Aircraft special situation training quality evaluation model training set design method for flight simulator

Abstract

The invention discloses an aircraft special situation training quality evaluation model training set design method for a flight simulator, and the method comprises the steps: 1, data collection, 2, data preprocessing, 3, calculation of an SHAP value, 4, feature weight calculation and weighted adjustment, 5, retraining of a model, 6, determination of aircraft special situation training feature parameters, and 7, training of an aircraft special situation training quality evaluation model. 7, establishing an evaluation model training set; according to the method, the training data of the flight simulator are collected and preprocessed, and the SHAP value is combined with the weighted random forest model to select the characteristic parameters, so that the scientific and effective training quality evaluation model training set is constructed, the data processing efficiency and accuracy can be improved, and the training quality evaluation accuracy is improved. And objectivity and reliability of special situation training quality evaluation can be remarkably improved, and a solid guarantee is provided for pilot training and flight safety.

Description

Technical Field

[0001] The invention relates to the technical field of aviation training and flight simulation, and in particular to a method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator. Background Art

[0002] With the rapid development of aviation technology, the training needs of pilots are becoming increasingly complex and diversified. Especially when dealing with aircraft special situations (special circumstances), the pilot's adaptability and control skills are crucial to ensuring flight safety; however, traditional training methods often find it difficult to fully cover various special situations, and cannot effectively quantify the training effect of pilots; therefore, relying on high-fidelity flight simulators to conduct pilot special situation training and conduct scientific and objective evaluation of its training quality has become one of the important research directions in the current aviation field.

[0003] As an advanced training tool, high-fidelity flight simulators can realistically simulate various flight conditions and special situations, providing pilots with a very realistic training environment. However, in special situation training, how to effectively extract characteristic parameters related to training quality from a large amount of flight data and establish an accurate and reliable training quality assessment model is still a difficult problem that needs to be solved urgently.

[0004] Current research usually relies on manual experience or simple statistical methods to evaluate the quality of special situation training, which has problems such as inconsistent evaluation standards and strong subjectivity, and is difficult to meet the high standards of modern aviation training. Therefore, the present invention proposes a training set design method for an aircraft special situation training quality evaluation model of a flight simulator to solve the problems existing in the prior art. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to propose a method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator collects and preprocesses flight simulator training data, and uses a SHAP value combined with a weighted random forest model to select feature parameters, thereby constructing a scientific and effective training quality assessment model training set, which can not only improve the efficiency and accuracy of data processing, but also significantly improve the objectivity and reliability of special situation training quality assessment, thereby providing a solid guarantee for pilot training and flight safety.

[0006] To achieve the purpose of the present invention, the present invention is implemented by the following technical scheme: a method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator, comprising the following steps:

[0007] Step 1: Data collection: The flight simulator training review system collects the training data of the flight trainees including aircraft status data, pilot control data and fault injection data when the aircraft encounters special conditions and when the aircraft is normal;

[0008] Step 2: Data preprocessing: first replace the infinite values ​​and maximum values ​​in the collected data with NAN and fill them with the average value of the column. Then use the Lagrange interpolation method to complete the missing data. Finally, combine the aircraft status data and pilot control data into one piece of data and add a label column.

[0009] Step 3: Calculate the SHAP value. Define a basic random forest model through "clf = RandomForestClassifier (random_state = 42)", and calculate the SHAP value after preliminary training through the Tree SHAP method.

[0010] Step 4: Calculate and adjust feature weights. Evaluate feature importance based on SHAP values, and complete feature weight calculation and weight adjustment.

[0011] Step 5: Retrain the model and calculate the SHAP value of each feature to get the corresponding weight w i , and apply the weights to the feature matrix X to generate the weighted feature matrix X weighted , through the weighted feature matrix X weighted Retrain the random forest model with the features in, and use the grid search method to find the optimal model parameters to obtain the optimized weighted random forest model;

[0012] Step 6: Determine the characteristic parameters of aircraft special situation training. For a certain aircraft special situation training subject, select n groups of training data and input them into the optimized weighted random forest model to select characteristic parameters, check the model output results, and determine the characteristic parameters of aircraft special situation training;

[0013] Step 7: Constructing the evaluation model training set. Based on the determined aircraft special situation training characteristic parameters, the corresponding characteristic parameters are extracted from the original data and combined into new data, and the aircraft special situation training quality evaluation model training set is further constructed.

[0014] A further improvement is that the Lagrange interpolation method in step 2 is used to supplement the missing data by the following formula:

[0015]

[0016] Where x represents the position of the independent variable, y represents the value of the corresponding function, and l j (x) represents the Lagrange polynomial.

[0017] A further improvement is that the label column adding method in step 2 is that when the aircraft has a special situation, the aircraft status data during the flight trainee training is labeled as 1, and when the aircraft is normal, the aircraft status data during the flight trainee training is labeled as 0; when the aircraft has a special situation, the flight trainee's control data is labeled as 1, and when the aircraft is normal, the flight trainee's control data is labeled as 0.

[0018] Further improvements are: the random forest model defined in step 3 includes the number of decision trees n_estimators, the maximum number of features max_features considered in each split, the maximum depth of the decision tree max_depth and the splitting criterion criterion; after calculating the SHAP value, the SHAP contribution value of each feature is obtained through "shap.TreeExplainer".

[0019] A further improvement is that the SHAP value calculation process in step 3 is as follows:

[0020] The goal of S1 and SHAP values ​​is to calculate the marginal contribution of each feature to the model prediction value, which is calculated by the following formula:

[0021]

[0022] Among them, θ i represents the SHAP value of feature i, N is the set of all features, S is the feature subset, and f(S) is the predicted output of the model under the feature subset S;

[0023] S2. For the decision tree model in the random forest model, use Tree SHAP to calculate the marginal contribution of the feature at each node of the model according to the segmentation condition of the feature on the path;

[0024] S3. For each path in each decision tree model, Tree SHAP uses a recursive algorithm to calculate the SHAP value of each feature. For each node, assuming that the probability of the decision path from the root node to the leaf node is p(z), and the contribution of each feature to the predicted value is v(z), Tree SHAP uses the following formula to recursively assign SHAP values:

[0025]

[0026] Among them, θ i (x) represents the SHAP value of feature i in sample x, p ij is the probability that sample x passes through feature i in path j, |S| is the size of the feature subset that has passed through the path, and f(z∪{i}) is the predicted value of feature i after adding feature subset z;

[0027] S4. For each path, from the root node to the leaf node, Tree SHAP needs to calculate the contribution of each feature to the leaf node output. The tree's predicted value f(x) is the value of the leaf node. The contribution of each path is calculated by the following formula:

[0028]

[0029] Where T represents the number of trees, w(z) is the weight of feature subset z, which represents the probability of the subset appearing on the path, and f(z∪{i})-f(z) is the incremental contribution after feature i is added;

[0030] S5. In the entire random forest model, Tree SHAP accumulates the SHAP values ​​of each decision tree to get the final SHAP value by the following formula:

[0031]

[0032] in, is the SHAP value of feature i in the tth decision tree.

[0033] A further improvement is that: the step 4 is specifically

[0034] The first step is to perform absolute value processing on the SHAP value calculated in step 3, using |θ i (x| represents the absolute value of the contribution of feature i in sample x;

[0035] Step 2: Calculate the absolute average of all sample SHAP values ​​using the following formula

[0036]

[0037] Where n is the number of samples in the dataset, |θ i (x j )| is the absolute value of the SHAP value of the jth sample on feature i;

[0038] The third step is to normalize the average SHAP value of each feature, which is expressed as follows:

[0039]

[0040] Among them, w i is the weight of feature i, is the average SHAP absolute value of feature i, d is the total number of features, and the sum of all feature weights after normalization is equal to 1;

[0041] Step 4: First assume a feature matrix X with a shape of n×d, where n is the number of samples and d is the number of features. Each column of this matrix represents a feature and each row represents a sample. X = [x1, x2, ..., xd].i Represents the value of the i-th feature. For each feature x i , using the calculated weight w i After weighted adjustment, the weighted feature matrix is ​​X weighted =X×W, where W is a diagonal matrix, W=diag(w1, w2, ..., w d ), weighted feature X weighted It is obtained by multiplying each feature column by its corresponding weight w i The obtained and weighted features are represented as X weighted [:, i] = X[:, i] × w i .

[0042] A further improvement is that in step six, the characteristic parameters of aircraft special situation training are determined specifically when a certain parameter appears in the output result greater than or equal to n / 2 times, then the parameter is determined as a characteristic parameter related to this aircraft special situation training subject.

[0043] The beneficial effects of the present invention are as follows: the present invention greatly improves the quality and processing efficiency of data by preprocessing the data collected by the flight simulator, making the training data more accurate and reliable, reducing the influence of noise and data deviation, and using the SHAP value combined with the weighted random forest model method to screen out key feature parameters related to special situation training subjects, making the feature selection process more objective and accurate;

[0044] At the same time, this method can be flexibly adjusted according to different aircraft special situation training subjects. By inputting and screening multiple sets of data, it ensures that the selected characteristic parameters are highly relevant to the specific training subjects, thereby improving the applicability and universality of the evaluation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a method for selecting characteristic parameters for aircraft special situation training according to Embodiment 1 of the present invention.

[0046] Figure 2 This is a diagram of aircraft single engine failure training data after preprocessing in Example 2 of the present invention.

[0047] Figure 3 This is a schematic diagram of SHAP values ​​of some parameters of single-shot failure training data in Example 2 of the present invention.

[0048] Figure 4 This is a diagram of some parameter weights of single-shot failure training data in Example 2 of the present invention.

[0049] Figure 5 This is a diagram showing the output results of the single-engine failure training subject model according to Example 2 of the present invention.

[0050] Figure 6This is a diagram of the characteristic parameters of aircraft single engine failure training selected for Example 2 of the present invention.

[0051] Figure 7 This is a diagram of a training set file of an aircraft single engine failure training quality assessment model that has been built in Example 2 of the present invention. DETAILED DESCRIPTION

[0052] In order to deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with examples. The examples are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0053] Example 1

[0054] according to Figure 1 As shown, this embodiment provides a method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator, comprising the following steps:

[0055] Step 1: When the flight trainees use the high-fidelity flight simulator for flight training, the various systems and subsystems of the aircraft will generate a large amount of diverse data, including but not limited to flight status data, environmental parameter data, pilot operation data, fault injection data, etc. The flight simulator training review system (BDS) collects the training data of the flight trainees when the aircraft is in a special situation and when the aircraft is normal, including aircraft status data, pilot operation data and fault injection data;

[0056] Step 2: Preprocess the collected aircraft status data, pilot control data and fault injection data. These data are a data sequence arranged in chronological order, also known as a time series. Due to the inherent reasons of the flight simulator training review system (BDS), the collected aircraft system data, pilot control data and fault injection data will inevitably have large deviations or garbled characters and data loss problems. Abnormal data needs to be deleted or corrected.

[0057] Replace the infinite values ​​and maximum values ​​in the collected data with NAN and fill them with the average value of the column; for the missing data, use the Lagrange interpolation method to complete the missing data. The Lagrange interpolation method is calculated by the following formula

[0058]

[0059] Where x represents the position of the independent variable, y represents the value of the corresponding function, and l j (x) represents the Lagrange polynomial;

[0060] The processed aircraft status data when the aircraft has a special situation and the aircraft is normal during flight student training are combined into one piece of data, and a label column is added to the combined data, wherein the aircraft status data when the aircraft has a special situation and the flight student training is labeled as 1, and the aircraft status data when the aircraft is normal is labeled as 0; similarly, the control data when the aircraft has a special situation and the flight student training is also collectively referred to as one piece of data, and a label column is added to the combined data, wherein the flight student control data when the aircraft has a special situation is labeled as 1, and the flight student control data when the aircraft is normal is labeled as 0.

[0061] Step 3: Define a basic random forest model through "clf = RandomForestClassifier (random_state = 42)", including the number of decision trees n_estimators, the maximum number of features considered in each classification max_features, the maximum depth of the decision tree max_depth and the splitting standard criterion, and calculate the SHAP value after preliminary training through the Tree SHAP method, and obtain the SHAP contribution value of each feature through "shap.TreeExplainer";

[0062] Tree SHAP is a specialization of the SHAP (SHapleyAdditive exPlanations) method. It is used to accelerate the calculation of SHAP values ​​for tree models (such as random forests, gradient boosted trees, etc.). Tree SHAP uses the structure of decision trees to avoid traversing all feature subsets, thereby greatly speeding up the calculation. Its computational complexity is linear and depends on the depth of the tree.

[0063] The SHAP value calculation process is as follows:

[0064] S1. Definition of SHAP value

[0065] The SHAP value is based on the concept of Shapley value. The goal is to calculate the marginal contribution of each feature to the model prediction value. The Shapley value is based on the average marginal contribution of the feature combination. The SHAP value calculation formula is as follows

[0066]

[0067] Among them, θ i represents the SHAP value of feature i, N is the set of all features, S is the feature subset, and f(S) is the predicted output of the model under the feature subset S;

[0068] S2. Decision tree structure and path

[0069] In the decision tree model, Tree SHAP uses the tree structure to optimize the calculation process of Shapley value. Each tree has multiple paths, and samples are predicted along the tree from the root node to the leaf node. For each node in the tree, the feature value determines which branch the sample goes down. The idea of ​​Tree SHAP is to calculate the marginal contribution of the feature at each node according to the segmentation conditions of the features on the path, without traversing all feature subsets.

[0070] S3. Allocation of path-dependent weights

[0071] For each path in each decision tree model, Tree SHAP calculates the SHAP value of each feature through a recursive algorithm. For each node, assuming that the probability of the decision path from the root node to the leaf node is p(z), and the contribution of each feature to the predicted value is v(z), Tree SHAP uses the following formula to recursively assign SHAP values

[0072]

[0073] Among them, θ i (x) represents the SHAP value of feature i in sample x, p ij is the probability that sample x passes through feature i in path j, |S| is the size of the feature subset that has passed through the path, and f(z∪{i}) is the predicted value of feature i after adding feature subset z;

[0074] S4. Accumulation of expectations and contributions

[0075] For each path, from the root node to the leaf node, Tree SHAP needs to calculate the contribution of each feature to the leaf node output. The tree prediction value f(x) is the value of the leaf node. The contribution of each path is calculated by the following formula

[0076]

[0077] Where T represents the number of trees, w(z) is the weight of feature subset z, which represents the probability of the subset appearing on the path, and f(z∪{i})-f(z) is the incremental contribution after feature i is added;

[0078] S5. Final accumulation of SHAP values

[0079] In the entire random forest model, Tree SHAP accumulates the SHAP values ​​of each decision tree to get the final SHAP value by the following formula:

[0080]

[0081] in, is the SHAP value of feature i in the tth decision tree, and the final SHAP value is the sum of the contributions of all trees.

[0082] Step 4: Feature weight calculation and weighting process is mainly based on the evaluation of feature importance based on SHAP value, and the feature importance is evaluated based on SHAP value to complete the feature weight calculation value and weighting adjustment;

[0083] The specific process is:

[0084] Step 1: Based on the SHAP value calculated in the preliminary model training in step 3, since SHAP values ​​can be positive or negative (which reflects the positive or negative contribution of the feature to the result); in order to calculate the importance weight of the feature, the absolute value of the SHAP value |θ is taken here i (x)| represents the absolute value of the contribution of feature i in sample x; the absolute value of the SHAP value calculated for each sample is used to measure the relative influence of the feature, regardless of whether it is positive or negative.

[0085] In the second step, in order to calculate the average contribution of each feature to the entire dataset, the absolute average of the SHAP values ​​of all samples is calculated by the following formula

[0086]

[0087] Where n is the number of samples in the dataset, |θ i (x j )| is the absolute value of the SHAP value of the jth sample on feature i, calculated by The average contribution of feature i can be obtained, indicating its influence on the entire data set;

[0088] Step 3: In order to convert the importance of features into weights, the average SHAP value of each feature is normalized, which is expressed as follows:

[0089]

[0090] Among them, w i is the weight of feature i, is the average SHAP absolute value of feature i, d is the total number of features, and after normalization, the sum of all feature weights is equal to 1, which makes the weights comparable and facilitates the subsequent feature matrix weighting processing;

[0091] Step 4: First assume a feature matrix X with a shape of n×d, where n is the number of samples and d is the number of features. Each column of this matrix represents a feature and each row represents a sample. X = [x1, x2, ..., xd]. i Represents the value of the i-th feature. For each feature x i , using the calculated weight wi After weighted adjustment, the weighted feature matrix is ​​X weighted =X×W, where W is a diagonal matrix, W=diag(w1, w2, ..., w d ), weighted feature X weighted It is obtained by multiplying each feature column by its corresponding weight w i The obtained and weighted features are represented as X weighted [:, i] = X[:, i] × w i .

[0092] After the above four steps, the value of each feature in the feature matrix will be adjusted according to the importance of its SHAP value; features with higher weights will have a greater impact on the model, reflecting its importance in the model, while features with lower weights will have less impact on the model.

[0093] Step 5: Retrain the model and calculate the SHAP value of each feature to get the corresponding weight w i , and apply the weights to the feature matrix X to generate the weighted feature matrix X weighted , for each sample x, the weighted eigenvalue X weighted [j, i] = X[j, i] × w i ;

[0094] Use the weighted feature matrix X weighted To replace the original feature matrix X, retrain the random forest model with the weighted features, and use the grid search method to find the optimal model parameters to obtain the optimized weighted random forest model; by subjectively setting a threshold (such as the top n importance rankings or the importance exceeding a certain value), select parameters above the threshold as the output result of the random forest model.

[0095] Step 6. For a certain aircraft special situation training subject (such as aircraft single engine failure training), the feature parameters selected by using the SHAP value combined with the weighted random forest model through a set of training data are not accurate enough. For a certain aircraft special situation training subject, select n sets of training data and use the SHAP value combined with the weighted random forest model to select feature parameters, check the model output results, and when the number of times a certain parameter appears in the output result is greater than or equal to n / 2, the parameter is determined as a feature parameter related to this aircraft special situation training subject; for a certain aircraft special situation training subject, the total number of feature parameters selected related to the aircraft special situation training can be determined subjectively.

[0096] Step 7: Based on the determined aircraft special situation training characteristic parameters, the corresponding characteristic parameters are extracted from the original data and combined into new data, and the aircraft special situation training quality assessment model training set is further constructed.

[0097] Example 2

[0098] according to Figure 2-Figure 7 As shown, this embodiment takes the aircraft single-engine failure training of flight trainees as an example to introduce in detail the process of building a training set of the aircraft single-engine failure training quality assessment model.

[0099] 1. Single-shot failure training data collection and preprocessing

[0100] Firstly, the training data of flight trainees during single-engine failure training and training data of flight trainees during normal aircraft conditions are collected through the flight simulator training review system (BDS). The collected training data include aircraft status data, pilot control data and fault injection data.

[0101] For infinite values ​​and maximum values ​​in the data, they are first replaced with NAN and then filled with the average value of the column; in the case of missing data, Lagrange interpolation is used to complete the data.

[0102] The processed aircraft status data during single-engine failure training and the aircraft status data during normal training are combined into one piece of data, and a label column is added to the combined data, wherein the aircraft status data during flight student training when a single-engine failure occurs is labeled 1, and the aircraft status data during flight student training when the aircraft is normal is labeled 0. Similarly, the control data during flight student training when a single-engine failure occurs and when the aircraft is normal are also combined into one piece of data, and a label column is added to the combined data, wherein the control data of the flight student when a special situation occurs is labeled 1, and the control data of the flight student when the aircraft is normal is labeled 0. The data after preprocessing is as shown in the attached manual Figure 2 shown.

[0103] 2. Selection of characteristic parameters for aircraft single engine failure training

[0104] First, define a basic random forest model through "clf = RandomForestClassifier (random_state = 42)", set "n_estimators = 100" (the number of decision trees), "max_features = sqrt" (the maximum number of features considered at each split), "max_depth = None" (the maximum depth of the decision tree) and "criterion = squard-error" (the splitting criterion), then input the single-shot failure training data after preprocessing into the random forest model for preliminary training and calculate the SHAP contribution value of each parameter. The SHAP values ​​of some parameters of the single-shot failure training data are shown in the attached manual. Figure 3 shown.

[0105] Secondly, the absolute value of the SHAP value of each parameter in the single-engine failure training data is summed and normalized, and then the weight of each parameter is calculated. The weights of some parameters in the single-engine failure training data are shown in the attached manual. Figure 4 As shown, the feature matrix is ​​weighted and adjusted by these weights.

[0106] Finally, the random forest model was retrained based on the weighted feature matrix, and the model hyperparameters were optimized using the grid search method to find the optimal "n_estimators", "max_features", "max_depth", and "criterion" parameters. The names of the top 20 parameters in terms of importance were output as the output results of the model through the subjectively set threshold. The output results of the single-shot failure training subject model are shown in the attached manual. Figure 5 shown.

[0107] For a certain aircraft special situation training subject, the feature parameters selected through a set of training data using the SHAP value combined with the weighted random forest model are not accurate enough. For a certain aircraft special situation training subject, n sets of training data can be selected to select feature parameters using the SHAP value combined with the weighted random forest model, and the output results of the random forest model can be checked. If the number of times a parameter appears in the output result is greater than or equal to n / 2, the parameter is determined as a feature parameter related to this aircraft special situation training subject.

[0108] This example uses five sets of training data to select aircraft single engine failure training feature parameters. If a parameter appears in the output results of the random forest model for more than or equal to 3 times, it is considered to be an aircraft single engine failure training feature parameter. Through experiments, the selected aircraft single engine failure training feature parameters are as shown in the attached manual. Figure 6 shown.

[0109] 3. Build a training set for the aircraft single engine failure training quality assessment model

[0110] According to the experimental results, the selected aircraft single-engine failure training feature parameters are extracted from the original data and combined into new data and further constructed as a training set for the aircraft single-engine failure training quality assessment model. The completed training set for the aircraft single-engine failure training quality assessment model is shown in the attached manual. Figure 7 shown.

[0111] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for designing a training set for a quality assessment model of aircraft special situation training for a flight simulator, characterized in that: The following steps are involved: Step 1: Data collection: The flight simulator training review system collects the training data of the flight trainees including aircraft status data, pilot control data and fault injection data when the aircraft encounters special conditions and when the aircraft is normal; Step 2: Data preprocessing: first replace the infinite values ​​and maximum values ​​in the collected data with NAN and fill them with the average value of the column. Then use the Lagrange interpolation method to complete the missing data. Finally, combine the aircraft status data and pilot control data into one piece of data and add a label column. Step 3: Calculate the SHAP value. Define a basic random forest model through "clf=RandomForestClassifier(random_state=42)" and calculate the SHAP value after preliminary training through the Tree SHAP method. Step 4: Calculate and adjust feature weights. Evaluate feature importance based on SHAP values, and complete feature weight calculation and weight adjustment. Step 5: Retrain the model and calculate the SHAP value of each feature to get the corresponding weight w i , and apply the weights to the feature matrix X to generate the weighted feature matrix X weighted , through the weighted feature matrix X weighted Retrain the random forest model with the features in, and use the grid search method to find the optimal model parameters to obtain the optimized weighted random forest model; Step 6: Determine the characteristic parameters of aircraft special situation training. For a certain aircraft special situation training subject, select n groups of training data and input them into the optimized weighted random forest model to select characteristic parameters, check the model output results, and determine the characteristic parameters of aircraft special situation training; Step 7: Constructing the evaluation model training set. Based on the determined aircraft special situation training characteristic parameters, the corresponding characteristic parameters are extracted from the original data and combined into new data, and the aircraft special situation training quality evaluation model training set is further constructed.

2. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator according to claim 1, characterized in that: The Lagrange interpolation method in step 2 to supplement the missing data is calculated by the following formula Where x represents the position of the independent variable, y represents the value of the corresponding function, and l j (x) represents the Lagrange polynomial.

3. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator according to claim 1, characterized in that: The label column adding method in step 2 is as follows: when a special situation occurs in the aircraft, the aircraft status data during the flight trainee training is labeled as 1, and when the aircraft is normal, the aircraft status data during the flight trainee training is labeled as 0; when a special situation occurs in the aircraft, the flight trainee's control data is labeled as 1, and when the aircraft is normal, the flight trainee's control data is labeled as 0.

4. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator according to claim 1, characterized in that: The random forest model defined in step 3 includes the number of decision trees n_estimators, the maximum number of features max_features considered in each split, the maximum depth of the decision tree max_depth, and the splitting criterion criterion; after calculating the SHAP value, the SHAP contribution value of each feature is obtained through "shap.TreeExplainer".

5. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator according to claim 1, characterized in that: The SHAP value calculation process in step 3 is as follows: The goal of S1 and SHAP values ​​is to calculate the marginal contribution of each feature to the model prediction value, which is calculated by the following formula: Among them, θ i represents the SHAP value of feature i, N is the set of all features, S is the feature subset, and f(S) is the predicted output of the model under the feature subset S; S2. For the decision tree model in the random forest model, use Tree SHAP to calculate the marginal contribution of the feature at each node of the model according to the segmentation condition of the feature on the path; S3. For each path in each decision tree model, Tree SHAP uses a recursive algorithm to calculate the SHAP value of each feature. For each node, assuming that the probability of the decision path from the root node to the leaf node is p(z), and the contribution of each feature to the predicted value is v(z), Tree SHAP uses the following formula to recursively assign SHAP values: Among them, θ i (x) represents the SHAP value of feature i in sample x, p ij is the probability that sample x passes through feature i in path j, |S| is the size of the feature subset that has passed through the path, and f(z∪{i}) is the predicted value of feature i after adding feature subset z; S4. For each path, from the root node to the leaf node, Tree SHAP needs to calculate the contribution of each feature to the leaf node output. The tree's predicted value f(x) is the value of the leaf node. The contribution of each path is calculated by the following formula: Where T represents the number of trees, w(z) is the weight of feature subset z, which represents the probability of the subset appearing on the path, and f(z∪{i})-f(z) is the incremental contribution after feature i is added; S5. In the entire random forest model, Tree SHAP accumulates the SHAP values ​​of each decision tree to get the final SHAP value by the following formula: in, is the SHAP value of feature i in the tth decision tree.

6. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator according to claim 1, characterized in that: The step 4 is specifically as follows: The first step is to perform absolute value processing on the SHAP value calculated in step 3, using |θ i (x)| represents the absolute value of the contribution of feature i in sample x; Step 2: Calculate the absolute average of all sample SHAP values ​​using the following formula Where n is the number of samples in the dataset, |θ i (x j )| is the absolute value of the SHAP value of the jth sample on feature i; The third step is to normalize the average SHAP value of each feature, which is expressed as follows: Among them, w i is the weight of feature i, is the average SHAP absolute value of feature i, d is the total number of features, and the sum of all feature weights after normalization is equal to 1; Step 4: First assume a feature matrix X with a shape of n×d, where n is the number of samples and d is the number of features. Each column of this matrix represents a feature and each row represents a sample. X = [x1, x2, ..., xd]. i Represents the value of the i-th feature. For each feature x i , using the calculated weight w i After weighted adjustment, the weighted feature matrix is ​​X weighted =X×W, where W is a diagonal matrix, W=diag(w1, w2, ..., w d ), weighted feature X weighted It is obtained by multiplying each feature column by its corresponding weight w i The obtained and weighted features are represented as X weighted [:, i] = X[:, i] × w i .

7. The method for designing a training set of an aircraft special situation training quality assessment model for a flight simulator according to claim 1, characterized in that: The step six of determining the characteristic parameters of aircraft special situation training specifically includes determining a certain parameter as a characteristic parameter related to the aircraft special situation training subject when the number of times a certain parameter appears in the output result is greater than or equal to n / 2.

Images