Method for monitoring state of health of aircraft turbine engine
By constructing the transient and steady-state data sets of the turbine engine and establishing a prediction model, the problem of missing steady-state data of the helicopter turbine engine is solved, and accurate monitoring of the turbine engine status and visual analysis of the health status are achieved.
Patent Information
- Application Number
- CN202380031114.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-03-28
- Filing Date
- 2023-03-27
- Publication Date
- 2024-11-12
AI Technical Summary
The prior art is difficult to effectively monitor the steady-state data of helicopter turbine engines, resulting in complex or impossible margin calculations, and the inability to accurately predict engine failure and overheating.
By building training databases, including transient and steady-state datasets, establishing transient and steady-state prediction models, estimating the steady-state behavior of the turbine engine, and calculating margins based on the set point matrix, providing a more comprehensive analysis point to monitor the engine Health status.
Accurate monitoring of turbine engine status under different conditions is achieved, estimation error is reduced, more flexible models are provided for multi-point analysis, and the ability to visualize and trend analysis of engine health status is enhanced.
Smart Images

Figure CN118946710A_ABST
Abstract
Description
Technical Field
[0001] The technical field of the invention is that of monitoring the condition of aircraft turbine engines.
[0002] The invention relates to a method for monitoring the health state of an aircraft turbine engine. Background Art
[0003] Health monitoring of a turbine engine of an aircraft, such as an airplane or a helicopter, makes it possible to monitor the condition of the turbine engine throughout its service life and to predict faults in the turbine engine based on data recorded during flights performed by the aircraft comprising the turbine engine.
[0004] A traditional indicator of health monitoring is a margin calculation, which is used to compare the expected value of at least one input parameter related to the turbine engine, the model theoretical value of at least one output parameter related to the turbine engine, and the actual value of the output parameter. For example, in order to obtain the expected power P emitted by the turbine engine, a physically modeled temperature T0 in the combustion chamber of the turbine engine is theoretically required. The temperature T0 is compared with the actual temperature T1 in the combustion chamber of the turbine engine to obtain the power P emitted by the turbine engine. The actual temperature T1 increases during the life cycle of the turbine engine, and the difference between the theoretical value T0 and the actual value T1 is a margin for research and analysis using a physical model to predict, for example, possible failures and overheating of the turbine engine. Margin calculations can be performed for several input parameter values (for example, the rotation speed of the first shaft of the turbine engine and the rotation speed of the second shaft of the turbine engine) and several output parameter values (for example, the power emitted by the turbine engine and the temperature of the turbine engine discussed above).
[0005] The physical model makes it possible to study the margins using parameter data related to the turbine engine recorded during the steady state of the turbine engine. However, a helicopter (e.g., one performing surveillance or lifting) may rarely or never operate in a steady state during a particular flight, which makes the margin calculation complicated or even impossible.
[0006] Patent application FR3028331A1 describes a monitoring algorithm based on regression. Unsupervised classification of exogenous variables (environmental parameters) distinguishes flight contexts. For each type of exogenous condition, a regression model is provided to normalize endogenous data (engine parameters) and offset the effects of exogenous changes. Therefore, all flights are of the same background and are in fact comparable. They find the steady-state phase of all flights based on the distance from a set of stable reference phases, and then calculate the average physical parameters of each phase that they can track over time. The application uses an algorithm for representing exogenous data (in the form of categories) with the aim of eliminating flight contexts and making it possible to distinguish between several related stable phases.
[0007] Patent EP2623747B1 provides an inspection device for automatically performing an engine health check on an aircraft turboshaft engine, comprising a step of acquiring (mechanical) parameters for monitoring the turboshaft engine through guidance control (the method automatically determines whether the current phase is suitable for data listing and warns the pilot that acquisition is in progress) and a step of evaluating the health of the engine relative to reference parameters (measured on a test bench or theoretical). The patent describes an invention based on a physical model of engine operation in predefined phases. The invention described in EP2623747B1 makes it possible to automatically detect the relevant phases and compensate for the absence of the relevant phases during flight.
[0008] In order to compensate for the lack of steady-state data of helicopter turbine engines, there are solutions in the prior art that monitor the engine status using indicators other than margin or other data processing methods that do not focus on physical modeling.
[0009] For example, patent EP2676176B1 provides a method for monitoring aircraft engines using endogenous data (engine related) and exogenous data (environment related) recorded during the flight. Each flight is standardized by exogenous variables to have a comparable background between flights (see, for example, FR3035232A1), and then compressed and projected onto an adaptive Kohonen map. For the flight of interest, the map is used to find similar flights in the past and statistically analyze any drift in engine behavior between the two. In patent EP2676176B1, the flight data representation is used to search for past data similar to the observed data and analyze the drift. The solution in EP2676176B1 can be distinguished by passively searching for drift in the projected data and using a single specific representation method. The solution is described in the paper "Aircraft engine health monitoring using Self-Organizing Maps" E. A more precise description is given in M. Cottrell, M. Verleysen and J. Lacaille, (2010).
[0010] In addition, patent application US20180297718A1 provides a method for monitoring the health of a turboshaft engine. The mechanism is as follows: in different phases of use, obtain initial data used as a reference for defining a healthy engine; obtain data during the service life of the device; compare using a module that calculates the difference between the "healthy" parameters and the new parameters. Set maintenance alarms. Use industry knowledge and the definition of health states. Use the drift of physical parameters for each flight phase.
[0011] An alternative to the above solution is also known in the prior art, which is based on estimating the value of an output parameter related to a steady-state turbine engine according to the value of an input parameter related to a transient turbine engine during a given flight. This solution enables two parameters related to the turbine engine to be converted independently of each other in order to obtain the value of the output parameter in the steady state for a given input parameter. However, the solution provided is only applicable to one pair of parameters (input, output) at a time, and each parameter is converted independently of the other parameter to obtain a pair of parameters (input, output) in the steady state. In addition, the margin can therefore only be calculated based on the converted parameters, but not for any value of the input parameter, which further limits the relevant margin calculation.
[0012] Therefore, there is a need to estimate the behavior of a steady state turbine engine based on one or more input parameters associated with a transient turbine engine without considering the input and output parameters independently of each other or two by two. Summary of the invention
[0013] The present invention provides a solution to the previously discussed problems by making it possible to more effectively estimate the steady-state behavior of a turbine engine for a particular flight.
[0014] One aspect of the present invention relates to a computer-implemented method for determining a set point matrix X according to the set point matrix X. SC The setpoint matrix X is used to monitor the health status of the aircraft's turbine engine TMI for the aircraft's flight VI of interest. SC Including at least one value of at least one input parameter related to a turbine engine TMI of interest, the method comprises the following steps:
[0015] -Build a training database, which includes:
[0016] o a so-called transient set T comprising transient learning data recorded during at least one flight V of at least one turbine engine TM operating in transient conditions and at least one flight recording instant of a flight of interest of a turbine engine of interest TMI;
[0017] o a so-called steady-state set S comprising steady-state learning data recorded at each flight recording instant of each flight V other than the flight VI of interest, for each turbine engine TM operating in steady-state conditions;
[0018] - Estimate the transient prediction model f based on the transient learning data of the transient set T;
[0019] - estimating a steady-state prediction model H based on the transient prediction model f and the steady-state learning data included in the steady-state set S;
[0020] - Based on the steady-state prediction model H, at least a portion of the transient prediction model f and the set point matrix X SC to estimate the matrix of interest Y representing the behavior of the steady-state turbine engine of interest TMI for the flight VI of interest SI ;
[0021] - According to the matrix of interest Y SI The margins are calculated to estimate the state of the turbine engine of interest.
[0022] Model f can be composed of one or more transient prediction sub-models, which may or may not be specific to a particular dataset.
[0023] The prediction model H may be formed from one or more steady-state prediction models, as described below.
[0024] By “input parameters related to the turbine engine” it is meant parameters related to the turbine engine for which setpoint values are required.
[0025] By “steady-state turbine engine behavior for flight V” is meant an estimation of at least one output parameter related to the turbine engine, the output parameter being a result of one or more input parameters of the turbine engine and the state of the turbine engine.
[0026] Margin is a physical trading indicator that uses empirical feedback and physical knowledge to inform after-sales service of the state of the engine. Therefore, the margins calculated (simulated / generated) by the present invention are either displayed in a viewing tool on a computer and analyzed by an operator, or analyzed in an algorithm pipeline in a computer to formally infer its trend and alert the operator to the health state of the monitored engine. Advantageously, the present invention allows for more available points for analysis, thereby compensating for the lack of steady-state data, and therefore enables the operator to better perform visual analysis or better calculate trends, thereby better estimating the health state of the monitored engine.
[0027] Thanks to the invention, it is possible to estimate the steady-state behavior of a turbine engine of interest for any flight of interest, using a steady-state prediction model H, based on transient and steady-state data related to at least one turbine engine in transient and steady states for a plurality of flights. Thus, unlike the prior art, the invention not only makes it possible to predict output data related to a steady-state turbine engine only under the conditions encountered, but also provides a model that makes it possible to predict output data according to any set point, so that the estimation of the model H makes it possible to calculate margins at different points, rather than being limited to the transformation of transient variables, thus making it possible to monitor the state of the turbine engine for any flight of interest. The calculation of margins at different points makes it possible to have more analysis points available and therefore to make it possible to carry out a more precise analysis with less noise.
[0028] Advantageously, the invention allows all data related to the turbine engine to be considered at once, rather than in pairs. Thus, considering all variables rather than in pairs results in a more realistic model and thus reduces estimation errors. The reduction in estimation errors means allowing margins to be more accurate and relevant for analysis.
[0029] Advantageously, the present invention enables the construction of a steady-state prediction model H VI , which enables the calculation of margins at different points, rather than being limited to the transformation of transient variables. Producing such a model gives greater flexibility in the use of the results. It is therefore not a simple transformation of data, but a complete modeling of the turbine engine available, allowing a multi-point analysis of the operation.
[0030] In addition to the features just discussed in the previous paragraph, the method according to one aspect of the present invention may have one or more of the following additional features, which may be considered individually or in any technically feasible combination:
[0031] For each flight V of each turbine engine TM, the transient learning data set T consists of a pair (χ VT, ψ VT )Build:
[0032] oMatrix ψ VT , comprising at least one row and at least one column, each row corresponding to an input parameter of the transient turbine engine TM, each input parameter of each row being different from other input parameters of other rows, each column corresponding to a recording moment of the flight V, each recording moment of each column being different from other recording moments of other columns,
[0033] oMatrix ψ VT , comprising at least one row and at least one column, each row corresponds to an output parameter of the transient turbine engine TM, each output parameter of each row is different from other input parameters of other rows, each column corresponds to a recording moment of the flight V, each recording moment of each column is different from other recording moments of other rows,
[0034] For each flight V of each turbine engine TM except the flight of interest, the steady-state learning data set S consists of a pair (χ VS ,ψ VS )Build:
[0035] oMatrix χ VS , comprising at least one row and at least one column, each row corresponding to an input parameter of the transient turbine engine TM, each input parameter of each row being different from other input parameters of other rows, each column corresponding to a recording moment of the flight V, each recording moment of each column being different from other recording moments of other columns,
[0036] oMatrix ψ VS , comprising at least one row and at least one column, each row corresponds to an output parameter of the transient turbine engine TM, each output parameter of each row is different from other input parameters of other rows, each column corresponds to a recording moment of the flight V, and each recording moment of each column is different from other recording moments of other columns.
[0037] -Turbine engine TM is the turbine engine of interest. Therefore, in the matrix pair (χ VT ,ψ VT ), at least one of these pairs represents a flight of the turbine engine of interest. This feature makes it possible to include data relating to the turbine engine of interest for at least one flight, making it possible to estimate a prediction model f making it possible to determine the behavior for one flight from the data recovered from the turbine engine of interest. This step makes it possible to incorporate the current behavior of the turbine engine under consideration and ensure that the model is not a simple abstract model, but a behavioral model of this specific machine, with all the available information.
[0038] - the flight V of the set T is the flight VI of interest of the turbine engine of interest, and the set T is further composed of a pair (χ VIT ,ψ VIT) is constructed. This feature comprises the parameters related to the turbine engine of interest for the flight of interest, making it possible to estimate a prediction model f making it possible to represent the turbine engine of interest for the flight of interest. The inclusion of these specific data makes it possible to obtain a model behavior that represents reality and therefore has a trading margin that corresponds to reality, that is to say the actual state of this turbine engine.
[0039] For each flight V of each turbine engine TM in the set T:
[0040] o The transient learning data of set T includes the pairs (χ VT, ψ VT ):
[0041] oThe transient prediction model f consists of at least one sub-model f v Formed so that f v (x VT )=ψ VT .
[0042] According to an embodiment that complements the first embodiment, the steady-state prediction model H is defined such that for each flight V of each turbine engine TM in the set S, H(f v (x VS ),x VS )=ψ VS .
[0043] According to one embodiment, the transient learning data of set T includes:
[0044] o is given by each matrix χ VT The matrix X obtained by horizontal concatenation of T ,
[0045] o is composed of each matrix ψ VT The matrix Y is obtained by horizontal concatenation of T ,
[0046] The transient prediction model f is defined such that f(X T )=Y T .
[0047] -The steady-state learning data of set S includes:
[0048] o is given by each matrix χ VS The matrix X obtained by horizontal concatenation of S ,
[0049] o is composed of each matrix ψ VS The matrix Y is obtained by horizontal concatenation of S ,
[0050] And the steady-state prediction model H is defined such that H(f(XS ),X S )=Y S .
[0051] - Estimating the steady-state prediction model H according to the transient prediction model f and the steady-state learning data included in the steady-state set S comprises the following sub-steps:
[0052] o For each flight V of each turbine engine TM, except for the flight of interest VI, estimate the steady-state prediction model H from the transient prediction model f and the steady-state learning data of each flight recording instant of the flight V included in the steady-state set S V ;
[0053] o Based on at least one steady-state prediction model H estimated for flight C C Construct a steady-state prediction model H for the flight VI of interest VI , the flight C is one of the one or more flights V;
[0054] o Steady-state prediction model H is the steady-state prediction model H VI .
[0055] For each flight V of each turbine engine TM in the stationary set S, except for the flight of interest VI of the turbine engine of interest TMI:
[0056] o The steady-state learning data of set S includes (χ VS ,ψ VS )
[0057] o Steady-state prediction model H V is defined such that Hv(f(χ VS ),x VS )=ψ VS
[0058] - Steady-state prediction model H VI constructed at least by the steady-state prediction model HC selected according to condition C1,
[0059] - Condition C1 is:
[0060] oX CT and χ VIT The distance between them is minimal;
[0061] o Steady-state prediction model H C The cost function is minimized; or
[0062] o Steady-state prediction model H C The generalization error is minimal.
[0063] According to one of the previous embodiments, the selection of the model H depends on the data, more specifically the noise level and the amount of available data.
[0064] - The method includes a margin calculation step, the margin and the matrix of interest Y SI and the matrix Y mp The difference between them is proportional to the matrix Y mp It means input X SC The output of the worst-case physical model.
[0065] - at least one input parameter related to the turbine engine is an operating parameter of the turbine engine, among the following parameters:
[0066] o the speed of the first shaft of the turbine engine;
[0067] o the speed of the second shaft of the turbine engine;
[0068] oThe temperature of the combustion chamber of the turbine engine;
[0069] oThe torque delivered by the turbine engine;
[0070] oAny other relevant thermodynamic parameters
[0071] or a parameter specific to a given environment in which the turbine engine is located, among the following parameters:
[0072] oThe temperature of a given environment;
[0073] oThe pressure of a given environment.
[0074] oRelease power
[0075] o any other parameters that influence the thermodynamic behavior of the turbine engine (e.g. valves);
[0076] - at least one output parameter related to the turbine engine is an operating parameter of the turbine engine, among the following parameters:
[0077] o the speed of the first shaft of the turbine engine;
[0078] o the speed of the second shaft of the turbine engine;
[0079] oThe temperature of the combustion chamber of the turbine engine;
[0080] oThe torque delivered by the turbine engine;
[0081] oAny other relevant thermodynamic parameters
[0082] or a parameter specific to a given environment in which the turbine engine is located, among the following parameters:
[0083] oThe temperature of a given environment;
[0084] oThe pressure of a given environment.
[0085] oRelease power
[0086] o Any other parameters that influence the thermodynamic behavior of the turbine engine (e.g., valves)
[0087] - Prediction model f and prediction model H are selected from conventional regression methods, such as but not limited to neural networks, decision trees, random forests, wide margin separators or random polynomial regression, wide margin separators or linear regression.
[0088] Another aspect of the invention relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to implement the method according to the invention.
[0089] Another aspect of the present invention relates to a computer-readable recording medium including instructions, which, when executed by a computer, cause the computer to implement the method according to the present invention.
[0090] The invention and its various applications will be better understood after reading the following description and viewing the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0091] The accompanying drawings are intended to illustrate, not to limit, the invention.
[0092] Figure 1 [ Figure 1 ] shows a schematic representation of a block diagram of a method according to the invention;
[0093] Figure 2 [ Figure 2 ] is an example of data recovered during the following flights of the turbine engine of interest: (Flight 0, Flight 1, Flight 2, Flight VI). DETAILED DESCRIPTION
[0094] The accompanying drawings are intended to illustrate, not to limit, the invention.
[0095] A first aspect of the present invention relates to a method for SC The method for monitoring the health status of the turbine engine TMI of the aircraft of interest for the flight VI of interest is as follows: SC At least one input parameter related to a steady-state turbine engine of interest TMI is included.
[0096] An aircraft is, for example, an airplane or a helicopter.
[0097] By "turbine engine" is meant a system utilizing a gas turbine in which energy is transferred between rotating components and gases.
[0098] The following features are applicable to any type of turbine engine within the scope of the present invention, including the turbine engine of interest.
[0099] A turbine engine may include a combustion chamber, a first shaft, and a second shaft.
[0100] By "shaft" is meant a mechanical element for transmitting power as torque and rotational movement.
[0101] The turbine engine is, for example, a turbojet engine, a turboprop engine or preferably a turboshaft engine.
[0102] By “input parameters related to the turbine engine” it is meant parameters related to the turbine engine for which setpoint values are required.
[0103] The input parameters related to the turbine engine may be operating parameters of the turbine engine among the following: the rotational speed of the first shaft of the turbine engine; the rotational speed of the second shaft of the turbine engine; the temperature of the combustion chamber of the turbine engine of interest; and the torque delivered by the turbine engine, the power delivered by the turbine engine. The input parameters related to the turbine engine may also be parameters specific to the external environment in which the aircraft including the turbine engine is located during a given flight, parameters among the following: the temperature of the external environment; the pressure of the external environment. The input parameters related to the turbine engine may also be any relevant thermodynamic or environmental parameter.
[0104] For example, two input parameters may be selected from the parameters previously associated with the turbine engine: a setpoint value for the power delivered by the turbine engine and a setpoint value for the rotational speed of a first shaft of the turbine engine that may be selected or required.
[0105] An output parameter of the turbine engine is a parameter that is a result of one or more input parameters of the turbine engine and a state of the turbine engine.
[0106] At least one output parameter associated with the turbine engine of interest may be an operating parameter of the turbine engine among: the rotational speed of a first shaft of the turbine engine; the rotational speed of a second shaft of the turbine engine; the temperature of a combustion chamber of the turbine engine, the torque delivered by the turbine engine and the power delivered by the turbine engine.
[0107] For example, if the input parameters are: the rotational speed of a first shaft of a turbine engine; the rotational speed of a second shaft of the turbine engine; the temperature of a given external environment; the pressure of a given external environment, the output parameters related to the turbine engine may be the temperature of a combustion chamber of the turbine engine and the torque delivered by the turbine engine.
[0108] In the following, the phrase “parameters related to the turbine engine” includes input parameter(s) related to the turbine engine and output parameters related to the turbine engine.
[0109] By "steady state of the turbine engine" is meant a state in which input and output parameters associated with the turbine engine vary little or not at all over time. A steady state may be referred to as a permanent state or a stationary state.
[0110] The steady state of the turbine engine is in contrast to the transient state of the turbine engine, in which input parameters and output parameters associated with the turbine engine vary with time.
[0111] Set point matrix X SC is a matrix including at least one value of at least one input parameter associated with a steady-state turbine engine of interest, each value of which may be randomly generated or selected. SC can have a number of rows greater than or equal to 1 and a number of columns greater than or equal to 1. The matrix X SC The coefficients are denoted by (x SC i,j ) i≥0,j≥0 .
[0112] Figure 1 [ Figure 1 ] is a block diagram of method 100 according to the present invention.
[0113] The method 100 may include a first step 101 of constructing a training database.
[0114] The construction step 101 may comprise a first sub-step 1011 of recovering, for at least one turbine engine TM, data D recorded during one or more flight recording moments during at least one flight V of said turbine engine TM.
[0115] According to one embodiment, the turbine engine TM is a turbine engine of interest TMI.
[0116] According to one embodiment, the data D include the following data: N ) N>0 The data recorded during one or more flight moments.
[0117] According to one embodiment, the data D include the following: 0-TM0 ,…V N-TM0 ,V 0-TMK …V P-TMK ) P>1 One or more flight record moments during which multiple turbine engines (TM0, ...TM K ) K>1The recorded data. For each turbine engine of the plurality of turbine engines, the number of flight moments for each flight of the plurality of flights may be different or equal from flight to flight and / or different or equal from turbine engine to turbine engine.
[0118] According to one embodiment, in which the turbine engine TM is the turbine engine of interest TMI, the flight V performed by the turbine engine of interest is performed before the flight of interest VI.
[0119] For each turbine engine TM and for each flight V of the turbine engine TM, the recovered data D includes transient data recorded during a transient state of the turbine engine TM during at least one flight moment of each flight V of the turbine engine TM, the flight V may be a flight VI of interest, and includes steady-state data recorded during a steady state of the turbine engine TM during at least one flight moment of each flight V of the turbine engine TM (except the flight VI of interest).
[0120] According to one embodiment, for each turbine engine TM and for each given flight V of the turbine engine TM, the transient data comprises a pair of matrices (χ VT , ψ VT ).
[0121] According to the aforementioned embodiment, each matrix x VT It comprises at least one row and at least one column, each row corresponding to an input parameter of the turbine engine TM, each input parameter of each row being different from the other input parameters of other rows, each column corresponding to a recording moment of the flight V during a transient phase, each recording moment of each column being different from the other recording moments of other columns.
[0122] Let x T V_ij is the matrix χ VT The coefficient of x T V_ij is equal to the value of the input parameter i at the recording instant j during the so-called transient phase for a given flight V. The coefficients i and j are respectively natural numbers greater than or equal to 0. The index V denotes the given flight V and the index T denotes the transient state.
[0123] For example, for a flight V of a transient turbine engine TM, the values of p input parameters related to the turbine engine TM are measured, p is an integer greater than 0, these values are measured for l records at the time of flight, l is greater than 0, and the matrix x VT as follows:
[0124]
[0125] According to the aforementioned embodiment, each matrix ψ VTIt comprises at least one row and at least one column, each row corresponding to a different output parameter of the transient turbine engine TM and each column corresponding to a different recording moment of the flight v during the transient phase.
[0126] Let y T V_ij is the matrix ψ VT The coefficient of y T V_ij is the value of output parameter i at recording instant j during the so-called transient phase for a given flight V. The coefficients i and j are respectively natural numbers greater than or equal to 0. The index V denotes a given flight V and the index T denotes a transient state.
[0127] For example, for a flight V of a steady-state turbine engine TM, the values of q input parameters related to the turbine engine TM are measured, q is an integer greater than 0, these values are measured for l records at the flight time, l is greater than 0, and the matrix ψ VT as follows:
[0128]
[0129] For each flight V of the turbine engine TM, χ VT The flight record time corresponds to ψ VT Recording moments.
[0130] According to one embodiment, for each flight V of each turbine engine TM, the transient data comprises a matrix (χ VT ) t and (ψ VT ) t , (χ VT ) t is the matrix χ VT is the transpose of VT ) T is the matrix ψ VT The transpose of .
[0131] According to an embodiment that complements the preceding one, for each turbine engine TM and for each given flight V of a turbine engine TM, except for the flight V1 of interest of the turbine engine of interest, the steady-state data comprises a pair of matrices (χ VS , ψ VS ).
[0132] Each matrix χ VS It comprises at least one row and at least one column, each row corresponds to an input parameter of the turbine engine TM, each input parameter of each row is different from other input parameters of other rows, each column corresponds to a recording moment of the flight V in a steady-state phase, and each recording moment of each column is different from other recording moments of other columns.
[0133] Let x S V_ij is the matrix χ VS The coefficient of x S V_ij is equal to the value of the input parameter i at the recording instant j during the so-called steady-state phase for a given flight V. The coefficients i and j are respectively natural integers greater than or equal to 0. The index V denotes the given flight V and the index T denotes the transient state.
[0134] For example, for a flight V of a steady-state turbine engine TM, the values of p input parameters related to the turbine engine TM are measured, p is an integer greater than 0, these values are measured for r records at the time of flight, r is greater than 0, and the matrix x VS as follows:
[0135]
[0136] Each matrix ψ VS , each matrix comprises at least one column and at least one row, each row corresponding to a different output parameter of the transient turbine engine TM, and each column corresponding to a different recording moment of the flight v.
[0137] Let y S V_ij is the matrix ψ VS The coefficient of y S V_ij is the value of output parameter i at recording instant j during the so-called transient phase for a given flight V. The coefficients i and j are respectively natural numbers greater than or equal to 0. The index V denotes a given flight V and the index T denotes a transient state.
[0138] For example, for a flight V of a steady-state turbine engine TM, the values of q input parameters related to the turbine engine TM are measured, q being an integer greater than 0, these values are measured for r records at the time of flight, r being greater than 0, and the matrix ψ VS as follows:
[0139]
[0140] For each flight V of each turbine engine TM, χ VS The flight record time corresponds to ψ VS Recording moments.
[0141] According to this embodiment, the data D include the data for the turbine engine TM of interest during at least one flight moment of the flight VI of interest. I The recorded transient data includes a matrix x VIT and ψ VIT .
[0142] According to an embodiment that complements the previous one, for each turbine engine TM and for each given flight V of a turbine engine TM, the matrix x VIT and χ VT The distance between them is less than the threshold S1.
[0143] According to an embodiment compatible with (but not exclusive of) the previous embodiment, the matrix x VIT and ψ VT The distance between them is less than the threshold S2.
[0144] The distance between matrices can be any relevant distance, such as Manhattan distance, Euclidean distance, Minkowski distance, or Chebyshev distance.
[0145] Figure 2 [ Figure 2 ] is an example of data recovered during the following flights of the turbine engine of interest: (Flight 0, Flight 1, Flight 2, Flight VI).
[0146] The first step 101 of the method according to the invention comprises a second sub-step 1012 for dividing the data D into two sets: a so-called transient set T and a so-called steady-state set S. The transient data for each turbine engine TM and for each flight V of the turbine engine TM are used to construct the transient learning data included in the set T, and the steady-state data for each turbine engine TM and for each flight V of the turbine engine TM are used to construct the steady-state learning data, distributed in the set S.
[0147] According to one embodiment, for each flight V of each turbine engine TM, by VT ,ψ VT ) constructs transient learning data for set T, flight V may be of interest flight VI.
[0148] According to the above embodiment, for each flight V of each turbine engine TM (except the flight VI of interest), the (x VS , ψ VS )Construct the transient learning data of set S.
[0149] According to a first sub-embodiment supplementing the above-mentioned embodiment, for each turbine engine TM and for each flight V, the transient learning data of the set T comprises a matrix pair (χ VT , ψ VT ), where flight V may be flight VI of interest.
[0150] According to a first sub-embodiment, the set S of steady-state learning data comprises, for each turbine engine TM and for each flight V (except the flight of interest VI of the turbine engine of interest), a matrix pair (χVS ,ψ VS ).
[0151] For example, if the recorded data D includes data from a flight (V1, V2) of a turbine engine TM, the set T may include the following matrix pair: (x V1T ,ψ V1T )、(χ V2T ,ψ V2T ), and the set S may include the following matrix pairs: (χ V1S ,ψ V1S )、(χ V2S ,ψ V2S ).
[0152] For example, if the recorded data D comprises data recorded during the flight (V1, V2, VI) of the turbine engine of interest TMI, the set T may comprise the following matrix pair: (x V1T ,ψ V1T )、(χ V2T ,ψ V2T ) and (χ VIT, ψ VIT ), and the set S may include the following matrix pairs: (χ V1S ,ψ V1S ),(χ V2S ,ψ V2S ).
[0153] According to a second sub-embodiment supplementing the above-mentioned embodiment, the transient learning data of the set T includes a pair of matrices (X T ,Y T ), the pair of matrices consists of each pair (χ VT ,ψ VT ) build.
[0154] According to the second sub-embodiment, the steady-state learning data of the set S includes a pair of matrices (X S ,Y S ), which is a pair of matrices consisting of each pair (χ VS ,ψ VS ) build.
[0155] According to a second sub-embodiment, the matrix X T From each matrix χ VT Therefore, when the data D includes data for multiple flights (V0, ..., V k ) k>0 When at least one record of the transient data recorded in the flight time is recorded, the matrix X THas the following form:
[0156] X T =[χ 0T … kT ]
[0157] To simplify the notation, each flight V k In the matrix X T is marked as k.
[0158] According to the aforementioned second sub-embodiment, the matrix Y T By each matrix ψ VT is produced by the horizontal cascade of.
[0159] Therefore, when the data D includes data for multiple flights (V0, ..., V k ) k>0 When at least one record of the transient data recorded in the flight time is T Has the following form:
[0160] Y T =[ψ 0T …ψ kT ]
[0161] To simplify the notation, each flight V k In the matrix Y T is marked as k.
[0162] X T and Y T Each column of represents the same flight V.
[0163] According to a second sub-embodiment, the matrix X S From each matrix χ VS Therefore, when the data D includes data for multiple flights (V0, ..., V N ) N>0 When the transient data recorded at least one time of flight, the matrix X S Has the following form:
[0164] X S =[χ 0S … NS ]
[0165] To simplify the notation, each flight V N In the matrix X S It is marked as N.
[0166] According to the second sub-embodiment, the matrix Y S By each matrix ψ VS is produced by the horizontal cascade of.
[0167] Therefore, when the data D includes data for multiple flights (V0, ..., V N ) N>0 When at least one record of the flight time records the transient data, the matrix Y S Has the following form:
[0168] Y S =[ψ 0S …ψ NS ]
[0169] To simplify the notation, each flight V N In the matrix Y S It is marked as N.
[0170] Matrix X T and X S Each row of represents the same input parameter associated with the turbine engine TM, and the matrix Y T and Y S Each row of represents the same output parameter associated with the turbine engine TM.
[0171] The method further comprises a second step 102 of estimating a transient prediction model f based on the learning data of the transient set T.
[0172] According to a first embodiment, for each flight V of each turbine engine TM, the transient learning data of the set T comprises a pair of matrices (χ VT ,ψ VT ), the transient prediction model consists of at least one transient prediction sub-model f v Formed so that f v (x VT )=ψ VT , and using each model f v The estimation step 102 is performed based on an estimate of
[0173] Therefore, when the transient learning data of the set T includes the flight (V0, ... V K ) k>0 Multiple matrix pairs ((χ V0T ,ψ V0T ),…(χ VKT ,ψ VKT )) k>0 When the sub-model (f V0 ,…f V K) k>0 The transient prediction model f is formed by the sub-models, which makes it possible to reduce the influence of noise in the data and utilize the similarities in the transient operation of different turbine engines TM. Specifically, according to the set T including a pair (χ VIT ,ψ VIT), the transient prediction model f is composed of at least a transient prediction sub-model f for the flight VI of interest. VI form.
[0174] According to a first sub-embodiment of the first embodiment, each transient prediction sub-model f for each flight V of the transient set T is estimated independently of other transient prediction sub-models. v Specifically, when the set T includes the turbine engine TM of interest I The matrix pair (χ VIT ,ψ VIT ), the transient prediction model f consists of at least sub-model f VI form.
[0175] According to the first embodiment, each transient prediction model f v This can be a model chosen from all the general regression models, such as neural networks, regression trees, random forests, wide margin separators, or linear regression.
[0176] According to a second sub-embodiment of the first embodiment, each transient prediction sub-model f for each flight V of the transient set T is estimated by relying on other transient prediction sub-models using a multi-task learning method. v This multi-task learning makes it possible to reduce the noise in the data and cover all parts of the data seen by each transient, rather than a reduced set. This increased coverage makes it possible to evaluate margins at multiple operating points and, therefore, to perform a more refined and precise analysis of the margins of the turbine engine under consideration.
[0177] Multi-task learning is a subfield of machine learning that enables solving several different tasks simultaneously, taking into account the dependencies between the tasks. A multi-task learning model includes a part common to each task and a part specific to each task. Multi-task learning makes it possible to improve the learning of a specific model by using features included in all tasks. Thus, according to a second embodiment, for each flight V of each turbine engine TM, each prediction submodel f v equal o F V ,function Each model f v common, and the function Fv is each sub-model f v Specific.
[0178] The multi-task learning model can be a model selected from a neural network, regression tree, random forest, wide margin separator, or linear regression.
[0179] For example, each prediction sub-model f vcan be associated with a set of training parameters, some of which are common to all other predictive sub-models and some of which are specific to the model.
[0180] According to the first embodiment, each transient prediction model f v The estimation of is performed by minimizing the cost function corresponding to the output data f provided by the transient prediction model f v (x VT ) and the expected real output data ψ VT The error between .
[0181] According to a first embodiment, for each flight V, the cost function is, for example, f v (x VT ) and ψ VT As a reminder, the matrix χ VT The coefficients are denoted by (x T v_ij ) i0,j≥0 , and the matrix ψ VT The coefficients are denoted by (y T V_ij ) i≥0,j≥0 The matrix f v (x VT ) will be labeled as (f(x T V_ij )) i≥0,j≥0 . Therefore, the cost function can be equal to:
[0182] Each submodel f can be optimized by using a gradient descent algorithm or a least squares algorithm or any other prior art method for optimizing the v The estimation method is used to perform the minimization of the cost function.
[0183] According to a second embodiment, wherein the set T comprises the matrix X T and the matrix Y T , the step 102 of estimating the transient prediction model f is to calculate the matrix X T and the matrix Y T Model the relationship between f(X T )=Y T , in this embodiment, f is called the overall transient prediction model.
[0184] According to the second embodiment, the transient prediction model f may be a model selected from all conventional regression models, such as a neural network, a regression tree, a random forest, a wide margin separator, or a linear regression.
[0185] According to a second embodiment, the transient prediction model f is estimated by minimizing a cost function corresponding to the output data f(X) provided by the transient prediction model f. T ) and the expected real output data Y T The error between .
[0186] According to the second embodiment, the cost function is, for example, f(X T ) and Y T As a reminder, the matrix X T can be equal to the column (χ VT ) V≥1 The matrix f(X T ) will be labeled as (f(x T V_ij )) i≥0,j≥0v≥0 . Therefore, the cost function can be equal to:
[0187] Therefore, the transient prediction model f can be defined such that for each flight V of each turbine engine TM, f consists of at least one sub-model f v Formed, the sub-model f v is defined such that f v (x VT )=ψ VT , or it can be defined such that f(X T )=Y T .
[0188] The method 100 further comprises a step 103 of estimating a steady-state prediction model H which is a function of the transient model f and the set S.
[0189] According to a first embodiment, the transient model f is composed of at least one transient sub-model f for each flight V of the transient set T v Formed, for each flight V in the set S, by the matrix pair (ψ VS ,x VS ) is defined as the steady-state prediction model H represented by H(f v (x VS ),x VS )=ψ VS .
[0190] According to a first embodiment, the step 103 of estimating the steady-state prediction model H is performed by minimizing a cost function, which is, for example, equal to: The index V represents each flight of the set S, the index i represents the flight time of the flight V, and the index j represents the value for the coefficient The input parameters of the turbine engine TM, and j represents the coefficient for Output parameters of the same turbine engine TM.
[0191] Minimization of the cost function may be performed by using a gradient descent algorithm or by a least squares algorithm or any optimization method known in the art.
[0192] According to a second embodiment, the transient set T comprises a matrix X T and Y T , and the steady-state set S includes the matrix X S and Y S , and where the transient prediction model f is defined such that f(X T )=Y T , the steady-state prediction model H is defined such that Y S =H(f(X S ),X S ).
[0193] According to a second embodiment, the steady-state prediction model H is estimated by minimizing a cost function, which is, for example, equal to: The index V represents each flight in the set S, the index i represents the flight time of flight V, and the index j represents the coefficient The input parameters of the turbine engine TM, j represents the coefficient for The minimization of the cost function can be performed by using a gradient descent algorithm or by a least squares algorithm or any state-of-the-art optimization method.
[0194] The prediction model H may be a model selected from a neural network, a regression tree, a random forest, a wide margin separator, or a linear regression, but is not limited thereto.
[0195] According to another embodiment, the transient set T comprises a matrix X T and Y T , and the steady-state set S includes the matrix X S and Y S , and where the transient prediction model f is defined such that f(X T )=Y T The step 103 of estimating the steady-state prediction model H comprises, for each flight V of each turbine engine TM, except for the flight of interest VI, estimating the steady-state prediction model H from the transient prediction model f and the steady-state learning data for each recorded instant of the flight V included in the steady-state set S. V steps.
[0196] According to a first embodiment, each steady-state prediction submodel H for each flight V in the steady-state set S is estimated independently of the other steady-state prediction models. V .
[0197] Each steady-state prediction model H V It can be a model selected from all the general regression models, such as neural network, regression tree, random forest, wide margin separator, or linear regression. For example, linear regression can be polynomial regression.
[0198] In the first embodiment, each steady-state prediction model H V The estimation of is performed by minimizing the cost function corresponding to the steady-state prediction model H V The output data provided is H V (f(χ VS ),x VS ) and the expected real output data ψ VS The error between .
[0199] According to an embodiment compatible with the aforementioned first embodiment, for each steady-state prediction model H V The cost function is, for example, H V (f(χ VS ),x VS ) and ψ VS As a reminder, the matrix χ VS The coefficients are denoted by (x T v_ij ) i0,j≥0 , and the matrix ψ VS The coefficients are denoted by (y T V_ij ) i≥0,j≥0 The matrix f v (x VS ) will be labeled as (f(x S V_ij )) i≥0,j≥0 . Therefore, the cost function can be equal to:
[0200] The minimization of the cost function can be performed by using a gradient descent algorithm or by a least squares algorithm or any optimization method of the prior art, allowing the estimation of each submodel H V .
[0201] In a second embodiment, a multi-task learning approach is used to estimate each steady-state prediction model H for each flight V of the transient set S by relying on other steady-state prediction models. V .
[0202] Multi-task learning is a subfield of machine learning that allows solving several different tasks simultaneously, taking into account the dependencies between the tasks. A multi-task learning model includes a part common to each task and a part specific to each task. Multi-task learning makes it possible to improve the learning of a specific model by using features included in all tasks. Thus, according to a second embodiment, for each flight V of each turbine engine TM, each prediction submodel H V equal oh V , function φ is each model H V common, and the function fv is each sub-model f v Specific.
[0203] The multi-task learning model can be a model selected from neural networks, regression trees, random forests, wide margin separators, or linear regression.
[0204] For example, each prediction model H V can be associated with a set of training parameters, some of which are common to all other predictive sub-models and some of which are specific to the model.
[0205] This second embodiment is advantageous because it allows reducing the number of steps used to estimate each steady-state prediction model H in the case where the steady-state data in the set S are noisy or sparse. V The cost function of .
[0206] In this embodiment, the step 103 of estimating the steady-state prediction model H may include: V Construct a steady-state prediction model H for the flight VI of interest VI steps, in this case, model H is model H VI .
[0207] According to one embodiment, the steady-state prediction model H VI At least one selected steady-state prediction model H C Construct,prediction model H C is estimated as flight C. When there are several flights V, flight C is selected from multiple flights V. When there is only one flight V, flight C is flight V. According to condition C1, from the estimated steady-state prediction model H V Select the steady-state prediction model H C .
[0208] Steady-state prediction model H VI The selected prediction model H C Construct such that H VI It can be compared with the steady-state prediction model H C Proportional or equal.
[0209] According to one embodiment, the transient learning data included in set T consists of multiple matrix pairs (χ VT , ψ VT )(including the pair (χ CT , ψ CT ) corresponding to flight C) and is constructed from the matrix pair (χ VIT , ψ VIT ) for selecting the prediction model H C to construct the model H VI , and the condition C1 can be the condition of the distance between the matrices χ VIT and χ CT .
[0210] For example, the distance between the matrices χ VIT and χ CT is less than the threshold S3, and S3 is, for example, a positive real number or zero.
[0211] For example, compared with the corresponding distances between each matrix χ VT and matrix χ VIT for each flight V, the distance between the matrices χ VIT and χ CT is the smallest.
[0212] According to one embodiment, the steady-state learning data included in set S consists of multiple matrix pairs (χ VS , ψ VS )(including the pair (χ CS , ψ CS ) corresponding to flight C) for constructing the steady-state prediction model H C to construct the model H VI , and the condition C1 can be the condition of the cost function that minimizes the deviation between the matrix Hv(f(χ CS ), χ VC ) and the matrix ψ CS for flight C. The condition C1 can be, for example: the selected steady-state prediction model H C is the model with the minimum cost function compared to the other cost functions of each steady-state prediction model H V of each flight V of each turbine engine TM in the steady-state set S.
[0213] According to one embodiment, the steady-state learning data included in set S is constructed from several pairs of matrices, such as the set of the following matrix pairs ((χ VS , ψ VS ), 1 < V ≤ N, where N is an integer strictly greater than 1, and where the steady-state prediction models (HV) 1 < V ≤ N have been estimated in the sub-steps of step 103 of method 100, and the steady-state prediction model HVI can be constructed from a weighted average of k steady-state prediction models out of N steady-state prediction models (H V ), where 1 < V ≤ N, k is an integer between 1 and N, and condition C1 is on the number k, such as k steady-state prediction models among the N steady-state prediction models (H V ) with 1 < V ≤ N that have k of the N smallest cost functions.
[0214] According to one embodiment, the steady-state learning data included in set S is constructed from multiple pairs of matrices (χ VS , ψ VS ), including the pair (χ CS , ψ CS ) corresponding to flight C, for selecting the steady-state prediction model HC to construct model H VI , and condition C1 can be the condition on the generalization error of model H C with respect to the data other than the data related to flight C included in set S.
[0215] Regarding the "generalization error of the model", it means the ability of the model to make robust predictions on new data not used during learning.
[0216] According to the foregoing embodiment, condition C1 can be, for example: the selected steady-state prediction model H C is the model with the smallest generalization error compared to the other corresponding generalization errors of each steady-state prediction model H V for each flight V of each turbine engine TM in the steady-state set S.
[0217] According to one embodiment, the steady-state learning data included in set S is constructed from several pairs of matrices, such as the set of matrix pairs ((χ VS , ψ VS ), where 1 < V ≤ N, N is an integer strictly greater than 1, and where the steady-state prediction models (H V ) with 1 < V ≤ N have been estimated in step 103 of method 100, and the steady-state prediction model H VI can be constructed from a weighted average of k steady-state prediction models out of N steady-state prediction models (H V ), where 1 < V ≤ N, k is an integer between 1 and N, and condition C1 is on the number k, such as k steady-state prediction models among the N steady-state prediction models (H V ) with 1 < V ≤ N that have k of the N smallest generalization errors.
[0218] Method 100 further includes step 104 of estimating matrix Y SI , the matrix Y SIis the value of the output parameter associated with the turbine engine of interest for at least one flight instant of the flight of interest, which is a matrix with the matrix X SC A vector of functions of the model H associated with at least a portion of the transient model f.
[0219] According to a first embodiment, the transient model f is composed of at least one sub-model f for each flight V of the set T v Formed, in particular, by the submodel f VI The transient model f is formed as a submodel f VI , and the matrix Y SI is as the matrix f VI (X SC ) and the matrix X SC The function associated with the model H is estimated so that Y SI =H(f VI (X SC ),X SC ).
[0220] According to a second embodiment, the transient model f is an overall transient prediction model such that f(X T )=Y T , and in particular the submodel f VI , the transient model f is the overall transient prediction model f, and the matrix Y SI is the matrix f(X SC ) and the matrix X SC The function associated with the model H is estimated so that Y SI =H(f(X SC ),X SC ).
[0221] According to a third embodiment, where H is the model H VI , matrix Y SI is defined such that Y SI =H VI (f(X SC ),X SC ).
[0222] The method 100 may further include a margin calculation step 105, wherein the margin is related to the matrix of interest Y SI and the matrix Y mp The difference between them is proportional to the matrix Y mp It means input X SC The output of the worst-case physical model is the worst possible engine before the problem occurs. Therefore, according to one embodiment, the margin calculation is performed as follows: Margin = Y SI -Y mp .
[0223] Depends on the matrix of interest Y SI , an estimation of the engine's state of health is performed through margin calculation.
[0224] Thus, monitoring of the health state of a turbine engine TMI of interest may be performed using margin calculations.
[0225] Another aspect of the invention relates to a computer configured to implement the method according to the first aspect of the invention.
Claims
1. A computer-implemented method (100) for determining a set point matrix X according to the set point matrix X. SC For the aircraft's flight VI of interest to monitor the health status of the aircraft's turbine engine of interest TMI, the setpoint matrix X SC Including at least one value of at least one input parameter related to a turbine engine TMI of interest, the method comprises the following steps: - construct (101) a training database, the training database comprising: o a so-called transient set T comprising transient learning data recorded during at least one flight V of at least one turbine engine TM operating in transient conditions and at least one flight recording instant of a flight of interest of a turbine engine of interest TMI; o a so-called steady-state set S comprising steady-state learning data recorded at each flight recording instant of each flight V other than the flight VI of interest, for each turbine engine TM operating in steady-state conditions; - estimating (102) a transient prediction model f based on transient learning data of a transient set T; - estimating (103) a steady-state prediction model H based on the transient prediction model f and the steady-state learning data included in the steady-state set S; - Based on the steady-state prediction model H, at least a portion of the transient prediction model f and the set point matrix X SC to estimate (104) the matrix of interest Y representing the behavior of the steady-state turbine engine of interest TMI for the flight of interest VI SI ; According to the matrix of interest Y SI The margins are calculated to estimate the state of the turbine engine of interest.
2. The method according to the preceding claim, wherein the turbine engine TM is a turbine engine of interest.
3. A method according to any one of the preceding claims, wherein: For each flight V of each turbine engine TM, the transient learning data set T consists of a pair (χ VT, ψ VT )Build: oMatrix χ VT , comprising at least one row and at least one column, each row corresponding to an input parameter of a turbine engine TM operating in a transient state, each input parameter of each row being different from other input parameters of other rows, each column corresponding to a recording moment of a flight V, each recording moment of each column being different from other recording moments of other columns, oMatrix ψ VT , comprising at least one row and at least one column, each row corresponding to an output parameter of the turbine engine TM operating in a transient state, each output parameter of each row being different from other input parameters of other rows, each column corresponding to a recording moment of the flight V, each recording moment of each column being different from other recording moments of other rows, For each flight V of each turbine engine TM except the flight of interest, the steady-state learning data of the set S consists of a pair (χ VS, ψ VS )Build: oMatrix χ VS , comprising at least one row and at least one column, each row corresponding to an input parameter of a turbine engine TM operating in a transient state, each input parameter of each row being different from other input parameters of other rows, each column corresponding to a recording moment of a flight V, each recording moment of each column being different from other recording moments of other columns, oMatrix ψ VS , comprising at least one row and at least one column, each row corresponds to an output parameter of the turbine engine TM operating in a transient state, each output parameter of each row is different from other input parameters of other rows, each column corresponds to a recording moment of flight V, and each recording moment of each column is different from other recording moments of other columns.
4. The method according to claim 2 and 3, wherein the flight V of the set T is the flight of interest VI, and the set T is further composed of the pair (x VIT , ψ VIT ) build.
5. The method according to any one of claims 2 to 4, wherein for each flight V of each turbine engine TM of the set T: -The transient learning data of set T includes pairs (χ VT ,ψ VT ): - The transient prediction model f consists of at least one sub-model f v Formed so that f v (x VT )=ψ VT .
6. The method according to claim 5, wherein the steady-state prediction model H is defined such that for each flight V of each turbine engine TM of the set S, H(f v (x VS ), χ VS )=ψ VS .
7. The method according to any one of claims 3 to 4: -The transient learning data of set T includes: o is composed of each matrix X VT The matrix X obtained by horizontal concatenation of T , o is composed of each matrix ψ VT The matrix Y is obtained by horizontal concatenation of T , And the transient prediction model f is defined as f(X T )=Y T .
8. The method according to the preceding claim, wherein the set S of steady-state learning data comprises: o is given by each matrix χ VS The matrix X obtained by horizontal concatenation of S , o is composed of each matrix ψ VS The matrix Y is obtained by horizontal concatenation of S , And the steady-state prediction model H is defined such that H(f(X S ), X S )=Y S .
9. The method according to any one of claims 1, 2, 3, 4 or 7, wherein estimating the steady-state prediction model H according to the transient prediction model f and the steady-state learning data included in the steady-state set S comprises the following sub-steps: - for each flight V of each turbine engine TM, except for the flight of interest VI, estimate the steady-state prediction model H from the transient prediction model f and the steady-state learning data for each flight recording instant of the flights V included in the steady-state set S V ; - Based on at least one steady-state prediction model H estimated for flight C C Construct a steady-state prediction model H for the flight VI of interest VI , flight C is one of the flights V or more; Steady-state prediction model H is the steady-state prediction model H VI .
10. The method (100) according to claim 7, claim 9 and one of claims 3 or 4, wherein for each flight V of each turbine engine TM in the steady-state set S, except for the flight of interest VI of the turbine engine of interest TMI: -The steady-state learning data of set S includes (χ VS , ψ VS ) - Steady-state prediction model H V is defined such that Hv(f(χ VS ), χ VS )=ψ VS .
11. The method (100) according to any one of claims 9 or 10, wherein the steady-state prediction model H VI At least the steady-state prediction model H selected according to condition C1 C And build.
12. The method (100) according to claim 4 and claim 11, wherein the condition C1 is: -x CT and χ VIT The distance between them is minimal; - Steady-state prediction model H C The cost function is minimized; or - Steady-state prediction model H C The generalization error is minimal.
13. A method according to any one of the preceding claims, comprising a margin calculation step (105), the margin being related to the matrix of interest Y SI and the matrix Y mp The difference between them is proportional to the matrix Y mp It means that for input X SC The output of the worst-case physical model.
14. A computer program product comprising instructions which, when said program is executed by a computer, cause said computer to implement the method (100) according to any one of the preceding claims.
15. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to implement the method according to any one of claims 1 to 13.
