A method and system for explaining decision-making of an aircraft control agent

Abstract

The present invention discloses a method and system for explaining the decision of an aircraft control intelligent agent, which relates to the technical field of the intersection of aircraft control and artificial intelligence, and is used to solve the technical problem that the existing model has poor interpretability of the intelligent agent's decision, resulting in the inability to determine the cause of the aircraft failure based on the decision result; the method for explaining the decision of an aircraft control intelligent agent of the present invention comprises: obtaining real-time working data of the aircraft; synchronizing and storing each type of working data and then preprocessing it; obtaining feature data that affects the decision of the aircraft based on hypergraph theory; determining the decision category made by the intelligent agent, and dividing the feature data obtained by the hypergraph theory into a training set and a test set; using a decision tree algorithm, obtaining the optimal partitioning features in the training set according to the information gain rate criterion, and determining the parameters of the decision tree; selecting a root node and constructing a decision tree model; evaluating and pruning the decision tree model, and visually displaying the decision basis corresponding to the decision rules.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of aircraft control and artificial intelligence, and more specifically, to a method and system for interpreting decision-making of an aircraft control intelligent agent. Background Art

[0002] With the continuous development of aerospace technology, aircraft mission requirements are becoming increasingly complex, and the requirements for aircraft control precision and reliability are also becoming increasingly higher. Existing aircraft control methods are gradually showing their limitations when facing complex environments and missions. The introduction of artificial intelligence technology can bring new solutions to aircraft control.

[0003] Intelligent agents, a key concept in the field of artificial intelligence, can perceive their environment and make decisions based on specific strategies, enabling more efficient and flexible aircraft control. However, existing intelligent agent decision-making models are mostly based on black-box models such as deep learning. The decision-making process of black-box models is a mystery, making it difficult for non-professionals to understand the decision-making basis. Even for pilots and engineers, it is very difficult to quickly grasp the decision-making logic, resulting in an inability to accurately interpret intelligent agent decisions, which undoubtedly brings great inconvenience to aircraft operation and maintenance. Given this situation, in actual flight, if an aircraft problem occurs, it is difficult to quickly identify the root cause of the decision from the complex calculations of the black-box model. Summary of the Invention

[0004] The present invention aims to provide a method and system for interpreting decisions made by an aircraft control agent. This method addresses the technical problem that existing models offer limited interpretability of agent decisions, i.e., the inability to accurately interpret agent decisions, leading to an inability to determine the cause of aircraft failures based on the decision results. In view of this, the present invention achieves this goal through the following solution.

[0005] In a first aspect, the present invention provides a method for interpreting decisions of an aircraft control agent, comprising:

[0006] Determine the type of aircraft and obtain real-time operating data of the aircraft;

[0007] After synchronizing and storing each type of work data, pre-processing is performed;

[0008] Evaluate the correlation between each feature data in the working data and the decision of the intelligent agent, and obtain the feature data that affects the decision of the aircraft based on hypergraph theory;

[0009] Determining the decision categories made by the agent and dividing the feature data obtained by the hypergraph theory into a training set and a test set;

[0010] Using a decision tree algorithm, obtaining optimal partitioning features in the training set according to an information gain rate criterion, and determining the parameters of the decision tree;

[0011] Determine the stopping condition of the decision tree, select the root node and build the decision tree model;

[0012] The decision tree model is evaluated and pruned for optimization. Based on the current aircraft status data, the decision rule corresponding to the status data is located in the rule set, and the decision basis corresponding to the decision rule is visualized.

[0013] Compared with the prior art, in the aircraft control agent decision interpretation method of the present invention, after preprocessing the working data, the correlation between each feature data in the working data and the agent decision is evaluated, and the feature data that affects the aircraft decision is obtained based on the hypergraph theory; further, the feature data obtained by the hypergraph theory is divided into a training set and a test set, and then the decision tree algorithm is used to obtain the optimal division features in the training set according to the information gain rate criterion, and the parameters of the decision tree are determined, and then the root node is selected and a decision tree model is constructed; further, the decision tree model is evaluated and pruned, and according to the current aircraft status data, the decision rule corresponding to the status data is located in the rule set, and the decision basis corresponding to the decision rule is visualized, thereby completing the interpretation of the aircraft control agent decision. Based on the above technical solution of the present invention, non-professionals, pilots and engineers can easily understand the decision basis of the aircraft control intelligent agent, quickly grasp the decision logic, facilitate the operation and maintenance of the aircraft, and quickly find the root cause of the decision when the aircraft has problems or failures; further, the present invention can reduce the instability of decisions caused by small fluctuations in data, reduce the risk of issuing erroneous instructions due to noise interference from sensor data in complex flight environments, and ensure flight safety and stability; further, technicians can trace back the decision steps of the intelligent agent based on the clear decision process after explanation, quickly identify abnormal factors, and thus adjust parameters or optimize rules in a targeted manner to improve the performance of the aircraft control system; further, the aircraft control intelligent agent decision interpretation method of the present invention is easy to connect with existing aircraft control systems and operating procedures, reduce system upgrade costs and technical implementation difficulties, and facilitate practical application. The above technical solution of the present invention can accurately interpret the intelligent agent decision, solving the technical problem that the existing model has poor interpretability of the intelligent agent decision, resulting in the inability to determine the cause of the aircraft failure based on the decision results.

[0014] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the working data is collected by the aircraft inertial measurement unit, the aircraft GPS module, and sensors deployed on the aircraft;

[0015] The operating data includes the aircraft's angular velocity, acceleration, and attitude angle collected by the aircraft's inertial measurement unit, the aircraft's position and velocity data collected by the aircraft's GPS module, and environmental data collected by sensors deployed on the aircraft; the sensors include an airspeed meter, a barometer, and a gyroscope; and the operating data also includes control command data for the aircraft.

[0016] After synchronizing and storing each type of work data, pre-processing is performed, including:

[0017] Use timestamp alignment technology to ensure the time synchronization between sensors and control instructions;

[0018] The working data is stored in matrix form as: ; Where X represents the working data matrix, R represents the real number space, T represents the total time step, and N represents the data feature dimension;

[0019] Based on the working data, the short-term missing data is supplemented by linear interpolation, and the formula is:

[0020] ;in, represents the interpolated data at time t, Indicates the starting time data, Indicates the end time data, Indicates the end time of the interpolation interval, Indicates the starting time of the interpolation interval;

[0021] Long-term missing segments are filled through time series model prediction, or the missing segments are directly discarded and invalid instruction segments of the corresponding time steps are simultaneously removed;

[0022] Normalizing the working data, wherein the normalization process includes:

[0023] Normalize bounded physical quantities using deviation standardization and map features to , expressed as:

[0024] ;in, represents the normalized value, x represents the original data point, Represents the minimum value in the data. Indicates the maximum value in the data;

[0025] The noise-sensitive features are normalized using standard scores, expressed as: ;in, represents the standard score, x represents the original observation value, represents the mean, and , n represents the number of samples, represents the i-th observation value, represents the standard deviation, and .

[0026] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the step of evaluating the correlation between each feature data in the working data and the agent decision includes:

[0027] S100, traverse all feature data and control decisions respectively, and obtain the Pearson coefficient of each feature data and each control decision, and retain the feature data with a Pearson coefficient greater than 0.3, which is expressed as:

[0028] ; Among them, r represents the Pearson coefficient, represents the i-th observation value of feature X, represents the mean of the feature data, represents the i-th observation value of the control decision Y, represents the mean of the control decision, n represents the number of observations, and , ;

[0029] S200, traverse the remaining feature data and control decisions respectively, and calculate the mutual information ; expressed as:

[0030] ;in, represents mutual information, express and The joint density probability function of express The marginal probability density function of express The marginal probability density function of represents the i-th observation value of feature X, represents the i-th observation value of the control decision Y, x represents the observation value of feature X, and y represents the observation value of the control decision Y;

[0031] S300, estimates the probability density function of the random variable, expressed as:

[0032] ;in, Represents the probability density estimate, n represents the number of sample points, x represents the input point for which the probability density is to be estimated, Represents the jth sample point, K represents the kernel function, for , then , represents the variable of kernel function K, and , represents pi, represents an exponential function with a natural constant as the base, h represents the bandwidth, and , T represents the number of feature samples, d represents the feature dimension;

[0033] S400, uses L1 regularization, selects regularization parameters through cross-validation, and retains features corresponding to non-zero coefficients;

[0034] In S500, the Pearson coefficient, mutual information, and regularization parameters are standardized and then weighted fused. The weights are set according to task requirements to obtain the comprehensive importance score of each feature.

[0035] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the step of obtaining feature data that affects the aircraft's decision based on hypergraph theory includes:

[0036] Determine a data set, construct a hypergraph network, use features in the data set as nodes of the hypergraph, and determine the connection of hyperedges in the hypergraph network based on the comprehensive importance score of each feature; the features refer to the common features of a type of data;

[0037] For features in highly correlated subgraphs in a hypergraph, if the combined importance scores of multiple features are higher than a set threshold, only one feature with the greatest impact on the agent's control decision is retained;

[0038] By analyzing the structure of the hypergraph, the number of hyperedges involved in each node feature, i.e., the node degree, is obtained. A threshold is set and features with a node degree greater than the threshold and a betweenness centrality higher than the average level are retained as features for constructing the decision tree. The calculation formula for the betweenness centrality is:

[0039] ;in, The betweenness centrality of node v, represents the number of paths passing through node v among all the shortest paths from node s to node t, represents the total number of shortest paths from node s to node t, V represents the set of all nodes, and s, v, and t are nodes in set V.

[0040] Furthermore, in the aircraft control agent decision interpretation method of the present invention, obtaining the optimal partitioning features in the training set according to the information gain rate criterion includes:

[0041] Obtain the information entropy of the data set; for the training set, obtain its information entropy according to different control decision categories in the training set, expressed as: ;in, represents the information entropy of the dataset D, n represents the number of categories in the dataset D, D represents the dataset, represents the proportion of samples of category i in the data set, and , represents the number of samples in the i-th category, Represents the total number of samples in the dataset D;

[0042] Get the information gain of the data set; for each feature A, set V values, divide the data set D into V subsets, and calculate the information entropy of each subset and the information gain of the corresponding feature, expressed as: ;in, represents the information gain of feature A in dataset D, represents the information entropy of the data set D, represents the total number of samples in the dataset D, represents the sample subset in the dataset D where the feature A takes the value V, express Information entropy;

[0043] Get the information gain rate of each feature, expressed as: ;in, represents the information gain rate of feature A in data set D, The information gain of feature A in dataset D, represents the inherent information of feature A, and , represents the sample subset in the dataset D where the feature A takes the value V, Represents a sample subset The total number of Represents the total number of samples in the dataset D;

[0044] In the decision tree algorithm, each time a split is made, the feature with the largest information gain rate is selected as the split feature of the current node, ensuring that each split step can reduce the uncertainty of the data while suppressing overfitting; for any node, the feature with the largest information gain rate is selected as the optimal split feature for the next layer of nodes, and the process is repeated until the stopping condition is met.

[0045] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the process of obtaining the optimal partitioning feature from the training set according to the information gain rate criterion, selecting the feature with the largest information gain rate as the optimal partitioning feature of the node, includes:

[0046] Set the feature candidate set A, and ,in Represents the nth subset in the feature candidate set A, and selects the feature with the largest information gain rate in the feature candidate set As the partition feature of the root node; set is the feature in dataset D The sample subset with value V, is removed After the remaining feature candidate set, continue to Select the remaining feature candidate set The feature with the largest information gain rate is used as the optimal partitioning feature for the next layer of nodes; the above process is repeated recursively until the stopping condition is met, and finally the decision tree algorithm is obtained.

[0047] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the step of constructing a decision tree model includes:

[0048] Compare the information gain rates of all features and select the feature with the largest information gain rate as the partitioning feature of the root node;

[0049] The data set is divided into multiple subsets according to the partitioning features of the root node. For each subset, its information entropy, information gain and information gain rate of each feature are obtained, and the feature with the largest information gain rate is selected for partitioning until the stopping condition is met to obtain a decision tree model.

[0050] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the evaluation and pruning optimization of the decision tree model includes:

[0051] Use the test set to comprehensively evaluate the performance of the constructed decision tree model, including:

[0052] Get the accuracy index of the decision tree model, expressed as: ;in, represents the accuracy index, represents the true positive examples, that is, the number of samples that are actually positive and predicted as positive by the model; represents false positives, that is, the number of samples that are actually negative but predicted as positive by the model;

[0053] Get the recall rate indicator of the decision tree model, expressed as: ;in, represents the recall rate indicator, represents the true positive examples, that is, the number of samples that are actually positive and predicted as positive by the model, represents false negative examples, that is, the number of samples that are actually positive but predicted as negative by the model;

[0054] Get the F1 value of the decision tree model, expressed as: ;in, represents the accuracy index, represents the recall rate indicator;

[0055] Based on the F1 value of the decision tree model, a comprehensive evaluation of the effect of the decision tree model is conducted;

[0056] Pruning optimization is performed on the constructed decision tree model based on the hypergraph, including:

[0057] Each node in the decision tree and the features and samples in the dataset are treated as nodes in a hypergraph; the relationships between features are connected with hyperedges; in the decision tree, if a decision node is partitioned based on any feature to generate child nodes, hyperedges are used to connect the decision node, the partition feature node, and the child node;

[0058] Evaluate the importance of nodes in the hypergraph: Analyze the betweenness centrality of nodes in the hypergraph and evaluate their importance to classification by calculating the error reduction of the subtree corresponding to each node on the training set or validation set;

[0059] For the local structure centered on any decision tree node in the hypergraph, if the connection between other parts of a node is weak and the subtree contributes little to the classification, the subtree is pruned to simulate the overall connectivity of the hypergraph and the association pattern between features after pruning, as well as the performance change of the decision tree on the validation set. If the key structure of the hypergraph is not destroyed after pruning and the F1 value of the decision tree increases, pruning is continued;

[0060] When it is determined that any subtree needs to be pruned, the subtree is replaced with a leaf node whose category label is determined according to the category of the sample in the subtree;

[0061] After the pruning operation is completed, the structure of the hypergraph is updated; at the same time, the performance of the pruned decision tree model is comprehensively evaluated, and pruning is completed when the performance of the decision tree no longer improves.

[0062] Furthermore, in the aircraft control agent decision interpretation method of the present invention, the visual display of the decision basis corresponding to the decision rule includes:

[0063] Extract all paths from the root node to the leaf node in the trained decision tree model, and each path corresponds to a decision rule;

[0064] Each rule is expressed as a logical expression:

[0065] ;

[0066] in, represents the nth feature, Indicates the value range of the nth feature or the corresponding rule condition, C represents the category, Indicates a specific category value. Indicates if, Indicates AND, that is, multiple conditions must be met at the same time, Represents the first feature Satisfy the corresponding value range or rule conditions , Means: Then the category For specific categories ;

[0067] Perform feature screening on the real-time collected aircraft working data, and only retain the feature data that has an impact on the aircraft's decision-making based on hypergraph theory;

[0068] Starting from the root node of the decision tree, according to the value of the corresponding feature in the real-time working data, traverse down the decision tree branches layer by layer until reaching the leaf node;

[0069] If there is a fuzzy match, the adjacent branch rules are activated at the same time, the confidence of each rule is calculated, and the results of multiple rules are output. The rule with the highest confidence and the shallowest depth is selected as the final decision;

[0070] Extract features that directly affect decision-making and their threshold conditions from the activated rules, and annotate the comprehensive importance scores of the features based on the hypergraph;

[0071] Generate a decision logic chain constructed by natural language description and generate a visual interactive interface design;

[0072] Manual override of automatic decisions is allowed, and the system records the reasons for override and subsequent results for subsequent model iterations. When multiple manual interventions occur continuously, dynamic updates of the rule set are triggered. Historical versions are saved after each rule update, supporting rollback to the stable version in abnormal situations.

[0073] In a second aspect, the present invention provides an aircraft control agent decision interpretation system, comprising:

[0074] A data acquisition module is used to determine the type of aircraft and obtain the real-time working data of the aircraft;

[0075] The data preprocessing module is used to synchronize and store each type of work data and then preprocess it;

[0076] a feature data acquisition module, configured to evaluate the correlation between each feature data in the working data and the agent's decision, and to acquire feature data that has an impact on the aircraft's decision based on hypergraph theory;

[0077] A feature data partitioning module is used to determine the decision category made by the intelligent agent and divide the feature data obtained by the hypergraph theory into a training set and a test set;

[0078] A decision tree parameter determination module, which uses a decision tree algorithm to obtain optimal partitioning features in the training set according to an information gain rate criterion and determines the parameters of the decision tree;

[0079] Decision tree model building module, used to determine the stopping condition of the decision tree, select the root node and build the decision tree model;

[0080] The visualization display module is used to evaluate and prune the decision tree model, locate the decision rule corresponding to the status data in the rule set according to the current aircraft status data, and visualize the decision basis corresponding to the decision rule.

[0081] Compared with the prior art, the beneficial effects of the aircraft control agent decision interpretation system of the present invention are the same as the beneficial effects of the aircraft control agent decision interpretation method described in the above technical solution, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0083] Figure 1 A schematic diagram of a hypergraph network constructed in the present invention;

[0084] Figure 2 A schematic diagram of a decision tree model constructed in the present invention;

[0085] Figure 3 Schematic diagram of a flow chart of an exemplary aircraft control agent decision interpretation method of the present invention;

[0086] Reference numerals:

[0087] exist Figure 1 In the equation, F1 represents the first feature; F2 represents the second feature; F3 represents the third feature; F4 represents the fourth feature; F5 represents the fifth feature; F6 represents the sixth feature; and F7 represents the seventh feature. DETAILED DESCRIPTION

[0088] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0089] It should be noted that when an element is referred to as being “fixed on” or “disposed on” another element, it may be directly on the other element or indirectly on the other element. When an element is referred to as being “connected to” another element, it may be directly connected to the other element or indirectly connected to the other element.

[0090] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined. "Several" means one or more, unless otherwise specifically defined.

[0091] Intelligent agents, a key concept in the field of artificial intelligence, can perceive their environment and make decisions based on specific strategies, enabling more efficient and flexible aircraft control. However, existing intelligent agent decision-making models are mostly based on black-box models such as deep learning. The decision-making process of black-box models is a mystery, making it difficult for non-professionals to understand the decision-making basis. Even for pilots and engineers, it is very difficult to quickly grasp the decision-making logic, resulting in an inability to accurately interpret intelligent agent decisions, which undoubtedly brings great inconvenience to aircraft operation and maintenance. Given this situation, in actual flight, if an aircraft problem occurs, it is difficult to quickly identify the root cause of the decision from the complex calculations of the black-box model.

[0092] In order to solve the above technical problems, the present invention provides a decision interpretation method for an aircraft control agent, comprising:

[0093] Determine the type of aircraft and obtain real-time operating data of the aircraft;

[0094] After synchronizing and storing each type of work data, pre-processing is performed;

[0095] Evaluate the correlation between each feature data in the working data and the decision of the intelligent agent, and obtain the feature data that affects the decision of the aircraft based on hypergraph theory;

[0096] Determining the decision categories made by the agent and dividing the feature data obtained by the hypergraph theory into a training set and a test set;

[0097] Using a decision tree algorithm, obtaining optimal partitioning features in the training set according to an information gain rate criterion, and determining the parameters of the decision tree;

[0098] Determine the stopping condition of the decision tree, select the root node and build the decision tree model;

[0099] The decision tree model is evaluated and pruned for optimization. Based on the current aircraft status data, the decision rule corresponding to the status data is located in the rule set, and the decision basis corresponding to the decision rule is visualized.

[0100] When adopting the above technical solution, in the aircraft control agent decision interpretation method of the present invention, after preprocessing the working data, the correlation between each feature data in the working data and the agent decision is evaluated, and the feature data that affects the aircraft decision is obtained based on the hypergraph theory; further, the feature data obtained by the hypergraph theory is divided into a training set and a test set, and then the decision tree algorithm is used to obtain the optimal partitioning features in the training set according to the information gain rate criterion, and the parameters of the decision tree are determined, and then the root node is selected and a decision tree model is constructed; further, the decision tree model is evaluated and pruned, and according to the current aircraft status data, the decision rule corresponding to the status data is located in the rule set, and the decision basis corresponding to the decision rule is visualized, thereby completing the interpretation of the aircraft control agent decision. Based on the above technical solution of the present invention, non-professionals, pilots and engineers can easily understand the decision basis of the aircraft control intelligent agent, quickly grasp the decision logic, facilitate the operation and maintenance of the aircraft, and quickly find the root cause of the decision when a problem occurs; further, the present invention can reduce the instability of decision-making caused by small fluctuations in data, reduce the risk of issuing erroneous instructions due to noise interference of sensor data in complex flight environments, and ensure flight safety and stability; further, technicians can trace back the decision steps of the intelligent agent based on the clear decision process after explanation, quickly identify abnormal factors, and thus adjust parameters or optimize rules in a targeted manner to improve the performance of the aircraft control system; further, the aircraft control intelligent agent decision interpretation method of the present invention is easy to connect with existing aircraft control systems and operating procedures, reduce system upgrade costs and technical implementation difficulties, and facilitate practical application. The above technical solution of the present invention can accurately interpret the intelligent agent decision, solving the technical problem that the existing model has poor interpretability of the intelligent agent decision, resulting in the inability to determine the cause of the aircraft failure based on the decision results.

[0101] In order to better understand the present invention, the content of the present invention is further explained below in conjunction with specific examples, but the content of the present invention is not limited to the following examples.

[0102] Example 1

[0103] This embodiment provides a method for interpreting decisions of an aircraft control agent, including:

[0104] S100, determining the type of the aircraft and obtaining real-time operating data of the aircraft;

[0105] S200, after synchronizing and storing each type of work data, pre-processing is performed;

[0106] S300, evaluating the correlation between each feature data in the working data and the agent's decision, and obtaining feature data that has an impact on the aircraft's decision based on hypergraph theory;

[0107] S400, determining the decision category made by the intelligent agent, and dividing the feature data obtained by the hypergraph theory into a training set and a test set;

[0108] S500, using a decision tree algorithm to obtain optimal partitioning features in the training set according to an information gain rate criterion, and determining parameters of the decision tree;

[0109] S600, determining the stopping condition of the decision tree, selecting a root node and building a decision tree model;

[0110] S700, evaluating and pruning the decision tree model, locating the decision rule corresponding to the status data in the rule set according to the current aircraft status data, and visually displaying the decision basis corresponding to the decision rule.

[0111] Example 2

[0112] See also Figures 1 to 3 This embodiment provides a method for explaining decision-making of an aircraft control agent, including:

[0113] S100, determining the type of the aircraft and obtaining real-time operating data of the aircraft;

[0114] The above-mentioned working data can be collected by the aircraft inertial measurement unit, the aircraft GPS module, and sensors deployed on the aircraft; the working data includes the aircraft angular velocity, aircraft acceleration and aircraft attitude angle collected by the aircraft inertial measurement unit, the aircraft position and speed data collected by the aircraft GPS module, and environmental data collected by the sensors deployed on the aircraft; the sensors include an airspeed meter, a barometer and a gyroscope; the working data also includes the control command data of the aircraft.

[0115] S200, after synchronizing and storing each type of work data, pre-processing is performed;

[0116] Furthermore, the content of this step includes:

[0117] S201, using timestamp alignment technology to ensure the time synchronization between sensors and control instructions;

[0118] S202, storing the working data in a matrix form as follows: ; Where X represents the working data matrix, R represents the real number space, T represents the total time step, and N represents the data feature dimension;

[0119] S203: Based on the working data, short-term missing data is supplemented using linear interpolation, and the formula is: ;in, represents the interpolated data at time t, Indicates the starting time data, Indicates the end time data, Indicates the end time of the interpolation interval, Indicates the starting time of the interpolation interval;

[0120] S204, filling in long-term missing segments of data through time series model prediction, or directly discarding the missing segments and simultaneously removing invalid instruction segments of the corresponding time steps;

[0121] S205: normalizing the working data. The normalization process includes:

[0122] S2051, normalize bounded physical quantities using deviation standardization and map features to , expressed as: ;in, represents the normalized value, x represents the original data point, Represents the minimum value in the data. Indicates the maximum value in the data;

[0123] S2052, the noise-sensitive features are normalized using standard scores, expressed as: ;in, represents the standard score, x represents the original observation value, represents the mean, and , n represents the number of samples, represents the i-th observation value, represents the standard deviation, and .

[0124] S300, evaluating the correlation between each feature data in the working data and the agent's decision, and obtaining feature data that has an impact on the aircraft's decision based on hypergraph theory;

[0125] Furthermore, in this step, evaluating the correlation between each feature data in the working data and the agent's decision includes:

[0126] S311, traverse all feature data and control decisions respectively, and obtain the Pearson coefficient of each feature data and each control decision, and retain the feature data with a Pearson coefficient greater than 0.3, which is expressed as:

[0127] ; Among them, r represents the Pearson coefficient, represents the i-th observation value of feature X, represents the mean of the feature data, represents the i-th observation value of the control decision Y, represents the mean of the control decision, n represents the number of observations, and , ;

[0128] S312, traverse the remaining feature data and control decisions respectively, and calculate the mutual information ; expressed as:

[0129] ;in, represents mutual information, express and The joint density probability function of express The marginal probability density function of express The marginal probability density function of represents the i-th observation value of feature X, represents the i-th observation value of the control decision Y, x represents the observation value of feature X, and y represents the observation value of the control decision Y;

[0130] S313, estimate the probability density function of the random variable, expressed as:

[0131] ;in, Represents the probability density estimate, n represents the number of sample points, x represents the input point for which the probability density is to be estimated, Represents the jth sample point, K represents the kernel function, for , then , represents the variable of kernel function K, and , represents pi, represents an exponential function with a natural constant as the base, h represents the bandwidth, and , T represents the number of feature samples, d represents the feature dimension;

[0132] S314, using L1 regularization, select the regularization parameter through cross-validation, and retain the features corresponding to non-zero coefficients;

[0133] S315: After standardization, the Pearson coefficient, mutual information, and regularization parameter are weighted and fused. The weights are set according to the task requirements to obtain the comprehensive importance score of each feature.

[0134] Furthermore, in this step, feature data that influences the aircraft's decision-making is obtained based on hypergraph theory, including:

[0135] S321, Determine the data set, see Figure 1 , construct as Figure 1The hypergraph network shown uses the features in the dataset as nodes of the hypergraph, and determines the connection of hyperedges in the hypergraph network by the comprehensive importance score of each feature; the features refer to the common features of a type of data;

[0136] S322, for features in highly correlated subgraphs in the hypergraph, if the comprehensive importance scores of multiple features are all higher than the set threshold, only one feature with the greatest impact on the agent's control decision is retained;

[0137] S323, by analyzing the structure of the hypergraph, the number of hyperedges involved in each node feature, i.e., the node degree, is obtained; a threshold is set, and features with a node degree greater than the threshold and a betweenness centrality higher than the average level are retained as features for constructing the decision tree; the betweenness centrality is calculated as follows:

[0138] ;in, The betweenness centrality of node v, represents the number of paths passing through node v among all the shortest paths from node s to node t, represents the total number of shortest paths from node s to node t, V represents the set of all nodes, and s, v, and t are nodes in set V.

[0139] S400, determining the decision category made by the intelligent agent, and dividing the feature data obtained by the hypergraph theory into a training set and a test set.

[0140] S500, using the C4.5 decision tree algorithm, according to the information gain rate criterion, to obtain the optimal partitioning features in the training set, and determine the parameters of the decision tree;

[0141] Furthermore, the content of this step includes:

[0142] S511, obtaining the information entropy of the data set; for the training set, obtaining its information entropy according to different control decision categories in the training set, expressed as: ;in, represents the information entropy of the dataset D, n represents the number of categories in the dataset D, D represents the dataset, represents the proportion of samples of category i in the data set, and , represents the number of samples in the i-th category, Represents the total number of samples in the dataset D;

[0143] S512, obtaining the information gain of the data set; for each feature A, set V values, divide the data set D into V subsets, and calculate the information entropy of each subset and the information gain of the corresponding feature, expressed as: ;in, represents the information gain of feature A in dataset D, represents the information entropy of the data set D, represents the total number of samples in the dataset D, represents the sample subset in the dataset D where the feature A takes the value V, express Information entropy;

[0144] S513, obtain the information gain rate of each feature, expressed as: ;in, represents the information gain rate of feature A in data set D, The information gain of feature A in dataset D, represents the inherent information of feature A, and , represents the sample subset in the dataset D where the feature A takes the value V, Represents a sample subset The total number of Represents the total number of samples in the dataset D;

[0145] In the C4.5 decision tree algorithm, S514 selects the feature with the largest information gain rate as the splitting feature of the current node at each split, ensuring that each split step can reduce the uncertainty of the data while suppressing overfitting. For any node, the feature with the largest information gain rate is selected as the optimal splitting feature for the next layer of nodes, and this process is repeated until the stopping condition is met.

[0146] S515: Select the feature with the largest information gain rate as the optimal partitioning feature for the node, specifically:

[0147] Set the feature candidate set A, and ,in Represents the nth subset in the feature candidate set A, and selects the feature with the largest information gain rate in the feature candidate set As the partition feature of the root node; set is the feature in dataset D The sample subset with value V, is removed After the remaining feature candidate set, continue to Select the remaining feature candidate set The feature with the largest information gain rate is used as the optimal partitioning feature for the next layer of nodes; the above process is repeated recursively until the stopping condition is met, and the decision tree algorithm is finally obtained;

[0148] Furthermore, determining the parameters of the decision tree includes:

[0149] S521, adjust the minimum number of sample splits for the decision tree algorithm based on the size and characteristics of the dataset. For the aircraft dataset, if the data volume is small, the minimum number of sample splits can be set to 5-10; if the data volume is large, the minimum number of sample splits can be appropriately increased, such as 20-50. You can try different values through cross-validation to select the parameters that give the best performance on the validation set.

[0150] S522, adjust the minimum number of sample leaves of the decision tree algorithm according to the size and characteristics of the data set; the minimum number of sample leaves can generally be tried from a smaller value, such as 2-5, and then the optimal value is determined through cross-validation; a larger minimum number of sample leaves can prevent overfitting, but may lead to underfitting;

[0151] S523, adjust the maximum tree depth of the decision tree algorithm according to the size and characteristics of the dataset. If the dataset contains many features and a small number of samples, a smaller maximum tree depth, such as 3-5, can be set to avoid overfitting. If the dataset is large and the relationships between features are complex, the value can be increased appropriately, such as 8-10. The optimal maximum tree depth can also be selected through cross-validation.

[0152] S524, further determine the appropriate information gain rate threshold through experiments and cross-validation. The information gain rate threshold is generally small, such as 0.01-0.1. A smaller threshold will make the decision tree more finely divided, which may lead to overfitting; a larger threshold will make the decision tree simpler, which may lead to underfitting.

[0153] S600, see Figure 2 , determine the stopping condition of the decision tree, select the root node and construct Figure 2 The decision tree model shown;

[0154] Furthermore, the content of this step includes:

[0155] S601, comparing the information gain rates of all features and selecting the feature with the largest information gain rate as the partitioning feature of the root node;

[0156] S602: Divide the data set into multiple subsets based on the partitioning features of the root node, obtain the information entropy, information gain and information gain rate of each feature of each subset, and select the feature with the largest information gain rate for partitioning until the stopping condition is met to obtain a decision tree model.

[0157] S700, evaluating and pruning the decision tree model, locating a decision rule corresponding to the current aircraft status data in a rule set, and visually displaying a decision basis corresponding to the decision rule;

[0158] Furthermore, the step of evaluating and pruning the decision tree model includes:

[0159] S711, use the test set to comprehensively evaluate the performance of the constructed decision tree model, including:

[0160] S7111, obtain the accuracy index of the decision tree model, expressed as: ;in, represents the accuracy index, represents the true positive examples, that is, the number of samples that are actually positive and predicted as positive by the model; represents false positives, that is, the number of samples that are actually negative but predicted as positive by the model;

[0161] S7112, obtain the recall rate indicator of the decision tree model, expressed as: ;in, represents the recall rate indicator, represents the true positive examples, that is, the number of samples that are actually positive and predicted as positive by the model, represents false negative examples, that is, the number of samples that are actually positive but predicted as negative by the model;

[0162] S7113, obtain the F1 value of the decision tree model, expressed as: ;in, represents the accuracy index, represents the recall rate indicator;

[0163] S7114, based on the F1 value of the decision tree model, comprehensively evaluate the effect of the decision tree model;

[0164] S712, pruning and optimizing the constructed decision tree model based on the hypergraph, including:

[0165] S7121: Treat each node in the decision tree and the features and samples in the dataset as nodes of a hypergraph; connect the relationships between features with hyperedges; in the decision tree, if a decision node is partitioned based on any feature to generate child nodes, use hyperedges to connect the decision node, the partition feature node, and the child node;

[0166] S7122, evaluate the importance of nodes in the hypergraph: analyze the betweenness centrality of nodes in the hypergraph and evaluate their importance to classification by calculating the error reduction of the subtree corresponding to each node on the training set or validation set;

[0167] S7123: For the local structure of a hypergraph centered on any decision tree node, if the connections between other parts of a node are weak and the subtree contributes little to the classification, prune the subtree. Simulate the overall connectivity of the hypergraph and the correlation pattern between features after pruning, as well as the performance changes of the decision tree on the validation set. If the key structure of the hypergraph is not destroyed after pruning and the F1 value of the decision tree increases, continue pruning.

[0168] S7124, when it is determined that any subtree needs to be pruned, the subtree is replaced with a leaf node whose category label is determined according to the category of the samples in the subtree;

[0169] S7125: After the pruning operation is completed, the structure of the hypergraph is updated; at the same time, the performance of the pruned decision tree model is comprehensively evaluated, and pruning is completed when the performance of the decision tree no longer improves;

[0170] Furthermore, the decision basis corresponding to the decision rule is visualized, including:

[0171] S721, extracting all paths from the root node to the leaf nodes in the trained decision tree model, where each path corresponds to a decision rule;

[0172] S722, express each rule as a logical expression:

[0173] ;

[0174] in, represents the nth feature, Indicates the value range of the nth feature or the corresponding rule condition, C represents the category, Indicates a specific category value. Indicates if, Indicates AND, that is, multiple conditions must be met at the same time, Represents the first feature Satisfy the corresponding value range or rule conditions , Means: Then the category For specific categories ;

[0175] S723, feature screening is performed on the aircraft working data collected in real time, and only feature data that is obtained based on hypergraph theory and has an impact on the aircraft's decision-making is retained;

[0176] S724 , starting from the root node of the decision tree, traverse down layer by layer along the branches of the decision tree according to the values of the corresponding features in the real-time working data until reaching a leaf node;

[0177] S725: If there is a fuzzy match, activate the adjacent branch rules at the same time, calculate the confidence of each rule, output the results of multiple rules, and select the rule with the highest confidence and the shallowest depth as the final decision; the specific expression is: ;

[0178] S726, extracting features and their threshold conditions that directly affect the decision from the activated rules, and marking the comprehensive importance scores of the features based on the hypergraph;

[0179] S727, generate a decision logic chain constructed by natural language description and generate a visual interactive interface design;

[0180] S728 allows manual override of automatic decisions. The system records the reasons for override and subsequent results for subsequent model iterations. When multiple manual interventions occur continuously, dynamic updates of the rule set are triggered. Historical versions are saved after each rule update, supporting rollback to the stable version in abnormal situations.

[0181] Further, see Figure 2 and Figure 3 , based on the above embodiment, combined again Figure 2 and Figure 3 The technical solution of the present invention is described. Figure 2 In the decision tree, the root node is "vertical wind speed". If the vertical wind speed is ≤-3.2m / s, the node "normal control mode" is entered. At this time, further decisions are made based on the "roll angular velocity". When the roll angular velocity is greater than 2.5rad / s, the decision is "full power climb + right rudder compensation 20°", and when the roll angular velocity is ≤2.5rad / s, the decision is "medium power climb + right rudder compensation 10°". If the vertical wind speed is greater than -3.2m / s, the node "emergency recovery mode" is entered. At this node, decisions are made based on other features. When the airspeed is less than 12m / s, the decision is "start the backup power system", and when the airspeed is ≥12m / s, the decision is "maintain current correction". At the same time, based on the "GPS altitude offset", when it is greater than 5m / s, the decision is "position fine-tuning pulse", and when it is ≤5m / s, the decision is "maintain hovering".

[0182] Further, in Figure 3The process shown in the figure can be summarized as follows: in the data preparation stage, according to the aircraft type, aircraft control command data, status data and environmental data are collected, and then data cleaning and normalization are performed; in the feature selection stage, the correlation between each feature and the intelligent agent's control decision is evaluated, and features that have a significant impact on the aircraft control decision are selected based on the hypergraph; in the decision tree construction stage, the control decision category, training set and test set are used to find the optimal partitioning features through the information gain rate criterion, determine the decision tree parameters and stop construction conditions, and then construct a C4.5 decision tree; in the decision tree optimization stage, the decision tree model is evaluated and pruned to determine whether it meets the performance requirements. If not, optimization is continued (marked as "N") until it meets the requirements (marked as "Y"); in the decision visualization stage, according to the current aircraft status data, the corresponding decision rule is quickly located in the rule set, the decision basis is explained and visualized.

[0183] In a second aspect, this embodiment further provides an aircraft control agent decision interpretation system, comprising:

[0184] A data acquisition module is used to determine the type of aircraft and obtain the real-time working data of the aircraft;

[0185] The data preprocessing module is used to synchronize and store each type of work data and then preprocess it;

[0186] a feature data acquisition module, configured to evaluate the correlation between each feature data in the working data and the agent's decision, and to acquire feature data that has an impact on the aircraft's decision based on hypergraph theory;

[0187] A feature data partitioning module is used to determine the decision category made by the intelligent agent and divide the feature data obtained by the hypergraph theory into a training set and a test set;

[0188] A decision tree parameter determination module, which uses a decision tree algorithm to obtain optimal partitioning features in the training set according to an information gain rate criterion and determines the parameters of the decision tree;

[0189] Decision tree model building module, used to determine the stopping condition of the decision tree, select the root node and build the decision tree model;

[0190] The visualization display module is used to evaluate and prune the decision tree model, locate the decision rule corresponding to the status data in the rule set according to the current aircraft status data, and visualize the decision basis corresponding to the decision rule.

[0191] In the description of the above embodiments, specific features, structures, materials or characteristics may be combined in an appropriate manner in any one or more embodiments or examples.

[0192] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for explaining decision making of an aircraft control agent, characterized in that: include: Determine the type of aircraft and obtain real-time operating data of the aircraft; After synchronizing and storing each type of work data, pre-processing is performed; Evaluate the correlation between each feature data in the working data and the decision of the intelligent agent, and obtain the feature data that affects the decision of the aircraft based on hypergraph theory; Determining the decision categories made by the agent and dividing the feature data obtained by the hypergraph theory into a training set and a test set; Using a decision tree algorithm, obtaining optimal partitioning features in the training set according to an information gain rate criterion, and determining the parameters of the decision tree; Determine the stopping condition of the decision tree, select the root node and build the decision tree model; The decision tree model is evaluated and pruned for optimization. Based on the current aircraft status data, the decision rule corresponding to the status data is located in the rule set, and the decision basis corresponding to the decision rule is visualized.

2. The aircraft control agent decision interpretation method according to claim 1, characterized in that: The working data is collected by the aircraft inertial measurement unit, the aircraft GPS module, and sensors deployed on the aircraft; The working data includes the aircraft angular velocity, aircraft acceleration and aircraft attitude angle collected by the aircraft inertial measurement unit, the aircraft position and speed data collected by the aircraft GPS module, and environmental data collected by sensors deployed on the aircraft; the sensors include an airspeed meter, a barometer and a gyroscope; The working data also includes control instruction data of the aircraft; After synchronizing and storing each type of work data, pre-processing is performed, including: Use timestamp alignment technology to ensure the time synchronization between sensors and control instructions; The working data is stored in matrix form as: ;in, X represents the working data matrix, R represents the real number space, T represents the total time step, N Represents the data feature dimension; Based on the working data, the short-term missing data is supplemented by linear interpolation, and the formula is: ;in, express t The interpolated data at time instants, Indicates the starting time data, Indicates the end time data, Indicates the end time of the interpolation interval, Indicates the starting time of the interpolation interval; Long-term missing segments are filled through time series model prediction, or the missing segments are directly discarded and invalid instruction segments of the corresponding time steps are simultaneously removed; Normalizing the working data, wherein the normalization process includes: Normalize bounded physical quantities using deviation standardization and map features to , expressed as: ;in, represents the normalized value, x represents the original data points, Represents the minimum value in the data. Indicates the maximum value in the data; The noise-sensitive features are normalized using standard scores, expressed as: ;in, represents the standard score, x represents the original observation value, represents the mean, and , n represents the number of samples, Indicates the i observations, represents the standard deviation, and .

3. The aircraft control agent decision interpretation method according to claim 2, characterized in that: The evaluating the correlation between each feature data in the working data and the agent decision includes: S100, traverse all feature data and control decisions respectively, and obtain the Pearson coefficient of each feature data and each control decision, and retain the feature data with a Pearson coefficient greater than 0.3, which is expressed as: ;in, r represents the Pearson coefficient, Representation characteristics X No. i observations, represents the mean of the feature data, Represents control decisions Y No. i observations, represents the mean of the control decision, n represents the number of observations, and , ; S200, traverse the remaining feature data and control decisions respectively, and calculate the mutual information ; expressed as: ;in, represents mutual information, express and The joint density probability function of express The marginal probability density function of express The marginal probability density function of Representation characteristics X No. i observations, Represents control decisions Y No. i observations, x Representation characteristics X The observed value of y Represents control decisions Y Observed values of S300, estimates the probability density function of the random variable, expressed as: ;in, represents the probability density estimate, n represents the number of sample points, x represents the input point whose probability density is to be estimated, Indicates the j Sample points, K Represents the kernel function, for , then , Represents the kernel function K variables, and , represents pi, represents an exponential function with a natural constant as the base, h represents bandwidth, and , T represents the number of feature samples, d Represents feature dimension; S400, uses L1 regularization, selects regularization parameters through cross-validation, and retains features corresponding to non-zero coefficients; In S500, the Pearson coefficient, mutual information, and regularization parameters are standardized and then weighted fused. The weights are set according to task requirements to obtain the comprehensive importance score of each feature.

4. The aircraft control agent decision interpretation method according to claim 3, characterized in that: The acquisition of feature data that affects the decision-making of the aircraft based on the hypergraph theory includes: Determine a data set, construct a hypergraph network, use features in the data set as nodes of the hypergraph, and determine the connection of hyperedges in the hypergraph network based on the comprehensive importance score of each feature; the features refer to the common features of a type of data; For features in related subgraphs in a hypergraph, if the comprehensive importance scores of multiple features are higher than the set threshold, only one feature with the greatest impact on the agent's control decision is retained; By analyzing the structure of the hypergraph, the number of hyperedges involved in each node feature, i.e., the node degree, is obtained. A threshold is set and features with a node degree greater than the threshold and a betweenness centrality higher than the average level are retained as features for constructing the decision tree. The calculation formula for the betweenness centrality is: ;in, node v The betweenness centrality of Representation node s To Node t All shortest paths passing through the node v The number of paths, Representation node s To Node t The total number of shortest paths, V represents the set of all nodes, s 、 v and t For collection V Nodes in .

5. The aircraft control agent decision interpretation method according to claim 4, characterized in that: The step of obtaining the optimal partitioning features in the training set according to the information gain rate criterion includes: Obtain the information entropy of the data set; for the training set, obtain its information entropy according to different control decision categories in the training set, expressed as: ;in, Representation dataset D The information entropy of n Representation dataset D The number of categories in , D represents a dataset, Indicates the i The proportion of class samples in the dataset, and , Indicates the i The number of class samples, Representation dataset D The total number of samples; Get the information gain of the data set; for each feature A , set with V value, the data set D Divided into V subsets, and calculate the information entropy of each subset and the information gain of the corresponding feature, expressed as: ;in, Representation dataset D Medium Features A The information gain of Representation dataset D The information entropy of Representation dataset D The total number of samples, Representation dataset D Medium Features A The value is V A subset of samples, express Information entropy of Get the information gain rate of each feature, expressed as: ;in, Representation dataset D Medium Features A The information gain rate, Dataset D Medium Features A The information gain of Representation characteristics A inherent information, and , Representation dataset D Medium Features A The value is V A subset of samples, Represents a sample subset The total number of Representation dataset D The total number of samples; In the decision tree algorithm, each time a split is made, the feature with the largest information gain rate is selected as the split feature of the current node, ensuring that each split step can reduce the uncertainty of the data while suppressing overfitting; for any node, the feature with the largest information gain rate is selected as the optimal split feature for the next layer of nodes, and the process is repeated until the stopping condition is met.

6. The aircraft control agent decision interpretation method according to claim 5, characterized in that: In the process of obtaining the optimal partition feature in the training set according to the information gain rate criterion, selecting the feature with the largest information gain rate as the optimal partition feature of the node includes: Set feature candidate set A ,and ,in Represents the feature candidate set A The n subsets, select the feature with the largest information gain rate from the feature candidate set As the partition feature of the root node; set For the dataset D Medium Features The value is V A subset of samples, Remove features After the remaining feature candidate set, continue to Select the remaining feature candidate set The feature with the largest information gain rate is used as the optimal partitioning feature for the next layer of nodes; the above process is repeated recursively until the stopping condition is met, and finally the decision tree algorithm is obtained.

7. The aircraft control agent decision interpretation method according to claim 6, characterized in that: The construction of the decision tree model comprises: Compare the information gain rates of all features and select the feature with the largest information gain rate as the partitioning feature of the root node; The data set is divided into multiple subsets according to the partitioning features of the root node. For each subset, its information entropy, information gain and information gain rate of each feature are obtained, and the feature with the largest information gain rate is selected for partitioning until the stopping condition is met to obtain a decision tree model.

8. The aircraft control agent decision interpretation method according to claim 7, characterized in that: The evaluating and pruning optimization of the decision tree model includes: Use the test set to comprehensively evaluate the performance of the constructed decision tree model, including: Get the accuracy index of the decision tree model, expressed as: ;in, represents the accuracy index, represents the true positive examples, that is, the number of samples that are actually positive and predicted as positive by the model; Represents false positive examples, that is, the number of samples that are actually negative but predicted as positive by the model; Get the recall rate indicator of the decision tree model, expressed as: ;in, represents the recall rate indicator, represents the true positive examples, that is, the number of samples that are actually positive and predicted as positive by the model, represents false negative examples, that is, the number of samples that are actually positive but predicted as negative by the model; Get the F1 value of the decision tree model, expressed as: ;in, represents the accuracy index, Represents the recall rate indicator; Based on the F1 value of the decision tree model, a comprehensive evaluation of the effect of the decision tree model is conducted; Pruning optimization is performed on the constructed decision tree model based on the hypergraph, including: Each node in the decision tree and the features and samples in the dataset are treated as nodes in a hypergraph; the relationships between features are connected with hyperedges; in the decision tree, if a decision node is partitioned based on any feature to generate child nodes, hyperedges are used to connect the decision node, the partition feature node, and the child node; Evaluate the importance of nodes in the hypergraph: Analyze the betweenness centrality of nodes in the hypergraph and evaluate their importance to classification by calculating the error reduction of the subtree corresponding to each node on the training set or validation set; For the local structure centered on any decision tree node in the hypergraph, if the connection between other parts of a node is weak and the subtree contributes little to the classification, the subtree is pruned to simulate the overall connectivity of the hypergraph and the association pattern between features after pruning, as well as the performance change of the decision tree on the validation set. If the key structure of the hypergraph is not destroyed after pruning and the F1 value of the decision tree increases, pruning is continued; When it is determined that any subtree needs to be pruned, the subtree is replaced with a leaf node whose category label is determined according to the category of the sample in the subtree; After the pruning operation is completed, the structure of the hypergraph is updated; at the same time, the performance of the pruned decision tree model is comprehensively evaluated, and pruning is completed when the performance of the decision tree no longer improves.

9. The aircraft control agent decision interpretation method according to claim 8, characterized in that: The visual display of the decision basis corresponding to the decision rule includes: Extract all paths from the root node to the leaf node in the trained decision tree model, and each path corresponds to a decision rule; Each rule is expressed as a logical expression: ； in, Indicates the n Features, Indicates the n The value range of each feature or the corresponding rule condition, C Indicates category, Indicates a specific category value. Indicates if, Indicates AND, that is, multiple conditions must be met at the same time, Represents the first feature Satisfy the corresponding value range or rule conditions , Means: Then the category For specific categories ; Perform feature screening on the real-time collected aircraft working data, and only retain the feature data that has an impact on the aircraft's decision-making based on hypergraph theory; Starting from the root node of the decision tree, according to the value of the corresponding feature in the real-time working data, traverse down the decision tree branches layer by layer until reaching the leaf node; If there is a fuzzy match, the adjacent branch rules are activated at the same time, the confidence of each rule is calculated, and the results of multiple rules are output. The rule with the highest confidence and the shallowest depth is selected as the final decision; Extract features that directly affect decision-making and their threshold conditions from the activated rules, and annotate the comprehensive importance scores of the features based on the hypergraph; Generate a decision logic chain constructed by natural language description and generate a visual interactive interface design; Manual override of automatic decisions is allowed, and the system records the reasons for override and subsequent results for subsequent model iterations. When multiple manual interventions occur continuously, dynamic updates of the rule set are triggered. Historical versions are saved after each rule update, supporting rollback to the stable version in abnormal situations.

10. An aircraft control agent decision interpretation system, characterized in that: include: A data acquisition module is used to determine the type of aircraft and obtain the real-time working data of the aircraft; The data preprocessing module is used to synchronize and store each type of work data and then preprocess it; a feature data acquisition module, configured to evaluate the correlation between each feature data in the working data and the agent's decision, and to acquire feature data that has an impact on the aircraft's decision based on hypergraph theory; A feature data partitioning module is used to determine the decision category made by the intelligent agent and divide the feature data obtained by the hypergraph theory into a training set and a test set; A decision tree parameter determination module, which uses a decision tree algorithm to obtain optimal partitioning features in the training set according to an information gain rate criterion and determines the parameters of the decision tree; Decision tree model building module, used to determine the stopping condition of the decision tree, select the root node and build the decision tree model; The visualization display module is used to evaluate and prune the decision tree model, locate the decision rule corresponding to the status data in the rule set according to the current aircraft status data, and visualize the decision basis corresponding to the decision rule.

Images