An intelligent identification method for abnormal flight state of an aerial vehicle

Abstract

The present application relates to the technical field of aviation flight data analysis, and particularly relates to an abnormal state detection method for an aviation aircraft. The method comprises the following steps: 1) selecting historical standard flight data of the aviation aircraft by an aviation aircraft flight expert, and marking flight stages of the historical standard flight data; 2) designing an optimized classification regression decision tree (CART) algorithm to construct a flight stage classification model (M) by using the marked flight stage data set D; 3) inputting flight data of an abnormal flight state to be identified; 4) dividing the flight data set X into flight stages by using the constructed classification model M; 5) identifying the abnormal flight state in the flight data set by using an optimized isolated forest algorithm (IF); and 6) outputting time sequence data of the abnormal flight state. The present application uses historical flight data of the aviation aircraft to train an algorithm to establish a model of a normal flight state, and uses the constructed model to identify an abnormal flight state.

Description

Technical Field

[0001] This invention relates to the field of aviation flight data analysis technology, and in particular to a method for detecting abnormal states of aircraft. Background Technology

[0002] The flight status of an aircraft refers to its motion at a specific instant. Normal flight status refers to the aircraft's state under the control of the flight control system, flying along a pre-programmed route or in a desired state. Abnormal flight status refers to a significant deviation between the aircraft's flight status and the desired state due to a component malfunction or external environmental interference. For example, a large deviation between the actual flight path and the set path, or a significant deviation between given attitude angles, speeds, and headings and their corresponding measured values ​​over a prolonged period. These abnormal flight statuses pose flight risks, and failure to identify and address them promptly could lead to major flight safety accidents. Therefore, identifying abnormal flight statuses is crucial for ensuring flight safety. Identifying abnormal flight statuses provides decision support for real-time aircraft control decisions and health management, facilitating early identification of the causes of abnormal events by aircraft maintenance personnel, preventing flight accidents, and ensuring flight safety.

[0003] To address the problem of automatic flight status identification, Xie Chuan et al. constructed a flight status identification method based on an expert knowledge base and a knowledge inference engine. However, this method is limited by the expert knowledge base. Meng Guanglei et al., in their literature, used flight parameter data corresponding to maneuvers from flight simulation training as the research object, constructed a dynamic Bayesian network model for maneuver recognition, and used an intelligent method based on recursive inference from the network model to identify flight status. Zhou Chao et al., mainly addressing the characteristics of strong randomness and varying lengths in tactical maneuver data, proposed a flight status identification method based on an improved dynamic time warping algorithm. This method uses different contribution levels for different feature parameters of the flight status, calculates the frame matching distance between flight parameter data and standard template data, and identifies the flight status based on the distance. These studies have all achieved automatic identification of certain flight statuses in specific application scenarios and are not applicable to scenarios involving the identification of abnormal flight statuses of aircraft.

[0004] Therefore, there is an urgent need for an intelligent identification method for abnormal flight status of aircraft that can solve the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an intelligent identification method for abnormal flight states of aircraft. This method defines multivariate feature parameters for attitude anomalies, speed anomalies, trajectory anomalies, and actuator control anomalies in different flight stages of aircraft. It introduces a grid search method to optimize the parameters of the classification decision tree, achieving accurate classification of flight stages. By adding the anomaly influence degree of feature parameters, an isolated forest is constructed to improve the accuracy of abnormal flight state identification. The intelligent method proposed in this invention has important application value in aircraft flight monitoring and is of great significance for improving the intelligent monitoring capabilities of flight safety.

[0006] The present invention provides an intelligent identification method for abnormal flight status of aircraft, characterized by comprising the following steps:

[0007] 1) Historical standard flight data of aircraft are selected by aircraft flight experts and their flight phases are marked;

[0008] Specifically, flight experts select historical standard flight data and label it with six basic flight phases: takeoff, climb, level flight, turn, descent and landing. The labeled flight data set is denoted as D = {D1, D2, ..., D6}.

[0009] 2) Using the labeled dataset D with the labeled flight phases, design and optimize the Classification and Regression Tree (CART) algorithm to construct the flight phase classification model M;

[0010] Specifically:

[0011] Step 1: For the standard maneuver template data D = {D1, D2, ..., D6}, calculate the standard flight state D. r The n-ary feature parameter vector at the u-th time point The formula (1) is used to calculate the Gini coefficient of dataset D by considering the rates of change of altitude Δh, pitch angle Δα, roll angle Δβ, yaw angle Δγ, and heading angle Δδ.

[0012]

[0013] In the formula p k This represents the proportion of the k-th category in dataset D, which is simpler than calculating the logarithm using the information gain ratio. For dataset D, the formula for calculating the Gini coefficient of feature parameter c is (2).

[0014]

[0015] In the formula, |D| represents the number of all time series vectors in dataset D, and E represents the total number of categories of the feature parameter c in dataset D. e | represents the total number of vectors in dataset D where all features c take the value e, Gini(D e The Gini coefficient is calculated according to formula (1).

[0016] Step 2: Based on the feature parameter C with the smallest Gini coefficient value, dataset D is divided into two parts to obtain its left and right nodes, denoted as Di. left and D right ;

[0017] Step 3: There are two conditions for stopping the construction of the decision tree. First, determine the relationship between the threshold and the current Gini coefficient. If the current Gini coefficient is less than the threshold, the current node stops building the decision tree subtree; otherwise, continue building. Second, check if there are any classification features in the feature parameter set D that can further decompose the set. If not, stop building the decision tree.

[0018] Step 4: Analyze the left and right child nodes of node D. left and D right Recursively execute steps 1 to 4 until the condition in step 3 is met, then exit and return to the decision tree to train model M;

[0019] Step 5: To avoid overfitting in the decision tree model, prune the decision tree. Pruning involves deleting non-leaf nodes {T1, T2, T3, ..., T...}. n The α value with the smallest surface error rate gain in} i Corresponding non-leaf node T i The left and right child nodes, α i The calculation formula (3) is as follows

[0020]

[0021] In the formula, R(i) represents the error caused by replacing the i-th non-leaf node with a leaf node, and is calculated using the following formula:

[0022] R(i)=r(i)p(i), (4)

[0023] In the formula, r(i) represents the error rate of node i, p(i) represents the percentage of samples at node i out of the total number of samples in the training set, and R(T) i ) indicates that the subtree T is not pruned when node i is not pruned. i The sum of the errors of all leaf nodes, i.e.

[0024]

[0025] Replace non-leaf nodes T with leaf nodes of the subtree. iRepeat this process until no non-leaf node can be replaced, pruning is complete, and return to the decision tree.

[0026] Step 6: Set the target loss function (accuracy, precision, recall, or F1 score), place all hyperparameters of the decision tree in the network, and use the grid search algorithm to adjust the decision tree parameters to make the decision tree model M classify optimally. At the same time, in order to avoid the impact of the training set and test set split, cross-validation is used to improve the classification performance of the decision tree model M.

[0027] 3) Input the flight data for the abnormal flight status to be identified;

[0028] Specifically: Input the set of aircraft flight data at t time points, denoted as X = {X1, X2, ..., X...} t}

[0029] 4) Use the constructed classification model M to divide the flight dataset X into flight phases;

[0030] Specifically, this involves using a classification model M to analyze the flight data set X = {X1, X2, ..., X...} at t time points. t The system automatically divides flight phases and labels the flight phases for each time period.

[0031] 5) Identify abnormal flight states using the optimized Isolation Forest (IF) algorithm;

[0032] Specifically:

[0033] Step 1: For X, which has been marked with flight phase, calculate the abnormal characteristic parameter values ​​for each flight phase according to Table 1. The set of characteristic parameters for each flight phase is denoted as S.

[0034] Taking the six most basic flight phases—takeoff, climb, level flight, turn, descent, and landing—as an example, each flight phase can be characterized by characteristic parameters such as flight speed, altitude, attitude angle, and heading angle. At the same time, each flight phase corresponds to a different flight control law, and the main characteristic parameters of the flight phase are established.

[0035] The abnormal characteristic parameters are categorized as follows: attitude anomaly, velocity anomaly, trajectory anomaly, and actuator control anomaly. The calculation methods for each type of parameter are as follows:

[0036] (1) Anomaly characteristics parameters

[0037] An abnormal flight attitude of an aircraft refers to a prolonged deviation of the measured values ​​of attitude angles (pitch angle θ, yaw angle ψ, roll angle φ) from given values ​​(given pitch angle θ_ref, given yaw angle ψ_ref, given roll angle φ_ref). The mean deviations of these attitude angles at N time points (E(Δθ), E(Δψ), and E(Δφ)) are calculated as characteristic parameters of the abnormal attitude.

[0038]

[0039] Where, θ i ψ i φ i θ_ref represents the measured values ​​of pitch angle, yaw angle, and roll angle at the i-th time point (i = 1, 2, ..., N), respectively. i ψ_ref i φ_ref i These represent the given values ​​of pitch angle, yaw angle, and roll angle at the i-th (i = 1, 2, ..., N) time point, respectively.

[0040] (2) Velocity anomaly characteristic parameters

[0041] An aircraft speed anomaly refers to a measured speed V deviating from a given speed value V_ref over a prolonged period. The average deviation E(ΔV) of the speed at N time points is calculated as a characteristic parameter of the speed anomaly.

[0042]

[0043] Among them, V i V_ref i These represent the measured and given values ​​of the aircraft's flight speed at the i-th (i = 1, 2, ..., N) time point, respectively.

[0044] (3) Track anomaly characteristic parameters

[0045] An aircraft track anomaly refers to a deviation of the actual flight path from the planned route. The route is composed of several waypoints. The average distance deviation E(Δd) of the waypoints at N time points is calculated as a characteristic parameter of the track anomaly.

[0046]

[0047] Wherein, the measured and given values ​​of the waypoint position's longitude, latitude, and altitude at time i (i = 1, 2, ..., N) are respectively represented as P. i (x i ,y i ,h i ) and P_ref i (x_ref i ,y_refi ,h_ref i ).

[0048] The mean heading angle deviation E(ΔPSI) of waypoints at N time points is calculated as a track anomaly characteristic parameter.

[0049]

[0050] Among them, PSI i PSI_ref i Let represent the measured and given values ​​of the heading angle at time i (i = 1, 2, ..., N), respectively.

[0051] (4) Abnormal control of the execution agency

[0052] An aircraft actuator control anomaly refers to the deviation of actuator displacements (aileron displacement dtx, rudder displacement dty, elevator displacement dtz, flaperon displacement dtjy, nose wheel displacement dtw) from actuator commands (aileron command dtxc, rudder command dtyc, elevator command dtzc, flaperon command dtjyc, nose wheel command dtwc). The mean deviations of the actuator commands at N time points (E(Δdtx), E(Δdty), E(Δdtz), E(Δdtjy), and E(Δdtw)) are calculated as characteristic parameters of the actuator control anomaly.

[0053]

[0054] Among them, dtx i dty i dtz i dtjy i dtw i Let dtxc represent the measured values ​​of aileron displacement, rudder displacement, elevator displacement, flaperon displacement, and nose wheel displacement at the i-th time point (i = 1, 2, ..., N). i dtyc i dtzc i dtjyc i dtwc i These represent the given values ​​for the aileron command, rudder command, elevator command, flaperon command, and nose wheel command at the i-th (i = 1, 2, ..., N) time point, respectively.

[0055] Step 2: Select the feature parameter with the greatest impact of the anomaly from the set of anomaly feature parameters S, and select the data point closest to the mean as the root node of the tree.

[0056] Step 3: Set the abnormal feature parameter values ​​that are less than the current split point as the left node of the binary tree, and set the values ​​that are greater than or equal to the current split point as the right node of the binary tree.

[0057] Step 4: Repeat steps 9 and 10 to recursively construct the tree for nodes Left and Right until only one data point cannot be constructed, or the tree has reached its height limit, at which point construction stops.

[0058] Step 5: Calculate the anomaly score s(x,n) for sample x:

[0059]

[0060] In the formula: h(x) is the path length of the tree containing sample x, and c(n) is the average path length of each tree.

[197] :

[0061]

[0062] In the formula: H(n-1)=ln(n-1)+ζ (ζ is Euler's constant, usually taken as 0.58). For data samples x whose outlier scores s(x,n) are close to 1, they are outlier data points.

[0063] 6) Output time series data of abnormal flight status.

[0064] Specifically, it outputs the flight data set of the aircraft, X = {X1, X2, ..., X...}. t Time series of abnormal flight states in}

[0065] The present invention provides an intelligent identification method for abnormal flight status of aircraft, which can identify abnormal flight status of aircraft, provide decision support for real-time control decisions and health management of aircraft, facilitate aircraft maintenance personnel to find the cause of abnormal events, avoid flight accidents, and ensure flight safety. Attached Figure Description

[0066] Figure 1 This is a flowchart of a method for intelligent identification of abnormal flight states of aircraft. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Example 1

[0069] This embodiment describes an intelligent identification method for abnormal flight states of aircraft. Please refer to [link to relevant documentation]. Figure 1 This includes the following steps:

[0070] 1) Historical standard flight data of aircraft are selected by aircraft flight experts and their flight phases are marked;

[0071] Specifically, flight experts select historical standard flight data and label it with six basic flight phases: takeoff, climb, level flight, turn, descent and landing. The labeled flight data set is denoted as D = {D1, D2, ..., D6}.

[0072] 2) Using the labeled dataset D with well-labeled flight phases, design and optimize the CART (Classification, Regression, Decision Tree) algorithm to construct a flight phase classification model M;

[0073] Specifically:

[0074] Step 1: For the standard maneuver template data D = {D1, D2, ..., D6}, calculate the standard flight state D. r The n-ary feature parameter vector at the u-th time point The formula (1) is used to calculate the Gini coefficient of dataset D by considering the rates of change of altitude Δh, pitch angle Δα, roll angle Δβ, yaw angle Δγ, and heading angle Δδ.

[0075]

[0076] In the formula p k This represents the proportion of the k-th category in dataset D, which is simpler than calculating the logarithm using the information gain ratio. For dataset D, the formula for calculating the Gini coefficient of feature parameter c is (2).

[0077]

[0078] In the formula, |D| represents the number of all time series vectors in dataset D, and E represents the total number of categories of the feature parameter c in dataset D. e | represents the total number of vectors in dataset D where all features c take the value e, Gini(D e The Gini coefficient is calculated according to formula (1).

[0079] Step 2: Based on the feature parameter C with the smallest Gini coefficient value, dataset D is divided into two parts to obtain its left and right nodes, denoted as Di. left and D right ;

[0080] Step 3: There are two conditions for stopping the construction of the decision tree. First, determine the relationship between the threshold and the current Gini coefficient. If the current Gini coefficient is less than the threshold, the current node stops building the decision tree subtree; otherwise, continue building. Second, check if there are any classification features in the feature parameter set D that can further decompose the set. If not, stop building the decision tree.

[0081] Step 4: Analyze the left and right child nodes of node D. left and D right Recursively execute steps 1 to 4 until the condition in step 3 is met, then exit and return to the decision tree training model Tree;

[0082] Step 5: To avoid overfitting in the decision tree model, prune the decision tree. Pruning involves deleting non-leaf nodes {T1, T2, T3, ..., T...}. n The α value with the smallest surface error rate gain in} i Corresponding non-leaf node T i The left and right child nodes, α i The calculation formula (3) is as follows

[0083]

[0084] In the formula, R(i) represents the error caused by replacing the i-th non-leaf node with a leaf node, and is calculated using the following formula:

[0085] R(i)=r(i)p(i), (4)

[0086] In the formula, r(i) represents the error rate of node i, p(i) represents the percentage of samples at node i out of the total number of samples in the training set, and R(T) i ) indicates that the subtree T is not pruned when node i is not pruned. i The sum of the errors of all leaf nodes, i.e.

[0087]

[0088] Replace non-leaf nodes T with leaf nodes of the subtree. i Repeat this process until no non-leaf node can be replaced, pruning is complete, and return to the decision tree.

[0089] Step 6: Set the target loss function (accuracy, precision, recall, or F1 score), place all hyperparameters of the decision tree in the network, and use the grid search algorithm to adjust the decision tree parameters to make the decision tree model M classify optimally. At the same time, in order to avoid the impact of the training set and test set split, cross-validation is used to improve the classification performance of the decision tree model M.

[0090] 3) Input the flight data for the abnormal flight status to be identified;

[0091] Specifically: Input the set of aircraft flight data at t time points, denoted as X = {X1, X2, ..., X...} t}

[0092] 4) Use the constructed classification model M to divide the flight dataset X into flight phases;

[0093] Specifically, this involves using a classification model M to analyze the flight data set X = {X1, X2, ..., X...} at t time points. t The system automatically divides flight phases and labels the flight phases for each time period.

[0094] 5) Use the optimized Isolation Forest algorithm (IF) to identify abnormal flight states;

[0095] Specifically:

[0096] Step 1: For X, which has been marked with flight phase, calculate the abnormal characteristic parameter values ​​for each flight phase according to Table 1. The set of characteristic parameters for each flight phase is denoted as S.

[0097] Taking the six most basic flight phases—takeoff, climb, level flight, turn, descent, and landing—as an example, each flight phase can be characterized by characteristic parameters such as flight speed, altitude, attitude angle, and heading angle. At the same time, each flight phase corresponds to a different flight control law. The main characteristic parameters of the flight phases are established as shown in Table 1.

[0098] Table 1. Characteristic parameters of aircraft during flight phases

[0099]

[0100]

[0101] The abnormal characteristic parameters in the table are categorized as follows: attitude anomaly, velocity anomaly, trajectory anomaly, and actuator control anomaly. The calculation methods for each type of parameter are as follows:

[0102] (1) Anomaly characteristics parameters

[0103] An abnormal flight attitude of an aircraft refers to a prolonged deviation of the measured values ​​of attitude angles (pitch angle θ, yaw angle ψ, roll angle φ) from given values ​​(given pitch angle θ_ref, given yaw angle ψ_ref, given roll angle φ_ref). The mean deviations of these attitude angles at N time points (E(Δθ), E(Δψ), and E(Δφ)) are calculated as characteristic parameters of the abnormal attitude.

[0104]

[0105] Where, θ i ψi φ i θ_ref represents the measured values ​​of pitch angle, yaw angle, and roll angle at the i-th time point (i = 1, 2, ..., N), respectively. i ψ_ref i φ_ref i These represent the given values ​​of pitch angle, yaw angle, and roll angle at the i-th (i = 1, 2, ..., N) time point, respectively.

[0106] (2) Velocity anomaly characteristic parameters

[0107] An aircraft speed anomaly refers to a measured speed V deviating from a given speed value V_ref over a prolonged period. The average deviation E(ΔV) of the speed at N time points is calculated as a characteristic parameter of the speed anomaly.

[0108]

[0109] Among them, V i V_ref i These represent the measured and given values ​​of the aircraft's flight speed at the i-th (i = 1, 2, ..., N) time point, respectively.

[0110] (3) Track anomaly characteristic parameters

[0111] An aircraft track anomaly refers to a deviation of the actual flight path from the planned route. The route is composed of several waypoints. The average distance deviation E(Δd) of the waypoints at N time points is calculated as a characteristic parameter of the track anomaly.

[0112]

[0113] Wherein, the measured and given values ​​of the waypoint position's longitude, latitude, and altitude at time i (i = 1, 2, ..., N) are respectively represented as P. i (x i ,y i ,h i ) and P_ref i (x_ref i ,y_ref i ,h_ref i ).

[0114] The mean heading angle deviation E(ΔPSI) of waypoints at N time points is calculated as a track anomaly characteristic parameter.

[0115]

[0116] Among them, PSI i PSI_ref iLet represent the measured and given values ​​of the heading angle at time i (i = 1, 2, ..., N), respectively.

[0117] (4) Abnormal control of the execution agency

[0118] An aircraft actuator control anomaly refers to the deviation of actuator displacements (aileron displacement dtx, rudder displacement dty, elevator displacement dtz, flaperon displacement dtjy, nose wheel displacement dtw) from actuator commands (aileron command dtxc, rudder command dtyc, elevator command dtzc, flaperon command dtjyc, nose wheel command dtwc). The mean deviations of the actuator commands at N time points (E(Δdtx), E(Δdty), E(Δdtz), E(Δdtjy), and E(Δdtw)) are calculated as characteristic parameters of the actuator control anomaly.

[0119]

[0120] Among them, dtx i dty i dtz i dtjy i dtw i Let dtxc represent the measured values ​​of aileron displacement, rudder displacement, elevator displacement, flaperon displacement, and nose wheel displacement at the i-th time point (i = 1, 2, ..., N). i dtyc i dtzc i dtjyc i dtwc i These represent the given values ​​for the aileron command, rudder command, elevator command, flaperon command, and nose wheel command at the i-th (i = 1, 2, ..., N) time point, respectively.

[0121] Step 2: Select the feature parameter with the greatest impact of the anomaly from the set of anomaly feature parameters S, and select the data point closest to the mean as the root node of the tree.

[0122] Step 3: Set the abnormal feature parameter values ​​that are less than the current split point as the left node of the binary tree, and set the values ​​that are greater than or equal to the current split point as the right node of the binary tree.

[0123] Step 4: Repeat steps 2 and 3 to recursively construct the tree for nodes Left and Right until only one data point cannot be constructed, or the tree has reached its height limit, at which point construction stops.

[0124] Step 5: Calculate the anomaly score s(x,n) for sample x:

[0125]

[0126] In the formula: h(x) is the path length of the tree containing sample x, and c(n) is the average path length of each tree.

[0127]

[0128] In the formula: H(n-1)=ln(n-1)+ζ (ζ is Euler's constant, usually taken as 0.58). For data samples x whose outlier scores s(x,n) are close to 1, they are outlier data points.

[0129] 6) Output time series data of abnormal flight status.

[0130] Specifically, it outputs the flight data set of the aircraft, X = {X1, X2, ..., X...}. t Time series of abnormal flight states in}

[0131] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of abnormal flight states of aircraft, characterized in that... Includes the following steps: 1) Historical standard flight data of aircraft are selected by aircraft flight experts and their flight phases are marked; Specifically, flight experts select historical standard flight data and label it with six basic flight phases: takeoff, climb, level flight, turn, descent and landing. The labeled flight dataset is denoted as D = {D1, D2, ..., D6}. 2) Using the labeled flight phase dataset D, design and optimize the classification, regression, and decision tree algorithm to construct the flight phase classification model M; 3) Input the flight data for the abnormal flight status to be identified; Specifically: Input the set of aircraft flight data at t time points, denoted as X = {X1, X2, ..., X...} t }; 4) Use the constructed classification model M to divide the flight dataset X into flight phases; Specifically, this involves using a classification model M to analyze the flight data set X = {X1, X2, ..., X...} at t time points. t The system automatically divides flight phases and labels the flight phases for each time period. 5) Utilize the optimized isolated forest algorithm to identify abnormal flight states; 6) Output time series data of abnormal flight states; Specifically, it outputs the flight data set of the aircraft, X = {X1, X2, ..., X...}. t Time series of abnormal flight states in}; Step 2) specifically includes the following steps: Step 1: For the dataset D = {D1, D2, ..., D6}, calculate the standard flight state D. r The n-ary feature parameter vector at the u-th time point The formula (1) is used to calculate the Gini coefficient of dataset D by considering the rates of change of altitude Δh, pitch angle Δα, roll angle Δβ, yaw angle Δγ, and heading angle Δδ. In the formula p k The Gini coefficient formula (2) represents the proportion of the k-th category in dataset D. It is simpler than calculating the logarithm using the information gain ratio. In the formula, |D| represents the number of all time series vectors in dataset D, and E represents the total number of categories of the feature parameter c in dataset D. e | represents the total number of vectors in dataset D where all features c take the value e, Gini(D e The Gini coefficient is calculated according to equation (1); Step 2: Based on the feature parameter C with the smallest Gini coefficient value, dataset D is divided into two parts to obtain its left and right nodes, denoted as Di. left and D right ; Step 3: There are two conditions for stopping the construction of the decision tree. First, it is necessary to determine the relationship between the threshold and the current Gini coefficient. If the current Gini coefficient is less than the threshold, the current node stops building the decision tree subtree; otherwise, it continues to build. Second, it is necessary to check whether there are any classification features in the feature parameter dataset D that can further decompose the set. If no such features are found, the construction of the decision tree is stopped. Step 4: Analyze the left and right child nodes D of the dataset D node. left and D right Recursively execute steps 1 to 4 until the condition in step 3 is met, then exit and return to the decision tree to train model M; Step 5: To avoid overfitting in the decision tree model, prune the decision tree. Pruning involves deleting non-leaf nodes {T1, T2, T3, ..., T...}. n The α value with the smallest surface error rate gain in} i Corresponding non-leaf node T i The left and right child nodes, α i The calculation formula (3) is as follows In the formula, R(i) represents the error caused by replacing the i-th non-leaf node with a leaf node, and is calculated using the following formula: R(i)=r(i)p(i), (4) In the formula, r(i) represents the error rate of node i, p(i) represents the percentage of samples at node i out of the total number of samples in the training set, and R(T) i ) indicates that the subtree T is not pruned when node i is not pruned. i The sum of the errors of all leaf nodes, i.e. Replace non-leaf nodes T with leaf nodes of the subtree. i Repeat this process until no non-leaf node can be replaced, pruning is complete, and return to decision tree M; Step 6: Set the target loss function, place all hyperparameters of the decision tree in the network, and use the grid search algorithm to adjust the decision tree parameters to make the decision tree model M optimal for classification. At the same time, in order to avoid the impact of the training set and test set partitioning, cross-validation is used to improve the classification performance of the decision tree model M.

2. The intelligent identification method for abnormal flight status of an aircraft according to claim 1, characterized in that... Step 5) specifically includes the following steps: Step 1: For X, which has been labeled with flight phases, calculate the abnormal characteristic parameter values ​​for each flight phase. The set of characteristic parameters for each flight phase is denoted as S. Taking the six most basic flight phases—takeoff, climb, level flight, turn, descent, and landing—as an example, each flight phase can be characterized by characteristic parameters such as flight speed, altitude, attitude angle, and heading angle. At the same time, each flight phase corresponds to a different flight control law, and abnormal characteristic parameters of the flight phase are established. The abnormal characteristic parameters are categorized as follows: attitude anomaly, velocity anomaly, trajectory anomaly, and actuator control anomaly. The calculation methods for each type of parameter are as follows: (1) Anomaly characteristics parameters An abnormal flight attitude of an aircraft refers to the deviation of the measured values ​​of attitude angles such as pitch angle θ, yaw angle ψ, and roll angle φ from given values. Given a given pitch angle θ_ref, a given yaw angle ψ_ref, and a given roll angle φ_ref, the average deviations E(Δθ), E(Δψ), and E(Δφ) of each attitude angle at N time points are calculated as characteristic parameters of the attitude anomaly. Where, θ i ψ i φ i Let θ_ref represent the measured values ​​of pitch angle, yaw angle, and roll angle at time points i = 1, 2, ..., N, respectively. i ψ_ref i φ_ref i These represent the given values ​​of pitch angle, yaw angle, and roll angle at time points i = 1, 2, ..., N, respectively. (2) Velocity anomaly characteristic parameters An aircraft speed anomaly refers to a deviation of the measured speed V from a given speed value V_ref. The mean deviation E(ΔV) of the speed at N time points is calculated as a characteristic parameter of the speed anomaly. Among them, V i V_ref i These represent the measured and given values ​​of the aircraft's flight speed at time points i = 1, 2, ..., N, respectively. (3) Track anomaly characteristic parameters An aircraft track anomaly refers to a deviation of the actual flight path from the planned route. The route is composed of several waypoints. The average distance deviation E(Δd) of the waypoints at N time points is calculated as a characteristic parameter of the track anomaly. Wherein, the measured and given values ​​of the waypoint position's longitude, latitude, and altitude at times i = 1, 2, ..., N are respectively represented as P. i (x i ,y i ,h i ) and P_ref i (x_ref i ,y_ref i ,h_ref i ); The mean heading angle deviation E(ΔPSI) of waypoints at N time points is calculated as a track anomaly characteristic parameter. Among them, PSI i PSI_ref i Let i represent the measured and given values ​​of the heading angle at time i = 1, 2, ..., N, respectively. (4) Abnormal control of the execution agency An anomalous control of aircraft actuators refers to deviations in actuator displacements (aileron displacement dtx, rudder displacement dty, elevator displacement dtz, flaperon displacement dtjy, nose wheel displacement dtw) from actuator commands (aileron command dtxc, rudder command dtyc, elevator command dtzc, flaperon command dtjyc, nose wheel command dtwc). The mean deviations E(Δdtx), E(Δdty), E(Δdtz), E(Δdtjy), and E(Δdtw) at N time points are calculated as characteristic parameters of the anomalous control. Among them, dtx i dty i dtz i dtjy i dtw i Let dtxc represent the measured values ​​of aileron displacement, rudder displacement, elevator displacement, flaperon displacement, and nose wheel displacement at time points i = 1, 2, ..., N. i dtyc i dtzc i dtjyc i dtwc i These represent the given values ​​for the aileron command, rudder command, elevator command, flaperon command, and nose wheel command at time points i = 1, 2, ..., N, respectively. Step 2: Select the feature parameters with the greatest impact from the abnormal feature parameter set S based on the experience of flight experts, and select the data point closest to the mean as the root node of the tree; Step 3: Set the abnormal feature parameter values ​​that are less than the current split point as the left node of the binary tree, and set the values ​​that are greater than or equal to the current split point as the right node of the binary tree. Step 4: Repeat steps 2 and 3 to recursively construct the tree for nodes Left and Right until only one data point cannot be constructed, or the tree has reached the limit height, then stop construction; Step 5: Calculate the anomaly score s(x,n) for sample x: In the formula: h(x) is the path length of the tree containing sample x, and c(n) is the average path length of each tree. In the formula: H(n-1)=ln(n-1)+ζ, where ζ is Euler's constant, with a value of 0.

58. For data samples x whose outlier scores s(x,n) are close to 1, they are outlier data points.

Images