluated based on the posterior distribution of the Gaussian process. Using Expected Improvement (EI) as the acquisition function, let the current optimal value be... The expected improvement is defined as:

[0102] ;

[0103] ;

[0104] In the formula, This indicates a desire for improvement. Acquisition function The optimal value is the optimal value among all true causal effects corresponding to the i-th sampling point. This is the optimal input point observed so far.

[0105] In Bayesian optimization, the next point to be evaluated is selected by maximizing the acquisition function, i.e., for features with high causal importance (brightness). Blue light ratio and refresh rate ), adapting more refined splitting strategies to Bayesian tree models; for features with low causal importance (controller age) and cumulative continuous working hours This simplifies the role of the tree and enables dynamic adjustment of feature weights driven by causality, as shown in the following formula:

[0106] ;

[0107] In the formula, For parameter combinations including intervention variables and hyperparameters, This indicates a search for the optimal combination of parameters;

[0108] The hyperparameters include the number of decision trees (n_estimators). Maximum tree depth (max_depth) Minimum number of split samples (min_samples_split) and the minimum number of leaf node samples (min_samples_leaf) .

[0109] Bayesian optimization can minimize the objective function by adjusting the combination of parameters. This enhances the robustness of random forests to noise. Compared with unoptimized or traditionally optimized models, it has stronger anti-interference capabilities and can maintain stable decision performance in complex environments. Furthermore, its rapid iterative optimization meets real-time requirements, providing accurate, efficient, and reliable technical support for ensuring regulatory safety.

[0110] Step 5: Accuracy assessment.

[0111] Using accuracy A cc The performance of the model is evaluated by the proportion of correctly judged sampling points out of the total number of sampling points, as shown in the following formula:

[0112] ;

[0113] In the formula, For the first The true label corresponding to the sampling point, i.e. the first sampling point. The estimated fatigue level corresponding to each sampling point ; For the first The judgment label corresponding to the sampling point, i.e. the first sampling point Fatigue level corresponding to each sampling point .

[0114] The accuracy of the random forest model before and after optimization was evaluated. The causal Bayesian network-optimized random forest model outperformed the unoptimized model in all evaluation metrics, with an accuracy improvement of 7.25%. This indicates that the causal Bayesian network significantly optimizes the random forest parameters and effectively improves the model's performance in determining the fatigue state of remote tower controllers. The confusion matrix before optimization is shown below. Figure 6 As shown, the optimized confusion matrix is ​​as follows: Figure 7 As shown.

[0115] It is evident that combining causal inference with Bayesian optimization to optimize the random forest model can integrate the advantages of all three in "causal relationship mining," "uncertainty quantification," and "complex pattern learning," forming a more robust, interpretable, and realistic judgment or decision-making model. At the end of the model, an early warning simulation is set up, outputting the fatigue level of the judgment through input screen parameters, thereby providing a fatigue-level warning. The process is as follows: Figure 8 As shown.

[0116] As can be seen from the above embodiments and figures, the present invention, by integrating causal inference, Bayesian methods and random forests, can effectively solve the limitations of traditional fatigue determination methods in determining the fatigue level of controllers.

[0117] First, while random forests excel at capturing correlations between variables, they cannot distinguish between causal relationships and spurious associations. Bayesian optimization, though effective at optimizing hyperparameters, still relies on optimizing the model based on data associations. Introducing causal inference, however, allows for the initial screening of features with genuine causal impact on the target variable. Then, Bayesian optimization adjusts the hyperparameters of the random forest, focusing the model on causal patterns rather than data noise. Addressing the inherent subjectivity of traditional feature selection methods, causal inference clarifies the impact paths of multi-dimensional factors, enhancing the objectivity of feature selection and laying the foundation for accurate judgment.

[0118] Second, while random forests are relatively fast to train, excessively long hyperparameter optimization times can affect real-time performance. Bayesian optimization, based on initial screening through causal inference and then further refined screening through iterative intelligent search, can reduce the number of iterations for hyperparameter optimization from hundreds in grid search to dozens, significantly shortening optimization time and improving the real-time performance of controller fatigue assessment.

[0119] Third: The tiered early warning module simulates an alarm mechanism in a remote tower scenario, avoiding either an overreaction or inadequate response in a "one-size-fits-all" manner. Clearly defined tiers allow decision-makers to quickly identify the urgency of risks, validating the real-time performance and reliability of the technology.

[0120] In summary, compared with the prior art, the technical solution provided by this invention uses multi-dimensional screen parameters as intervention variables, which can maintain high interpretability and anti-interference ability under the interference of confusing factors, has strong scene adaptability and robustness, and can realize real-time determination of controller fatigue level.

[0121] Furthermore, Bayes' dynamic update capability and stable causal inference relationship enable adaptive parameter adjustment, which can cope with the dynamic fluctuations of fatigue, thereby enhancing noise resistance and robustness; the hierarchical early warning mechanism can achieve more accurate, real-time and stable controller fatigue judgment, providing strong protection for aviation safety.

Claims

1. A method for determining air traffic controller fatigue based on random forest classification, characterized in that, Includes the following steps: Step S1: Collect screen parameters; Step S2: Construct a DAG graph, input the screen parameters, and calculate the causal Markov property and the true causal effect; Step S3: Based on the true causal effect, perform Bayesian optimization on the random forest classification model; Step S4: Use a random forest classification model to determine the controller's fatigue level; The screen parameters include brightness. Blue light ratio and refresh rate ; ； ； ； In the formula, For sampling point sequence The serial number in The number of sampling points. , For the first The brightness of each sampling point For the first The proportion of blue light at each sampling point For the first Refresh rate of each sampling point; Step S2 specifically includes: Step S21: Define the causal problem and construct the DAG graph; Among them, brightness Blue light ratio and refresh rate As an intervention variable, the controller's age and cumulative continuous working hours To obfuscate variables, fatigue level For the outcome variable; Construct a set of variables ; ； ； ； In the formula, For the first Age of controllers at each sampling point For the first The cumulative continuous working time of each sampling point; For the first The fatigue level at each sampling point; where fatigue level includes mild fatigue, moderate fatigue and severe fatigue; Construct S22: Calculate the causal Markov property using the following formula: ； In the formula, Represents the prior probability. When the intervention variable is known, the degree of fatigue The conditional probability of occurrence; Step S23: Calculate the true causal effect of the intervention variables combined. The formula is as follows: ； ； ； In the formula, Expressing expectations, Indicates causal intervention, For the first The brightness of each sampling point For the first The proportion of blue light at each sampling point For the first Refresh rate of each sampling point; For the first The brightness of each sampling point For the first The proportion of blue light at each sampling point For the first Refresh rate of each sampling point; express , , , , At that time, the fatigue level was The conditional probability, This represents the predefined distribution probability of the confounding variable.

2. The method for determining air traffic controller fatigue based on random forest classification as described in claim 1, characterized in that, The random forest classification model includes A decision tree, taking feature vectors as input and fatigue level as input. For output; Among them, the The eigenvector is derived from the eigenvector of the th ? Brightness of each sampling point Blue light ratio Refresh rate ,age and cumulative continuous working hours It is constructed by splicing.

3. The method for determining air traffic controller fatigue based on random forest classification as described in claim 2, characterized in that, Step S3 specifically includes: Step S31: Perform Bayesian optimization on the random forest classification model; Step S311: Construct the proxy model, with the following formula: ； ； ； In the formula, Indicates including the first The combination of intervention and confounding variables for each sampling point. Let be the objective function. Represents a Gaussian process. For kernel function, It is a mean function; Step S312: Construct the acquisition function, the formula is as follows: ； ； In the formula, This indicates a desire for improvement. Acquisition function The optimal value; To maximize the acquisition function, select the next set of variable combinations, as shown in the following formula: ； In the formula, For parameter combinations including intervention variables and hyperparameters, This indicates a search for the optimal combination of parameters; Among them, the hyperparameters include the number of decision trees in the random forest classification model. Maximum tree depth Minimum number of split samples and minimum number of leaf node samples ; Step S313: Minimize the objective function by adjusting the combination of hyperparameters. Complete Bayesian optimization.

4. A method for determining air traffic controller fatigue based on random forest classification as described in claim 2 or 3, characterized in that, Step S4 includes: Input the feature vectors respectively Calculate the fatigue level by using the output of each decision tree. The formula is as follows: ； In the formula, For indexes of a single decision tree, According to the fatigue level category, For the first The output of each decision tree is the classification label determined. For indicator functions, if ,but ,otherwise, .

Images