Method for estimating a state of health of an aircraft engine

Abstract

The invention relates to a method for estimating a state of health of an aircraft engine, comprising a training phase during which an agent is trained by reinforcement to estimate hyperparameters for a Bayesian estimator, and a deployment phase during which the Bayesian estimator estimates a state of health of the aircraft engine using hyperparameters provided by the agent.

Description

[0001] DESCRIPTION

[0002] TITLE: Computer-based method for estimating the health status of an aircraft engine. SCOPE OF THE INVENTION

[0003] The present invention relates to a computer method for estimating the health status of an aircraft engine.

[0004] STATE OF THE ART

[0005] To identify the maintenance tasks to be performed on an aircraft engine, it is important to assess its condition. There are many indicators of an aircraft engine's health, but such indicators are difficult to measure over the engine's lifetime.

[0006] One known method to overcome this problem is to produce an estimate of the engine's health status made up of these inaccessible parameters, using a Bayesian estimator that takes as input measurements of the engine, which are accessible.

[0007] The Bayesian estimator is parameterized with hyperparameters. The accuracy of the estimate produced by the Bayesian estimator depends on the values ​​chosen for these hyperparameters. In some cases, setting the values ​​of these hyperparameters is feasible a priori with some technical knowledge. However, in many other cases, this setting can prove very difficult. In particular, it is difficult to find hyperparameter values ​​that allow for the accurate estimation of sudden changes in an engine's health status.

[0008] DESCRIPTION OF THE INVENTION

[0009] One problem to be solved is to estimate the health status of an aircraft engine in an easy, automatic and dynamic way (i.e. applicable to different engines).

[0010] This problem is solved by a computer-based method for estimating the health status of an aircraft engine, the method comprising:

[0011] • a training phase including:

[0012] to obtain a degradation scenario including the health states of a reference engine and reference measurements on the reference engine at reference times,

[0013] o calculate estimates of the health status of the reference engine at reference times, using a first Bayesian estimator taking as input the reference measurements at reference times; o determine reference degradation levels at reference times from the estimates of the health status of the reference engine at reference times, each reference degradation level representing a difference between health parameters from consecutive estimates of the health status of the reference engine,

[0014] o train an agent, via a reinforcement learning method, to select values ​​for hyperparameters of a second Bayesian estimator, the training being carried out from the degradation scenario and the reference degradation levels;

[0015] • a deployment phase including:

[0016] to obtain previously acquired measurements on the aircraft engine at measurement times,

[0017] to calculate initial estimates of the aircraft engine's health status at measurement times, using the first Bayesian estimator which takes as input the previously acquired measurements,

[0018] to determine aircraft engine degradation levels at measurement times based on initial estimates, each aircraft engine degradation level representing a difference between health parameters derived from consecutive initial estimates of the aircraft engine's health status,

[0019] to calculate second estimates of the aircraft engine health status at measurement times using the second Bayesian estimator taking as input the previously acquired measurements and using as hyperparameters deployment values ​​selected by the agent from the previously acquired measurements and the aircraft engine degradation levels.

[0020] In the proposed method, the reference degradation levels occurring during the training period and the aircraft engine degradation levels estimated by two estimators during the deployment period allow the agent to select hyperparameter values ​​that account for abrupt changes in engine health. Their subsequent use by the second Bayesian estimator thus leads to more accurate estimates of the aircraft engine's health when the engine has undergone these abrupt changes. Furthermore, compared to a supervised learning method, it is not necessary to define a database before testing; the reinforcement learning agent constructs its own experiments by maximizing a reward function.It is particularly noteworthy that some data used for reinforcement learning of the agent is generated progressively iteration by iteration, which represents a saving of time and memory space.

[0021] The proposed process may also include the following optional features, taken alone or in combination whenever technically possible.

[0022] Preferably, an iteration of the reinforcement learning method, associated with a current time point among the reference time points, includes:

[0023] • obtain an agent state at the current time from:

[0024] o a previous estimate of the health status of the reference engine calculated by the second Bayesian estimator during a previous iteration of the reinforcement learning method,

[0025] o the reference measurement at the current time,

[0026] o the reference level of degradation at the current time,

[0027] • Have the agent select values ​​for the hyperparameters based on the agent's current state.

[0028] • calculate a current estimate of the health status of the reference engine using the second Bayesian estimator, taking as input the reference measurement at the current time and using the selected values ​​as hyperparameters,

[0029] • calculate a reward based on the difference between the current estimated health status of the reference engine and the current-time health status of the reference engine in the degradation scenario.

[0030] • Update the agent from the reward.

[0031] Preferably, the deployment phase includes successive deployment iterations, and a deployment iteration associated with a current measurement instant among the measurement instants includes:

[0032] • obtain an agent state at the current measurement time from:

[0033] o a second previous estimate of the aircraft engine's health status calculated by the second Bayesian estimator during a previous deployment iteration,

[0034] o the measurement previously acquired on the aircraft engine at the current measurement time,

[0035] o the level of degradation of the aircraft engine at the current measurement time, • have the agent select deployment values ​​for the hyperparameters from the agent state at the current measurement time, • calculate a second current estimate of the health state of the reference engine using the second Bayesian estimator taking as input the reference measurement at the current measurement time and using the selected deployment values ​​as hyperparameters.

[0036] Preferably, at least one of the first Bayesian estimator and the second Bayesian estimator is a Kalman filter that treats reference measurements and previously acquired measurements on the aircraft engine as observations.

[0037] Preferably, the hyperparameters include at least one of the following: a noise bias occurring in a transition equation of the second Bayesian estimator, and a noise bias occurring in an observation equation of the second Bayesian estimator. Preferably, the first Bayesian estimator uses fixed hyperparameters during both the training and deployment phases.

[0038] Preferably, at least one of the reference degradation levels represents a deviation between selected parameters in consecutive estimates of the health status of the reference engine according to an observability criterion, and / or at least one of the aircraft engine degradation levels represents a deviation between selected parameters in successive initial estimates of the health status of the aircraft engine according to an observability criterion.

[0039] Preferably, at least one of the reference degradation levels is: non-zero when the difference between the parameters from consecutive estimates of the health status of the reference engine is greater than a reference threshold, and zero when the difference between the parameters from consecutive estimates of the health status of the reference engine is not greater than the reference threshold, and / or at least one of the aircraft engine degradation levels is: non-zero when the difference between the parameters from the first consecutive estimates of the health status of the aircraft engine is greater than a threshold, and zero when the difference between the parameters from the first consecutive estimates of the health status of the aircraft engine is not greater than the threshold.

[0040] Preferably, the agent state at the current time includes a measurement gap between: the reference measurement at the current time, and a measurement predicted by the second Bayesian estimator at the current time from the reference measurement at the current time using as hyperparameters previous values ​​calculated during the previous iteration of the reinforcement learning method, and / or the agent state at the current measurement time includes a measurement gap between: the measurement acquired on the aircraft engine at the current time, and a measurement predicted by the second Bayesian estimator at the current measurement time from the measurement acquired on the aircraft engine at the current time using as hyperparameters previous deployment values ​​calculated during the previous deployment iteration.

[0041] Preferably, the learning phase is repeated for several different degradation scenarios.

[0042] Preferably, the aircraft engine health status includes a flow rate or efficiency of a component through which a flow passes when the aircraft engine is in operation, the component preferably being one of: a fan, a compressor, a combustion chamber and a turbine.

[0043] A maintenance procedure for an aircraft engine is also proposed, the procedure comprising: applying the estimation procedure described above to the aircraft engine, determining modifications to be made to the aircraft engine based on second estimates of the engine's health status obtained during the estimation procedure, and making the modifications to the aircraft engine.

[0044] A computer program product is also proposed, comprising program code instructions for executing the steps of the proposed estimation process, when this program is executed by a computer.

[0045] DESCRIPTION OF THE FIGURES

[0046] Other features, purposes and advantages of the invention will become apparent from the following description, which is purely illustrative and not limiting, and which should be read in conjunction with the accompanying drawings on which:

[0047] Figure 1 schematically illustrates components of a system for estimating the health status of an aircraft engine, according to one embodiment.

[0048] Figure 2 illustrates the principle of applying reinforcement learning to an agent.

[0049] Figure 3 is a flowchart of steps of a learning phase that is part of a process for estimating the health status of an aircraft engine, according to one embodiment.

[0050] Figure 4 is a flowchart representing an iteration carried out during a reinforcement learning step of the learning phase of Figure 3, according to one embodiment. Figure 5 is a flowchart of steps of a deployment phase that is part of a process for estimating the health status of an aircraft engine, according to one embodiment.

[0051] Figure 6 is a flowchart representing a deployment iteration carried out during a step of the deployment phase of Figure 5, according to one embodiment.

[0052] Across all figures, similar elements bear identical references.

[0053] DETAILED DESCRIPTION OF THE INVENTION

[0054] 1) System for estimating the health status of an aircraft engine

[0055] With reference to Figure 1, an estimation system, more simply called a system in the following, comprises a first Bayesian estimator A, a second Bayesian estimator B, an RL agent and a control module C.

[0056] These components can take the form of circuits dedicated to specific functions, which will be described below. Alternatively, these components are computer programs comprising code instructions for performing these specific functions when these code instructions are executed by one or more processors. Thus, the system includes at least one memory for storing the programs and at least one processor for executing their code instructions. The processor(s) can be any number of processors.

[0057] These components can be grouped together in a single package. In this case, the components can constitute portions of the same computer program stored in memory. Alternatively, these three components are geographically separated and communicate via a network; the system is then a distributed system.

[0058] As is well known, a Bayesian estimator calculates the probability of a phenomenon based on observations on the one hand and known information (also called priors) on the other. A Bayesian estimator relies on Bayes' theorem and the calculation of so-called posterior probabilities, that is, probabilities that incorporate observations into the prior probabilities. For example, these methods can be used to calculate a probabilistic estimate of a certain unknown parameter based on a prior about its behavior and real-time observed information.

[0059] 1.1) First Bayesian estimator

[0060] The first Bayesian estimator A is configured to estimate the health states of an aircraft engine, denoted

[0061]

[0062] -,x A T , from measurements acquired on this aircraft engine, denoted y^y^, -,yT, from time 1 to time T Here, the hat notations indicate that the health states produced by the first Bayesian estimator A are only estimates of the health state of the aircraft engine.

[0063] The measurements may include, for example, temperatures or pressures measured in the aircraft engine.

[0064] The estimated x A t The engine health state at time t, calculated by the first Bayesian estimator A, comprises a plurality of health parameters. These health parameters may relate to at least one of the following engine components, or all of them:

[0065] • an engine fan (whose function is to generate thrust to move an aircraft including the engine),

[0066] • a low-pressure compressor (sometimes also called a "booster"), arranged downstream of the blower,

[0067] • a high-pressure compressor, arranged downstream of the low-pressure compressor, • the combustion chamber, whose function is to carry out combustion of a mixture of air from the high-pressure compressor and fuel,

[0068] • a high-pressure turbine, arranged downstream of the combustion chamber, to drive the rotation of a rotor of one or the other of the compressors,

[0069] • a low pressure turbine, arranged downstream of the high pressure turbine, to drive the rotation of the blower or a rotor of one of the compressors.

[0070] For a given component of the aircraft engine under consideration, health parameters may include the efficiency of that component and / or the flow rate of that component.

[0071] Thus, in one embodiment, the estimated x A t of the health status at time of measurement t includes 10 health parameters, namely a yield and a flow rate for each of the five components listed above.

[0072] More specifically, the first Bayesian estimator A is configured to calculate the estimated x A t at the measurement time t using a calculation function taking the following data as input: the acquired measurement y t * acquired at the measurement time t, and a previous one estimated x A t -i calculated by the first estimator for the previous measurement time t-1.

[0073] To perform this estimation, the calculation function is parameterized using initial hyperparameters that are fixed.

[0074] In one embodiment, the first Bayesian estimator A is a Kalman filter that treats the measures as observations. The first hyperparameters then include the following hyperparameters: a transition function f A , an observation function h A noise biases ^i A and v A , and noise covariance matrices, Q A and R A These hyperparameters are linked together by the following equations, which are known to those skilled in the art:

[0075] x t = f A (x t _^ +^^ A Q A )

[0076]

[0077] y t = h A (x t} + N(v A , R A )

[0078] The first of these two equations is usually called the "transition equation," and the second is usually called the "observation equation." The notation JV denotes a normal distribution.

[0079] 1.2) Second Bayesian estimator

[0080] The second Bayesian estimator B is configured to estimate the health states of an aircraft engine, denoted x B 1,x B 2,...,x B T , based on the measures y,y2, -,yT discussed previously.

[0081] The measurements may include, for example, temperatures or pressures measured in the aircraft engine.

[0082] The estimated x B t The engine health status at time t, calculated by the second Bayesian estimator B, includes a plurality of health parameters that may be identical to those of the estimated x A t or different.

[0083] In one embodiment, the estimated x B t of the health status at measurement time t includes 10 health parameters, namely a yield and a flow rate for each of the five components listed above for the section dedicated to the first Bayesian estimator A.

[0084] More specifically, the second Bayesian estimator B is configured to calculate the estimated x B t at the measurement time t using a calculation function taking the following data as input: the acquired measurement y t * acquired at the measurement time t, which is used as the observation, and a previous estimated x B t- calculated by the second Bayesian estimator B for the previous measurement instant t-1.

[0085] To perform this estimation, the calculation function is parameterized using second hyperparameters, which are conventionally denoted P B .

[0086] Unlike the first hyperparameters, the second hyperparameters depend at least partially on time. In other words, at least one of the second hyperparameters varies over time. Note that only some of the second hyperparameters can depend on time; this means that at least one of the second hyperparameters can be fixed (but not all).

[0087] In one embodiment, the second Bayesian estimator B is a Kalman filter that treats the measures as observations. The second hyperparameters then include the following hyperparameters: a transition function f B , an observation function h B noise biases

[0088]

[0089] and v B , and noise covariance matrices, Q B and R B .

[0090] In one variant, these second hyperparameters are linked together by the following transition and observation equations:

[0091] x t = / B (x t -i ) +^( / Aç t B )

[0092]

[0093] y t = h B Çx t ) + H(y B , R B )

[0094] It should be noted that in this variant, only the Q parameters B and R B are allowed to change when t varies, while the other hyperparameters remain fixed.

[0095] In another variant, the noise biases are also varied as a complementary measure. The transition and observation equations are then rewritten as follows:

[0096] x t = / B (% t -i ) +^( / if,Ç t B )

[0097]

[0098] y t = h B (x t) + W(y B , R B )

[0099] Varying these noise biases over time allows the second estimator to produce more accurate health status estimates.

[0100] 1.3) Environmental reinforcement learning agent

[0101] The RL agent is designed to undergo reinforcement learning. As a reminder, reinforcement learning is one of the three main machine learning methods, alongside supervised and unsupervised learning. Reinforcement learning involves the RL agent learning actions to perform, based on experience, in order to optimize a quantitative reward over time. To this end, the RL agent is immersed in an environment E and makes decisions based on its current state, denoted S. t In return, the environment provides the agent with a reward R t+1, which can be positive or negative. The RL agent seeks, through iterations, an optimal decision-making behavior (called a strategy or policy, which is a function associating the current state with the action to be executed), in the sense that it maximizes the sum of rewards over time. With reference to Figure 2, the RL agent perceives its state S t at time t, which is a random variable. The agent is configured, during reinforcement learning, to perform action A t from its state S t (the content of which will be discussed later). The agent receives a new state S from the environment t+1 and an R reward t+1 The policy applied by the agent is a function that advocates for any state S t an action A t to be carried out with the aim of maximizing rewards.

[0102] Here, the action that the RL agent learns to perform is a selection of values ​​for the second set of hyperparameters.

[0103]

[0104] intended to parameterize the second Bayesian estimator B.

[0105] The environment includes degradation scenarios which will be described below.

[0106] Finally, the R award t attributed to the agent during his training depends on a difference between an estimate produced by the second Bayesian estimator B and a health state that is part of one of the scenarios.

[0107] 1.4) Control Module

[0108] Returning to Figure 1, the control module C is responsible for controlling the other components. It is capable of launching the first Bayesian estimator A and the second Bayesian estimator B (by providing them with data to process).

[0109] The C command module also has the function of participating in training the RL agent, via a reinforcement learning method.

[0110] 2) Method for estimating the health status of an aircraft engine

[0111] A method for estimating the health status of an aircraft engine, implemented by the system described above, comprises two distinct and successive phases: a training phase and then a deployment phase.

[0112] 2.1) Training Phase

[0113] The training phase consists of training the RL agent to perform the aforementioned action by reinforcement (selecting values ​​for the second hyperparameters of the second Bayesian estimator B).

[0114] With reference to Figure 3, the training phase includes the following steps.

[0115] In step 102, the system obtains a degradation scenario (

[0116]

[0117] (yf,xf), (yi,x£), —, (y, Xr) comprising:

[0118] • health conditions (

[0119]

[0120] xf,%2, -, Xr) of a reference engine at respective reference times ranging from 1 to T, and reference measurements (

[0121]

[0122] y^yj, -'VT) SUR reference engine at the respective reference times.

[0123] The degradation scenario thus represents a potential degradation trajectory over a given lifespan of the reference engine. For example, T = 10000 cycles.

[0124] Preferably, the system generates the reference scenario using an h-engine simulator, which takes the health parameters x as input and returns the measurements y:

[0125] y = h(x)

[0126] In step 104, the first Bayesian estimator A calculates estimates

[0127]

[0128] -,x A Tof the health status of the reference engine, based on the reference measurements y^yj, -,yr- The first hyperparameters P A factors are involved in this calculation, as previously indicated.

[0129] For example, the health parameters included in the reference engine health status include the efficiency and airflow of the blower, both compressors, and both turbines, so for each t =

[0130]

[0131] x A t is a 10-dimensional vector and is denoted by

[0132] X t — — f {X ^> A

[0133]

[0134] t>1 ,..., X.cA t>10 )

[0135] It is possible that an estimated x A tThe estimate calculated by the first Bayesian estimator A may not be entirely correct. However, it is possible to select certain relevant components of these estimates based on an observability criterion. To do this, the control module C evaluates the observability of each of the health parameters in the estimate x A t by applying a method known to a person skilled in the art.

[0136] This observability can be deduced from business knowledge. Some health parameters are known to be observable because sensors allow their values ​​to be calculated or because the estimates are very reliable.

[0137] Alternatively, the control module C calculates an observability index for each health parameter. When the first Bayesian estimator A is a Kalman filter, the control module C calculates an invertibility of the observation process consisting of h at the point of interest x. A tThis invertibility can be measured by the eigenvalues ​​of the observability matrix, namely the Jacobian of h at the point x A t multiplied by the covariance of measurement noise R A .

[0138] In step 106, the control module C selects, for each t = 1, a subset of highly observable parameters x select t constituting a portion of the estimated x A t , and excludes another subset of comparatively less observable parameters.

[0139] If, for example, the subset of highly observable health parameters of the estimated x A t consists of the j-th health parameter of the estimated x A t So we have:

[0140] x ^select _ (vA

[0141]

[0142] t — t,j)

[0143] The C control module can perform the same selection for all t, but not necessarily.

[0144] In step 108, the control module C calculates a degradation level z A of the reference engine at each t = 1,

[0145]

[0146] The level of degradation z A is representative of a discrepancy between health parameters derived from

[0147]

[0148] the estimate and health parameters derived from the estimate x A t The data

[0149]

[0150] and x A t constitute two estimates of the health status of the reference engine, calculated consecutively by the first Bayesian estimator A for times t-1 and t.

[0151] More precisely, the parameters derived are those selected by the C control module in the previous state; the level of degradation therefore depends on the difference x select t - x ^select

[0152] t-1- In a preferred embodiment, the degradation level is a boolean whose value depends on this deviation and a predefined threshold e, chosen according to business rules. We have: z A = 1 if ^select ^select

[0153]

[0154] e and z; 4 = 0 otherwise. For example, it can be chosen e = 0.5% = 0.005. In a learning step 110, the control module C reinforces the RL agent from the degradation scenario (

[0155]

[0156] (yf,xf), (y£,x£), —, (yr, x r) et calculated degradation levels z A .

[0157] The fact that the calculated degradation levels can only have two values ​​(0 or 1) has the advantage of simplifying reinforcement learning. The inventors found that the agent can converge more quickly to a satisfactory state than with degradation levels whose values ​​vary within a continuous range.

[0158] The reinforcement learning method takes place through successive iterations for times t ranging from 0 to T.

[0159] With reference to Figure 4, the iteration of the reinforcement learning method associated with time t comprises the following substeps. In a substep 1102, in accordance with reinforcement learning, the RL agent applies its policy of determining action A t based on state S t from agent RL and an R reward t We remind you that action A tis a selection of new values ​​for the second hyperparameters, with the goal of maximizing rewards; the product of this first substep is therefore the new values ​​of the second hyperparameters. These new values ​​are denoted P t B .

[0160] State S t of the RL agent at time t depends on the data [x B t -i,yt,z^]. We can thus see that the learning process involves using the second Bayesian estimator B to produce the estimated x B t _i of the state of health of the reference engine at time t-1.

[0161] For example, in one embodiment, substep 1102, the control module C can determine the state of agent S t by first launching the second Bayesian estimator B with previous values ​​P B _ ± for the second hyperparameters, so as to produce a so-called "test" estimate xp ti t from dey B The estimated "trial" x p ti t will only serve as intermediate data, just like the estimate given by the first Bayesian estimator A, which is only useful for obtaining z. Then, the control module C calculates:

[0162] 9t = h

[0163]

[0164] B (x p ?-^ - y B

[0165] Here, g B represents a measurement difference between the reference measurement y B and a measure predicted by the second estimator B configured with the previous values ​​of the second hyperparameters P B _ ± (left-hand term). The state of agent S t is then made up of (g B , Zt )- The RL agent takes this S into account t and sort out the action P t B which are the new values ​​of the second hyperparameters.

[0166] We note that the above equality relates to an embodiment in which the observation function h B constitutes a fixed hyperparameter.

[0167] In substep 1104, the control module C launches the second Bayesian estimator B so that it produces an estimated x B t of the health status of the reference engine, using the reference measurement y B as an observation, and this using the new values ​​of the second hyperparameters P B (given by agent RL).

[0168] In substep 1106, the C command module calculates the reward R t+1 for the action performed (consisting of having produced the values ​​of second hyperparameters P) B ) based on a difference between the estimated x B t and health status x B provided in the degradation scenario. For example, we have:

[0169] R t+1 = - || x Bt - x L t ||2where ||·||2 is the L2 standard.

[0170] In substep 1108, the command module C updates the RL agent policy from the R reward t+1 . For this, the C command module can use known prior art methods working in a continuous state and action space, such as DDPG (Deep Deterministic Policy Gradient) or PPO (Proximal Policy Optimization).

[0171] Finally, the module updates t <— t + 1, and repeats the previously described substeps in the form of a subsequent iteration (currently t+1). During the next iteration, the estimated x B t The health state of the reference engine obtained at iteration t is used to form the next agent state S t+1 .

[0172] The preceding steps (starting with step 102) are repeated for a plurality of different degradation scenarios. This repetition enriches the reinforcement learning to which the RL agent is subjected; in other words, at the end of this repetition, the RL agent is configured to apply a so-called "trained" policy, selecting more relevant values ​​for the second set of hyperparameters than if the RL agent were trained on the basis of a single degradation scenario.

[0173] 2.2) Deployment Phase

[0174] The deployment phase uses the now-trained RL agent to parameterize the second Bayesian estimator B, which is then used to perform health status estimates for an aircraft engine. This aircraft engine, more simply called the "target" engine hereafter, is in principle different from the reference engines to which the degradation scenarios used as input data during the training phase refer. Referring to Figure 5, the deployment phase comprises the following steps.

[0175] In step 202, the control module C obtains measurements

[0176]

[0177] previously acquired on the target engine at respective measurement times. The measurements y yl -'VT may have been stored during the life of the target engine by a memory, such as an on-board memory in the aircraft including the target engine, and then subsequently transmitted to the control module C by appropriate communication means.

[0178] Furthermore, the C control module obtains or determines an initial estimate x B *0 of the target engine's health status. This estimate has the letter B as a superscript because it is intended to be processed by the second Bayesian estimator B, as will be seen later. It also has an asterisk * as a superscript to indicate that this estimate relates to the target engine, unlike the estimates calculated during the training period.

[0179] In step 204, the control module C asks the first Bayesian estimator A to calculate estimates X J4*1, X J4 *2,...,x A * T of the health status of the target engine, based on the measurements -,yT*, which are used as observations by the first Bayesian estimator. The first hyperparameters P A are involved in this calculation as previously indicated. This step 204 corresponds to step 104 except that the measurements yl,y2, -'yT* are used instead of the reference measurements y,y2, -,yr of any of the scenarios of the learning phase.

[0180] In step 206, the control module C selects, for each t = 1,

[0181]

[0182] a subset of highly observable parameters x select t constituting a portion of the estimated x A * t and excludes another subset of comparatively less observable parameters. This step 206 corresponds to step 106.

[0183] In step 208, the control module C calculates a degradation level z A * of the target engine at each t = 1,..., T. The degradation level z A * is representative of a discrepancy between health parameters from

[0184]

[0185] of the estimate and health parameters derived from the estimate x A * t The data

[0186]

[0187] and x A * t constitute two estimates of the health status of the target engine, calculated consecutively by the first Bayesian estimator A for times t-1 and t at step 206.

[0188] Step 208 can implement calculations identical to those used in step 108. In particular, we can have: z A * = 1 if x select t - x select t- ) > e and z A* = 0 otherwise. It should be noted that the threshold e already used during step 108 of the training phase can be reused during the deployment phase; alternatively, two different thresholds can be used respectively for the calculations of steps 108 and 208.

[0189] In step 210, the control module C jointly uses the RL agent and the second Bayesian estimator B to produce x estimates B* 1,x B* 2,...,x B* T of the health status of the target engine. Step 210 includes successive iterations for times t from 1 to T, which are conventionally called "deployment iterations" to distinguish them from the iterations of the reinforcement learning method described previously.

[0190] With reference to Figure 6, the deployment iteration associated with measurement time t comprises the following substeps. In a substep 2102, the control module C requests the RL agent to calculate new values ​​for the second hyperparameters, applying its trained policy to an agent state based on [x B * t-1 By convention, these values ​​are called "deployment values" and are denoted P t B * in order to distinguish them from P values B calculated during the reinforcement learning method.

[0191] In one embodiment, substep 2102, the agent state processed by the RL agent has the same format as the agent state used in substep 1102 of the reinforcement learning method. Thus, when the term g Bis part of the agent state during reinforcement learning; in the agent state used in the deployment iteration associated with measurement time t, we find a term g B * constructed in a similar way, as follows:

[0192] = fi" - y t '

[0193]

[0194] The RL agent takes into account the RL S state t = (g Bt ,z^) et performs the action of calculating P B * (the new deployment values ​​of the second hyperparameters).

[0195] In substep 2104, the control module C requests the second Bayesian estimator B to calculate an estimated x B * t of the health status of the target engine, using the y measure as observations, and using the deployment values ​​P B * as second hyperparameters.

[0196] The deployment iteration associated with time t is then completed, and we move on to the next deployment iteration (t+1).

[0197] During this step 212, the calculations implemented by the second Bayesian estimator B are done from deployment values ​​for the second hyperparameters which progress over time, taking into account the past.

[0198] The estimated x B* 1,x B* 2,...,x B* T the health status of the target engine constitute the output data of the target engine health status estimation process.

[0199] 3) Aircraft engine maintenance procedure

[0200] A previously described target engine maintenance process includes exploiting the estimated x B* 1,x B* 2,...,x B* T of the health status of the target engine.

[0201] More specifically, these estimated x B* 1,xB* 2,...,x B* T can be used to make decisions related to the target engine, such as sending the target engine for maintenance and / or determining the work required to maintain or repair the target engine (workscoping). This includes assessing the current condition of the engine, determining the specific inspections and repairs to be carried out, and planning the resources and time required to complete these tasks.

[0202] 4) Other embodiments

[0203] Embodiments in which the first Bayesian estimator A and the second Bayesian estimator B are Kalman filters have been detailed above. These filters have the advantage of being resource-efficient, as they are essentially based on linear computations. However, the estimation method is applicable to other Bayesian estimators. For example, at least one of the first Bayesian estimator A and the second Bayesian estimator B can be a particle filter (implementing a sequential Monte Carlo method). The particle filter hyperparameters that the agent selects can then include a specific number of particles to be generated, or particle resampling parameters. Alternatively, Bayesian estimators implementing a Markov chain Monte Carlo method, or an Expectation Maximization method, etc., can be used.

[0204] Although advantageous, the steps of selecting highly observable parameters during the learning phase and the deployment phase remain optional.

Claims

DEMANDS 1. A method for computer-based estimation of the health status of an aircraft engine, the method comprising: • a training phase including: • obtain a degradation scenario including health states (x,...,x) of a reference engine and reference measurements (y,...,yr) on the reference engine at reference times, • calculate estimates of the health status of the reference engine (x,...,x A T ) at reference times, using a first Bayesian estimator taking as input the reference measurements (y,...,yr) at reference times, • determine reference degradation levels (z A ,...,z A) at reference times based on estimates of the health status of the reference engine at reference times, each of the reference degradation levels representing a difference between health parameters derived from consecutive estimates of the health status of the reference engine, • train an agent, via a reinforcement learning method, to select values ​​for hyperparameters of a second Bayesian estimator, the training being carried out from the degradation scenario and reference degradation levels; • a deployment phase including: • obtain measurements (y^,...,yr) previously acquired on the aircraft engine at measurement times, • calculate initial estimates of the aircraft engine's health status (x A ,..., X A * T) at the measurement times, using the first Bayesian estimator taking as input the previously acquired measurements (y^,...,yr), • determine aircraft engine degradation levels (z A *,...,z A *) at the measurement times from the first estimates, each of the aircraft engine degradation levels representing a difference between health parameters from consecutive first estimates of the aircraft engine's health status, • calculate estimated seconds (x B ,...,x B * T ) of the aircraft engine health status at measurement times using the second Bayesian estimator taking as input the previously acquired measurements (yl,...,?•) and using deployment values ​​(P) as hyperparameters B *) selected by the agent from previously acquired measurements and aircraft engine degradation levels.

2. A method according to the preceding claim, wherein an iteration of the reinforcement learning method, associated with a current time among the reference times, comprises: • obtain an agent state at the current time (t) from: • a previous estimate of the health status of the reference engine calculated by the second Bayesian estimator (B) during a previous iteration of the reinforcement learning method, • the reference measurement (y B ) at the current moment, • the reference degradation level (z) at the current time, • have the agent select values ​​(P t B ) for hyperparameters based on the agent state at the current time, • calculate a current estimate (x B t) of the health status of the reference engine using the second Bayesian estimator (B) taking as input the reference measure (y B ) at the current time and using the values ​​(P t B ) selected as hyperparameters, • calculate a reward based on the difference between the current estimate (x B t ) of the health status of the reference engine and the health status (x B ) of the reference engine of the degradation scenario at the current time, • Update the agent from the reward.

3. A method according to the preceding claim, in which • the agent's state at the current time includes a measurement gap ( ) between: • the reference measurement (y B ) at the current moment, and • a measure predicted by the second Bayesian estimator at the current time from the reference measure (y B) at the current time using as hyperparameters previous values ​​(Pt-i) calculated during the previous iteration of the reinforcement learning method.

4. A method according to any one of the preceding claims, wherein the deployment phase comprises successive deployment iterations, and wherein a deployment iteration associated with a measurement instant currently among the measurement instants comprises: • obtain an agent state at the current measurement time from: • a previous second estimate of the aircraft engine health state calculated by the second Bayesian estimator during a previous deployment iteration, • the measurement (y) previously acquired on the aircraft engine at the current measurement time, • the aircraft engine degradation level (z^*) at the current measurement time, • have the agent select deployment values ​​(P t B*) for hyperparameters from the agent state at the current measurement time, • calculate a second current estimate (x B t ) of the health status of the reference engine using the second Bayesian estimator (B) taking as input the reference measurement (y) at the current measurement time and using the deployment values ​​(P t B *) selected as hyperparameters.

5. A method according to the preceding claim, in which • The agent state at the current measurement time includes a measurement difference between: • the measurement (y) acquired on the aircraft engine at the current time, and • a measure predicted by the second Bayesian estimator at the current measurement time from the measurement (y) acquired on the aircraft engine at the current time using previous deployment values ​​(P) as hyperparameters B_*i) calculated during the previous deployment iteration.

6. A method according to any one of the preceding claims, wherein the first Bayesian estimator is a Kalman filter that treats the reference measures (y,...,yr) as observations.

7. A method according to any one of the preceding claims, wherein the second Bayesian estimator is a Kalman filter that treats the measurements (y^,...,yr) previously acquired on the aircraft engine as observations.

8. A method according to any one of the preceding claims, wherein the hyperparameters comprise at least one of: • a noise bias occurring in a transition equation of the second Bayesian estimator, and • a noise bias occurring in an observation equation of the second Bayesian estimator.

9. Method according to any one of the preceding claims, wherein the first Bayesian estimator uses fixed hyperparameters during the training phase and during the deployment phase.

10. A method according to any one of the preceding claims, wherein • at least one of the reference degradation levels (z) represents a deviation between selected parameters in consecutive estimates of the health status of the reference engine according to an observability criterion, and / or • at least one of the aircraft engine degradation levels (z^*) represents a deviation between selected parameters in successive initial estimates of the aircraft engine health status according to an observability criterion.

11. A method according to any one of the preceding claims, wherein • at least one of the reference degradation levels (z) is: • non-zero when the difference between the parameters from consecutive estimates of the health status of the reference engine is greater than a reference threshold, and • zero when the difference between the parameters from consecutive estimates of the health status of the reference engine is not greater than the reference threshold, and / or • at least one of the aircraft engine degradation levels (z*) is: • not zero when the difference between the parameters from the first consecutive estimates of the aircraft engine's health status is greater than a threshold, and • zero when the difference between the parameters from the first consecutive estimates of the aircraft engine's health status is not greater than the threshold.

12. A method according to any one of the preceding claims, wherein the learning phase is repeated for several different degradation scenarios.

13. A method according to any one of the preceding claims, wherein the health status of the aircraft engine includes a flow rate or efficiency of a component through which a flow passes when the aircraft engine is in operation, the component preferably being one of: a fan, a compressor, a combustion chamber and a turbine.

14. A method for maintaining an aircraft engine, the method comprising: • applying the estimation method according to any one of the preceding claims to the aircraft engine, • determine modifications to be made to the aircraft engine based on the second estimates of the aircraft engine's health status obtained during the estimation process, • perform the modifications on the aircraft engine.

15. Product computer program comprising program code instructions for the execution of the steps of the process according to any one of claims 1 to 13, when this program is executed by a computer.

Images