EP3941826A1

EP3941826A1 - System and method for monitoring an aircraft engine

Info

Publication number: EP3941826A1
Application number: EP20727833.4A
Authority: EP
Inventors: Sébastien Philippe RAZAKARIVONY
Original assignee: Safran SA
Current assignee: Safran SA
Priority date: 2019-04-23
Filing date: 2020-04-21
Publication date: 2022-01-26
Also published as: CN113748066A; US20220242592A1; FR3095424A1; WO2020217004A1

Abstract

A system for monitoring an aircraft engine, comprising: an acquisition module (3) for acquiring current measurements of physical size, referred to as input (ME) and output (MS) physical sizes, relative to the aircraft engine (15), a module (5) for simulating the physical behavior of the aircraft engine, in order to simulate output physical sizes (VS) as a function of the current measurements of input physical sizes (ME), a processor for calculating physical margins, referred to as actual physical margins (MR), between the simulated values of output physical sizes (VS) and the corresponding current measurements of output physical sizes (MS), a learning module (7) for predicting margins, said margins (MP) being predicted using current measurements of input physical sizes (ME) and in that the processor (9) is configured to calculate surveillance residuals (R) showing an indication of the status of the aircraft engine.

Description

DESCRIPTION

TITLE: System and method for monitoring an aircraft engine

Technical area

The present invention relates to the field of monitoring an aircraft engine. In particular, the invention relates to a method and a monitoring system for monitoring the state of the engine in order to anticipate maintenance operations.

The term “aircraft engine” designates the set of turbine engines fitted to flying devices, in particular helicopters and airplanes.

State of the prior art

Currently, monitoring of the condition of aircraft engines and in particular, helicopter turbine engines is carried out by calculating margins or indicators during specific maneuvers known as “CSM” Engine Health Check (or “EPC” Engine Power Check). , in English). Margins are calculated using a physical model simulating the thermodynamic behavior of the engine. More particularly, the model takes as input data relating to the engine and the flight conditions and outputs output data estimated from the input data. The variables of the output data are also recorded in real time by sensors so that the margins are calculated by subtracting between the estimates and the records corresponding to the same variables. These margins are then displayed as a function of the flight dates allowing them to be analyzed visually or by statistical techniques in order to detect anomalies and trends in margins.

However, before launching an automatic data analysis procedure, it is first necessary for the pilot to place the aircraft in particular flight conditions in stable phases. This problem is solved by a method described in patent FR3003032 which allows automatic detection of these stable phases. This method consists in detecting the conditions in which the validity constraints of the thermodynamic model are verified, before calculating the margins. However, these margins remain valid only on the parts for which the physical model remains valid, and therefore suffer from the limitation of the operating domain of the surveillance. In addition, because of a small amount of data in the stable phases, it is sometimes necessary for a pilot to perform flights entirely dedicated to the calculation of the margins, which represents a workload and additional costs.

The object of the present invention is therefore to provide a system for monitoring and tracking an aircraft engine which overcomes the aforementioned drawbacks, in particular by making the best use of the data measured during flights to determine precise indicators. on the state of health of the engine.

Presentation of the invention

The invention relates to a monitoring system suitable for monitoring an aircraft engine, said system comprising:

an acquisition module configured to acquire, during a flight time of the aircraft, current measurements of physical quantities, called physical input quantities and physical output quantities, relating to said aircraft engine and to its environment ,

a module for simulating the physical behavior of said aircraft engine, configured to simulate values of output physical quantities as a function of said current measurements of input physical quantities,

a processor configured to calculate physical margins, called real margins, between said simulated values of physical output quantities and said corresponding current measurements of physical output quantities,

a learning module configured to predict margins, called predicted margins, from current measurements of input physical quantities, and in that said processor is further configured to calculate monitoring residuals between said real margins and said predicted margins, said monitoring residues presenting an indication of the condition of the aircraft engine.

This system consists of a hybrid combination (and not a simple juxtaposition) between the simulation module and the learning module, thus creating a synergy which allows the physical model to be used outside of its normal operating mode, in addition of course to its use in its usual operating mode. This makes it possible to better exploit the data in the validity space of the physical model used by the simulation module, but also to obtain information outside this validity space. It will also be noted that the learning module would have needed to take into account a very large number of physical variables if it were used alone. This would have drastically complicated the control of learning because the physical laws connecting these variables can be non-linear and therefore very sensitive to the initial conditions so that tiny differences can lead to very different results, making the prediction very noisy.

Advantageously, said current measurements of physical input quantities and physical output quantities are acquired during the stable and transient phases of said flight of the aircraft.

By taking into account the transient or unstable phases, we obtain many more points of residuals and consequently, the results are much less noisy, allowing to automatically carry out better trend following and consequently, to be able to react much more quickly in case anomaly. This also makes it possible to monitor unstable or transient phases (for example, during take-off and landing). With fewer points it would have been much more difficult to determine the trend.

Advantageously, the system comprises an interaction and / or display interface for viewing graphical representations of said monitoring residues.

This provides information on trends, anomalies or failures relating to the aircraft engine.

Advantageously, the learning module is based on a learning model built beforehand using a reference aircraft engine during a predetermined number of learning flights, the measurements of input physical quantities relating to the reference engine as well. that the real margins generated by the simulation module being injected during each learning flight into the learning module, allowing the latter to build the learning model. Thus, by using a reference aircraft engine in the same contexts, the learning module is allowed to learn to predict the margins accurately.

Advantageously, said number of learning flights is chosen to ensure a compromise between precision and stability of the learning model and in that only the first elements in the series of flights are taken into account. This makes it possible to increase the learning efficiency while maintaining high precision.

It will be noted that this is not the selection of any number of learning flights, but the selection of the first elements in the series of flights. Indeed, if these flights are not among the first flights of engine use, the relationship to be learned (the physical model / real system margins) changes too much to be learned. The modeling error would no longer be constant but would depend on time, and could therefore no longer be learned.

According to one embodiment, said learning model is constructed according to a statistical technique of linear regression, neural networks, or random forests.

The physical input quantities comprise at least one input parameter relating to the aircraft engine and / or to the flight conditions of the aircraft, comprising at least one parameter chosen from the speed of rotation of the engine, the outside temperature, the external pressure, the fuel flow, the air flow taken from within the engine, the electrical energy taken from the engine, the position of the blades, the flight altitude, the absence or presence of filters, and in that the physical output quantities comprise at least one output parameter representative of the operating state of the aircraft engine, comprising at least one parameter chosen from the internal temperature of the engine and the torque of an engine shaft.

Advantageously, the monitoring residuals are aggregated in the form of means or modes for a synthetic representation.

According to one embodiment of the present invention, the aircraft engine is a helicopter turbine engine.

The invention also relates to a method for monitoring an aircraft engine, comprising the following steps:

- acquire, during a flight time of the aircraft, current measurements of quantities physical quantities, called physical input quantities and physical output quantities, relating to said aircraft engine and its environment,

- simulate the values of physical output quantities as a function of said current measurements of physical input quantities,

- calculating physical margins, called real margins, between said simulated values of physical output quantities and said corresponding current measurements of physical output quantities,

- predict margins, known as predicted margins, from current measurements of physical input quantities, and

- calculating monitoring residues between said real margins and said predicted margins, said monitoring residues presenting an indication of the state of the aircraft engine.

It will be noted that according to the present invention a first (physical) model is used to build a better model (empirical + physical).

It will also be noted that in the present invention, the predicted margins are defined at the same time as the data used. In addition, the term “prediction” is used here in the sense of statistical learning and therefore of an output of the empirical model on input data.

It will also be noted that in the present invention the monitoring residue is defined between the actual margins and the predicted margins. Thus, the monitoring residues are distinct from the “margins” usually used by those skilled in the art.

Furthermore, it will be noted that the learning module according to the present invention does not predict the state of the system, but the residue between a physical model and the system. Thus, the present invention relates to the use of two waterfall models: the first model (physical) gives margins, and the second model (empirical / learning) predicts the usual deviation between the physical model and the real system . Note that it is not easy to use the physical model in areas in which it is not valid according to physical theories. However, by construction, the system according to the present application learns to correct these errors. Brief description of the figures

Other characteristics and advantages of the invention will become apparent on reading a preferred embodiment of the invention, described with reference to the appended figures, among which:

[Fig. 1] schematically illustrates a system for monitoring an aircraft engine according to one embodiment of the invention;

[Fig. 2] schematically illustrates a method for monitoring an aircraft engine according to one embodiment of the invention;

[Fig. 3A] schematically illustrates the prior learning phase of the learning module, according to one embodiment of the invention;

[Fig. 3B] schematically illustrates the operational phase, according to one embodiment of the invention; and

[Fig. 4] illustrates a graph representing monitoring margins and residuals, according to one embodiment of the invention.

Description of embodiments

The principle of the invention consists in coupling a physical model of the behavior of the aircraft engine with a learning model so that the use of the measurements acquired during flights is maximized, thus allowing complete and optimal monitoring of the engine.

FIG. 1 schematically illustrates a system for monitoring an aircraft engine according to one embodiment of the invention.

The monitoring system 1 comprises an acquisition module 3, a simulation module 5, a learning module 7 and a processor 9.

The monitoring system 1 can be entirely included in an aircraft 11 or shared between the aircraft 11 and a maintenance center 13.

The aircraft 11 (here represented by a helicopter but can be an airplane) comprises an engine 15, an on-board computer 17 and sensors 19. The sensors 19 measure physical quantities, called physical input quantities and physical output quantities, relating to the aircraft engine 15 and to its environment.

By way of example, the input physical quantities can comprise at least one input parameter relating to the engine 15 of the aircraft and / or to the flight conditions of the aircraft 11, comprising at least one parameter chosen from the speed of rotation of the engine 15, the outside temperature, the outside pressure, the fuel flow, the air flow taken from the engine 15, the electrical energy taken from the engine 15, the position of the blades, the altitude theft, and the absence or presence of filters. The physical output quantities may include at least one output parameter representative of the operating state of the aircraft engine, comprising at least one parameter chosen from the internal temperature of the engine 15 and the torque of a shaft of the engine 15.

The on-board computer 17 comprises the processor 9, the acquisition 3, simulation 5 and learning modules 7 as well as a recording memory 21 and an interface 23. It will be noted that the maintenance center 13 also comprises a computer 117 which can include the same elements (ie processor 109 and acquisition 103, simulation 105 and learning 107 modules, a memory 121 and an interface 123) as the on-board computer 17.

The memory 21 (and / or 121) forms a recording medium, readable by the processor 9 (and / or 109) and on which is recorded one or more computer program (s) comprising instruction codes for the execution of the monitoring method described below with reference to Figure 2.

In fact, FIG. 2 schematically illustrates a method for monitoring an aircraft engine according to one embodiment of the invention.

In step E1, the acquisition module 3 (and / or 103) is configured to acquire, during a flight time of the aircraft 11, the current measurements of physical input quantities ME and of physical quantities of MS output.

In step E2, the simulation module 5 (and / or 105) is adapted to simulate the physical behavior of the aircraft engine 15. Indeed, the simulation module 5 comprises a thermodynamic model of the relationships between physical quantities relating to the engine 15 and it acts as a solver which calculates output data according to the input data. Thus, the simulation module 5 is configured to simulate the values of physical output quantities VS as a function of the current measurements of physical input quantities ME retrieved from the acquisition module B. Thus, the values of physical output quantities VS are simulated from the input values observed in real time.

In step E3, the processor 9 (and / or 109) is configured to calculate physical margins, called real margins, between the simulated values of physical output quantities VS and the corresponding current measurements of physical output quantities MS. The actual margins represent the differences between the actual engine data observed in real time and the outputs calculated by the simulation module. In other words, the real margins define the real errors between the physical model and the observed physical quantities.

Advantageously, the current measurements of physical input quantities ME and physical output quantities MS are acquired by the acquisition module 3 during the stable and transient phases of the flight of the aircraft 11.

Indeed, the physical model has a specific domain of validity, but outside this domain of validity, the modeling can be considered as a reasonable approximation of the behavior of the engine 15 and can be used as support by the learning module 7. Note, however, that this approximation is not good enough to be used on its own.

In step E4, the learning module 7 (and / or 107) is configured to predict margins, called predicted margins MP, from current measurements of physical input quantities ME (ie the observed input values ). The learning module 7 comprises a statistical learning model based on a known technique of the linear regression type, neural networks, or random forests.

The learning module 7 is based on a statistical learning model built beforehand to learn the margins according to classical techniques of linear regression type, or random forests. This learning phase is described later with reference to FIG. 3.

Thus, a coupling is built between the simulation module 5 and the learning module 7, allowing the latter to learn to predict precise margins on all phases of the flight from the margins generated by the simulation module 5, by correcting the approximate margins generated by the simulation module 5 in the transient phases. This coupling makes it possible to better use the data in the validity space of the physical model used by the simulation module 5 and to obtain, in addition, information outside this validity space.

In step E5, processor 9 (and / or 109) is further configured to calculate monitoring residues R between the real margins and the predicted margins. The monitoring residues provide an indication of the condition of the aircraft engine. The fact of following the monitoring residuals makes it possible to improve the calculated margins and to be able to obtain precise results outside the domain of validity of the physical model.

The monitoring residues R can be represented by point clouds or graphs in order to be visualized on the interaction and / or display interface 23 (or 123) of the computer 17 (or 117) thus giving information on trends, anomalies or failures relating to the aircraft engine. Advantageously, for better readability, the monitoring residues are aggregated in the form of means, modes, or any other dimension reduction technique for a synthetic representation.

Advantageously, the operational data collected during the flight of the aircraft are downloaded at the end of the flight. The computer 117 of the maintenance center 13 performs the steps according to FIG. 2 and the results are displayed on the interface 123. The maintenance experts can thus look at the residue curves, in order to alert in the event of abnormal behavior, that either breaks in the curves, abnormal tendencies, or anomalies.

FIGS. 3A and 3B schematically illustrate the prior learning phase of the learning module as well as the operational phase, according to one embodiment of the invention. More particularly, FIG. 3A represents the learning phase during which the learning module constructs a learning model based on a very stable reference aircraft engine 115.

In order to increase the precision of the learning, it is advantageous to have a reference motor 115 used in the same contexts as the motors to be monitored. It is also advantageous to take the first elements in the series of recorded flights, and not randomly chosen flights from the database, as is usually done in statistical learning.

The learning takes place over a time window [t-k; t] defined as relevant from about a few hours to a few tens of flight hours, with a sufficient number of examples taken in the first flights of a predetermined number of learning flights. Indeed, the learning model works on the margins and not directly on the output variables. By definition, the margins are small and therefore noisy, which requires good control of the learning part, in particular the number of examples to be used. The larger this number, the more precise the learning model. However, if it is too big, the motor will change state and we lose the stability of the model. Thus, the number of training flights is chosen to ensure a compromise between precision and stability of the training model and it is consequently advantageous to choose a reasonable number of flights. This number depends on the learning technique and can be, by way of example, between three and ten flights.

Advantageously, the time window encompasses the physical input quantities relating to the aircraft engine and to the flight conditions of the aircraft, until the variable (s) of physical output quantities is predicted, of the time step. tk at time step t.

Indeed, given that outside the areas of validity of the physical model, the input variables relating to the engine and to the external conditions are not in a steady state, the learning technique according to the invention takes into account the past over a certain time window to further reduce noise. In other words, one uses not only the preceding time steps, from tk to t-1, but also the values of the variables at time t. It will be noted that in the classical methods learning, only the previous time steps are taken into account according to a predictive logic.

As in the operational phase, the acquisition module of the reference motor 115 collects during each learning flight, the current measurements of physical input quantities ME and of physical output quantities MS relating to the reference motor 115 during the stable and transient phases of flight.

The physical input quantities ME are injected into the simulation module 5. The latter estimates values of physical output quantities VS as a function of the current measurements of physical input quantities ME.

The processor then calculates the differences between the simulated values of physical output quantities VS and the corresponding current measurements of physical output quantities MS. These differences produce the actual MR margins.

Then, the measurements of physical input quantities ME (relating to the reference motor and to the external conditions) over several time steps as well as the real margins MR generated by the simulation module 5 are injected into the learning module 7 allowing the latter to build the learning model.

Thus, the learning module 7 learns, thanks to the input physical quantities affecting the reference engine over several time steps and not only on the instantaneous values, the relations between the real margins MR and the predicted margins MP (ie the difference between the margins generated by the physical model and those generated by the learning model).

The learning according to the invention is thus facilitated by the physical model which absorbs the impact of parameterization data and fills a large part of the data. More specifically, the monitoring system merges the physical model and the training model where the results of the physical model are used by the training model to predict the margins. Furthermore, learning is based on historical data and uses the entire recorded time series (ie stable and unstable). FIG. 3B illustrates the operational phase, according to the steps described with reference to FIG. 2, which consists in predicting the values of the margins obtained and comparing them with the margins actually output by the reference engine.

More particularly, the acquisition module collects, during each operational flight, the current measurements of physical input quantities ME and physical output quantities MS relating to the engine 15 under surveillance during all the stable and transient phases of the flight.

The simulation module 5 then estimates the values of physical output quantities VS as a function of current measurements of physical input quantities ME. The processor then calculates the actual margins MR by making the difference between the simulated values of physical output quantities VS and the corresponding current measurements of physical output quantities MS.

Then, the learning module 7 determines the predicted margins MP from the current measurements of physical input quantities ME. Finally, the processor calculates the monitoring residuals R between the real margins MR and the predicted margins MP.

FIG. 4 illustrates a cloud of points representative of the monitoring margins and residuals, according to one embodiment of the invention.

According to this example, the ordinate axis represents the internal temperature margins of the engine 15 (or 115) on a scale of 10 ° C. and the abscissa axis represents the dates or periods of the flights subdivided into quarters.

Each black point represents the average of the actual MR margins per flight and each gray point represents the average of the monitoring residuals R per flight.

At the start, there is a preliminary learning phase (PO period) where there are only black dots (ie real margins). Then, we enter an operational phase (periods P1-P8) where the monitoring residuals R (ie the gray points) are calculated in addition to the real margins MR. It will be noted that there are many more gray points than black points because the latter are calculated only on the stable phases while the gray points are determined on all the phases of flight. Thus, by taking into account the transient or unstable phases, we obtain many more monitoring points and consequently, the results are much less noisy, allowing to automatically carry out better trend following and suddenly to be able to react much faster in case of anomaly. This also makes it possible to monitor unstable or transient phases (for example, during take-off and landing). With fewer points it would have been much more difficult to determine the trend.

Note that an engine is considered old when it no longer has any margin (i.e. when the internal temperature margin according to the diagram in figure 4 is close to zero). This diagram also shows that in period P6, the engine underwent maintenance which increased the margin.

Thus, the surveillance system according to the invention uses the physical model as a first order approximation, which makes it possible to facilitate learning. In addition, it uses the first flights of a benchmark engine to learn the model rather than taking random flights. It also uses the history of the variables and not just the instantaneous value of these variables, to properly model the dynamic parts.

The monitoring method and system according to the invention is suitable for all aircraft engines and in particular for an aircraft engine which has many transient or unstable phases such as the helicopter.

1. System for monitoring an aircraft engine, comprising: - an acquisition module (S) configured to acquire, during a flight time of the aircraft, current measurements of physical quantities, called physical quantities d 'input (ME) and output physical quantities (MS), relating to said aircraft engine (15) and to its environment,

- a simulation module (5) of the physical behavior of said aircraft engine, configured to simulate values of output physical quantities (VS) as a function of said current measurements of input physical quantities (ME),

- a processor (9) configured to calculate physical margins, called real margins

(MR), between said simulated values of physical output quantities (VS) and said corresponding current measurements of physical output quantities (MS),

- a learning module (7) configured to predict margins, called predicted margins (MP), from current measurements of input physical quantities (ME), and in that said processor (9) is further configured for calculating monitoring residuals (R) between said actual margins (MR) and said predicted margins (MP), said monitoring residuals presenting an indication of the condition of the aircraft engine.

2. System according to claim 1, characterized in that said current measurements of physical input quantities and physical output quantities are acquired during the stable and transient phases of said flight of the aircraft.

B. System according to claim 1 or 2, characterized in that it comprises a display interface (23) for viewing graphical representations of said monitoring residues.

4. System according to any one of claims 1 to 3, characterized in that the learning module (7) is based on a learning model previously constructed using a reference aircraft engine (115) during a predetermined number of learning flights, the measurements of input physical quantities relating to the reference engine as well as the actual margins generated by the

Claims

simulation being injected during each learning flight into the learning module allowing the latter to build the learning model.

5. System according to claim 4, characterized in that said number of learning flights is chosen to ensure a compromise between precision and stability of the learning model and in that only the first elements in the series of flights are taken into account. account.

6. System according to claim 4 or 5, characterized in that said learning model (7) is constructed according to a statistical technique of linear regression or of random forests.

7. System according to any one of the preceding claims, characterized in that the input physical quantities comprise at least one input parameter relating to the aircraft engine and / or to the flight conditions of the aircraft, comprising at least at least one parameter chosen from among the engine speed, outside temperature, outside pressure, fuel flow, air flow taken from within the engine, electrical energy taken from within the engine, position of the blades , the flight altitude, the absence or presence of filters, and in that the physical output quantities include at least one output parameter representative of the operating state of the aircraft engine, comprising at least one chosen parameter among the internal temperature of the engine and the torque of an engine shaft.

8. System according to any one of the preceding claims, characterized in that the monitoring residues are aggregated in the form of means or modes for a synthetic representation.

9. System according to any one of the preceding claims, characterized in that the aircraft engine (15) is a helicopter turbine engine.

10. A method of monitoring an aircraft engine, comprising the following steps:

- acquire, during the flight time of the aircraft, current measurements of physical quantities, called physical input quantities (ME) and physical output quantities (MS), relating to said aircraft engine and its environment,

- simulate the values of physical output quantities (VS) as a function of said current measurements of physical input quantities (ME), - calculate physical margins, known as real margins (MR), between said simulated values of physical output quantities (VS) and said corresponding current measurements of physical output quantities (MS),

- predict margins, known as predicted margins (MP), from current measurements of physical input quantities, and

- calculating monitoring residues (R) between said real margins and said predicted margins, said monitoring residues presenting an indication of the state of the aircraft engine.