Method and system for detecting and classifying maneuvers performed by an aircraft based on measurements acquired during a flight of the aircraft

Abstract

A method implemented by a computer (12) for detecting that an aircraft (3) is performing a maneuver belonging to a macro category among a plurality of macro categories (MC) comprises: receiving a plurality of time series of data structures (205) having values of quantities related to a flight of the aircraft (3); for each duration (TW' among a plurality of predetermined durations (TW p ), selecting (310) a corresponding subset of the data structures (205) and extracting (320) a corresponding feature vector (FVX kp ); generating (330, 335; 500, 510; 720; 820) a corresponding input macro vector (MPVX' k ; MPVX" k , MPVX’’' k ; MFVX k ) based on the feature vector (FVX kp ); and applying (338; 520; 730; 830) an output classifier (151; 251; 561; 751) to the input macro vector (MPVX' k ; MPVX" k , MPVX’’' k ; MFVX k ) to generate an estimate of a probability that the aircraft (3) is performing a maneuver belonging to the macro category (MC).

Description

[0001] Cross-references to related applications

[0002] This patent application claims priority to European Patent Application No. 20425059.1, filed on December 18, 2020, the entire disclosure of which is incorporated herein by reference. Technical Field

[0003] The present invention relates to a method and system for detecting and classifying maneuvers performed by an aircraft based on measurements acquired during the flight of the aircraft. Background Technology

[0004] As is well known in aeronautics, there is a particular need to monitor the fatigue condition of aircraft components, and more generally, the health condition, in order to accurately estimate the remaining life of each component and thus optimize maintenance activities without compromising flight safety.

[0005] In particular, it is known that the fatigue state of aircraft components depends on the manipulations the aircraft has undergone during its service life, as the load on each component depends on the manipulations performed by the aircraft. Therefore, it is perceived that there is a need to correctly identify the manipulations performed by the aircraft so that the so-called "actual usage spectrum" can then be determined. For this purpose, it is known to equip aircraft with monitoring systems suitable for detecting the temporal trends of quantities relative to flight; this allows for the acquisition of a large number of measurements, which can be analyzed to study the history of manipulations performed by the aircraft. However, the applicant has observed that even with such measurements, the correct identification of performed manipulations requires advanced data processing techniques and is further hampered by the fact that different manipulations typically have different durations, complicating the aforementioned analysis of temporal trends.

[0006] EP2384971 discloses a method for determining a maneuver performed by an aircraft having sensors for monitoring motion data, the method comprising: periodically sampling the sensors to electronically determine segments of motion data of the aircraft; aggregating a sequence of motion data segments; comparing the aggregated motion data segments with a model of a specific maneuver; and determining the maneuver performed by the aircraft.

[0007] EP2270618 discloses a method for fault determination of an aircraft, the method comprising: generating predicted control based on a model of aircraft performance; determining actual control of the aircraft using information obtained from an inertial measurement system; and comparing the predicted control with the actual control.

[0008] EP3462266 discloses a method for maintaining an aircraft based on multiple maintenance messages generated during aircraft operation. Summary of the Invention

[0009] The object of the present invention is to provide a method for detecting the type of manipulation performed during the flight of an aircraft, the method at least partially satisfying the above-mentioned requirements.

[0010] According to the present invention, methods and systems for detection and classification as defined in the appended claims are provided. Attached Figure Description

[0011] To better understand the present invention, embodiments thereof are now described by way of non-limiting example only with reference to the accompanying drawings. Wherein:

[0012] Figure 1A This is a schematic diagram of an aircraft equipped with a monitoring system;

[0013] Figure 1B An example is shown illustrating the trend of the value of a quantity acquired through a monitoring system over time;

[0014] Figure 2 A block diagram illustrating a set of maneuvers that can be performed by an aircraft is shown, which can be further subdivided into macro categories.

[0015] Figure 3 and Figure 6 A block diagram is shown that relates to the training steps according to the first strategy;

[0016] Figure 4 A block diagram of the training data structure is shown;

[0017] Figure 5A The diagram schematically illustrates a portion of the training data structure and the arrangement of the time windows;

[0018] Figure 5B and Figure 5C They are shown schematically respectively. Figure 5A The training data structure shown is part of the arrangement of time windows with two different durations;

[0019] Figure 7 Two block diagrams illustrating training operations according to a first policy and a second policy are shown.

[0020] Figure 8 An example is shown. Figure 3 The table shown is a part of the operation;

[0021] Figure 9 A block diagram related to the analysis operations according to the first strategy is shown;

[0022] Figure 10 A portion of an unlabeled data structure and the arrangement of windows with different durations are schematically shown;

[0023] Figure 11Two block diagrams illustrating the analysis operations based on the first and second strategies are shown;

[0024] Figure 12 and Figure 13 A block diagram related to the training steps according to the second strategy is shown;

[0025] Figure 14 schematically shown Figures 5A-5C The diagram shows a portion of the training data structure and the arrangement of time windows used during training according to the second strategy;

[0026] Figure 15 An example is shown. Figure 12 and Figure 13 The table shown is a part of the operation;

[0027] Figure 16 A block diagram related to the analysis operations according to the second strategy is shown;

[0028] Figure 17 A block diagram related to training and analysis operations based on the third strategy is shown;

[0029] Figure 18 This is a block diagram illustrating the training operation based on the third strategy;

[0030] Figure 19 This is a flowchart illustrating the analysis operations based on the third strategy;

[0031] Figures 20A-20B A block diagram related to the classification operation performed according to the first strategy is shown;

[0032] Figure 21 A block diagram related to training and analysis operations according to the fourth strategy is shown;

[0033] Figure 22 Two block diagrams are shown, illustrating the training and analysis operations according to the fourth strategy, respectively;

[0034] Figure 23 and Figure 26 Block diagrams are shown respectively, relating to the operations performed according to the first and second variations of this method;

[0035] Figure 24 A block diagram illustrating the analysis operation according to the second variation is shown;

[0036] Figure 25 and Figure 27 Tables relating to the first and second variations of this method are shown respectively; and

[0037] Figure 28 A block diagram related to the operation according to this method is shown. Detailed Implementation

[0038] This method is based on the fact that aircraft (e.g., helicopters) can currently be equipped with numerous sensors that allow for the determination of trends in corresponding quantities characterizing the aircraft's flight (i.e., during a series of maneuvers). In other words, values ​​assumed by these characteristic quantities can be monitored during flight. For example, Figure 1A A helicopter 1 equipped with a monitoring system 2 is shown. The monitoring system 2 includes, for example, sensors for measuring the tilt angle of the helicopter 1, and thus provides the tilt angle value (i.e., a sample) at a given sampling frequency (e.g., approximately 10 Hz) in use. Continuing, the quantities described above will be referred to as principal quantities, and it is assumed that their number is equal to N; by way of example only, principal quantities may include: aircraft kinematic variables (such as pitch angle, roll angle, yaw angle, heading angle, vertical acceleration, vertical velocity, longitudinal acceleration, lateral acceleration, roll rate, pitch rate, yaw rate, northward velocity, eastward velocity, main rotor speed); aircraft control variables (such as, for example, collective control position, lateral cycle control position, longitudinal cycle control position, pedal position); environmental variables (e.g., airspeed, radar altitude, barometric altitude, wind speed, wind direction, total temperature, takeoff weight); and variables related to the energy system (e.g., motor torque, motor turbine speed, electric generator speed).

[0039] For example only, Figure 1B The trends of five main quantities (indicated as quantities 1-5, therefore n=5) over time are shown. These trends are monitored by corresponding sensors in monitoring system 2, which periodically provide corresponding samples. For simplicity, and without loss of generality, it is assumed that the sensors operate at the same sampling frequency f. c (For example, equal to 12.5Hz) Synchronous operation; in fact, even if the sensor operates locally at a different sampling frequency, it will be possible to return to the same sampling frequency in any case, for example, through oversampling, double sampling and interpolation operations.

[0040] Still as an example, Figure 1B The diagram shows a first time interval T1 and a second time interval T2, during which the first test maneuver M1 and the second test maneuver M2 of the first test flight of helicopter 1, signaled by the pilot, respectively occur. Additionally, in Figure 1B In this context, NMP1, NMP2, and NMP3 indicate three time intervals separated from the first time interval T1 and the second time interval T2, during which the pilot does not issue signals to notify controls, as described in more detail below. In the following text, time intervals NMP1, NMP2, and NMP3 are referred to as unmarked periods.

[0041] like Figure 2As shown, the set of maneuvers that can be performed by an aircraft (indicated by 5) can be further subdivided into multiple subsets, which are referred to as macro categories (indicated by MC) in the following description. Each macro category MC corresponds to a subcategory of maneuvers with similar characteristics (in... Figure 2 Grouping is done using the SC indicator. For example, Figure 2 The diagram illustrates macro categories (MCs) associated with horizontal flight, banked turns, vertical takeoff, and climb. Furthermore, macro categories associated with horizontal flight can include multiple maneuvers (in... Figure 2 Only four are shown, indicated by "40-node horizontal flight", "60-node horizontal flight", "90-node horizontal flight", and "150-node horizontal flight". In other words, each macro category MC represents the corresponding maneuver category (or type). Furthermore, as described in more detail below, given a macro category MC, maneuvers belonging to this macro category MC can be distinguished from each other in a deterministic manner, for example, based on the corresponding trends of one or more major quantities (e.g., based on average speed or average roll angle values, etc.).

[0042] In the following text, it is assumed that the number of macro categories MC is equal to NUM_MC.

[0043] Considering the foregoing, such as Figure 3 As shown, this method provides a way to obtain (box 100) training data structure 10, an example of which is shown in... Figure 4 It is qualitatively shown. For example... Figure 4 As shown again, the training data structure 10 can be stored in the computer 12.

[0044] In detail, training data structure 10 stores a time series (intended as a continuation of samples linked to corresponding moments) of values ​​of key quantities detected during test flights by monitoring system 2 of helicopter 1 and by monitoring systems (not shown) of other aircraft (not shown) during corresponding test flights. Furthermore, as referenced above... Figure 1B As shown in the first test maneuver M1 and the second test maneuver M2, the training data structure 10 stores the initial and final times of each test maneuver signaled by the aircraft pilot. Additionally, the training data structure 10 stores the macro category MC to which each test maneuver belongs.

[0045] For example, suppose the first manipulation M1 and the second manipulation M2 belong to the same macro category MC1, then Figure 4 The training data structure 10 shown stores the values ​​of principal quantities 1-5 during the first time interval T1 and the second time interval T2, and in both cases, they are concatenated to the macro category MC1. In other words, the training data structure 10 stores a first data cluster formed by the assumed values ​​of quantities 1-5 during the first time interval T1 (in... Figure 1B and 4The middle indicator is DG1), and the second data cluster is formed by the values ​​of quantities 1-5 assumed during the second time interval T2 (in Figure 1B and 4 (Indicated by DG2); Furthermore, for either the first data cluster DG1 or the second data cluster DG2, the connection of data structure 10 stored in macro category MC1 is detected. Figure 4 In this context, the storage connected to the macro category of the test manipulation is qualitatively represented by connecting each data cluster to the corresponding (test manipulation, macro category).

[0046] Referring again to the first test flight, training data structure 10 also stores values ​​assumed by principal quantities 1-5 during the aforementioned unmarked periods NMP1, NMP2, and NMP3, as previously stated, where the unmarked periods NMP1, NMP2, and NMP3 represent the periods during which the pilot has not yet specified the maneuvers to be performed.

[0047] Refer again Figure 4 The training data structure 10 shown also stores a time series of values ​​of key quantities detected by corresponding sensors during a second test flight, performed, for example, by an aircraft other than the helicopter 1 that performed the first test flight, and wherein two other test maneuvers occurring in a third time interval T3 and a fourth time interval T4 are signaled, respectively. Typically, the values ​​acquired during each test flight have a corresponding time axis, with the origin coinciding with the start of the flight; the alignment of the time axes for different flights can be managed in a manner known per se and is irrelevant to the purpose of the method. The first time interval T3 and the fourth time interval T4 alternate with three other unlabeled time intervals NMP4, NMP5, and NMP6. Furthermore, in the third and fourth time intervals T3 and T4, third and fourth test maneuvers belonging to, for example, macro categories MC2 and MC3, respectively, occur.

[0048] In fact, the training data structure 10 comprises multiple sub-blocks SB, which are referred to below as training substructures SB. Each training substructure SB stores a time series of hypothetical values ​​of quantities during the corresponding test flight. Therefore, each data cluster DG belongs to a single corresponding training substructure SB.

[0049] Refer again Figure 3 Computer 12 processes (box 110) the data structure 10 for manipulation, for example, to remove any anomalous data or incorrectly labeled data. This processing is performed in a manner known per se and is optional; for simplicity, it is assumed below that this processing does not alter the contents of the training data structure 10; otherwise, the operations described below are performed on the processed training data structure.

[0050] Subsequently, for each test operation, computer 12 extracts (box 120) a vector of statistics calculated as follows from training data structure 10, where, for simplicity, reference is made to the m-th test operation M. m It belongs to the macro category MC m And connected to the m-th training substructure SB m The test flight time interval T m This occurs during the period. Furthermore, it is assumed that during the time interval T... m During this period, the data cluster DG is formed by the assumed values ​​of the main quantities. m Considering the above, computer 12 is based on data cluster DG. m Extract the vector, which will be referred to below as the feature vector FV of the entire manipulation training. m .

[0051] Specifically, the feature vector FV of the entire manipulation training m like Figure 5A The computation is performed as shown (i.e., based on the entire data cluster DG). In other words, computer 12 employs a first-type time window TW. m Its relationship with the time interval T m Consistent and therefore consistent with test manipulation M m The duration is consistent. For example, the time interval T m At time t 开始 and t 结束 Extending between, it refers to the time indicated as flight time, which has an equal sampling frequency f. c The reciprocal of (i.e., equal to the sampling period, referred to as Δ below) c Instructions. (In) Figure 5A The discretization is qualitatively shown in the figure.

[0052] More specifically, the feature vector FV of the entire manipulation training m It consists of multiple elements, each of which is equal to the value based on the time window TW. m In the middle (and therefore, the entire test manipulates M) m The values ​​of the statistics are calculated based on the values ​​of the corresponding principal quantities assumed during the period. This is just an example; the entire manipulated training feature vector FV... m It can be formed by NUM_Ftot = NUM_F * N elements (with integers NUM_F), in which case, for example, the entire manipulated training feature vector FV occurs. m The first N elements are respectively (for example) equal to the time windows TW m The main quantity assumed during the period is the time average of the value, while the feature vector FV of the entire manipulation training is... m The second Nth element (for example) is equal to the time window TW respectively. mThe variance of the values ​​of the main quantities assumed during the period, etc. For example, in addition to the mean and variance, other statistics can be calculated, such as: maximum, minimum, median, the mean of the first derivative, the mean of the second derivative, the angle coefficient of the trend line, etc. As explained, statistics are used to manipulate M. m It is calculated over the entire duration. Furthermore, the feature vector FV of the entire manipulation training... m Connecting to the corresponding test manipulation M involved in the vector m MC macro category m ,like Figure 5A As shown qualitatively.

[0053] In fact, assuming, for example, that a number of test manipulations, NUM_M, have been performed, the operation in box 120 allows the generation of a number of feature vectors FV equal to the total number of manipulations trained, each feature vector being concatenated to the corresponding macroclass MC involved in the test manipulation.

[0054] Then, computer 12 trains a classifier 131 (e.g., box 130) based on the feature vector FV trained for the entire manipulation as determined for the test manipulation and the macro-class MC connected to the latter. Figure 7 As shown in the diagram, classifier 131 is referred to hereinafter as first-level classifier 131. In other words, computer 12 performs supervised and multi-class training; furthermore, first-level classifier 131 is of a known type, such as a random forest classifier.

[0055] Then, computer 12 extracts (box 140) multiple additional feature vectors for each test manipulation, which are referred to below as the extended manipulation training feature vector FV'. mpj The index 'm' indexes the test operations, while indices 'p' and 'j' are explained below. Furthermore, computer 12 uses a number of durations TW' shared among all test operations, NUM_TW (e.g., NUM_TW = 4), meaning they do not change with the test operations under consideration.

[0056] In detail, for each test operation, computer 12 executes... Figure 6 It was mentioned in Figure 5B The operation shown in the figure, the latter figure again involves the aforementioned test operation M. m And the first duration TW′1.

[0057] More specifically, computer 12 has its own periodic Δ clk The time base (indicated by time_clk) is used by computer 12 to determine ( Figure 6 (frame 200) a series of moments t clk (like Figure 5B (As shown in the diagram). For example, the period Δclk Equal to the sampling period Δ c Multiples of; for example, the period Δ clk Equal to the sampling period Δ c Twenty-five times that.

[0058] In addition, computer 12 detects (box 210) falling within time interval T. m And therefore fall within the test manipulation period M m The moment t clkj For example, in Figure 5B In the diagram, j = 0, ..., 4 is shown. For simplicity, the above will fall within the time interval T. m Time t within clkj The intermediate time t clkj In addition, refer to NUM_J to indicate the intermediate time t. clkj The quantity, where the quantity is the time interval T m and period Δ clk It is a function and is independent of the duration TW′ under consideration. For example, in Figure 5B The example shows NUM_J = 5. The value of NUM_J depends on the duration of the test manipulation under consideration and its alignment with time_clk.

[0059] Subsequently, for each duration TW′ P (where p = 1, ..., NUM_TW, used to index the duration TW′), computer 12 for each intermediate time t clkj Select (box 220) and place it in t clkj -(TW′ p / 2) and t clkj +(TW' p The m-th training substructure Sbm within the time window between / 2) m A subset of the values. In other words, for each duration TW' P Identify the corresponding time windows of equal duration, at each intermediate time t. clkj The values ​​of the training substructure SB, which fall within the translation time window and are associated with the test manipulation under consideration, are selected. Figure 5B It can be noted that, according to the intermediate time t clkj And the corresponding subset of selected values ​​for the training substructure SB, along with the considered duration TW', may include portions of the time series of the principal quantities adjacent to the time interval during which the considered test manipulation has occurred; that is, it includes values ​​assumed by the principal quantities during the waiting period before / after the test manipulation or during the test manipulation before / after the considered test manipulation. This is merely an example. Figure 5B In the above, it is assumed that the m-th test manipulation Mm There is an unmarked period NMP at the beginning. x And then immediately following is the (m+1)th test manipulation M m +1. Furthermore, when the extended time window 1 has an extension equal to the first duration TW' and the intermediate time t... clk0 Or at the intermediate time t clk1 When aligned (centered), the training substructure SB appears. m The corresponding subset of the selected values ​​also includes those from the previous unlabeled NMP period. x The values ​​of the main quantities assumed during a certain period. Furthermore, when an extended time window 1 with a duration equal to the first duration TW' is at the intermediate time t... clk3 At or at the intermediate time t clk4 When centered, the training substructure SB appears. m The corresponding subset of the selected values ​​also includes M manipulated by subsequent tests. m+1 The values ​​of the main quantities assumed during a certain period.

[0060] Subsequently, based on the training substructure SB m For each selected subset of values, the computer 12 extracts (box 230) the corresponding feature vector FV' of the extended manipulation training. mpj Therefore, for each test manipulation, compute NUM_TW*NUM_J extended manipulation training feature vectors FV'. mpj All feature vectors are connected to the corresponding test manipulation M. m MC macro category m .

[0061] Extend the manipulated feature vector FV' mpj Having the feature vector FV of the entire manipulation training m They have the same dimensions and are calculated in the same way—that is, they involve the same statistics. However, these statistics are calculated based on a subset of the values ​​of the principal quantity assumptions during each test manipulation period, rather than based on the values ​​of the principal quantity assumptions during the entire test manipulation period. In particular, each feature vector FV' trained by the extended manipulation is... mpj Each element is equal to the element based on the time window t. clkj -(TW' p / 2) and t clkj +(TW' p / 2) The corresponding statistic is calculated by assuming the value of the corresponding main quantity during the period.

[0062] As an example, Figure 5C Again, the m-th test manipulation M is involved. m Furthermore, the training substructure SB associated with the second duration TW'2 is shown. mThe selection of a subset, such that the second duration TW'2 has twice the duration of the first duration TW'1, allows for the computation of the five feature vectors FV' trained by the extended manipulation. m20 -FV' m24 , and similar to Figure 5B The extended manipulation training feature vector FV' shown m10 -FV' m14 Similarly, it is connected to the m-th test manipulation M. m MC macro category m .

[0063] Refer again Figure 3 For each test operation, computer 12 performs a test for each corresponding intermediate time t. clkj The first-level classifier 131 is applied (box 150) to the feature vector FV' trained by the extended manipulation. mpj Each of them, so as to obtain a training prediction vector PV' equal to the first policy. mpj The number of NUM_TW, the first policy training prediction vector PV' mpj The feature vector FV' trained by the extended manipulation can be used. mpj Index in the same way.

[0064] Each first strategy trains a prediction vector PV' mpj It has multiple elements equal to the number of macro categories MC, NUM_MC, where each element indicates the corresponding feature vector FV' trained by the extended manipulation. mpj The probability of connecting to the macro category MC corresponding to this element.

[0065] Furthermore, for each test operation, computer 12 performs tests at each intermediate time t. clkj Aggregate (box 160) the first policy to train the prediction vector PV' mpj (Equal to NUM_TW in quantity) to form the corresponding macro vector, which is still connected to the macro category MC of the test manipulation, and the corresponding macro vector is referred to below as the corresponding first trained prediction macro vector MPV'. mj .

[0066] For example, Figure 7 Involves the m-th test manipulation M m And referencing a single intermediate time t clkj (In the example shown, the intermediate time t) clk0 The diagram shows a graphical example of the operation of boxes 150 and 160. Specifically, Figure 7 The first-level classifier 131 is shown to be applied to the four feature vectors FV′ during extended manipulation training. m10 -FV′ m40How can we generate the four corresponding first-strategy training prediction vectors PV′ respectively? m10 -PV′ m40 Four first-strategy training prediction vectors PV′ m10 -PV′ m40 Aggregated into the first training prediction macro vector MPV′ m0 The first training predicts the macro vector MPV′ m0 Connect to test manipulation M m MC macro category m .

[0067] also, Figure 8 Involves the m-th test manipulation M m It also shows how to divide the five intermediate times t for each of the four durations TW′1-TW′4. clk0 -t clk4 Each of these is connected to a corresponding first policy training prediction vector PV′, as described above, where each first policy training prediction vector PV′ comprises multiple elements equal to the number of macro classes NUM_MC. Furthermore, Figure 8 Five intermediate times t are shown. clk0 -t clk4 Each of these is connected to the corresponding first trained prediction macrovector MPV′. m0 -MPV′ m4 The corresponding first trained prediction macro vector MPV′ m0 -MPV′ m Also connected to the macro category MC m .

[0068] Refer again Figure 3 After the operation of execution box 160, computer 12 calculates the first trained prediction macro vector MPV' based on the test manipulation. mj And based on the macro-classes MC trained (box 170), a first secondary classifier 151 is trained (box 170). Figure 7 (as shown in the image).

[0069] In practice, the first and second-level classifiers 151 are trained in a supervised manner. Furthermore, by way of example only, the first and second-level classifiers 151 could be logistic regression type classifiers.

[0070] Once the first-level classifier 131 and the first-level classifier 151 have been trained, the macro category to which the unknown maneuver (and therefore, the one with an unknown duration) belongs can be determined, for example, by an unknown helicopter 3 (an example of which is in...). Figure 1AAs shown in the diagram, the unknown helicopter 3 is equipped with a corresponding monitoring system 4 including sensors (not shown) suitable for monitoring the aforementioned key quantities. More specifically, during the unknown flight of the unknown helicopter 3, that is, during the flight of unknown maneuvers, the occurrence of maneuvers belonging to one of the aforementioned macro categories MC used during the training of the first-level classifier 131 and the first-second-level classifier 151 can be identified. For this purpose, the execution... Figure 9 The operation shown is illustrated.

[0071] In detail, obtain (box 300) the new data structure (in Figure 10 As shown in the diagram (where it is indicated by 205), this new data structure is referred to hereinafter as unlabeled data structure 205 because it is formed by time series of values ​​of the main quantities assumed to be measured by the monitoring system 4 equipped with the unknown helicopter 3 during unknown flight. In fact, unlabeled data structure 205 is formed by a single training substructure SB. Furthermore, since the manipulations performed during unknown flight are unknown, unlabeled data structure 205 does not store any connections to the macro category MC.

[0072] Similar to the case of training data structure 10, in the unlabeled data structure 205, the values ​​of each principal quantity are also concatenated to the corresponding sampling time; that is, they are concatenated along a path that is still equal to the sampling period Δ. c The discretized flight time distribution described above. In fact, by assuming unknown flights along time intervals W... tot Extending, this time interval W tot The total time interval W will be referred to below. tot Each time series connected to the corresponding principal quantity includes equal to W tot *f c Multiple values.

[0073] Computer 12 still has a periodic Δ clk The time base time_clk, whose generation time t _clk Furthermore, given the generality at time k t... clkk (like Figure 10 In the case shown (box 310), computer 12 selects (box 310) the number of NUM_TW values ​​that are equal to the subset of values ​​in the unlabeled data structure 205. Specifically, for the duration TW' p For each of the numbers (where p = 1, ..., NUM_TW), computer 12 selects the range falling within t. clkk -(TW′ p / 2) and t clkk +(TW′ p A subset of the values ​​of the unlabeled data structure 205 within a time window between / 2). If the range is within t clkk -(TW′NUM_TW / 2) and t clkk +(TW′ NUM_TW When the time window between / 2 (i.e., the widest time window) extends beyond the boundary of the unlabeled data structure 205, the NUM_TW time window relative to time t can be used. clkk Different centering methods ensure that the time window falls entirely within the unlabeled data structure 205, or that time t can be discarded. clkk Alternatively, the corresponding subset of the unlabeled data structure 205 may include only the values ​​that are actually available; these details are irrelevant to the purpose of performing this method.

[0074] Then, for each selected subset of the values ​​of the unlabeled data structure 205, the computer 12 extracts (box 320) the corresponding feature vector, which is referred to below as the input feature vector FVX. kp Therefore, for the general k-th time t provided by the time base time_clk clkk The input feature vector FVX is calculated to be equal to any connection to any macroclass MC. kp The number of NUM_TW.

[0075] Input feature vector FVX kp Having the feature vector FV′ of extended manipulation training mpj and the feature vector FV of the entire manipulation training m They share the same dimensions and are calculated in the same way—that is, they involve the same statistics; however, these statistics are calculated based on a subset of values ​​assumed by the main quantities during unknown flight periods. Specifically, the input feature vector FVX... kp Each element is equal to the element based on the time window t. clkk -(TW′ p / 2) and t clkk +(TW′ p The value of the corresponding statistic is calculated from the value of the corresponding principal quantity assumption in / 2).

[0076] Then, for example, refer to the general k-th time t clkk Computer 12 applies the first-level classifier 131 (box 330) to the corresponding input feature vector FVX. kp In order to obtain an equal to Figure 11 Qualitatively, the first policy input prediction vector PVX′ is illustrated. kp The number of NUM_TW, in Figure 11 The diagram shows four first-policy input prediction vectors PVX′. k1 -PVX′ k4 It outputs four input feature vectors FVX from the first-level classifier 131. k1 -FVX k4Export the application.

[0077] Each first strategy input prediction vector PVX′ kp It has multiple elements equal to the number of macro categories MC, NUM_MC, where each element indicates the corresponding input feature vector FVX. kp Connect the probability to the macro category MC corresponding to the element, and thus indicate the unknown helicopter 3 at the corresponding k-th time t. clkk The probability of performing manipulations belonging to this macro category (MC).

[0078] Referring again to time k t clkk Computer 12 aggregation (box 335) first strategy input prediction vector PVX′ kp (Equal to NUM_TW in quantity) to form the corresponding macro vector, which is referred to below as the first strategy input macro vector MPVX′. k .

[0079] Referring again to time k t clkk Computer 12 then applies the first secondary classifier 151 (box 338) to the first policy input macro vector MPVX′. k To obtain the first output vector OUT_A k It has multiple elements equal to the number of macro categories, NUM_MC, where each element indicates the corresponding input feature vector FVX. kp Connect the probability to the macro category MC corresponding to the element, and thus indicate the unknown helicopter 3 at time k t. clkk The probability of performing manipulations belonging to the macro category MC. In fact, the first output vector OUT_A k This indicates that the prediction vector PVX′ contained in the first policy input vector is included. kp The improvement in probability is explained in the following references to all the second-level classifiers mentioned.

[0080] As a replacement or supplement to what has been described so far, computer 12 can implement different strategies, which now refer to Figure 12 The description is provided, and multiple second-level classifier classes and second-level classifier classes 251 are trained starting from training data structure 10, such as... Figure 7 This is qualitatively illustrated above. Figure 3 The content already described in boxes 100 and 110 also applies to this strategy.

[0081] Initially, for each test manipulation, computer 12 extracts (box 340, Figure 12 Multiple feature vectors, referred to below as the "partially manipulated training feature vector FV"; for this purpose, computer 12 performs Figure 13 The description in Figure 14 The operation illustrated in the figure below, and the subsequent figure again involve the aforementioned test operation M. m As mentioned above, the m-th test manipulates M m At the moment known as flight time t 开始 and time t 结束 Extending between.

[0082] More specifically, computer 12 utilizes a synchronization time base (hereinafter referred to as time_clk_sync) because it has a period Δ. clk The period, and relative to time t 开始 Synchronize so as to have time t 开始 The coincident origins. In other words, computer 12 determines ( Figure 13 (400) A series of moments t clk sync (like Figure 14 (as shown in the image).

[0083] In addition, computer 12 detection (box 410, Figure 13 ) falls within time interval T m And therefore fall under the test manipulation M m The time t during the period clk_sync_u For example, in Figure 14 The diagram shows u = 1, ..., 5; no t was detected. clk_sync_0 Because it is related to time t 开始 Overlap. For the sake of brevity, the time interval T will be used in the following text. m The above time t within clk_sync_u The intermediate time t is called the synchronization intermediate time. clk_sync_u Furthermore, NUM_U is referred to as the indicator of the intermediate time t during synchronization. clk_sync_u The quantity, NUM_U, is the time interval T. m and period Δ clk The function.

[0084] Subsequently, for each duration TW′ p (where p = 1, ..., NUM_TW, used to index the duration TW′), computer 12 selects multiple subsets (or possibly the entire subset, as described below) of the data cluster DG corresponding to each test operation, as referred again below to the general p-th duration TW′. p and the mth test manipulation M m As stated and as Figure 14 The first duration TW'1 is shown in the reference.

[0085] In detail, computer 12 checks (box 420) where the m-th test operation M mThe time interval T that has already occurred m Does it have a duration TW' less than or equal to the p-th duration? p The duration, in this case (the output "Yes" in box 420), computer 12 selects (box 430) the entire data cluster DG m And then from the entire data cluster DG m Extracting (box 440) a single feature vector FV from the manipulated portion. mp0 The feature vector FV” mp0 The above eigenvector FV is equal to the entire manipulation. m And connected to the test manipulation M m MC macro category m .

[0086] Conversely, if the m-th test manipulates M m The time interval M that has already occurred m Having a duration greater than p-th TW' p If the duration (output "No" in box 420) is specified, then computer 12 selects (box 450) each synchronization intermediate time t. clk_sync_u This makes the time window t clk_sync_u -(TW p ' / 2) and t clk_sync_u +(TW p ' / 2) Falls completely within time interval T m Inside. For example, see reference. Figure 14 Where it is assumed that the first duration TW'1 is equal to the period Δ clk Four times that of the time interval, therefore computer 12 only selects the intermediate time t for synchronization. clk_sync_2 and t clk_sync_3 In the continuation, the synchronization intermediate time t selected during the operation of box 450 will be... clk_sync_u These are called active synchronization moments; they vary with the duration TW' under consideration. p And change.

[0087] Furthermore, for each active synchronization intermediate time t clk_sync_u Computer 12 selects (box 460) and falls into t clk_sync_u -(TW' p / 2) and t clk_sync_u +(TW' p Data cluster DG within the time window between / 2) m A subset of the values. In other words, for each active synchronization intermediate time t clk_sync_u , equal to the duration TW' p The duration of the time window is centered on it, and this window is used to select the data cluster DG. m The value of .

[0088] Subsequently, for the data cluster DG m For each selected subset of values, the computer extracts (box 470) the corresponding feature vector FV from the manipulated training. mpu It is connected to the corresponding test control M m MC macro category m It has the feature vector FV' of the extended manipulation training. mpj It uses the same dimensions and is calculated in the same way; that is, it involves the same statistics calculated based on a subset of values ​​that include the main quantity assumptions during the sub-period manipulated by the test. As an example, in Figure 14 The text indicates two feature vectors FV that were partially manipulated during training. m12 and FV″ m13 .

[0089] Refer again Figure 12 At the end of the operation in box 340, computer 12 has a corresponding set FV of feature vectors for each duration TW' that were partially manipulated during training. mpu For example, consider the p-th duration TW' P Computer 12 has a corresponding subset FV of feature vectors trained for each test manipulation, including parts of the manipulation. mpu Set p ; when the duration is less than the duration TW' p In the case of test manipulation, the latter subset includes the unique feature vector FV″ of the partial manipulation. mp0 (equal to the corresponding feature vector FV of the entire manipulation) m Otherwise, it includes partially manipulated training of multiple feature vectors FV″. mpu This depends on the duration of the test manipulation and the duration of TW' P In any case, for a given p-th duration TW' P In the case of partially manipulating the trained feature vector FV” mpu The corresponding set SET p Includes TW' with a duration not exceeding the same p The extended corresponding data cluster DG partially computes the feature vector.

[0090] As an example, Figure 15 This shows a set of feature vectors (indicated as SET1) of partial manipulation training associated with the first duration TW'1, and associated with the m-th manipulation M. m The relevant subset (indicated as SET1) is formed by the two feature vectors mentioned above that were partially manipulated during training, thus forming FV”. m12 and FV” m13 These two feature vectors form FV” m12and FV” m13 These refer to the intermediate time t of active synchronization. clk_sync_2 and t clk_sync_3 .

[0091] Considering the foregoing and referring again Figure 12 For each duration TW', computer 12 trains a feature vector FV″ based on partial manipulation. mpu The corresponding set SET and the macro-class MC connected to these vectors are trained (box 350) to the corresponding second-level classifier class CLASS. In other words, computer 12 performs supervised and multi-class training to obtain, for example, Figure 7 The qualitative representation shows multiple second-level classifiers equal to NUM_TW, in Figure 7 The diagram shows four second-level classifiers, CLASS1-CLASS4, corresponding to durations TW1'-TW4' respectively.

[0092] The second-level classifier class CLASS is a known type, such as the random forest classifier. Furthermore, as mentioned above, refer to, for example, the p-th second-level classifier class CLASS. p It has been trained based on feature vectors computed on a portion of the data cluster DG associated with test manipulation, the portion having a duration no higher than the p-th time TW'. p Time extension.

[0093] Once the second-level classifier class has been trained, the computer 12 applies (box 360) the second-level classifier class to the aforementioned feature vector FV' trained by the extended manipulation in the following manner: mpj .

[0094] In detail, for each test operation, and for each corresponding intermediate time interval t clkj Computer 12 will classify the p-th second-level classifier as CLASS. p (where p = 1, ..., NUM_TW) is applied to the corresponding feature vector FV' of the extended manipulation training. mpj That is, applied by applying a method with a duration equal to TW' (i.e., equal to the duration used to train the same second-level classifier CLASS). p The corresponding feature vector FV' obtained from the extended manipulation training is the time window duration. mpj And it obtains the corresponding second policy training prediction vector PV, which is connected to the macro category MC of the test manipulation. mpj .

[0095] Following the operation of box 360, computer 12 performs the operation for each corresponding intermediate time t of each test manipulation. clkjHaving an equal number of prediction vectors PV trained by the second strategy mpj The number of NUM_TW.

[0096] As an example, Figure 7 Four second-policy training prediction vectors PV are shown. m10 -PV” m40 Its relationship with the m-th test manipulation M m The midpoint t clk0 The related features are generated by four second-level classifier classes CLASS1-CLASS4, respectively, and are connected to durations TW1'-TW4', respectively, from the four feature vectors FV' trained by the extended manipulation. m10 -FV' m40 start.

[0097] Then, for each test operation, computer 12 performs tests at each intermediate time t. clkj Aggregation (box 370, Figure 12 The second strategy trains the prediction vector MPV. mj (Equal to NUM_TW in quantity) to form the corresponding macro vector, which is then concatenated to the macro category MC of the test manipulation, and this corresponding macro vector is referred to below as the corresponding second trained prediction macro vector MPV″. mj .

[0098] For example, refer to Figure 7 Four second-policy training prediction vectors PV″ m10 -PV″ m40 Aggregated into a second training prediction macro vector MPV″ m0 It is connected to the test control M m MC macro category m .

[0099] Refer again Figure 12 Computer 12 is based on the intermediate time t for each test operation. clkj The second training prediction macro vector (MPV) is calculated. mj Training (box 380) Second-level classifier 251 ( Figure 7 (as shown in the image).

[0100] In fact, the macro vector MPV″ is predicted in a supervised manner based on the second training. mj The second-level classifier 251 is trained by connecting its macro-classes. Furthermore, by way of example only, the second-level classifier 251 could be a logistic regression type classifier.

[0101] Once the second-level classifier CLASS and the second-level classifier 251 have been trained, the second output vector OUT_B can be determined to identify the macro-class to which the indication of unknown maneuvering related to unknown flight belongs. k This is at time k t clkk This occurs. For this purpose, computer 12 executes... Figure 16 The and shown Figure 11 The operations illustrated in .

[0102] In detail, for each k-th time t provided by the time base time_clk clkk Computer 12 processes the above input feature vector FVX in the following manner kp (Equal to NUM_TW in quantity) Application (box 500, Figure 16 (This is related to the second-level classifier CLASS.)

[0103] Specifically, for each k-th time t clkk Computer 12 will classify the p-th second-level classifier as CLASS. p Applied to the p-th input feature vector FVX kp In order to obtain the corresponding p-th second policy input prediction vector PVX” kp In other words, each input feature vector FVX kp Classification is performed using a second-level classifier class CLASS, which is based on an input feature vector FVX that is equal to or less than the same. kp The training is performed on a subset of the data cluster DG involving the duration of a subset of the unlabeled data structure 205.

[0104] More specifically, the second strategy input prediction vector PVX” kp Having the first policy input prediction vector PVX′ kp Same dimensions; furthermore, any second policy input prediction vector PVX kp Each element of the expression indicates the corresponding input feature vector FVX. kp Connected to the macro category MC corresponding to the same element, and therefore the unknown helicopter 3 at the corresponding k-th time t clkk The probability of performing an operation belonging to the macro category MC.

[0105] Referring again to the general k-th time t clkk Computer 12 aggregation (box 510) second strategy input prediction vector PVX” k (Equal to NUM_TW in quantity), in order to form the corresponding macro vector, which is referred to below as the second strategy input macro vector MPVX. k .

[0106] Subsequently, computer 12 applies the first secondary classifier 151 (box 520) to the second policy input macro vector MPVX. k To obtain the second output vector OUT_B k Each element indicates the corresponding input feature vector FVX. kp Connect to the macro category MC corresponding to this element, and therefore the unknown helicopter 3 at time k t clkk The probability of performing an operation belonging to the macro category MC. In fact, the second output vector OUT_B k This indicates that the prediction vector PVX is included in the second policy input. kp The improvement in probability is explained in the following references to all the second-level classifiers mentioned.

[0107] Typically, it is included in the first output vector OUT_A k Second output vector OUT_B k The information in the data can be used as an alternative to identify the corresponding k-th time t. clkk The macro category of unknown manipulation performed in the process. Furthermore, the applicant has observed that it is contained in the second output vector OUT_B. k The indication in is usually greater than that contained in the first output vector OUT_A k The indications in the whole manipulation are more accurate, especially in cases where the manipulation is characterized by a relatively constant trend of the main quantity during the manipulation. However, in some cases, and particularly in cases where the manipulation is characterized by very distinctive initial and final parts (so-called "entry" and "recovery" steps), the opposite occurs. In fact, in cases where there are distinctive parts arranged at the beginning and end of the manipulation, training based on the parts of the manipulation may not be very effective compared to training based on the whole manipulation, if such distinctive parts have durations that are very different from the duration of the training window. Conversely, in cases where there is a gradual change in a quantity (e.g., velocity) between, for example, an initial value and a final value, training based on the parts of the manipulation tends to be more effective; in fact, in this case, a classifier trained on the whole manipulation tends to only identify manipulations in which the quantity precisely exhibits such initial and final values, while a classifier based on the parts of the manipulation has the potential to appropriately weight the trend (change) of the quantity in each part of the manipulation.

[0108] According to another variation, the first and second strategies can be combined by using a third secondary classifier 651, as shown in... Figure 17 The and Figure 18 They were trained as illustrated in the example.

[0109] In detail, for each test operation, computer 12 performs tests for each corresponding intermediate time t. clkk Aggregation (box 700, Figure 17 The first policy training prediction vector PV′ generated by the first-level classifier 131 mpj (Equal in quantity to NUM_TW) and the second-policy training prediction vector PV, both generated by the corresponding second-level classifier CLASS. mpj In order to obtain the third training prediction macro vector MPV”′ mj It is still connected to the intermediate time t. clkj The test manipulation occurs in the macro class MC.

[0110] For example, Figure 18 The first policy training prediction vector PV′ is qualitatively shown. m10 -PV′ m40 With the second strategy, the predicted vector PV is trained. m10 -PV” m40 The aggregation of these vectors is used to form a third training prediction macro vector, MPV”′. m0 .

[0111] Subsequently, computer 12 is based on the intermediate time t of the test manipulation. clkj The relevant third training prediction macro vector MPV”′ mj and connected to the third trained prediction macrovector MPV”′ mj The macro-class MC is used to train (box 710) the third secondary classifier 651. The third secondary classifier 651 is then trained in a supervised manner and can be of the same type as, for example, the first secondary classifier 151 and the second secondary classifier 251.

[0112] Once the third-level classifier 651 has been trained, the computer 12 can analyze unknown flights. For this purpose, for example, referring to the general k-th time t... clkk The first policy input prediction vector PVX′ corresponding to the computer 12 aggregation (box 720) kp With the second policy input prediction vector PVX” kp To form the corresponding third strategy input macro vector MPVX”′ k Computer 12 applies the third secondary classifier 651 (box 730) to the third policy input macro vector MPVX”′. k In order to obtain the third output vector OUT_C k .

[0113] For example, Figure 19 This illustrates how the first policy input prediction vector PVX′ is aggregated. k1 -PVX′ k4Second policy input prediction vector PVX” k1 –PVX” k4 To generate the third strategy input macro vector MPVX”′ k .

[0114] Typically, for the sake of simplicity, only the first strategy is referred to above, and therefore the first output vector OUT_A is referenced. k The reason explained is that the first output vector OUT_A k Second output vector OUT_B k and the third output vector OUT_C k Both benefit from the actions of the corresponding second-level classifier, which allows for improvements to the classification provided by the first-level classifier. Specifically, see reference below. Figure 20A and Figure 20B This involves a simplified case where NUM_MC equals 2 (only the first macro category MC1 and the second macro category MC2 are available), and furthermore, NUM_TW equals 2 (only the first duration TW′1 and the second duration TW′2 are available). Additionally, in Figure 20A and Figure 20B In the middle, referencing the operation of the m-th manipulator M m The midpoint t clk1 Related extended manipulation training of feature vector FV′ m11 and FV′ m21 To train the first and second level classifiers 151, the m-th manipulation M m It belongs to the first macro category MC1, and its time interval M is assumed to have an extension falling between the first duration TW′1 and the second duration TW′2. The intermediate time t clk1 Falling on time interval T m About half of it.

[0115] Considering the above, the first-level classifier 131 generates feature vectors FV′ trained from the extended manipulation. m11 and FV′ m21 The first two first-policy training prediction vectors PV′ m11 and PV′ m21 The first policy-trained prediction vector PV′ associated with the macro-class MC1 m11 The elements correctly have high values ​​(0.9), while the elements associated with macro category MC2 correctly have low values ​​(0.1). Conversely, the first policy-trained prediction vector PV′ associated with macro category MC1... m21 The elements incorrectly have low values ​​(0.1), while the elements associated with the macro category MC2 incorrectly have high values ​​(0.9). This is due to the fact that although the extended manipulation of the trained feature vector FV′ m11 Involves having a time interval Tm The duration is similar to the time window of the duration, but the extended manipulation of the trained feature vector FV′ m21 The important part of the time window involved lies in the time interval T. m In addition. However, due to the expansion manipulation of the two feature vectors FV′ trained. m11 and FV′ m21 Connecting to the macro-class MC1, the training of the first secondary classifier 151 causes it to assign greater weights to the first policy training prediction vector PV′ for the macro-class MC1. m11 The corresponding element, instead of assigning it to the first policy training prediction vector PV′. m21 The corresponding element. Therefore, as... Figure 20B As shown in the figure, when applied to the first trained prediction macro vector MPV' m1 The possible output vector OUT_TRAIN1 generated by the first and second classifiers 151 is reported (by aggregating the predicted vector PV′ trained by the first policy). m11 and PV' m21 To obtain M21), the elements associated with the macro categories MC1 and MC2 of the output vector OUT_TRAIN1 are correctly equal to 1 and 0, respectively.

[0116] In addition, the third output vector OUT_C k Typically, it is more than the first output vector OUT_A k Second output vector OUT_B k More accurate, because the relative generation mechanism is suitable for manipulations where the primary quantity is relatively constant (i.e., static), or for manipulations where the primary quantity is rapidly variable.

[0117] according to Figure 21 As shown and in Figure 22 In another variant illustrated, computer 12 trains a fourth secondary classifier 751 as described below in order to implement a fourth strategy.

[0118] In detail, for each test operation, computer 12 performs tests for each corresponding intermediate time t. clkj Aggregate (box 800) the corresponding feature vector FV' of the extended manipulation training mpj In order to obtain the training feature macro vector (MFV) mj It is connected to the macro category MC of the test manipulation. For example, Figure 22 The four feature vectors FV' of the extended manipulation training are shown. m10 -FV' m40 Aggregates into training feature macro vectors (MFV) m0 It is connected to the m-th test manipulation M m MC macro category m .

[0119] Subsequently, computer 12 uses the trained feature macro vector (MFV) mj The fourth secondary classifier 751 is trained (box 810) and connected to its macro-class MC. In fact, the fourth secondary classifier 751 is trained in a supervised manner and can be, for example, a random forest type classifier.

[0120] Regarding the unknown flight, for each k-th time t clkk Computer aggregation (box 820) input feature vector FVX kp In order to obtain the corresponding input feature macro vector MFVX k Computer 12 applies the fourth secondary classifier 751 (box 830) to the corresponding input feature macro vector MFVX. k In order to obtain the fourth output vector OUT_D k . Figure 22 This illustrates, for example, the input feature vector FVX. k1 -FVX k4 The aggregation of.

[0121] Moreover, in this case, the presence of the fourth secondary classifier 751 allows for the same benefits described with reference to the first secondary classifier 151, the second secondary classifier 251, and the third secondary classifier 651. Furthermore, this strategy is characterized by lower complexity because it provides classification at a single level.

[0122] Typically, the applicant has already noticed that, due to the inclusion in the output vector OUT_A k OUT_B k OUT_C k and OUT_D k The probability estimates in the data are generated by the fact that they are generated by a classification algorithm that starts by selecting a portion of the unlabeled data structure 205 via time windows with different dimensions. These estimates are generally satisfactory and are independent of the duration of the manipulation.

[0123] In practice, the aforementioned variant allows for the identification of maneuvers belonging to one of the aforementioned macro categories (MCs) during unknown flight with considerable accuracy. However, the applicant has observed that if two or more maneuvers are performed simultaneously during unknown flight, the accuracy of identification may be reduced. To eliminate this drawback, the applicant has observed that the following can be implemented.

[0124] like Figure 23 As shown, the computer trains (box 900) multiple classifiers 910, which are referred to below as single-class classifiers SCC (e.g., ...). Figure 24 (as shown in the image).

[0125] In detail, for each macro category MC, the computer 12 uses the feature vector FV of the entire manipulation. m To train the corresponding uniclass classifier SCC. Specifically, consider, for example, the i-th uniclass classifier SCC. I (where i = 1, ..., NUM_MC), it is based on the feature vector FV of the entire manipulation connected to the test manipulation. m To perform training, if the entire manipulated feature vector FV m With MC belonging to the i-th macro category i If the manipulation is related, then they are connected to the first label (e.g., single), otherwise they are connected to the second label (e.g., invalid); in other words, the label indicates the macroclass MC of the tested manipulation and the corresponding single-class classifier SCC. I MC macro category i Matching / not matching between.

[0126] Subsequently, consider the unknown flight and the general k-th time t. clkk Computer 12 will process each input feature vector FVX kp (where p = 1, ..., NUM_TW) is applied (box 910) to the single-class classifier SCC so that for each input feature vector FVX kp Obtain the corresponding single-class probability vector SCV kp , where the i-th element represents the input feature vector FVX kp Involves, for example, the i-th single-class classifier SCC i The calculation of the i-th macro category MC i The probability of.

[0127] For example, Figure 24 The four input feature vectors FVX to the four single-class classifiers SCC1-SCC4 are shown. k1 -FVX k4 Applications and Single Class Probability Vector (SCV) k1 -SCV k4 The subsequent generation. In Figure 25 An example of the latter is shown in the figure.

[0128] Then, computer 12 generates (box 920) the update vector OUT_UPDATE. k This makes it have multiple elements equal to the number NUM_MC, and generally the i-th element is equal to the single-class probability vector SCV. kp The maximum value among the values ​​of the i-th element, that is, equal to the value of the i-th single-class classifier SCC when it is applied to the t-th element. clkk Input feature vector FVX at time step kpThe maximum value among those provided when (equal to the number of NUM_TW). Although not further described, this is for generating the update vector OUT_UPDATE. k The i-th element of the latter can be set to be equal to the single-class probability vector SCV. kp The i-th element is used instead of the statistic (e.g., the average) calculated from the maximum value mentioned above.

[0129] Subsequently, computer 12 updates the vector OUT_UPDATE. k Detection (box 930) at time k t _clkk Whether two or more operations belonging to different macro categories (MC) are performed (i.e., it detects conditions of multiple macro categories), for example by detecting the update vector OUT_UPDATE. k Whether two or more elements exceed a predetermined threshold. For example, refer to Figure 25 Assuming a threshold of 0.5, computer 12 detects at time t clkk Two manipulations belonging to macro category MC2 and macro category MC4 are executed simultaneously. Typically, although not further described, the detection of multiple macro categories can also provide the ability to change the value of the aforementioned threshold, for example by connecting the corresponding threshold to each macro category MC and / or by adding additional control downstream of the detection of multiple macro categories, in order to exclude possible combinations of multiple macro categories that have no physical meaning (e.g., to exclude multiple macro categories that include macro categories associated with opposite manipulations).

[0130] In the continuation, refer to the multi-class policy output vector OUT_Y k The value is alternately equal to the first output vector OUT_A, depending on the program implemented by computer 12. k Second output vector OUT_B k The third output vector OUT_C k Or the fourth output vector OUT_D k The vector. Considering the above, if no manipulation belonging to a different macro category MC is detected (the output of box 930 is "No"), computer 12 will output the final vector OUT_FIN. k Set (box 940) to be equal to the multi-class policy output vector OUT_Y k Because at time t clkk Manipulations belonging to different macro categories (MC) did not occur, and therefore the multi-policy output vector OUT_Y was not generated. k The probability is reliable. Conversely, in the case of detecting manipulation belonging to a different macro category MC (the output of box 930 is "Yes"), computer 12 will output the final vector OUT_FIN. k Set (box 950) to equal the update vector OUT_UPDATEk This is because in this specific case (simultaneous execution of manipulations belonging to different macro categories), the probability contained in the latter tends to be higher than that in the multi-class policy output vector OUT_Y. k The probability is more accurate.

[0131] Based on the final vector OUT_FIN k Computer 12 detects (box 951) at time t clkk The macro category MC of the manipulation performed in the process, for example, by selecting to connect to the final vector OUT_FIN when no manipulation belonging to a different macro category MC is detected. k The macro category MC containing the highest value, or, if manipulation belonging to a different macro category MC is detected, the final vector OUT_FIN is selected for connection. k The macro category MC with the highest value contained therein, or the macro category most relevant from the perspective of fatigue of the components of the unknown helicopter 3, is selected from among the multiple macro categories detected during the operation involved in box 930.

[0132] according to Figure 26 In another variation shown, computer 12 trains (box 960) a single-class classifier SCC for each duration TW', equal to the number of macro-classes NUM_MC. Figure 27 As shown in the diagram, which is referred to below as the single-class single-window classifier SCC'. Therefore, the number of training operations for calculator 12 is equal to the number of NUM_TW*NUM_MC of the single-class single-window classifier SCC', which is then indexed by means of index "p" and index "i".

[0133] In detail, considering the single-class single-window classifier SCC′ pi It is based on the duration TW' of the p-th time. p The relevant part of the manipulated feature vector FV” mpu To perform training, if the feature vector FV is partially manipulated. mpu With MC belonging to the i-th macro category i If the manipulation is related, then they are connected to the first label (e.g., single), otherwise they are connected to the second label (e.g., invalid). That is, they are connected to the macro category MC that indicates the test manipulation and the corresponding single-class single-window classifier SCC'. pi MC macro category i The matching / non-matching tags between them.

[0134] General single-class single-window classifier SCC' pi Therefore, based on having a duration not exceeding TW' P The feature vectors computed on a portion of the time-extended data cluster DG are used for training.

[0135] Subsequently, consider the unknown flight and the general k-th time t. clkk Computer 12 will be connected with the p-th duration TW' p For each input feature vector FVX kp (where p = 1, ..., NUM_TW) is applied (box 970) to TW' with the same p-th duration. p Related single-class single-window classifier SCC' pi So that for each input feature vector FVX kp Obtain the corresponding single-class single-window probability vector SCV′ kp , where the i-th element represents the input feature vector FVX kp Involves, for example, a single-class single-window classifier SCC' pi The calculation of the i-th macro category MC i The probability of.

[0136] For example, Figure 27 The following single-class single-window classifiers SCC' are shown respectively. 11 -SCC' 14 SCC' 21 -SCC' 24 SCC' 31 -SCC' 34 and SCC' 41 -SCC' 44 The four input feature vectors FVX for classification k1 -FVX k4 In this way, a single-class, single-window probability vector SCV' is generated. k1- SCV' k4 For the sake of simplicity only, Figure 27 China has already adopted Figure 25 The same values ​​are used, but there may be differences in reality.

[0137] Then, computer 12 generates a single-window update vector OUT_UPDATE' (box 980) such that it has multiple elements equal to the number NUM_MC, and generally the i-th element is equal to the single-class single-window probability vector SCV'. kp The maximum value among the values ​​of the i-th element of (where p = 1, ..., NUM_TW). Although not described further, to generate the single-window update vector OUT_UPDATE′, the i-th element of the latter can be set to be equal to the single-class single-window probability vector SCV′. kp The i-th element is used instead of the statistic (e.g., the average) calculated from the maximum value mentioned above.

[0138] Subsequently, computer 12 detects (box 985) at time k t based on the single-window update vector OUT_UPDATE′. _clkk Whether two or more operations belonging to different macro categories (MC) are performed (i.e., it detects conditions of multiple macro categories), for example by detecting whether two or more elements of the single-window update vector 'OUT_UPDATE' exceed a predetermined threshold. Furthermore, although not further described, the detection of multiple macro categories can also provide further threshold control / modification mechanisms as described in reference box 930.

[0139] If no manipulation belonging to a different macro category MC is detected (box 985 outputs "No"), computer 12 executes the operation in box 940. Otherwise, if manipulation belonging to a different macro category MC is detected (box 985 outputs "Yes"), computer 12 will output the final vector OUT_FIN. k Set (box 990) to equal the single-window update vector OUT_UPDATE' k Then execute the operation in box 951.

[0140] Compared to operations in boxes 900-950, operations in boxes 960-990 are characterized by a greater computational burden and the creation of a large number of classifiers; however, they can guarantee good performance, especially when manipulating quantities with relatively constant trends.

[0141] Regardless of the strategy employed, computer 12 can perform detection (identification) of the corresponding manipulation being performed using the macro categories detected by the operation indicated in box 951. This identification can be performed deterministically based on the values ​​of one or more principal quantities, and by… Figure 23 or Figure 26 Box 991 in the diagram represents this.

[0142] Specifically, given that it is detected at box 951 and at time k t clkk In the case of a macro category related to the manipulation performed, computer 12 can, based on the detected macro category and at least one value of at least one major quantity of unlabeled data structure 205 (e.g., with respect to time k t), determine the appropriate macro category. clkk In or at other times t clk The speed-related value in the data is used to identify the manipulation.

[0143] Typically, information about the detected macro categories and (if needed) about the detected manipulations can be used to determine the operational status of the unknown helicopter 3, for example, in order to effectively plan the maintenance of the unknown helicopter 3.

[0144] For example, such as Figure 28As shown, computer 12 can store (box 992) information related to the loads experienced by one or more components of helicopter 1 when helicopter 1 has performed a corresponding maneuver belonging to the corresponding macro category MC; for this purpose, helicopter 1 is equipped with a load detection system 17 during the maneuver, which is configured to detect the loads experienced by the components.

[0145] Furthermore, for the sake of simplicity and conciseness, referring to a single component of the unknown helicopter 3, and assuming that the unknown helicopter 3 is equal to helicopter 1, computer 12 determines (box 993) the load that the component of the unknown helicopter 3 has experienced during that operation based on the stored load corresponding to that operation for each maneuver identified by the operation via box 991. Based on the determined load, computer 12 determines (box 994) the fatigue state of the component, and thus determines the remaining fatigue life of the component.

[0146] Based on the remaining fatigue life of the components of the unknown helicopter 3 as determined in this way, any maintenance operations of the unknown helicopter 3 can also be planned.

[0147] The advantages that this method allows for are clearly evident from the previous description.

[0148] In particular, this system allows for the precise detection of macro categories of maneuvers performed by the aircraft, regardless of the type of macro category and therefore its duration. Thus, such measurements can be reliably used to estimate the fatigue state of aircraft components and, consequently, their remaining fatigue life; and can therefore be used, for example, to optimize a fleet of aircraft maintenance operations to comply with safety requirements.

[0149] Obviously, however, changes can be made to the methods and systems described and shown herein without departing from the scope of the invention as defined in the appended claims.

[0150] For example, first-level classifiers and second-level classifiers can be different types from those already described.

[0151] The time window can be aligned differently from a specific moment, rather than being centered relative to a moment. For example, a reference might involve each duration TW′. P The operation in box 220 selects a subset of values ​​from the unlabeled data structure 205, which can be selected from values ​​falling within the range t. clkk and t clkk +TW′ p The values ​​of the unlabeled data structure 205 are formed within the time window between them.

[0152] Finally, typically, at least some macro categories can include a limited number of manipulations (in extreme cases, only the corresponding manipulations).

Claims

1. A computer-implemented method for detecting aircraft performing maneuvers belonging to a macro category among multiple macro categories, comprising the following steps (a1): (a1) Receive a data structure, the data structure comprising multiple time series of values ​​of quantities related to the flight of the aircraft, the values ​​having been acquired by a monitoring system coupled to the aircraft; The method further includes performing the following steps (a2) to (a4) for each of the first series of times: (a2) For each of a predetermined plurality of durations, select a corresponding subset of the data structure, the corresponding subset having a time extension equal to the duration, and the selected subset of the data structure having the same time distance from the time in the first series of times; (a3) Extract the corresponding feature vector from each selected subset of the data structure; (a4) Based on the feature vector, generate the corresponding input macro vector, alternatively through aggregation of the feature vector or through the steps (a41) to (a42): (a41) The classification of the feature vectors is performed to generate a plurality of input prediction vectors, each of the plurality of input prediction vectors indicating the probability that the aircraft is performing a maneuver belonging to the macro category at the time in the first series of times; as well as (a42) Subsequent aggregation of the multiple input prediction vectors; The method further includes performing the following steps (a5) and (a6) for each of the first series of times: (a5) Apply an output classifier to the input macro vector, the output classifier being configured to generate a corresponding output vector, the corresponding output vector including, for each of the macro categories, a corresponding estimate of the probability that the aircraft is performing a maneuver belonging to the macro category at the time in the first series of time steps; as well as (a6) Based on the output vector, detect the macro category to which the maneuver performed by the aircraft at the time in the first series of times belongs.

2. The method of claim 1, wherein the output classifier has been generated from a training data structure comprising multiple time series of values ​​of the quantity associated with aircraft flight, wherein during corresponding time intervals, a test maneuver belonging to the macro category has been performed, the values ​​of the training data structure associated with the time interval of each test maneuver are labeled with a label indicating the macro category to which the test maneuver belongs, and the output classifier has been generated by performing the following steps (b1) to (b4) for each time moment in a second series of time moments falling within the time interval of the test maneuver: (b1) For each of the predetermined plurality of durations, select a corresponding subset of the training data structure having a time extension equal to the duration, wherein the selected subset of the training data structure has the same time distance from the time in the second series of times; (b2) Extract the corresponding first training feature vector from each selected subset of the training data structure; (b3) Based on the first training feature vector, generate the corresponding training macro vector by aggregating the first training feature vector or by following steps (b31) and (b32): (b31) The classification of the first training feature vector is performed to generate a plurality of training prediction vectors, each of the plurality of training prediction vectors indicating the probability that the test manipulation belongs to the corresponding macro category for each of the macro categories; as well as (b32) Subsequent aggregation of the multiple training prediction vectors; (b4) The output classifier has also been generated by performing the following step (b41) for each of the second series of times falling within the time interval of the test manipulation: (b41) The output classifier is trained in a supervised manner based on the corresponding training macro vector and the label indicating the macro category to which the test manipulation belongs.

3. The method of claim 2, wherein the step of generating the corresponding input macro vector comprises, for each of the first series of times, the following steps (c1) to (c3): (c1) Apply the first classifier to each of the corresponding feature vectors, wherein the first classifier is configured to generate a corresponding first policy input prediction vector, and the first policy input prediction vector forms the plurality of input prediction vectors; (c2) Aggregate the first policy input prediction vector into a first policy input macro vector, and the first policy input macro vector forms the input macro vector; (c3) wherein the output classifier is formed by a first output classifier; and wherein the step of applying the output classifier to the corresponding input macro vector for each of the first series of times includes applying the first output classifier to the corresponding first policy input macro vector.

4. The method of claim 3, wherein the first classifier has been generated by performing the following steps (d1) to (d3): (d1) For each of the test manipulations, starting from the time series portion of the training data structure extending over the time interval of the test manipulation, extract the corresponding entire manipulation feature vector; as well as (d2) The first classifier is trained in a supervised manner based on the entire manipulation feature vector and the label of the macro category to which the test manipulation belongs. (d3) and wherein the first output classifier has been generated by performing the following steps (d31) to (d33) for each of the second series of times falling within the time interval of the test manipulation: (d31) Apply the first classifier to the corresponding first training feature vector to obtain a plurality of first policy training prediction vectors equal to the number of durations, the first policy training prediction vectors forming the plurality of training prediction vectors; and (d32) Aggregate the multiple first strategy training prediction vectors to form a corresponding first training prediction macro vector, and the first training prediction macro vector forms the training macro vector; (d33) The first output classifier is trained in a supervised manner based on the corresponding first training prediction macro vector and the label indicating the macro category to which the test manipulation belongs.

5. The method of claim 2, wherein the step of generating the corresponding input macro vector comprises, for each of the first series of times, the following steps (e1) to (e3): (e1) For each of the predetermined multiple durations, a corresponding second classifier from a plurality of different second classifiers equal to the number of predetermined durations is applied to a feature vector associated with the time and the duration in the first series of times, the corresponding second classifier being configured to generate a corresponding second policy input prediction vector, the second policy input prediction vector forming the plurality of input prediction vectors; (e2) The second policy input prediction vector is aggregated into a second policy input macro vector, and the second policy input macro vector forms the input macro vector; (e3) and among them, The output classifier is formed by the second output classifier; and the step of applying the output classifier to the corresponding input macro vector for each time in the first series of time steps includes applying the second output classifier to the corresponding second policy input macro vector.

6. The method of claim 5, wherein each second classifier has been generated by performing the following steps (f1) to (f4) for each of the test operations: (f1) Select a plurality of corresponding subsets of the training data structure that extend in the time interval of the test manipulation, each of the selected subsets having a time extension no higher than the duration corresponding to the second classifier; (f2) Extract the corresponding second training feature vector from each selected subset of the portion of the training data structure that extends during the time interval of the test manipulation; as well as (f3) The second classifier is trained in a supervised manner based on the extracted second training feature vector and the label indicating the macro category to which the test manipulation belongs; (f4) and wherein the second output classifier has been generated by performing the following steps for each of the second series of times falling within the time interval of the test manipulation: (f41) to (f43) (f41) For each duration, the second classifier corresponding to the duration is applied to a first training feature vector that corresponds to the time in the second series of times and has been extracted from a selected subset of training data structures having a time extension equal to the duration, so as to obtain a corresponding second policy training prediction vector, the second policy training prediction vector forming the plurality of training prediction vectors; (f42) Aggregate the second strategy training prediction vectors to form a corresponding second training prediction macro vector, and the second training prediction macro vector forms the training macro vector; (f43) The second output classifier is trained in a supervised manner based on the corresponding second training prediction macro vector and the label indicating the macro category to which the test manipulation belongs.

7. The method of claim 2, wherein the step of generating the corresponding input macro vector comprises, for each of the first series of times, the following steps (g1) to (g4): (g1) Apply a first classifier to each of the corresponding feature vectors, wherein the first classifier is configured to generate a corresponding first policy input prediction vector, and the first policy input prediction vector forms the plurality of input prediction vectors; (g2) For each of the predetermined plurality of durations, a corresponding second classifier from a plurality of different second classifiers equal to the number of predetermined durations is applied to the feature vector associated with the time in the first series of times and the duration, the corresponding second classifier being configured to generate a corresponding second policy input prediction vector, the first policy input prediction vector and the second policy input prediction vector forming the plurality of input prediction vectors; (g3) The first policy input prediction vector and the second policy input prediction vector are aggregated into a third policy input macro vector, and the second policy input macro vector forms the input macro vector; (g4) and wherein the output classifier is formed by a third output classifier; and wherein the step of applying the output classifier to the corresponding input macro vector for each of the first series of times includes applying the third output classifier to the corresponding third policy input macro vector.

8. The method of claim 7, wherein the first classifier is generated by performing the following steps (h1) to (h3): (h1) For each of the test operations, extract the corresponding whole operation feature vector starting from the time series portion of the training data structure extending over the time interval of the test operation; as well as (h2) The first classifier is trained in a supervised manner based on the entire manipulation feature vector and the label of the macro category to which the test manipulation belongs. (h3) and each of the second classifiers has been generated by performing the following steps (h31) to (h34) for each of the test operations: (h31) Select a plurality of corresponding subsets of the training data structure that extend in the time interval of the test manipulation, each of the selected subsets having a time extension no higher than the duration corresponding to the second classifier; (h32) Extract the corresponding second training feature vector from each selected subset of the portion of the training data structure that extends during the time interval of the test manipulation; as well as (h33) The second classifier is trained in a supervised manner based on the extracted second training feature vector and the label indicating the macro category to which the test manipulation belongs; (h34) and wherein the third output classifier has been generated by performing the following steps (h341) to (h344) for each of the second series of times falling within the time interval of the test manipulation: (h341) The first classifier is applied to the corresponding first training feature vector to obtain a plurality of first policy training prediction vectors equal to the number of predetermined durations; (h342) For each duration, the second classifier corresponding to the duration is applied to a first training feature vector that corresponds to each moment in the second series of moments and has been extracted from a selected subset of training data structures having a time extension equal to the duration, in order to obtain a corresponding second policy training prediction vector, the first policy training prediction vector and the second policy training prediction vector forming the plurality of training prediction vectors. (h343) Aggregate multiple first policy training prediction vectors and multiple second policy training prediction vectors to form a corresponding third training prediction macro vector, wherein the third training prediction macro vector forms the training macro vector; (h344) The third output classifier is trained in a supervised manner based on the corresponding third training prediction macro vector and the label indicating the macro category to which the test manipulation belongs.

9. The method of claim 1, further comprising performing steps (i1) to (i4) for each of the first series of times: (i1) For each duration, a plurality of single-class classifiers equal to the number of macro categories are applied to the feature vector corresponding to the duration and the time in the first series of times, each single-class classifier being configured to generate a corresponding probability value indicating the probability that the aircraft is performing a maneuver belonging to the macro category corresponding to the single-class classifier at the time in the first series of times; (i2) For each single-class classifier, calculate the corresponding statistic based on the probability value generated by the single-class classifier; (i3) Detection of manipulations belonging to multiple macro categories based on computational statistics; as well as (i4) If no execution of an operation belonging to multiple macro categories is detected, perform the step of detecting the macro category to which the operation performed by the aircraft belongs based on the output vector; otherwise, perform the step of detecting the macro category to which the operation performed by the aircraft at the time in the first series of time periods belongs based on the calculated statistics.

10. The method of claim 9, wherein the output classifier has been generated from a training data structure comprising multiple time series of values ​​of the quantity associated with aircraft flight, wherein during corresponding time intervals, a test maneuver belonging to the macro category has been performed, the values ​​of the training data structure associated with the time interval of each test maneuver are labeled with a label indicating the macro category to which the test maneuver belongs, and the output classifier has been generated by performing the following steps (j1) to (j4) for each time moment in a second series of time moments falling within the time interval of the test maneuver: (j1) For each of the predetermined plurality of durations, select a corresponding subset of the training data structure having a time extension equal to the duration, wherein the selected subset of the training data structure has the same time distance from the time in the second series of times; (j2) Extract the corresponding first training feature vector from each selected subset of the training data structure; (j3) Based on the first training feature vector, generate the corresponding training macro vector, or alternatively by aggregating the first training feature vector or by the following steps (j31) to (j1): (j31) The classification of the first training feature vector is performed to generate a plurality of training prediction vectors, each of the plurality of training prediction vectors indicating the probability that the test manipulation belongs to the corresponding macro category for each of the macro categories; as well as (j32) Subsequent aggregation of the multiple trained prediction vectors; (j4) The output classifier has also been generated by performing the following steps (j41) to (j44) for each of the second series of times falling within the time interval of the test manipulation: (j41) The output classifier is trained in a supervised manner based on the corresponding training macro vector and the label indicating the macro category to which the test manipulation belongs; (j42) and each single-class classifier has been generated by performing the following steps for each of the test operations: (j43) Starting from the time series portion of the training data structure extending over the time interval of the test manipulation, extract the corresponding entire manipulation feature vector; as well as (j44) The single-class classifier is trained in a supervised manner based on the extracted whole manipulation feature vector and the labels indicating the matching between the macro category of the test manipulation and the macro category corresponding to the single-class classifier.

11. The method of claim 10, further comprising performing steps (k1) to (k4) for each of the first series of times: (k1) For each duration, a plurality of corresponding single-class classifiers corresponding to the duration are applied to the feature vector corresponding to the duration and the time, each of the plurality of corresponding single-class classifiers being configured to generate a corresponding probability value, the probability value indicating the probability that the aircraft is performing a maneuver belonging to the macro category corresponding to the single-class classifier at the time of the first series of times; (k2) For each macro category, the corresponding statistic is calculated based on the probability values ​​generated by the single-class classifier that correspond to the macro category and respectively to the duration; (k3) Statistical detection based on computation to identify manipulations belonging to multiple macro categories; as well as (k4) If no execution of a maneuver belonging to multiple macro categories is detected, perform the step of detecting the macro category to which the maneuver performed by the aircraft belongs based on the output vector; otherwise, perform the step of detecting the macro category to which the maneuver performed by the aircraft at the time in the first series of time based on the calculated statistics.

12. The method of claim 11, wherein the output classifier has been generated from a training data structure comprising multiple time series of values ​​of the quantity associated with aircraft flight, wherein during corresponding time intervals, a test maneuver belonging to the macro category has been performed, the values ​​of the training data structure associated with the time interval of each test maneuver are labeled with a label indicating the macro category to which the test maneuver belongs, and the output classifier has been generated by performing the steps (l1) to (l4) for each time moment in a second series of time moments falling within the time interval of the test maneuver: (l1) For each of the predetermined plurality of durations, select a corresponding subset of the training data structure having a time extension equal to the duration, wherein the selected subset of the training data structure has the same time distance from the time in the second series of times; (l2) Extract the corresponding first training feature vector from each selected subset of the training data structure; (l3) Based on the first training feature vector, generate the corresponding training macro vector, or alternatively by aggregating the first training feature vector or by the following steps (l31) and (l32): (l31) The classification of the first training feature vector is performed to generate a plurality of training prediction vectors, each of the plurality of training prediction vectors indicating the probability that the test manipulation belongs to the corresponding macro category for each of the macro categories; as well as (l32) Subsequent aggregation of the multiple training prediction vectors; (l4) The output classifier has also been generated by performing the following steps (l41) and (l42) for each of the second series of times falling within the time interval of the test manipulation: (l41) The output classifier is trained in a supervised manner based on the corresponding training macro vector and the label indicating the macro category to which the test manipulation belongs; (l42) and each single-class classifier has been generated by performing the following steps (l421) to (l423) for each of the test operations: (l421) Select a plurality of corresponding subsets of the training data structure that extend in the time interval of the test manipulation, each of the selected subsets having a time extension no higher than the duration corresponding to the single class classifier; (l422) Extract the corresponding second training feature vector from each selected subset of the portion of the training data structure that extends during the time interval of the test manipulation; as well as (l423) The single-class classifier is trained in a supervised manner based on the extracted second training feature vector and the label indicating the matching between the macro category of the test manipulation and the macro category corresponding to the single-class classifier.

13. The method according to any one of claims 1 to 8, further comprising the step (m): (m) Based on the detected macro categories, and at least based on the value of at least one of the quantities related to the flight of the aircraft, identify the maneuver performed by the aircraft at the time in the first series of times.

14. The method according to claim 9 or 11, further comprising the steps (n1) and (n2): (n1) In the absence of detection of the execution of an operation belonging to multiple macro categories, based on the macro categories, identify the operation performed by the aircraft at the times in the first series of times, the macro categories being detected based on at least one value of at least one of the output vector and the flight-related quantities of the aircraft; (n2) In the event that the execution of an operation belonging to multiple macro categories is detected, the operation performed by the aircraft at the time in the first series of times is identified based on the macro categories, wherein the macro categories are detected based on at least one value of at least one of the calculated statistics and the flight-related quantities of the aircraft.

15. A computer-implemented method for determining the operational status of an aircraft, comprising the steps (o1) to (o3): (o1) Store information that associates payload data with manipulation; (o2) Perform the method according to claim 13 or 14; (o3) Based on the identified manipulation, determine at least one quantity indicating the state of use.

16. A processing system, comprising: An apparatus configured to perform the method of any one of claims 1 to 15.

17. A computer program product comprising a computer program that, when executed by a computer, implements the method of any one of claims 1 to 15.

18. A computer-readable storage medium storing a computer program that, when executed by a computer, implements the method of any one of claims 1 to 15.

Images