A data-driven method for identifying aircraft energy anomalies

Through sliding window technology and single-classified SVM model, the balance of false alarm rate and missing alarm rate in aircraft energy abnormality detection is solved, the generalization and accuracy of the identification model are improved, and more accurate energy abnormality judgment and airspace management support are provided. .

CN113869379BActive Publication Date: 2024-11-12NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202111079489.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-15
Publication Date
2024-11-12
Estimated Expiration
2041-09-15

AI Technical Summary

Technical Problem

The prior art is difficult to balance false alarm rate and missed alarm rate in aircraft energy abnormality detection, and lacks consideration of the correlation of energy data events, resulting in insufficient general usage of the identification model.

Method used

Using classification-based anomaly detection method, the aircraft flight data is recombined in chronological order through sliding window technology, energy characteristic vectors in the time window are generated, and energy abnormality detection model is constructed using a single classification support vector machine (SVM) and combined with data-driven The method improves the generality and recognition capabilities of the model.

Benefits of technology

It improves the generality and recognition accuracy of the aircraft energy abnormality recognition model, can more accurately judge aircraft energy abnormalities, provide timely and effective auxiliary decision-making information, and enhances the operational status analysis capabilities of airspace management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113869379B_ABST
    Figure CN113869379B_ABST
Patent Text Reader

Abstract

The present invention discloses a data-driven aircraft energy anomaly identification method. First, the aircraft trajectory related data is read and the data is analyzed for quality; second, the aircraft trajectory information data is preprocessed and divided into a training set, a test set and a verification set; then, according to the energy formula, the data in the ENU coordinate system is converted into energy data, and data in units of time windows are generated according to the time data and the time window length; finally, a single-classification SVM is used to construct an energy anomaly detection model, and the energy anomaly detection model is trained through historical flight data to enable it to have the ability to identify abnormal energy. The present invention improves the versatility of the recognition model, and can obtain relevant statistical data in different terminal airspaces, provide timely and effective auxiliary decision-making information for controllers, and provide airspace management departments with airspace operation status analysis to help airspace planning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of civil aviation technology, and in particular relates to a data-driven aircraft energy anomaly identification method. Background Art

[0002] In the past ten years, all aspects of my country's civil aviation industry have been growing rapidly, and the selection rate of airports, airlines and passengers has been significantly improved. Behind the rapid development of the civil aviation industry, the problems of safety and efficiency have also emerged. Limited airspace resources, the accuracy of ground or airborne navigation equipment, the command ability of controllers, etc. cannot meet the efficient operation of flights under large traffic. Therefore, in order to ensure safety, some busy airspaces adopt flow restriction, rerouting, waiting and other methods to alleviate the traffic flow pressure in the airspace, which easily leads to the emergence of problems such as flight delays and low operating efficiency. To this end, in view of the above problems, the present invention develops a system for assisting controllers in commanding aircraft in the control area, providing controllers with danger warnings and auxiliary decision-making information, and improving the ability of controllers and the operating efficiency and safety of air traffic flows.

[0003] Aircraft energy management and energy situational awareness are two important concepts in safety analysis at this stage. Many aircraft out-of-control accidents and incidents are caused by the aircraft's failure to do a good job of real-time energy management. Energy indicators can also be an objective indicator for evaluating safety. Therefore, we define abnormal energy that may cause accidents or incidents as aircraft energy indicators or energy characteristics that are significantly different from the energy data of most other flight behaviors. At present, there are some studies on aircraft abnormal behavior detection at home and abroad. Aircraft can be divided into four main parts: 1) Abnormal behavior detection based on distance; 2) Abnormal behavior detection based on classification; 3) Abnormal behavior detection based on reconstruction; 4) Abnormal behavior detection based on statistics.

[0004] The present invention adopts a classification-based detection method by comparing different anomaly detection methods to better solve the problem of imbalance between normal and abnormal aircraft energy data. At present, the classification-based anomaly detection methods all use a single data point as the detection target, which will lack consideration of the data around the detection point, making it difficult for the model to balance the false alarm rate and the missed alarm rate and ignore the event correlation of energy data. The present invention uses sliding window technology to reassemble the original aircraft flight data in chronological order to generate a time window, calculates the energy indicators of different windows to generate feature vectors, and determines whether there is an energy anomaly at the observation time point by making a comprehensive anomaly judgment on the data in the window. Summary of the invention

[0005] Purpose of the invention: In view of the above-mentioned problems existing in the dynamic time warping method, the present invention proposes a data-driven aircraft energy anomaly identification method to improve the versatility of the identification model and obtain relevant statistical data in different terminal airspaces.

[0006] Technical solution: A data-driven aircraft energy anomaly identification method according to the present invention comprises the following steps:

[0007] (1) Read aircraft trajectory data and perform quality analysis on the data;

[0008] (2) Preprocess the aircraft trajectory information data and divide it into training set, test set and validation set;

[0009] (3) According to the energy formula, the data in the ENU coordinate system is converted into energy data, and data in units of time windows is generated according to the time data and the length of the time window;

[0010] (4) A single-class SVM is used to build an energy anomaly detection model, which is then trained using historical flight data to enable the model to identify abnormal energy.

[0011] Furthermore, the step (1) comprises the following steps:

[0012] (11) Reading aircraft trajectory data including time, secondary radar number, aircraft model, aircraft longitude, aircraft latitude, aircraft altitude, aircraft speed, and aircraft angle;

[0013] (12) Perform data quality analysis from three aspects: whether there are duplicate values, missing values, and outliers in the data: compare all the attributes of all track points in the aircraft track information dataset. When all the attributes between two track points are the same, the two track points are judged to be duplicate values; check the attribute values ​​of the track points. When there are null values, the track information is missing values; outliers are single data points that obviously deviate from the flight trajectory. This point may be caused by sensor anomalies or data transmission anomalies.

[0014] Furthermore, the step (2) comprises the following steps:

[0015] (21) Data cleaning of aircraft trajectory information dataset: duplicate values ​​in the data are deleted by using the deletion method, and data records of duplicate trajectory information are eliminated; missing values ​​in the data are deleted and interpolated by using the deletion method;

[0016] (22) Perform coordinate transformation on the aircraft trajectory information dataset, converting the geographic coordinate system into the ECEF coordinate system with the Earth's center of mass as the origin;

[0017] (23) The ECEF coordinate system is converted into the ENU coordinate system with the airport reference point as the origin. The ZUp axis direction of the ENU coordinate system is perpendicular to the earth surface where the airport reference point is located, the XEast axis direction is due east, and the YNorth direction is due north;

[0018] (24) Interpolate, smooth, and extrapolate the trajectory to obtain smooth, equal-length data:

[0019]

[0020] (D 2 ) i =e i -2e i+1 +e i+2

[0021] (D 3 ) i =(-e i-1 +2e i -2e i+2 +e i+3 )

[0022] Where P∈R N×M Reconstruct the track, the length is N, the data dimension is M, Represents the original trajectory, with a length of K and a data dimension of M, A∈R K×N represents a diagonal sampling matrix, if There are missing values ​​in , and the corresponding value in matrix A is 0. represents the F norm, β 1 ,β 2 represents the regularization parameter, D 2 ∈R (N-2)×N is the second-order difference matrix representing the acceleration of the aircraft trajectory, D 3 ∈R (N-4)×N Indicates suppressing small high-frequency fluctuations in the fitting trajectory, e i Represents the standard unit vector at time i=1,2,...,N.

[0023] Furthermore, the step (3) comprises the following steps:

[0024] (31) selecting energy indicators, and converting the preprocessed data into energy indicators through an energy calculation formula, and generating data in units of time windows from the processed energy indicators in chronological order; the energy indicators include unit mechanical energy, unit potential energy, unit kinetic energy, unit mechanical energy error, and unit mechanical energy rate;

[0025] (32) Generate aircraft energy data in time windows:

[0026] The sliding window is used to reassemble the original aircraft flight data in chronological order to generate a time window, and the energy index of different windows is calculated to generate a feature vector:

[0027]

[0028] The flight data sequence {x t-k ,...,x t ,...,x t+k} to observe the aircraft flight data at time t; each time window is used as an analysis unit, and the corresponding energy formula eq is used j Calculate the energy index corresponding to each time point E j represents the jth energy characteristic of the aircraft in the time window, Represents the jth energy feature of the aircraft at time t, calculates different energy features, and constructs the energy feature vector y of the i-th energy window i =[E 1 i ,E 2 i ,E 3 i ,...,E N i ]; perform feature scaling on the energy feature vectors within the energy window so that the mean and unit variance of each energy index are zero, so that all energy indexes can be compared equally and avoid limiting the model performance due to dimension or unit reasons.

[0029] Furthermore, the step (4) comprises the following steps:

[0030] (41) Use single-class SVM to build an energy anomaly detection model to improve the matching degree between the model and the dataset. Define the training dataset {x 1 ,x 2 ,...,x n}∈R, R is the feature space where the data originally exists, and the mapping function φ 1 Map all training sets from feature space R to feature space F, and split the hyperplane F 1 , the slack variable ζ i , hyperplane intercept ρ, hyperplane normal vector w, the ratio of allowed abnormal data to all data υ, Lagrange multiplier λ 1 ,λ 2 , k(·) kernel function, used to calculate the inner product of the input mapping to the energy features of the high-dimensional space, SV is the support variable, and the variable set defines the separating hyperplane;

[0031] (42) The single-class support vector machine projects all the data into a high-dimensional feature space F through a kernel function, and solves the hyperplane F that is farthest from the origin of a normal data set. 1 ; Single-class SVM model adds slack variable ζ i To prevent overfitting during training, the parameter υ is set to allow a small number of samples to be outside the segmentation hyperplane; the parameter υ is a variable that matches the data set and the model. When the parameter υ is small, the sensitivity of the model is lower, and it is easy to identify abnormal energy data as normal; conversely, the higher the sensitivity of the model, the easier it is to mistakenly identify normal energy data as abnormal;

[0032] (43) The energy anomaly detection model calculates the normal area of ​​the training set data in the feature space. If the data of the input test set is outside the segmentation hyperplane, that is, outside the normal area, then -1 is returned and the sample is judged as abnormal. If it is within the segmentation hyperplane, then 1 is returned and the sample is judged as normal.

[0033] When the data cannot be directly separated by a "linear" hyperplane in the original feature space R, the mapping function φ 1 Map the data to a higher-dimensional feature space F, and construct a segmentation hyperplane in F by adjusting the distance and angle of the sample in F, y i =[E 1 i ,E 2 i ,E 3 i ,...,E N i ] indicates the energy characteristic data after preprocessing:

[0034] w·φ(y)-ρ=0

[0035] By optimizing the values ​​of w and ρ, we obtain a plane farthest from the origin and transform the problem into a quadratic optimization problem:

[0036]

[0037] w·φ(y i )-ρ+ζ i ≥0,ζ i ≥0

[0038] The conditional extreme value is calculated by Lagrange multiplication:

[0039]

[0040]

[0041]

[0042]

[0043] By using the Lagrangian technique and the kernel function of the dot product calculation, the decision function is:

[0044]

[0045]

[0046] when When , the sample data is outside the separation hyperplane and is considered to be abnormal data; when The sample data is on the separating hyperplane and is considered normal data; when λ 1 = 0, the sample data is in the separating hyperplane;

[0047] According to the λ corresponding to the support vector on the separating hyperplane 1 Reset the intercept ρ:

[0048]

[0049] y j ∈SV

[0050] Then use the sgn function to judge the abnormality:

[0051]

[0052] When the return value is +1, the sample data is normal data; when the return value is -1, the sample data is abnormal data.

[0053] Furthermore, the size of the sliding window in step (32) is 5.

[0054] Furthermore, the kernel function mapped in the single-class support vector machine is an RBF radial basis kernel function.

[0055] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention takes into account the time correlation of flight data, accurately characterizes data features, and improves the versatility of the recognition model; in view of the imbalance problem in which the data set is mainly composed of normal data, a single-classification SVM model is built, and the model is trained by training the training set data to find the optimal separation hyperplane to determine whether the data is normal; 2. The present invention has data-driven characteristics, can generate models from historical data, and automatically create generation models for any terminal airspace; and can obtain relevant statistical data in different terminal airspaces, provide controllers with timely and effective auxiliary decision-making information, and provide airspace management departments with airspace operation status analysis to assist in airspace planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 Schematic diagram of coordinate system, where (a) is the geographic coordinate system; (b) is the ECEF coordinate system; (c) is the ENU coordinate system;

[0058] Figure 3 Schematic diagram of time window, where (a) shows the time window including time t and k time points before and after; (b) shows the time window division of real-time analysis, which is different from post-analysis in that the flight data of the aircraft at the last flight data at the analysis time point is analyzed;

[0059] Figure 4 Schematic diagram for calculating the separating hyperplane;

[0060] Figure 5 Schematic diagram of kernel function projection onto hyperplane;

[0061] Figure 6 This is a comparison chart of the effectiveness of different energy detection models. DETAILED DESCRIPTION

[0062] The present invention is further described in detail below with reference to the accompanying drawings.

[0063] The present invention proposes a data-driven method for identifying aircraft energy anomalies. Figure 1 As shown, the following steps are included:

[0064] Step 1: Read the aircraft trajectory data and perform quality analysis on the data. The specific steps are as follows:

[0065] Step 1.1: Read the aircraft trajectory data, including time, secondary radar number, aircraft model, aircraft longitude, aircraft latitude, aircraft altitude, aircraft speed and aircraft angle.

[0066] Step 1.2: Perform quality analysis on the read data, and analyze whether there are duplicate values, missing values, and outliers in the data. Compare all the attributes of all track points in the aircraft track information data set. When all the attributes between two track points are the same, the two track points are judged to be duplicate values; check the attribute values ​​of the track points. When there are null values, the track information is missing values; outliers are single data points that obviously deviate from the flight trajectory. This point may be caused by sensor anomalies or data transmission anomalies.

[0067] Step 2: Preprocess the aircraft trajectory information data and divide it into training set, test set and validation set. The specific steps are as follows:

[0068] Step 2.1: Clean the aircraft trajectory information data set. The duplicate values ​​in the data are deleted by deletion, and the data records of duplicate track information are eliminated; the missing values ​​in the data are deleted and interpolated. When the missing attribute of the data is greater than 85%, the information data of the track point is deleted, otherwise the missing values ​​are interpolated by regression method based on other non-missing variables; the outliers in the data are replaced by regularization algorithm, and the flight trajectory is made smoother. The cleaned data is reordered in chronological order to obtain the aircraft trajectory information data in the time series.

[0069] Step 2.2: If Figure 2 As shown in the figure, the aircraft trajectory information dataset is transformed into the geographic coordinate system, such as Figure 2 (a) Convert to the ECEF coordinate system with the Earth's center of mass as the origin, such as Figure 2 (b) Convert the ECEF coordinate system into the ENU coordinate system with the airport reference point as the origin, such as Figure 2 (c).

[0070] The specific calculation formula is as follows:

[0071]

[0072]

[0073] X=(N+h)×cosφ×cosλ (3)

[0074] Y=(N+h)×cosφ×sinλ (4)

[0075] Z=[N×(1-e 2 )+h]×sinφ (5)

[0076] In the formula, a represents the radius of the earth's major axis, which is 6378137m, b represents the radius of the earth's minor axis, which is 6356752.3142m. e represents the eccentricity of the earth, N represents the radius of curvature when the latitude is φ, λ represents the longitude, in degrees, and x, y, z are the coordinate values ​​of the X, Y, and Z axes corresponding to the ECEF coordinate system.

[0077] Step 2.3: Convert the ECEF coordinate system to the ENU coordinate system with the airport reference point as the origin. The ZUp axis of the ENU coordinate system is perpendicular to the earth's surface where the airport reference point is located, the XEast axis is due east, and the YNorth axis is due north. The conversion formula is as follows:

[0078] x o ,y o ,z o =ECEF(φ o,λ o ,h o ) (6)

[0079] x,y,z=ECEF(φ,λ,h) (7)

[0080] XEast=-sinλ o ×(xx o )-cosλ o ×(yy o ) (8)

[0081] YNorth=-cosλ o ×sinφ o ×(xx o )-sinφ o ×sinλ o ×(yy o )+cos(φ o ×(zz o ))(9)

[0082] ZUp=cosλ o ×cosφ o ×(xx o )-sinφ o ×cosλ o ×(yy o )+sin(φ o ×(zz o )) (10)

[0083] Where λ o ,φ o ,h o Indicates the latitude, longitude, and altitude of the airport reference point. o ,y o ,z o It represents the coordinates of the airport reference point in the ECEF coordinate system, ECEF(·) represents the coordinate conversion to the ECEF coordinate system, and XEast, YNorth, and ZUp represent the three-dimensional coordinates in the ENU coordinate system.

[0084] Step 2.4: Reconstruct the trajectory. This patent uses a regularization algorithm to reconstruct the trajectory, making the trajectory smoother and closer to the real trajectory.

[0085] By solving the (least squares method) optimization problem of equation (25), the trajectory is interpolated, smoothed, and extrapolated to obtain smooth, equal-length data.

[0086]

[0087] Where P∈R N×MReconstruct the track with length N and data dimension M. Represents the original trajectory, with a length of K and a data dimension of M. A∈R K×N represents a diagonal sampling matrix, if If there are missing values ​​in , the corresponding values ​​in matrix A are 0. Denotes the F norm. β 1 ,β 2 represents the regularization parameter. D 2 ∈R (N-2)×N is the second-order difference matrix representing the acceleration of the aircraft trajectory. 3 ∈R (N-4)×N Indicates the suppression of small high-frequency fluctuations in the fitting trajectory. i Represents the standard unit vector at time i=1,2,...,N.

[0088] (D 2 ) i =e i -2e i+1 +e i+2 (12)

[0089] (D 3 ) i =(-e i-1 +2e i -2e i+2 +e i+3 ) (13)

[0090] Step 3: According to the energy formula, the data in the ENU coordinate system is converted into energy data, and data in units of time windows is generated according to the time data and the length of the time window.

[0091] Step 3.1: First, select the energy index, and convert the data processed in step 2 into the energy index through the energy calculation formula, and generate data in time windows according to the time sequence of the processed energy index.

[0092] Energy indicator selection:

[0093] 1) Specific Mechanical Energy (SME), that is, the mechanical energy per unit mass of the aircraft (the sum of potential energy and kinetic energy):

[0094]

[0095] Among them, m represents the mass of the aircraft, W represents the weight of the aircraft, h represents the corrected sea pressure altitude of the aircraft, v represents the true airspeed of the aircraft, and g represents the local gravitational acceleration.

[0096] The physical meaning of unit mechanical energy is: if the drag and thrust are balanced during variable speed climb, the theoretical altitude that the aircraft can reach when all its kinetic energy is converted into potential energy. If no external force does work on the aircraft, the energy altitude will not change. If the drag is too large or a sudden external force does negative work, the unit mechanical energy will decrease. Therefore, its change can reflect the flight status of the aircraft and can also serve as an important indicator of whether the aircraft is flying normally.

[0097] 2) Specific Potential Energy (SPE), that is, the potential energy contained in the unit mass of the aircraft:

[0098]

[0099] Where m represents the mass of the aircraft, W represents the weight of the aircraft, h represents the corrected sea pressure altitude of the aircraft, and g represents the local acceleration of gravity.

[0100] Unit potential energy represents the altitude of the aircraft. For an approaching aircraft, since its flying altitude is relatively low, the normal change of altitude is the most important part of the normal operation of the aircraft.

[0101] 3) Specific Kinetic Energy (SKE), that is, the kinetic energy contained in the unit mass of the aircraft:

[0102]

[0103] Among them, m represents the mass of the aircraft, W represents the weight of the aircraft, and g represents the local gravitational acceleration.

[0104] Unit kinetic energy represents the ratio of the aircraft's speed to twice the acceleration of gravity. For an approaching aircraft, if the unit kinetic energy is too large, it may easily run off the runway, while if the unit kinetic energy is too low, it may easily cause unsafe incidents and accident symptoms such as the aircraft landing before the runway.

[0105] 4) Specific Mechanical Energy Error (SMEE), which is the difference between the mechanical energy per unit mass of the aircraft and the mechanical energy corresponding to the typical approach profile:

[0106]

[0107] Among them, SME act It is the unit mechanical energy corresponding to the aircraft during the actual flight process, SME ref It is the unit mechanical energy corresponding to the aircraft's theoretical approach flight profile.

[0108] The unit mechanical energy error intuitively represents the difference between the aircraft's mechanical energy data during actual flight and the pre-designed approach and landing procedure. It provides a typical comparison data to determine whether the aircraft is abnormal.

[0109] 5) Specific Mechanical Energy Rate (SMER), that is, the rate of change of unit mechanical energy over time:

[0110]

[0111] Among them, d(·) represents the first-order derivative, t represents time, h' represents the rate of change of the corrected sea pressure altitude, and v' represents the rate of change of velocity, i.e., acceleration.

[0112] Under the condition of stable external environment, the unit mechanical energy rate is proportional to the size of the work done by the external force of the aircraft. Therefore, for an aircraft in the final approach phase, the throttle is in the slow position, and the unit mechanical energy rate is only related to the drag. Therefore, the unit mechanical energy rate fully reflects the changing trend of the aircraft's drag, which provides a method for abnormal energy detection and real-time prediction of abnormal trajectories of aircraft.

[0113] The energy index ignores the individual differences of the aircraft itself, such as the same model, different fuel or passenger load, etc. In the calculation process, the aircraft weight is not used, and only the energy index abnormality detection of the corresponding aircraft model classification is required. Therefore, the application of energy index is more extensive and convenient, so the invention adopts energy index as an important feature of aircraft energy detection.

[0114] Step 3.2: Generate aircraft energy data in time windows.

[0115] During the flight of an aircraft, different flight missions are performed at different stages. Therefore, the flight data features corresponding to different flight stages are also different, so the flight data features have a strong time correlation. In order to better detect and extract data features, the present invention uses sliding window technology to reassemble the original aircraft flight data in chronological order to generate time windows, and calculates the energy indexes of different windows to generate feature vectors.

[0116]

[0117] The data conversion process is:

[0118]

[0119] For post-analysis, the flight data sequence {x t-k ,...,x t ,...,xt+k} to observe the aircraft flight data at time t, such as Figure 3 The time window shown in (a) includes time t and k time points before and after it. Each time window is used as an analysis unit, and the energy index corresponding to each time point is calculated. E j represents the jth energy characteristic of the aircraft in the time window, Represents the jth energy feature of the aircraft at time t, calculates different energy features, and constructs the energy feature vector y of the i-th energy window i =[E 1 i ,E 2 i ,E 3 i ,...,E N i ]. We can get a complete trajectory from historical data. The analysis of a time point can refer to the previous and next time points. Based on the flight data of a short time period, the energy state of the aircraft at the analysis time point is evaluated, ensuring the time correlation of the aircraft data. The difference between real-time analysis and post-analysis is that the analysis time point is based on the last flight data to analyze the flight data of the aircraft at the time point, such as Figure 3 (b), the time corresponding to the leftmost data point in the figure is the observation time point, and the time window is 2k time points after this time point. The energy feature vectors corresponding to different time windows are calculated, where the dimensions and units of different energy indicators are different and vary greatly, such as the units and dimensions of unit mechanical energy and unit mechanical energy rate. Therefore, in order to make all energy indicators equally comparable and avoid limiting the model performance due to dimensions or units, feature scaling is required to make the mean and unit variance of each energy indicator zero.

[0120] Step 4: Use a single-class SVM to build an energy anomaly detection model, and train the energy anomaly detection model through historical flight data to enable it to have the ability to identify abnormal energy.

[0121] Aircraft operation anomaly detection can be simply regarded as a classification problem, which classifies normal and abnormal data. However, unlike previous studies, the size of the normal data set in the aviation data set is much larger than the size of the abnormal data set. In view of the extreme imbalance between normal and abnormal data sets, the present invention uses a single-class vector machine to construct an energy anomaly detection model to improve the matching degree between the model and the data set, thereby improving the detection accuracy and efficiency.

[0122] Step 4.1: Use single-class SVM to build an energy anomaly detection model to improve the matching degree between the model and the dataset. Define the training dataset {x 1 ,x2 ,...,x n}∈R, R is the feature space where the data originally exists, and the mapping function φ 1 Map all training sets from feature space R to feature space F, and split the hyperplane F 1 , the slack variable ζ i , hyperplane intercept ρ, hyperplane normal vector w, the ratio of allowed abnormal data to all data υ, Lagrange multiplier λ 1 ,λ 2 , k(·) kernel function, used to calculate the inner product of the input mapping to the high-dimensional space energy feature, y i =[E 1 i ,E 2 i ,E 3 i ,...,E N i ] represents the energy feature data after preprocessing, SV is the supporting variable, and this variable set defines the separating hyperplane.

[0123] Step 4.2: First, the single-class support vector machine projects all the data into a high-dimensional space F through the kernel function, and solves the hyperplane F farthest from the origin of a normal data set. 1 ,like Figure 4 The model adds a slack variable ζ i To prevent overfitting during training, the parameter υ is set to allow a small number of samples to be outside the segmentation hyperplane. The larger the parameter υ, the more sensitive the model is but the less normal data it contains. If υ is too large, normal data may be mistakenly identified as abnormal.

[0124] Step 4.3: The energy anomaly detection model calculates the normal area of ​​the training set data in the feature space. If the data of the input test set is outside the segmentation hyperplane, that is, outside the normal area, a -1 sample is returned to be judged as abnormal. If it is within the segmentation hyperplane, a 1 sample is returned to be judged as normal.

[0125] When the data cannot be directly separated by a "linear" hyperplane in the original feature space R, we can use the mapping function φ 1 Map the data to a higher-dimensional feature space F, and construct a segmentation hyperplane in F by adjusting the distance and angle of the sample in F, such as Figure 5 .

[0126] The formula of the hyperplane is:

[0127] w·φ(y)-ρ=0 (20)

[0128] y i =[E 1i ,E 2 i ,E 3 i ,...,E N i ] represents the energy characteristic data after preprocessing. By optimizing the values ​​of w and ρ, a plane farthest from the origin is obtained, so the problem is transformed into a quadratic optimization problem. In formula (21), m = n-2k+1 represents the number of windows:

[0129]

[0130] w·φ(y i )-ρ+ζ i ≥0,ζ i ≥0 (21)

[0131] The conditional extreme value is calculated by Lagrange multiplication:

[0132]

[0133]

[0134]

[0135]

[0136] By using the Lagrangian technique and the kernel function of the dot product calculation, the decision function becomes (26)

[0137]

[0138] From formula (26), we can conclude that When , the sample data is outside the separation hyperplane and is considered to be abnormal data; when The sample data is on the separating hyperplane and is considered normal data. 1 = 0 means the sample data is in the separating hyperplane.

[0139] According to the λ corresponding to the support vector on the separating hyperplane 1 Reset the intercept (offset) ρ:

[0140]

[0141] Substitute (27) into (26), and then use the sgn function to determine the anomaly:

[0142]

[0143] When the return value of (28) is +1, the sample data is normal data; when the return value of (28) is -1, the sample data is abnormal data.

[0144] Step 4.3: Parameter setting and optimization.

[0145] The present invention fully considers the abnormal conditions caused by the data transmitted by sensors in the actual operation of the aircraft. For example, the pitot tube assembly (ADM) may be intermittently unstable, resulting in unreliable airspeed. Through a quick query, it is checked that the single airspeed is unreliable for more than five consecutive seconds before switching to the backup system. In addition, a large sliding window dimension will affect the model detection performance, resulting in an increase in the model data leakage rate. Therefore, the size of the sliding window is set to 5.

[0146] In the single-classification SVM, it is also necessary to define parameter υ and the kernel function of the selected mapping. The RBF radial basis kernel function can reduce the computational difficulty of the numerical value and has a strong anti-interference ability to the noise of the data, so the present invention selects the radial basis kernel function as the mapping function. Parameter υ is the slack variable of the single-classification SVM. The smaller the parameter υ, the lower the sensitivity of the model, and the easier it is to identify the abnormal energy data as normal. On the contrary, the higher the sensitivity of the model, the easier it is to misidentify the abnormality of the normal energy data. The present invention is based on the development principle of civil aviation with safety first. In the selection of parameters, the false alarm rate is first guaranteed to be 0, and then the parameters with the minimum false alarm rate are selected.

[0147] Based on the existing ADS-B data of Guangzhou Baiyun Airport, the departing aircraft, approaching aircraft, and aircraft flying over the terminal area with a radius of 50km and an altitude below 6000m with the airport reference point as the center are screened and classified. The data includes aircraft registration number, time, longitude, latitude, pressure altitude, speed, and heading angle.

[0148] Since ADS-B data is missing and repeated during the transmission process, the missing data in the aircraft data is supplemented by difference and the repeated data is directly deleted. The aircraft position data in ADS-B is based on the earth coordinate system. Using the earth coordinate system data will increase the complexity of the formula and is not conducive to visual display. The aircraft position in the earth coordinate system is converted into the ECEF coordinate system with the center of mass of the earth as the origin, and then the ECEF coordinate system is converted into the ENU coordinate system with the airport reference point as the origin. In the ENU coordinate system, the positive direction of the ZUp axis is perpendicular to the earth surface where the airport reference point is located, the XEast axis direction is due east, and the YNorth direction is due north. The regularization algorithm is used to reconstruct the trajectory, making the trajectory smoother and closer to the real trajectory.

[0149] Generate energy data with time window as the smallest unit. Substitute the above energy formula into the energy data at each moment. i, i∈1,2,...,n} is combined in chronological order with a time window length K as the step length to obtain the data combination {X i ,X i+1 ,...,X i+k-1},i∈1,2,...,n-k+1.

[0150] The trajectory is matched with the nearest runway based on the distance from the first data point at takeoff or the last data point at landing to the center point. Since the single-classification support vector machine is an unsupervised algorithm, 60% of the flight data is selected as the training set, 20% of the flight data is used as the test set, and the last 20% of the flight data is used as the validation set. 50 flight data samples containing abnormal energy are added to the validation set and the test set, respectively. The abnormal data samples are obtained by calibration by air traffic controllers.

[0151] In the single-classification SVM, only the parameter υ and the time window length K need to be determined. The present invention balances the drawbacks of short-term data anomalies caused by abnormal sensor data and increased false alarm rate caused by excessive time window dimension, and selects the time window length 5 according to the time of switching the backup system quickly and simply based on the abnormal aircraft flight data. The parameter υ is related to the quality of the data. When the noise of the training sample data is large, a larger υ is selected to filter the abnormal data. On the contrary, when the noise of the training sample data is small, a smaller υ is selected to avoid reducing the ability of the model to identify normal data, resulting in reduced practicality of the model. However, based on the basic idea of ​​safety first, the parameter υ must first ensure that all abnormalities in the test set are detected, that is, the false alarm rate is 0, and then consider other parameter indicators. Finally, after the validation set data test, if the model performance indicators all reach the preset threshold and are optimal, the model parameters are determined to complete the model training.

[0152] Taking the terminal area of ​​Guangzhou Baiyun Airport as an example, flight data from September 2018 to February 2019 were used, including a total of 65,000 trajectory data, of which 55,000 trajectories were used to train the model and 10,000 trajectories were used for experimental comparison. The present invention uses the same training set, test set, and validation set for training and parameter selection. Finally, the radar chart is drawn with the mean of the four runway model detection indicators AUC, false alarm rate, missed alarm rate, precision rate, recall rate, and F score, as shown in the figure. Figure 6 .from Figure 6 It can be seen that the single-classification support vector machine used in the present invention performs outstandingly in detecting aircraft energy anomalies, and its missed alarm rate and false alarm rate are lower than those of other detection models, and its AUC value, precision rate, recall rate, and F score are higher than those of the other three detection models. Therefore, the abnormal energy detection model proposed in the present invention has the best detection performance.

Claims

1. A data-driven aircraft energy anomaly identification method, characterized in that: The following steps are involved: (1) Read aircraft trajectory data and perform quality analysis on the data; (2) Preprocess the aircraft trajectory information data and divide it into training set, test set and validation set; (3) According to the energy formula, the data in the ENU coordinate system is converted into energy data, and data in units of time windows is generated according to the time data and the length of the time window; (4) Use single-class SVM to build an energy anomaly detection model, and train the energy anomaly detection model through historical flight data to enable it to have the ability to identify abnormal energy; The step (2) comprises the following steps: (21) Data cleaning of aircraft trajectory information dataset: duplicate values ​​in the data are deleted by using the deletion method, and data records of duplicate trajectory information are eliminated; missing values ​​in the data are deleted and interpolated by using the deletion method; (22) Perform coordinate transformation on the aircraft trajectory information dataset, converting the geographic coordinate system into the ECEF coordinate system with the Earth's center of mass as the origin; (23) The ECEF coordinate system is converted into the ENU coordinate system with the airport reference point as the origin. The ZUp axis direction of the ENU coordinate system is perpendicular to the earth surface where the airport reference point is located, the XEast axis direction is due east, and the YNorth direction is due north; (24) Interpolate, smooth, and extrapolate the trajectory to obtain smooth, equal-length data: (D2) i =e i -2e i+1 +e i+2 <h2 style=";text-align:left;direction:ltr">(D3)<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> (-e)<h2 style=";text-align:left;direction:ltr"> i-1 <h2 style=";text-align:left;direction:ltr"> +2e<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> -2e<h2 style=";text-align:left;direction:ltr"> i+2 <h2 style=";text-align:left;direction:ltr"> +e<h2 style=";text-align:left;direction:ltr"> i+3 <h2 style=";text-align:left;direction:ltr"> ) Where P∈R N×M Reconstruct the track, the length is N, the data dimension is M, Represents the original trajectory, with a length of K and a data dimension of M, A∈R K×N represents a diagonal sampling matrix, if There are missing values ​​in , and the corresponding value in matrix A is 0. represents the F norm, β1, β2 represent regularization parameters, D2∈R (N-2)×N is the second-order difference matrix representing the acceleration of the aircraft trajectory, D3∈R (N-4)×N Indicates suppressing small high-frequency fluctuations in the fitting trajectory, e i Represents the standard unit vector at time i=1,2,...,N.

2. The data-driven aircraft energy anomaly identification method according to claim 1, characterized in that: The step (1) comprises the following steps: (11) Reading aircraft trajectory data including time, secondary radar number, aircraft model, aircraft longitude, aircraft latitude, aircraft altitude, aircraft speed, and aircraft angle; (12) Perform data quality analysis from three aspects: whether there are duplicate values, missing values, and outliers in the data: compare all the attributes of all track points in the aircraft track information dataset. When all the attributes between two track points are the same, the two track points are judged to be duplicate values; check the attribute values ​​of the track points. When there are null values, the track information is missing values; outliers are single data points that obviously deviate from the flight trajectory. This point may be caused by sensor anomalies or data transmission anomalies.

3. The data-driven aircraft energy anomaly identification method according to claim 1, characterized in that: The step (3) comprises the following steps: (31) selecting energy indicators, and converting the preprocessed data into energy indicators through an energy calculation formula, and generating data in units of time windows from the processed energy indicators in chronological order; the energy indicators include unit mechanical energy, unit potential energy, unit kinetic energy, unit mechanical energy error, and unit mechanical energy rate; (32) Generate aircraft energy data in time windows: The sliding window is used to reassemble the original aircraft flight data in chronological order to generate a time window, and the energy index of different windows is calculated to generate a feature vector: The flight data sequence {x t-k ,...,x t ,...,x t+k } to observe the aircraft flight data at time t; each time window is used as an analysis unit, and the corresponding energy formula eq is used j Calculate the energy index corresponding to each time point E j represents the jth energy characteristic of the aircraft in the time window, Represents the jth energy feature of the aircraft at time t, calculates different energy features, and constructs the energy feature vector y of the i-th energy window i =[E1 i ,E2 i ,E3 i ,...,E N i ]; perform feature scaling on the energy feature vectors within the energy window so that the mean and unit variance of each energy index are zero, so that all energy indexes can be compared equally and avoid limiting the model performance due to dimension or unit reasons.

4. The data-driven aircraft energy anomaly identification method according to claim 1, characterized in that: The step (4) comprises the following steps: (41) A single-class SVM is used to build an energy anomaly detection model to improve the matching degree between the model and the dataset. The training dataset {x1, x2, ..., x n }∈R, R is the feature space where the data originally exists, the mapping function φ1 maps the entire training set from the feature space R to the feature space F, the segmentation hyperplane F1, and the relaxation variable ζ i , hyperplane intercept ρ, hyperplane normal vector w, the ratio of allowed abnormal data to all data υ, Lagrange multipliers λ1, λ2, k(·) kernel function, used to calculate the inner product of input mapping to high-dimensional space energy features, SV is the support variable, and the variable set defines the separating hyperplane; (42) The single-class support vector machine projects all the data into a high-dimensional feature space F through a kernel function, and solves the hyperplane F1 that is farthest from the origin of a normal data set; the single-class SVM model adds a slack variable ζ i To prevent overfitting during training, the parameter v is set to allow a small number of samples to be outside the segmentation hyperplane; the parameter v is a variable that matches the data set and the model. When the parameter v is small, the sensitivity of the model is lower, and it is easy to identify abnormal energy data as normal; conversely, the higher the sensitivity of the model, the easier it is to mistakenly identify normal energy data as abnormal; (43) The energy anomaly detection model calculates the normal area of ​​the training set data in the feature space. If the data of the input test set is outside the segmentation hyperplane, that is, outside the normal area, then -1 is returned and the sample is judged as abnormal. If it is within the segmentation hyperplane, then 1 is returned and the sample is judged as normal. When the data cannot be directly separated by a "linear" hyperplane in the original feature space R, the data is mapped to a higher-dimensional feature space F through the mapping function φ1, and a segmentation hyperplane y is constructed in F by adjusting the distance and angle of the sample in F. i =[E1 i ,E2 i ,E3 i ,...,E N i ] indicates the energy characteristic data after preprocessing: w·φ(y)-ρ=0 By optimizing the values ​​of w and ρ, we obtain a plane farthest from the origin and transform the problem into a quadratic optimization problem: w·φ(y i )-p+z i ≥0.ζ i ≥0 The conditional extreme value is calculated by Lagrange multiplication: By using the Lagrangian technique and the kernel function of the dot product calculation, the decision function is: when When , the sample data is outside the separation hyperplane and is considered to be abnormal data; when The sample data is on the separating hyperplane and is considered normal data; when λ1 = 0, the sample data is in the separating hyperplane; Reset the intercept ρ according to λ1 corresponding to the support vector on the separating hyperplane: and j ∈SV Then use the sgn function to judge the abnormality: When the return value is +1, the sample data is normal data; when the return value is -1, the sample data is abnormal data.

5. The data-driven aircraft energy anomaly identification method according to claim 3, characterized in that: The size of the sliding window in step (32) is 5.

6. The data-driven aircraft energy anomaly identification method according to claim 4, characterized in that: The kernel function mapped in the single-class support vector machine is the RBF radial basis kernel function.

Citation Information

Cited By

  • Vehicle tracking

    US20230064859A1