Method of monitoring the health of an aircraft engine

Abstract

One aspect of the invention relates to a method (1) for monitoring a health x of an aircraft engine comprising training (40) a neural network, the neural network being configured to provide as output a prediction of parameters of the health x of the aircraft engine and taking as input the measurement data Y of the aircraft engine, the training being further carried out from a training data set comprising the noisy measurement data Y and generated from a simulation of the aircraft engine. Figure to be published with the abstract: Figure 1

Description

Title of the invention: Method for monitoring the health of an aircraft engine TECHNICAL FIELD OF THE INVENTION

[0001] The technical field of the invention is that of aeronautics.

[0002] The present invention relates to a method for monitoring the health of an aircraft engine and in particular such monitoring using a neural network. TECHNOLOGICAL BACKGROUND OF THE INVENTION

[0003] Methods for monitoring the health of an aircraft engine in aeronautics use measurement data from sensors and / or analysis data. This data may be data from an aircraft or data from an aircraft engine. These monitoring methods then use this data to generate indicators and alerts on the health and usage status of the engine. These indicators and alerts can be sent to an operator, in order to better manage the use of the aircraft engine. For example, it is possible to anticipate the future maintenance needs of an aircraft or an aircraft system. Such monitoring makes it possible in particular to optimize maintenance programs and / or reduce unplanned downtime and / or improve the availability and reliability of aircraft.

[0004] The aircraft engine is for example a turbomachine and the method for monitoring the health of the turbomachine comprises the monitoring of modular parameters such as the efficiencies and flow rates of different compressors and / or turbines of the turbomachine. Several conventional methods for monitoring the health of the turbomachine exist such as thermodynamic analysis, "gas path analysis" in English, Bayesian filtering on variables of interest or even dimension reduction and nearest neighbor methods. Neural networks are also used to monitor the health of an aircraft engine. Conventionally, a method for monitoring the health of an aircraft engine based on a neural network can comprise three steps: - Generation of a training data set consisting of values ​​of the efficiencies and airflows of the aircraft engine modules and measurements from a set of sensors simulated by a thermodynamic simulator at a single predetermined operating point. As a reminder, an operating point, also called an operational point, corresponds to an operation of the aircraft engine for predetermined inputs supplied to the aircraft engine, - Training the neural network using the training dataset, and - use the neural network trained in real conditions to find the efficiencies and air flow rates of the engine modules from the observed measurements.

[0005] Engine health monitoring methods using neural networks have several advantages over other prior art methods. For example, engine health monitoring methods using neural networks allow for faster execution in real conditions than thermodynamic analysis since they do not require optimization during the processing of the real data. In addition, engine health monitoring methods using neural networks do not rely on strong assumptions of statistical laws unlike Bayesian filtering methods on variables of interest.Finally, engine health monitoring methods using neural networks do not require discretization of space, which introduces additional limitations in accuracy, as is the case with methods such as dimension reduction and nearest neighbor methods, which require discretization of space.

[0006] When an aircraft engine, having given health parameters, is observed via sensors, it is possible that the system of equations modeling the relationship between the health of the aircraft engine and the measurement data for a given observation, does not allow to differentiate between several sets of health parameters of the aircraft engine. In other words, the same observation by sensors can lead to several solutions for the health parameters of the aircraft engine. This scenario is known as the underdetermined case, and appears when few sensors are available compared to the numbers of health parameters to be determined.

[0007] None of the conventional methods for monitoring the health of the turbomachine can solve this problem satisfactorily. For example, conventional neural network methods fail to provide acceptable performance in underdetermined cases.

[0008] One solution to this problem is to increase the number of sensors used. However, this solution has several drawbacks. First, adding measurement sensors presents a significant financial cost and adds additional constraints on the system such as additional weight to the system and / or additional maintenance tasks to manage these sensors.

[0009] There is therefore a need to provide a method for monitoring the health of an aircraft engine which limits, at least partially, the problems associated with the methods of the prior art. Summary of the invention

[0010] The invention provides a solution to the problems mentioned above, by proposing a method for monitoring the health of an aircraft engine based on a neural network combined with the use of a multi-point perspective, which consists of imposing one or more appropriate hypotheses of link between the health of the aircraft engine and the operating points. In addition, the generation of the training data set of the neural network includes the addition of noise to the measurement data used in order to improve the prediction accuracy of the health state of the engine in real conditions.

[0011] One aspect of the invention relates to a method for monitoring a health x of an aircraft engine comprising training a neural network, the neural network being configured to provide as output a prediction of parameters of the health x of the aircraft engine and taking as input measurement data Y of the aircraft engine, the training being further carried out from: - a training dataset comprising noisy Y measurement data generated from an aircraft engine simulation performed: • according to at least one business hypothesis relating to knowledge of the aircraft engine, and • according to one or more operating points u of the set of operating points U of the aircraft engine and the health parameters x of the aircraft engine, - a function penalizing a deviation between the aircraft engine health parameters x corresponding to the measurement data Y and the prediction of the health parameters x obtained by providing the trained neural network with the generated and noisy measurement data Y.

[0012] Thanks to the invention, the method for monitoring the health of an aircraft engine has a high speed of execution in real conditions, that is to say during the inference phase of the neural network. In addition, the health of the aircraft engine is estimated accurately even in underdetermined cases.

[0013] In addition to the characteristics which have just been mentioned in the preceding paragraph, the method according to one aspect of the invention may have one or more complementary characteristics among the following, considered individually or according to all technically possible combinations: - The method is implemented by a monitoring device comprising a set of sensors adapted to carry out measurements of an operation of the aircraft engine and being configured to comprise a simulator s of the aircraft engine, the simulator s taking as input health parameters x of the aircraft engine and an operating point u of the aircraft engine, and providing corresponding generated measurement data Y as output, The simulation of the aircraft engine allowing the generation of the training data set is carried out with the simulator s of the aircraft engine and the training data set is obtained by further performing, after the acquisition step and before the training step, additional steps of: • Obtaining at least one business hypothesis, the at least one business hypothesis being a hypothesis of the link between the health x of the aircraft engine and parameter values ​​of all operating points, • Generation of the training data set, each training data including generated and noisy measurement data Y as well as parameters of the health x of the aircraft engine and an operating point u of the aircraft engine, the generation being carried out respecting the link hypothesis and includes for each training data: • Obtaining the measurement data Y generated by providing the aircraft engine simulator s with the health parameters x of the aircraft engine and an operating point u of the aircraft engine, the operating point u of the aircraft engine being included in the set of operating points U, • Adding noise to the generated measurement data Y, training the neural network is further performed using the training data set, the neural network further taking as input the operating point u of the aircraft engine, the training function penalizing a deviation, for each training data, between the parameters of the health x of the aircraft engine of the training data and the parameters of the health x obtained by providing the neural network with the generated measurement data Y and the operating point of the aircraft engine of the training data, and Aircraft Engine Health Diagnosis using Trained Neural Network, the business hypothesis includes at least one hypothesis from among a multi-point hypothesis and a temporal hypothesis, with: • the multi-point hypothesis consisting of considering that the health x of the aircraft engine has an influence on the values ​​of the parameters of the set of operating points and that the relationship between the health x of the aircraft engine and the value of the operating points is predetermined, and • the temporal hypothesis consisting of considering that the health x of the aircraft engine evolves over time and that the evolution over time of health x is predetermined. the neural network is of type: • A multi-layer perceptron, or • A recurrent neural network, or • A convolutional neural network, or • A transformer. adding noise to the generated measurement data Y comprises, for each measurement data, adding a sample from a noise distribution having a predetermined mean and variances, the method further comprises a step of validating the trained neural network, carried out between the training of the neural network and the diagnosis of the health x of the aircraft engine, and the validation step comprising: • Generation of a validation data set, each validation data including Y measurement data and x health parameters of the aircraft engine and the operating point of the aircraft engine, the generation being carried out in compliance with the link hypothesis and includes for each validation data: • Obtaining Y measurement data by providing the aircraft engine simulator § with the aircraft engine health x parameters and the aircraft engine operating point, • Addition of noise to the measurement data Y including, for each measurement data y, the addition of a sample respecting a Gaussian noise distribution having a mean equal to zero and a variance equal to a predetermined variance matrix of the noises of all the sensors. • Validation of the trained neural network using the validation dataset and verifying that the deviation, for each validation data, between the aircraft engine health x parameters of the training data and the obtained health x parameters is less than a predetermined threshold.

[0014] Another aspect of the invention relates to a system comprising a set of sensors adapted to carry out measurements on an aircraft engine and a computer configured to implement the method according to the invention.

[0015] A further aspect of the invention relates to an aircraft comprising the system according to the invention.

[0016] The invention also relates to a computer program product comprising instructions which, when the program is executed on a computer, cause the latter to implement the steps of the method according to the invention.

[0017] Finally, the invention relates to a computer-readable recording medium comprising instructions which, when executed by a computer, cause the latter to implement the steps of the method according to the invention.

[0018] The invention and its various applications will be better understood by reading the following description and examining the accompanying figures. BRIEF DESCRIPTION OF THE FIGURES

[0019] The figures are presented for information purposes only and in no way limit the invention. - [Fig.l] shows a block diagram illustrating the steps of an example of the method according to the invention. DETAILED DESCRIPTION

[0020] Unless otherwise specified, the same element appearing in different figures has a single reference.

[0021] [Fig. 1] is a block diagram illustrating the steps of an example of method 1 according to the invention. The mandatory steps of the example of method 1 are indicated by a solid rectangle and the optional steps are indicated by a dotted rectangle.

[0022] The method 1 is implemented by a monitoring device comprising a set of sensors adapted to carry out measurements of an operation of the aircraft engine. In addition, the monitoring device is configured to comprise a simulator § of the aircraft engine. For example, the device comprises a computer or by a processor configured to provide a simulator s of the aircraft engine. Thus, steps of the method 1 can be implemented by a processor included in a system for monitoring the health of an aircraft engine. The monitoring system preferably has the structure of a computer, in this case an on-board computer, and / or a computer. It comprises an electronic circuit, in one or more parts, equipped with at least one non-volatile type memory, and a processor for executing logic operations.It may also include one or more other memories, such as RAM or another type, and one or more other processors.

[0023] By "computer-implemented" is meant that the steps, or substantially all of the steps of the method 1, are executed by at least one computer or processor or any other similar system. Thus, steps are performed by the computer, possibly fully automatically, or semi-automatically. In examples, the triggering of at least some of the steps of the method may be performed by user-computer interaction. The level of user-computer interaction required may depend on the intended level of automation and balanced against the need to implement the user's wishes. In examples, this level may be user-defined and / or predefined.

[0024] A typical example of a computer implementation of a method is to execute the method with a system adapted for this purpose. The system may comprise a processor coupled to a memory and a graphical user interface (GUI), a computer program comprising instructions for implementing the method being stored in the memory. The memory may also store a database. Memory is any hardware adapted for such storage, possibly comprising several distinct physical parts.

[0025] Method 1 comprises simulations of the aircraft engine in order to generate a training dataset. The training dataset is used to train a neural network. The simulations can be carried out with a simulator $ of the aircraft engine. In one example, the aircraft engine is a turbomachine and the method for monitoring the health of the turbomachine comprises the monitoring of modular parameters such as the efficiencies and flow rates of different compressors and / or turbines of the turbomachine. The simulator s can for example be previously stored on the computer implementing method 1. The aircraft engine can be modeled by a system S as described below:

[0026] y = S(x,u)

[0027] With: - y GP' - the observation of the system, i.e. the measurement data y from at least one sensor, - xe PJ- the health of the system, so for example the modular efficiencies and flow rates of the aircraft engine, or for example the differences to a reference of modular efficiencies and flow rates of the aircraft engine, - HE P*”, an operating point.

[0028] The simulator s of the aircraft engine is therefore governed by the following equation:

[0029] yS(xu)

[0030] The aircraft engine simulator § therefore takes as input the health x of the aircraft engine, or parameters of the health x of the aircraft engine, and a point of operation u of the aircraft engine. The aircraft engine simulator § thus provides corresponding measurement data y as output, i.e. measurement data y corresponding to the health x of the aircraft engine and the operating point u of the aircraft engine provided as input. For example, a simulator S compatible with the invention is the simulator s described in French patent application FR2202472. It is also possible to use other types of simulators § such as those distributed: - in the TURBO® libraries distributed by EcosimPro® and in the commercial software PROOSIS®, or - in the GasTurb® software, or - in the commercial software NPSS®, or - by Elettronicavenata Veneta®, for example the STG / EV model which is a turbine simulator.

[0031] A first advantageous step 10 of the method 1 comprises the acquisition of measurement data Y of an operation of the aircraft engine. The measurement data Y comprises parameters of the health x of the aircraft engine and a set of operating points U. This set of measurement data Y may originate from the measurement of an operation of the aircraft engine. This set of measurement data Y may be provided by at least one sensor, having carried out the measurements during the operation of the aircraft engine. The set of measurement data Y may originate from the measurement of several days or several weeks of use of the aircraft engine. The set of measurement data may also originate from a single complete flight of the aircraft. The set of operating points U acquired may for example have been determined using the method described in French patent application FR2202472.These measurement data y can be used to configure the simulator s, i.e., train the simulator s to accurately provide aircraft engine measurement data y corresponding to the aircraft engine health x and the aircraft engine operating point u provided as input.

[0032] A second advantageous step 20 of the method 1 comprises obtaining a business hypothesis relating to knowledge of the aircraft engine. The business hypothesis may be a hypothesis of a link between the health x of the aircraft engine and parameter values ​​of all the operating points.

[0033] In one example, the business hypothesis used in method 1 is the multipoint hypothesis. The multipoint hypothesis consists of considering that the health x of the aircraft engine has an influence on the values ​​of the parameters of all the operating points and that the link between the health x of the aircraft engine and the value of the operating points is predetermined. For example, the multipoint hypothesis consists of considering that when an observed degradation of the values ​​of the parameters is 1% on a first operating point (for example during the cruise phase during a flight), then we can deduce a degradation of these same parameters of 0.5% on a second operating point (for example during the takeoff phase of the same flight). Thus, with this multipoint hypothesis, it is for example possible to consider that when a degradation of the values ​​of the aircraft engine parameters due to engine wear occurs during a flight, the efficiency curves in relation to the corrected airflow in the performance maps are displaced linearly along a fixed direction, and the amounts of displacements are the same for all operating points included in the set of operating points U.

[0034] Alternatively, the business hypothesis used in method 1 is the temporal hypothesis. The temporal hypothesis consists of considering that the health x of the aircraft engine evolves over time and that the evolution over time of the health x is predetermined or at least partially known. For example, the temporal hypothesis consists of considering that when a degradation of the health x of the engine of 1% is observed at a first operating time, then the degradation of the health x of the engine is 1.1% at a second operating time, the evolution of the state of health of the aircraft engine can for example be modeled by a Gaussian function, which will model the uncertainty due to the partial aspect of the knowledge of this relationship.Thus, with this temporal hypothesis, it is for example possible to consider that given an operating pointM, we assume the existence of a function f which governs the changes in values ​​of the health parameters of an engine during a first flight to its new values ​​when the same engine is operated during a second flight, the second flight taking place after the first flight. An equation describing this example, with the function f in a form of a linear function, can for example be in the form of: .

[0035] x2 = / (Xj) = x, - 4*v,

[0036] With: - xi, the value of health parameters during the first flight, - x2 the value of health parameters during the second flight, - , a direction of degradation of the values ​​of the engine parameters aircraft, - At > 0 is a scalar reflecting the rate of degradation of aircraft engine parameter values.

[0037] Alternatively, the business hypothesis used is the combination of the multipoint hypothesis and the temporal hypothesis. For example, for a set of operating points U containing several operating points from two different flight missions of an aircraft engine, it is possible to impose the multipoint hypothesis for the changes in parameter values ​​at the different operating points of the same flight, and to impose the temporal hypothesis for the changes in parameter values ​​at the different operating points not belonging to the same flight.

[0038] These business hypotheses therefore make it possible to add information to the problem to be solved, i.e. to carry out the diagnosis of the health x of the aircraft engine, thus they make it possible to solve at least in part the problem of indeterminacy. It should be noted that this addition of information results in an increase in the dimension of the problem considered. It is possible to consider that the addition of the hypotheses to method 1 fulfills its objective if the accuracy of the method is better with these hypotheses than without its use, i.e. as is done in the state of the art: using a single operating point. The accuracy of method 1 can be quantified by the error measured between the real value to be estimated and the value estimated by method 1.This precision of method 1 can for example be calculated with the mean square error between the real value to be estimated and the value estimated by method 1 or the mean absolute error between the real value to be estimated and the value estimated by method 1 or any other metric weighted by sample or by dimension.

[0039] A third advantageous step 30 of the method 1 comprises generating a training data set for an artificial neural network. Different types of neural network can be used in the method 1. For example, it is possible to use a multilayer perceptron or a recurrent neural network or a transformer or a convolutional neural network. A multilayer perceptron is a type of neural network organized into several layers. A multilayer perceptron has at least three layers: an input layer, at least one hidden layer, and an output layer. Each layer consists of a potentially different number of neurons. Information flows from the input layer to the output layer only. The neurons of the last layer are the outputs of the overall system.A recurrent neural network is a network of neurons with recurrent connections. A recurrent neural network consists of interconnected neurons that interact nonlinearly and for which there is at least one cycle in the structure. Neurons are connected by edges that have a weight. The output of a neuron is a nonlinear combination of its inputs. A transformer, or self-attentive model, is a deep learning architecture designed to handle sequential data. The transformer uses an encoder-decoder architecture. The encoder consists of a set of encoding layers that process the input sequence iteratively layer by layer and the decoder consists of a . A set of decoding layers that performs the same thing on the encoder output sequence. A convolutional neural network is a type of acyclic artificial neural network, in which the connection pattern between neurons is inspired by the visual cortex of animals. A convolutional neural network consists of two types of artificial neurons, arranged in "layers" that successively process information: - processing neurons, which process a limited portion of the image through a convolution function. - the output pooling neurons called total or partial pooling.

[0040] Each training data of the training data set may comprise generated and noisy measurement data Y as well as parameters of the health x of the aircraft engine and an operating point u of the aircraft engine, the generation being carried out in compliance with the business hypothesis(es) obtained in step 20.

[0041] The generation 30 of the training data set may comprise, for each training data, two sub-steps. The first sub-step is obtaining the measurement data Y generated by providing the aircraft engine simulator $ with the health parameters x of the aircraft engine and an operating point u of the aircraft engine.

[0042] The second sub-step is adding noise to the generated measurement data Y. In an example compatible with the previous examples, adding noise to the generated measurement data Y comprises, for each generated measurement data y, adding a sample from a noise distribution having a predetermined mean and variances.

[0043] For example, step 30 can be implemented by following this protocol when the multipoint hypothesis is used: - (a) determine a range, denoted P, of possible values ​​in which engine health parameters can vary. For example, each deformation coefficient can vary in the range [0.95, 1.00]ct p — [ 0.95. 1.00 where is the number of health parameters to be estimated. - (b) determine the number, noted Ntraia, of samples included in the set of training data, - (c) generate N train vectors included in the range P, for example, by following a Latin hypercube sampling scheme. The set of these vectors is denoted Xtrain and represents the values ​​of the engine health parameters. The set of these Xtrajn vectors will act as labeled outputs of this dataset. - (d) Generate the measurements corresponding to the values ​​of the health parameters of motor in Xtnjin and the operating points included in the set of operating points U. The generation of the measurements is carried out by the simulator s while respecting the business hypothesis(es) obtained in step 20. In particular, it is possible to calculate: V=(j r ($(*, «O,

[0044] With: • xe Xtraill, a vector representing a value of the engine health parameters, • ..., uv e U, the operating points of the set of points of U operation, * • • • » yv e Yraw the generated measurements included in the set of generated measures Yrm>. • Enrich the set of generated measurements Yrmv by adding a sample respecting a noise distribution, representing for example the variance of artificial noises of sensor measurements, for example a Gaussian noise distribution. In particular, for any jp e Yraw, it is possible to calculate: ÿ=y+ (?r rv)

[0045] With: • ÿ, the generated and noisy measurements included in the set of generated and noisy measurements Y treated * ~ Pi v“' 1 corresponding with each operating point vtrain' train] in which * ^raîn' 'c vector representing the bias, otherwise called normal law of expectation, of the Gaussian function N • ep^', |a variance of artificial noises of sensor measurements.

[0046] It can be noted that when the temporal hypothesis is used, the step (c) of the protocol consists of generating N train vectors in the range P, for example, following a Latin hypercube sampling scheme. The set of these vectors is denoted ^i. Then, for each vector Xj e X} a degradation vector ye P^ is sampled from the Gaussian distribution N(v, tj) with ye P^, the degradation direction in the temporal hypothesis and ep^ is a predetermined variance matrix. Finally, the calculation x2 = x}-À*v is performed in order to calculate with 2 a deterministic value dependent on the time spent between the two instants considered, for example, 0.01 for 100 flights. The Xtrain set of these generated vector pairs will act as the labeled outputs of this dataset.

[0047] Thus, in this implementation mode, a training data set (Ytrain) is obtained at the end of step 30 of generating the training dataset. Here, Ytrain serves as input in the training, and Xtrain serves as labeled output in the training. In addition, the neural network used in this implementation mode will have a first layer having a size corresponding to the size of Ytrain and a last layer having a size corresponding to the size and Xtrain. Advantageously, the addition of noise during this generation of training data makes the network more robust during its training, and counterbalances the increase in the dimension to be noted in paragraph

[0038] .

[0048] A fourth step 40 of the method 1 comprises training the neural network using the training data set. Thus, the neural network is configured to take as input the generated and noisy measurement data Y obtained in step 30 and an operating point u of the aircraft engine. As output, the neural network is configured to provide as output a prediction of the health parameters x of the aircraft engine. The training of the neural network is performed from a function penalizing a deviation, for each training data item, between the health parameters x of the aircraft engine of the training data item and the health parameters x obtained by providing the neural network with the generated measurement data Y and an operating point of the aircraft engine. The training of the neural network can for example be performed using an optimization method such as stochastic gradient descent.

[0049] A fifth advantageous step 50 of the method 1 comprises the validation of the trained neural network. This validation step 50 can be carried out after the training 40 of the neural network and before the diagnosis 60 of the health x of the aircraft engine, and the validation step (50) comprises two sub-steps.

[0050] A first sub-step of the fifth advantageous step 50 of the method 1 is the generation of a validation data set. Each validation data comprising measurement data Y and parameters of the health x of the aircraft engine and the operating point of the aircraft engine. The generation of the validation data set can be carried out in a similar manner to the generation 30 of the training data set except that for the addition of the noises to the generated measurements, noises corresponding to the noises of the sensors that can be used are used, these noises of the sensors can be different from the noises used to generate the training data. Preferably, the validation data set does not include any data common to the training data set.

[0051] For example, the generation of the validation data set may comprise, for each validation data, the sub-steps of: - obtaining the measurement data Y by providing the aircraft engine simulator $ with the aircraft engine health parameters x and the aircraft engine operating point, - adding noise to the measurement data Y comprising, for each measurement data y, the addition of a sample respecting a Gaussian noise distribution having a mean equal to zero and a variance equal to a predetermined variance matrix of the noises of all the sensors, for example, a variance matrix obtained from the analysis of real measurements during flights or on a test bench.

[0052] As an alternative to the previous example, this first sub-step of generating the validation data set may comprise obtaining data from measurements of an operation of an aircraft engine in real conditions. Thus, in this alternative, the validation data set therefore comprises these data from measurements of an operation of an aircraft engine in real conditions.

[0053] This first sub-step of generating the validation data set can be carried out before the training 40 of the neural network, for example at the same time as the generation of the training data set.

[0054] A second sub-step is the validation of the trained neural network using the validation data set and verifying that the deviation, for each validation data, between the parameters of the health x of the aircraft engine of the training data and the parameters of the health x obtained by the prediction of the neural network is less than a predetermined threshold.

[0055] A sixth step 60 of the method 1 comprises diagnosing the health x of the aircraft engine using the trained neural network.

[0056] A seventh advantageous step 70 of the method 1 comprises a modification of the conditions of use of the aircraft engine. This step 70 can be implemented when the health x of the aircraft engine is lower than a predetermined health level x. This step 70 can for example comprise one of the following actions: - carry out maintenance on the aircraft engine, and / or - change the operating conditions of the aircraft engine so that its health deteriorates less quickly.

Claims

Claims

1. Method (1) for monitoring a health x of an aircraft engine comprising training (40) a neural network, the neural network being configured to provide as output a prediction of parameters of the health x of the aircraft engine and taking as input measurement data Y of the aircraft engine, the training being further carried out from: - a training data set comprising noisy measurement data Y and generated from a simulation of the aircraft engine carried out: • according to at least one business hypothesis relating to the knowledge of the aircraft engine, and • according to one or more operating points u of the set of operating points U of the aircraft engine and parameters of the health x of the aircraft engine,- a function penalizing a deviation between the aircraft engine health parameters x corresponding to the measurement data Y and the prediction of the health parameters x obtained by providing the trained neural network with the generated and noisy measurement data Y.,

2. Monitoring method (1) according to claim 1 wherein: - The method (1) is implemented by a monitoring device comprising a set of sensors adapted to carry out measurements of an operation of the aircraft engine and being configured to comprise a simulator § of the aircraft engine, the simulator s taking as input parameters of the health x of the aircraft engine and an operating point u of the aircraft engine, and providing as output corresponding generated measurement data Y, - The simulation of the aircraft engine allowing the generation of the training data set is carried out with the simulator s of the aircraft engine and the training data set is obtained by further carrying out, before the training step (40), additional steps of: • Obtaining (20) at least one business hypothesis, the at least one business hypothesis being a hypothesis of the link between the health x of the aircraft engine and parameter values ​​of the set of operating points, • Generation (30) of the training data set, each training data item comprising generated and noisy measurement data Y as well as parameters of the health x of the aircraft engine and an operating point u of the aircraft engine, the generation being carried out in compliance with the link hypothesis and comprising for each training data item: • Obtaining the measurement data Y generated by providing the aircraft engine simulator s with the health parameters x of the aircraft engine and an operating point u of the aircraft engine, the operating point u of the aircraft engine being included in the set of operating points U, • Adding noise to the generated Y measurement data, the training (40) of the neural network is further performed using the training data set, the neural network further taking as input the operating point u of the aircraft engine, the training function penalizing a deviation, for each training data, between the parameters of the health x of the aircraft engine of the training data and the parameters of the health x obtained by providing the neural network with the generated measurement data Y and the operating point of the aircraft engine of the training data, and Diagnosis (60) of aircraft engine health using trained neural network.

3. Method (1) according to claim 1 or 2 wherein the business hypothesis comprises at least one hypothesis from among a multipoint hypothesis and a temporal hypothesis, with: - the multipoint hypothesis consisting in considering that the health x of the aircraft engine has an influence on the values ​​of the parameters of all the operating points and that the link between the health x of the aircraft engine and the value of the operating points is predetermined, and - the temporal hypothesis consisting in considering that the health x of the aircraft engine evolves over time and that the evolution over time of the health x is predetermined.

4. Method (1) according to any one of the preceding claims in which the neural network is of the type: - A multi-layer perceptron, or - A recurrent neural network, or - A convolutional neural network, or - A transformer.

5. A method (1) according to claim 2 and any one of claims 3 to 4 wherein adding noise to the generated measurement data Y comprises, for each measurement data, adding a sample of a noise distribution having predetermined mean and variances.

6. Method (1) according to claim 2 and any one of claims 3 to 5 further comprising a step (50) of validating the trained neural network, carried out between the training (40) of the neural network and the diagnosis (60) of the health x of the aircraft engine, and the validation step (50) comprising: - Generation of a validation data set, each validation data item comprising measurement data Y and parameters of the health x of the aircraft engine and the operating point of the aircraft engine, the generation being carried out in compliance with the link hypothesis and comprises for each validation data item: • Obtaining the measurement data Y by providing the simulator s of the aircraft engine the parameters of the health x of the aircraft engine and the operating point of the aircraft engine, • Adding noise to the measurement data Y comprising, for each measurement data y, the addition of a sample respecting a Gaussian distribution of noise having a mean equal to zero and a variance equal to a predetermined variance matrix of the noises of all the sensors. - Validating the trained neural network using the validation data set and verifying that the difference, for each validation data, between the parameters of the health x of the aircraft engine of the training data and the parameters of the health x obtained is less than a predetermined threshold.

7. System comprising a set of sensors adapted to carry out measurements on an aircraft engine and a computer configured to implement the method (1) according to one of the preceding claims.

8.

9. Aircraft comprising the system according to the preceding claim. Computer program product comprising instructions which, when the program is executed on a computer, cause the latter to implement the steps of the method (1) according to one of claims 1 to 6.

10. A computer-readable recording medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method (1) according to one of claims 1 to 6.

Images