Method and device for monitoring the health of a vehicle's mechanical system using variable thresholds based on system operating parameters

Abstract

The present invention relates to a method for monitoring a mechanical system (10) for a vehicle (1). During an initial phase, vibration values ​​from vibration sensors (20) and context parameters from context sensors (25) are measured to determine several values ​​of a health indicator for the mechanical system (10) and a model for the variation of a threshold for this health indicator, which varies according to the context parameters. Then, during a monitoring phase, vibration values ​​and context parameters are again measured to determine an operational value for the health indicator and a threshold specific to this health indicator, using the threshold variation model and the measured context parameters. A fault alarm can then be triggered if there is a risk of a fault in the mechanical system (10).Abbreviated figure: figure 1.

Description

Title of the invention: Method and device for monitoring the health of a mechanical system of a vehicle using variable thresholds according to operating parameters of the system

[0001] The present invention lies in the field of health monitoring systems for mechanical systems.

[0002] The present invention relates to a method and a device for monitoring the health of a mechanical system equipping a vehicle using variable thresholds according to operating parameters of the system.

[0003] A mechanical system may comprise several moving members, for example several rotating members, such as an input shaft and an output shaft.

[0004] For example, a mechanical system may comprise one or more bearings in order to guide the rotation of an input shaft and an output shaft or even one or more intermediate rotating members. A bearing comprises, for example, a rolling bearing provided with one or more rows of rolling elements such as balls, rollers or others.

[0005] A mechanical system can also be equipped with at least one toothed wheel, a pinion, a toothed crown.

[0006] The operation of a mechanical system can be monitored to detect an operating anomaly, or even anticipate the occurrence of a breakdown or malfunction. An aircraft can thus include an on-board monitoring system for monitoring the health of a mechanical system in order to increase flight safety.

[0007] Such a monitoring system can, regardless of the mechanical system being monitored, measure and record operational data of the mechanical system, in particular of a vibration nature. This monitoring system can then determine health indicator values ​​based on this operational data and compare them to predefined thresholds to detect a potential anomaly linked to an emerging mechanical defect. The value of each threshold can be obtained by experience or by a statistical analysis of a history of measurements from several similar mechanical systems acquired during testing or validation phases or even during the first hours of use of such a mechanical system. The thresholds can for example be estimated from a parametric modeling of the statistics of the health indicator with the assumption that the mechanical system is healthy, namely without anomaly, malfunction or defect.

[0008] The value of each threshold may be constant over the entire operating range of the mechanical system. Each detection threshold may alternatively comprise several constant values ​​relating respectively to specific phases. operating characteristics of the mechanical system.

[0009] A health indicator may take the form of a vibration indicator evaluated using at least one sensor, this sensor being able to comprise at least one accelerometer, one tachometer. Such a health indicator may be defined using a signal provided by a single sensor, for example being equal to the maximum amplitude of a temporal vibration signal provided by an accelerometer or by combining the signals from several sensors. This is for example the case of health monitoring systems designated by the acronym HUMS for “Health and Usage Monitoring System” in English.

[0010] A health indicator may allow monitoring of a mechanical system as a whole while other health indicators may be specific to a particular type of fault that may appear on a particular component of the mechanical system.

[0011] Documents US 7684936, US 2017 / 0023438, US 8442778, US 8131420, CN 102963533 and US 2011 / 0166799 describe HUMS monitoring systems. Documents US 8442778 and US 2017 / 0023438 are intended in particular for monitoring wind turbines. Documents US 8131420, CN 102963533 and US 2011 / 0166799 are dedicated to rotary wing aircraft.

[0012] In particular, document CN 102963533 presents a monitoring system. This monitoring system comprises an on-board system for identifying a risk in real time and generating an alert, as well as a remote ground-based system capable of downloading and analyzing the data acquired in flight in order to plan the necessary maintenance operations.

[0013] Document US 2011 / 0166799 describes a monitoring system determining a health indicator whose value is adjusted as a function of an operating parameter of the monitored mechanical system, and in particular as a function of a torque exerted on a rotating member. This health indicator can then be compared to a threshold to estimate a risk of a fault occurring.

[0014] Such monitoring systems are interesting but the usual health indicators are dependent on the operating conditions of the mechanical system. These operating conditions are, for example, characterized by context parameters. These context parameters are, for example, the altitude or the forward speed of the aircraft, the torque exerted within a transmission box, the rotation speed of a rotor, etc. This dependence can result in a non-stationarity of the health indicators over the entire operating range of the mechanical system and / or the aircraft. Therefore, these health indicators may only be usable during particular operating regimes. Consequently, monitoring the health of a mechanical system may be restricted to certain periods of use, excluding any transient regime in particular, which may delay the detection of a fault, or on the contrary cause the detection of false failures.

[0015] Specific methods applied to health indicators then aim to improve fault detection, in particular by taking into account the operating conditions of the mechanical system.

[0016] For example, document EP 2345894 describes a method for indicating a fault propagating in a rotating shaft. After measurements of data corresponding to a health indicator of the shaft and measurements of an input factor, for example a torque exerted on the shaft, a relationship between the health indicator and the input factor is determined. A filtered health indicator is then determined based on this relationship and an effect of the input factor on the health indicator, this effect being estimated using a Kalman filter, recursive least squares estimation and / or a particle filter.

[0017] Document EP 4091945 describes a method for detecting faults in a transmission system of an aircraft. This method proposes to acquire, during several time intervals, a signal relating to the dynamic behavior of the transmission system as well as values ​​of flight parameters of the aircraft. Groups of indicators relating to each time interval are determined from these signals. An estimate of the probability that the values ​​of the indicators of a group of indicators are abnormal compared to a predefined learning condition is calculated for each indicator of the group of indicators, taking into account the measured flight parameters. Health indicators are determined according to this estimate of the probability, then compared with a threshold to identify the presence of a fault.

[0018] Document WO 2011 / 023596 describes a method and a system for controlling vibrations in a wind turbine. A range of characteristic vibration values ​​is predicted from real-time operating data of an electrical generator of the wind turbine according to rules extracted from a set of rules. A real-time comparison of the characteristic operating values ​​of the generator with a calculated threshold is carried out to detect and signal a possible presence of a fault on the electrical generator.

[0019] Document CN 107218997 describes a method for detecting faults on a hydroelectric group based on the identification of the operating conditions of the hydroelectric group. Historical vibration data is used to estimate a threshold. An operating status signal of the hydroelectric group is acquired in order to determine the operating conditions of the hydroelectric group. An anomaly can thus be sensitively detected based on the historical vibration data and the operating status identified using the status signal when the vibration values ​​exceed the threshold.

[0020] The present invention therefore aims to propose an innovative method and device aimed at monitoring a mechanical system of a vehicle in order to detect as early as possible the appearance of a fault on the mechanical system while limiting the risk of false detection of a fault.

[0021] The present invention relates, for example, to a method for monitoring the health of a mechanical system equipping a vehicle, the mechanical system comprising at least one moving member as well as at least one vibration sensor emitting a vibration signal, the mechanical system or the vehicle comprising at least one context sensor emitting a context signal relating to at least one context parameter, said at least one context parameter being chosen from a list comprising at least one or more functional parameters of the mechanical system, one or more navigation parameters of the vehicle and one or more atmospheric parameters.

[0022] This method is composed of an initial phase of defining a threshold variation model as a function of said at least one context parameter, this initial phase being followed by a phase of operational monitoring of the mechanical system.

[0023] The initial phase includes the following steps: - measurements of successive initial vibration values ​​from said at least one vibration sensor and successive initial context values ​​from said at least one context sensor, - determination, using a calculator, of several initial health values ​​of at least one health indicator CI relating to the mechanical system as a function of the initial vibration values, each initial health value being associated with one of the initial context values ​​of said at least one context parameter, - definition for each CI health indicator, of a threshold variation model, the threshold variation model being defined as a function of the initial health values ​​of the CI health indicator and the initial context values ​​of said at least one context parameter, by partitioning a domain formed by the initial context values ​​of said at least one context parameter into several ranges of values ​​for which the threshold relating to the CI health indicator is statistically constant over each range of values, each range being associated with several of the initial context values, and by determining threshold values ​​with a method using regressions on quantile distribution parameters conditioned on context parameters and domain decompositions of said at least one context parameter using at least one decision tree.

[0024] The operational monitoring phase includes the following steps during operation of the mechanical system: - measurements of successive operational vibration values ​​from said at least one vibration sensor, and of successive operational context values ​​from said at least one context sensor, - determination, with the calculator, of an operational health value of at least one health indicator Cl as a function of the vibration operational values, the operational health value of a health indicator CI being associated with one of the context operational values ​​relating to one or more context parameters, - determination of the threshold specific to the operational health value of said at least one health indicator Cl using the threshold variation model and said at least one associated context operational value, - triggering an alert signaling a risk of the presence of a fault on the mechanical system if the operational health value of said at least one CI health indicator is greater than the determined threshold.

[0025] According to this method, one or more CI health indicators are determined during the initial phase from measurements from said at least one vibration sensor, then a threshold variation model is defined for each CI health indicator as a function of the values ​​of this CI health indicator and the measurements from said at least one context sensor on this initial phase using regressions on quantile distribution parameters. To define this variation model, at least one decision tree is used to define one or more ranges of the context parameter(s) for which the threshold relating to the CI health indicator is statistically constant.A threshold for a CI health indicator is considered to be statistically constant over a range of contiguous values ​​of a context parameter if, for any division of said range of values ​​into two sub-ranges of contiguous values, the thresholds estimated respectively over the two sub-ranges of values ​​are considered identical according to a statistical test. Such a statistical test may be, for example, a Welch test.

[0026] The term "each" is used for convenience, both in the presence of a single CI health indicator and CI health indicators to facilitate reading.

[0027] The initial phase can thus be a learning period making it possible to define the threshold variation model for each health indicator CI of the mechanical system. This initial phase is carried out following the first use of the vehicle or following a significant maintenance operation, for example following the replacement of the mechanical system. During this learning period, the mechanical system is considered to be fault-free. This initial phase thus makes it possible to determine the threshold variation model of the mechanical system considered to be healthy in the vehicle.

[0028] Following this learning period and the definition of the threshold variation model, the health of the mechanical system is monitored using the threshold variation model during the operational monitoring phase. The method according to the invention thus makes it possible to adapt the threshold value to which a health indicator Cl is compared to the context parameters, which makes it possible to improve the detection of a mechanical defect, or of an anomaly linked to an emerging mechanical defect, while reducing the false detection rate.

[0029] The reliability and availability of the mechanical system and of the vehicle equipped with this system are thus improved thanks to the method according to the invention, compared to traditional monitoring methods whose threshold is generally algebraically constant over the entire area of ​​use of the vehicle.

[0030] Furthermore, the context parameters may include operating parameters of the mechanical system, such as the rotation speed of one or more rotating shafts, the torque(s) at this or these shafts, the temperature of the mechanical system for example.

[0031] The context parameters may include operating parameters of the vehicle, and in particular of an aircraft, equipped with the mechanical system, such as the forward speed of this vehicle, its possible vertical speed, its altitude, the rotation speed of a rotor for example.

[0032] The context parameters may also include atmospheric parameters, such as the temperature and atmospheric pressure outside the mechanical system, for example.

[0033] Thus, the context sensor(s) make it possible to measure context characteristics relating to the operation of the mechanical system, and / or the vehicle as such, as well as relating to the environment, and to transmit a temporal context signal, carrying information relating to these context characteristics, to a computer for example. The temporal context signal carries the initial context values ​​during the initial phase of defining a threshold variation model, and the operational context values ​​during the operational monitoring phase of the mechanical system.

[0034] The vibration signal and the context signal can be generated simultaneously, or even synchronously.

[0035] A computer may be present in the vehicle or present on a station separate from the vehicle. The vibration and context signals may be stored in a memory of the computer or in a memory present in the vehicle and possibly connected to the computer by a wired or wireless connection for example.

[0036] The method according to the invention may comprise one or more of the following characteristics, taken alone or in combination.

[0037] According to one example, the definition of the threshold variation model may be performed using a single decision tree. Alternatively, the definition of the threshold variation model may be performed by aggregating at least two decision trees.

[0038] According to another example compatible with the previous ones, the definition of the threshold variation model can be carried out with a predetermined minimum false alarm rate and a predetermined confidence rate in the minimum false alarm rate. The use of this minimum false alarm rate, corresponding to a probability of emission of false alarms, thus makes it possible to reduce the false detection rate, and consequently, the false alarm emission rate, thus limiting the immobilizations of the mechanical system and the vehicle while no fault is present.

[0039] According to another example compatible with the previous ones, the initial phase may comprise an analysis step for analyzing a sensitivity of each health indicator Cl to the context parameters and for selecting influential context parameters. This analysis step is carried out after determining several initial health values ​​of each health indicator CI and before defining a threshold variation model.

[0040] In this way, the context parameters most influential on the variations of each CI health indicator are determined and selected during this analysis. Consequently, only these selected context parameters are taken into account in defining the threshold variation model for each CI health indicator and, subsequently, in determining the variable threshold.

[0041] Alternatively, the initial phase may comprise a preselection step for selecting context parameters taken into account by the method from a multitude of context parameters measured by context sensors. This preselection step is carried out before the measurements of the initial vibration and context values.

[0042] In this case, the context parameters most influential on the variations of each health indicator are selected prior to carrying out the initial phase, for example by an empirical analysis of measurements on other mechanical systems, for example equivalent or similar. Consequently, only these previously selected context parameters are taken into account in defining the threshold variation model for each health indicator CI and, subsequently, to determine the variable threshold.

[0043] According to another example compatible with the previous ones, said initial values ​​of said at least one CI monitoring indicator can be calculated from a decomposition into Fourier coefficients of the successive initial vibrational values ​​modeled with a hypothesis of cyclostationarity of order one and two of the vibrational signal, which follows a Chi-square law, generalized by a GAMMA distribution.

[0044] The present invention also relates to a computer program comprising instructions which, when the program is executed, lead to the implementation of the various examples of the monitoring method described above. This computer program can be stored for example in a memory of a computer present in the vehicle, or in a memory located in the vehicle and connected to the computer, or even in a memory of a station separate from the vehicle.

[0045] This computer program can be implemented entirely by a calculator present in the vehicle or a vehicle computer, or jointly by such a calculator or computer present in the vehicle and by a station separate from the vehicle.

[0046] The present invention also relates to a monitoring device configured to monitor a mechanical system equipping a vehicle, the mechanical system comprising at least one moving member, the monitoring device comprising at least one vibration sensor emitting a vibration signal, at least one context sensor emitting at least one context signal relating to at least one context parameter and a computer.

[0047] The monitoring device is configured to implement the various examples of the health monitoring method previously described.

[0048] Furthermore, the computer may comprise a single computing unit embedded in the vehicle. Therefore, the entire monitoring process may be carried out in the vehicle, whether the vehicle is stationary or in operation.

[0049] Alternatively, the calculator may comprise a single calculation unit separate from the vehicle, and present for example in a station external to the vehicle. In this case, the step of measuring the initial vibration values ​​and the initial context values ​​is carried out in the vehicle in operation during the initial phase of defining a threshold variation model, then the station receives these initial vibration values ​​and these context values, the calculation unit of the station carrying out the steps of determining the initial health values ​​of each health indicator C / , and of defining the threshold variation model.During the operational phase, the step of measuring the vibration operational values ​​and the context operational values ​​is carried out in the vehicle in operation, while the steps of determining the operational health value of each health indicator Cl, determining the threshold specific to the operational health value of each health indicator CI, and triggering an alert of a risk of the presence of a fault on the mechanical system if the operational health value of at least one health indicator Cl is greater than the determined threshold are carried out by the station's calculation unit.

[0050] Alternatively, the calculator may comprise a calculation unit embedded in the vehicle and a computing unit separate from the vehicle, and present for example in a station external to the vehicle. In this case, the step of measuring the initial vibration values ​​and the initial context values ​​of the initial phase and the step of measuring the operational vibration values ​​and the operational context values ​​of the operational phase are carried out in the vehicle in operation. The other steps of the method can be carried out either in the vehicle in operation or when stopped, or in the station external to the vehicle which receives these initial vibration values ​​and these context values ​​as well as these operational vibration values ​​and these operational context values.

[0051] The present invention also relates to a mechanical system comprising at least one moving member and a monitoring device as previously described. This mechanical system may for example be a power transmission box of a vehicle, and of an aircraft in particular.

[0052] The present invention also relates to a vehicle and in particular an aircraft, comprising such a mechanical system.

[0053] The invention and its advantages will appear in more detail in the context of the description which follows with examples given for illustrative purposes with reference to the appended figures which represent: - [Fig.l], a schematic side view of an aircraft, - [Fig.2], a diagram illustrating a system monitoring process mechanics according to the invention, - [Fig.3], a diagram illustrating a method of monitoring a system mechanics according to the invention, - [Fig.4], a diagram illustrating a histogram of a health indicator conditioned on a context parameter, - [Fig.5], a diagram illustrating a single decision tree associated with two pa contextual records, - [Fig.6], a diagram illustrating the separation of a data block relating to the values ​​of a health indicator in a space of two context parameters, - [Fig.7], a curve illustrating the variation of Welch's statistics in function of a context parameter, - [Fig.8], a diagram illustrating a two-dimensional discretization of a block of data based on two context parameters, - [Fig.9], a diagram illustrating an aggregation of decision trees, - [Fig. 10], a diagram illustrating a threshold variation model, and - [Fig. 11], a graph representing curves of a variable threshold.

[0054] Elements present in several distinct figures are assigned a single and same reference.

[0055] [Fig.l] represents a vehicle 1, and a rotary wing aircraft of the rotorcraft type in particular. This vehicle 1 comprises a mechanical system 10 provided with one or more moving members 15. These members 15 are for example rotating around an axis AX of rotation.

[0056] The mechanical system 10 may for example comprise an output shaft or an input shaft. The mechanical system 10 may also comprise one or more rotational guide bearings, for example for the rotational guidance of at least one rotating member. A bearing comprises for example a rolling bearing provided with rolling elements.

[0057] The mechanical system 10 may also comprise, for example, at least one toothed wheel, a pinion, a toothed crown, fixed or mobile.

[0058] A mechanical system 10 may be, for example, a gearbox or a power transmission box of the vehicle 1. This mechanical system 10 may be connected, for example, to one or more engines 2, via respectively one or more input shafts and may drive a rotor in rotation via an output shaft, such as for example a main rotor 3 or an auxiliary rotor 4 of a rotorcraft as shown in [Fig.l]. According to another example, a mechanical system 10 may be a rotary wing.

[0059] Alternatively, such a mechanical system 10 may, for example, be arranged in a gearbox or a power transmission box of a vehicle 1 or any other mechanical equipment.

[0060] These examples are simply given to illustrate the invention.

[0061] Whatever its arrangement, the mechanical system 10 also comprises a monitoring device 9 intended to monitor the mechanical system 10 in order to detect and identify the presence of a fault, or a risk of a fault occurring.

[0062] The monitoring device 9 comprises one or more vibration sensors 20, one or more context sensors 25 and a computer 50. The computer 50 can be dedicated to the monitoring device 9 or be shared with other devices of the vehicle 1.

[0063] Each vibration sensor 20 can emit a temporal vibration signal relating to a vibration behavior of the mechanical system 10 as a whole, or to a particular vibration behavior of a member 15 of the mechanical system 10, for example a shaft, a bearing or even a gear. The temporal vibration signal carries successive vibration values ​​relating to the vibrations of the mechanical system 10, according to for example a sampling frequency. The vibration sensor(s) 20 can comprise, for example, an accelerometer, a tachometer or others.

[0064] Each context sensor 25 can emit a temporal context signal relating to a context characteristic linked to the operation of the mechanical system 10 and / or the vehicle 1 as well as to the environment. The temporal context signal carries successive context values ​​relating to a context parameter linked to the mechanical system 10, to the vehicle 1 and / or to the environment, according to for example a sampling frequency.

[0065] The context parameters may comprise operating parameters of the mechanical system 10, such as the rotational speed of one or more rotating shafts, the torque(s) at this or these shafts, the temperature of the mechanical system 10 for example. One or more context sensors 25 may then comprise different types of sensors associated with each operating parameter of the mechanical system 10, such as for example an accelerometer, a tachometer, a torque meter, or others. One or more context sensors 25 may also comprise an angular sensor, such as an encoder type sensor, making it possible to measure a variation in the angular position of a rotating member 15 of the mechanical system 10 in order, for example, to synchronize the vibration signals and the context signals with the rotation and the angular positions of this rotating member 15.In the case where vehicle 1 is an aircraft, the context parameters can be managed by an on-board system designated by the acronym UMS corresponding to the English designation “Usage Monitoring System”.

[0066] The context parameters may include operating parameters of the vehicle 1, such as its forward speed, its vertical speed, its altitude, the rotation speed of a rotor 3, 4 for example. One or more context sensors 25 may then include different types of sensors associated with each operating parameter of the vehicle 1, such as for example an accelerometer, a tachometer, an altimeter, or others.

[0067] The context parameters may also include atmospheric parameters linked to atmospheric conditions, such as the temperature outside the mechanical system 1 and the atmospheric pressure for example. One or more context sensors 25 may then include a thermometer and / or a barometer for example.

[0068] Furthermore, the computer 50 may comprise a single calculation unit 51 on board the vehicle 1.

[0069] Alternatively, the computer 50 may comprise a single calculation unit 55 separate from the vehicle 1, and present for example in a station 40 external to the vehicle 1.

[0070] Alternatively, the computer 50 may comprise a calculation unit 51 embedded in the vehicle 1 and a calculation unit 55 separate from the vehicle 1, and present for example in a station 40 external to the vehicle 1.

[0071] Each calculation unit 51,55 may comprise at least one processor and at least one memory 6', at least one integrated circuit, at least one programmable system or at least one logic circuit, these examples not limiting the scope given to the expression "computing unit". The computer 50 can also be connected to a memory 6', 56 present in the vehicle 1 or in the station 40, by a wired connection or a wireless connection. The computer 50 can also include a memory 6'.

[0072] The vibration and context signals can thus be stored in the memory 6,6' after having been transmitted to this memory 6,6', by a wired link or a wireless link, the vibration and context signals being able to be analog or digital, electrical or optical signals.

[0073] The memory 6,6',56 can also store a computer program comprising instructions which, when the program is executed, lead to the implementation of a method for monitoring the health of the mechanical system 10. This method for monitoring the health of the mechanical system 10 is dedicated to monitoring the health, for example vibration, of the mechanical system 10.

[0074] Furthermore, each vibration sensor 20 and each context sensor 25 can continuously emit a vibration signal and a context signal respectively.

[0075] A vibration signal and a context signal may be signals formed by raw measurements emitted respectively by a vibration sensor and a context sensor or by measurements obtained by more or less complex signal processing carried out by a processing unit, for example integrated into the corresponding sensor, from such raw measurements, for example via standard filtering or sampling, or even the application of transformations.

[0076] The method for monitoring the health of the mechanical system 10 according to the invention comprises two phases 100, 200 as shown in the diagram of [Fig.2].

[0077] During an initial phase 100, a threshold variation model H is defined, the threshold varying as a function of the context parameter(s) and being intended to be compared to a health indicator for monitoring the health of the mechanical system 10. Then, an operational monitoring phase 200 is executed during which monitoring of the mechanical system 1 is carried out using the threshold determined with the model as a function of the context parameter(s). During the initial phase 100, the mechanical system 10 is considered healthy, and therefore without anomaly or fault.

[0078] The initial phase 100 comprises several steps as shown in [Fig.3].

[0079] The initial phase 100 may firstly comprise a preselection step 110 for selecting context parameters taken into account by the method from a multitude of context parameters measured by the context sensor(s) 25. Indeed, a very large number of context parameters, for example more than 1000 context parameters, may be measured on complex systems, and in particular on a vehicle 1 such as an aircraft. In fact, in order to limit the context parameters used by the method according to the invention, only a part of these measured context parameters are taken into account. This selection can be carried out by feedback on the basis of similar mechanical systems or vehicles, the selected context parameters being the most influential parameters for these similar mechanical systems or vehicles. The selected context parameters can be stored in the memory 6,6',56.

[0080] This preselection step 110 is not essential to carrying out the method according to the invention and is therefore optional.

[0081] During a measurement step 120, initial vibration values ​​and successive initial context values ​​originating respectively from the vibration sensor(s) 20 and the context sensor(s) 25 are measured and transmitted to the on-board memory 6, 6' to be stored there and / or are transmitted directly to the calculation unit 51.

[0082] Then, during a determination step 130, several initial health values ​​of each health indicator CI are determined in the usual manner as a function of the initial vibration values ​​measured during the measurement step 120, using the calculator 50. The different initial health values ​​of each health indicator CI correspond respectively to different times of the initial phase 100.

[0083] An initial health value of a health indicator CI is for example equal to an initial value from a single vibration sensor 20 or the result of a law involving initial values ​​respectively from several vibration sensors 20. In the case where the vehicle 1 is an aircraft, these initial health values ​​of health indicators CI are for example determined using a system designated by the acronym HMS corresponding to the English designation “Health Monitoring System”.

[0084] In addition, each initial health value of a CI health indicator is then associated with the initial context value of the one or more context parameters measured substantially simultaneously with the initial vibration value(s) used to determine the initial health value of this CI health indicator. For example, an initial health value of a CI health indicator is associated with an initial context value of three distinct context parameters.

[0085] This determination step 130 can be carried out by the calculation unit 51 of the calculator 50 which is on board the vehicle 1. Each initial value of a health indicator CI can then be stored in the on-board memory 6, 6'. Each initial value of the health indicator(s) CI can thus be calculated during the use of the vehicle 1, substantially in real time, or else when the vehicle 1 is stopped.

[0086] Alternatively, this determination step 130 can be carried out by the calculation unit 55 of the calculator 50 of the station 40, once the vehicle 1 is stopped, and by example placed on the ground in the case of an aircraft. For this purpose, the initial context values ​​and the initial vibration values ​​are transmitted from the memory 6,6' present in the vehicle 1 to the calculation unit 55, by a wired link or a wireless link, in the form of analog or digital, electrical or optical signals, for example, or even using a memory card, during a transmission step 135, once the vehicle 1 is stationary. The initial context values ​​and the initial vibration values ​​can then be stored in the memory 56 of the station 40. Each initial value of the health indicator(s) CI can thus be calculated after use of the vehicle 1, by the calculation unit 55, each initial value of a health indicator Cl being able to be stored in the memory 56.

[0087] The initial phase 100 may then comprise, as an alternative or in addition to the preselection step 110, an analysis step 140 for analyzing a sensitivity of each health indicator to the context parameters and for selecting influential context parameters. This analysis step makes it possible to determine and select, from among the context parameters used, the context parameters that are most influential on the variations of each health indicator CI. This analysis step 140 may be carried out by the calculation unit 51 on board the vehicle 1 or by the calculation unit 55 of the computer 50 of the station 40.

[0088] During a definition step 150, a threshold variation model H is defined for each health indicator CI as a function of the initial health values ​​of this health indicator CI, considered as characterizing a healthy mechanical system 10, and the initial context values ​​of the context parameter(s) previously determined and stored. This definition step 150 requires a large number of initial health values ​​for each health indicator CI so that this threshold variation model H is reliable and robust. This large number of initial health values ​​CI must cover a large part of the area of ​​use of the vehicle 1 in order to subsequently make it possible to determine a reliable and safe threshold throughout the area of ​​use of the vehicle 1.This large number of initial health values ​​CI is obtained during a minimum learning time ta of use of vehicle 1, for example equal to 50 hours, corresponding substantially to the duration of the initial phase 100.

[0089] This definition step 150 can be carried out by the calculation unit 51 of the computer 50 on board the vehicle 1, with the initial values ​​of the CI health indicator(s) and the initial context values ​​being stored in the memory 6,6'. The initial context values ​​and the initial values ​​of the CI health indicator(s) are transmitted from the memory 6,6' to the computer 50, for example by a wired link or a wireless link, in the form of analog or digital, electrical or optical signals, during a transmission sub-step 155 by example during operation of vehicle 1 or once vehicle 1 is stationary. The threshold variation model H can then be stored in the memory 6.6' present in vehicle 1.

[0090] Alternatively, this definition step 150 can be carried out by the calculation unit 55 of the calculator 50 of the station 40, with the initial values ​​of the health indicator(s) C1 and the initial context values ​​stored in the memory 6,6' of the vehicle 1, after the vehicle 1 has stopped. The initial context values ​​and the initial values ​​of the health indicator(s) C1 are in this case transmitted from the memory 6,6' present in the vehicle 1 to the calculation unit 55 of the station 40, for example during a transmission sub-step 155. The threshold variation model H can then be stored in the memory 56 of the station 40.

[0091] Alternatively, this definition step 150 can still be carried out by the calculation unit 55 of the station 40, with the initial values ​​of the health indicator(s) CI and the initial context values ​​stored in the memory 56 of the station 40, after the vehicle 1 has stopped. The threshold variation model H can then be stored in the memory 56 of the station 40.

[0092] The threshold model H is dependent on the context parameters Ç and can be written as H(T,C,r} = r'(O- with r, a probability level of the quantile distribution, corresponding to a requested false alarm rate, y, a level of uncertainty on the estimated threshold H, corresponding to a level of confidence on this requested false alarm rate, and C) ' an inverse cumulative function of a lognormal distribution function of C

[0093] As a reminder, a variable X is said to follow a log-normal distribution if the variable Y = log (X) follows a normal distribution.

[0094] The function / '(£) can be estimated from a method using decision tree regressions on the quantile distribution parameters via an aggregation of decision trees often referred to in English as a "Bagging Decision Tree".

[0095] A quantile Qr is a value that separates a data set according to a proportion defined by a probability t. For a set of context parameters Ç, the function is a probability density that characterizes the quantile distribution p^q (C) ) as a function of the context parameters Ç. The quantile distribution p^q (Ç) ) is defined by the mean p*(Ç) and the variance E*(f). The function thus characterizes the relationship between the quantile distributions and the parameters of context C.

[0096] The function can then be written: / (f) =p(«,(f)) =

[0097] The mean ( ^) and variance E*( f ) can be estimated from the image of the function f( £ ) for a context parameter space Ç of dimension p, where p is also the number of context parameters Ç used. Thus, if two distinct context parameters Ç are used, they form a two-dimensional context parameter space.

[0098] The method according to the invention can use statistical modeling of the distribution of the health indicator CL. Furthermore, it is known that the health indicators CI estimated from the Fourier coefficients of a vibration signal modeled under the hypothesis of cyclostationarity of order one or two, follow, in the case where the mechanical system 10 does not have a defect, a chi-square law. The generalization of this law is a gamma law T with two parameters, a shape parameter a and a scale parameter [3. The statistical distribution of the health indicator CI is thus modeled by a mixture of gamma distributions. Such modeling advantageously makes it possible to consider more complex data distributions and therefore offers greater flexibility. The probability density of such a mixture of gamma distributions for the value x of the health indicator CI can be written: P ( * ) = ( a; an, Pn ) ' with a mixing coefficient ir, and where the distribution ) is a gamma distribution such that: \ ■G,®»,) with FO, the Gamma function, and K, the number of gamma distributions.

[0099] The parameters of the mixture model fi are estimated with an expectation-maximization algorithm known as “Expectation Maximization” while the number of gamma distributions K is estimated from a Bayesian information criterion called BIC for “Bayesian Information Criterion”, with the condition K < Ny where NY is the maximum number of components of the gamma distribution to be estimated.

[0100] Once the mixture model has been estimated, it is possible to estimate the quantile at a probability level r, by defining a cumulative distribution F^t) such that: k PX ( 0 = J 0n ) dx

[0101] Estimating the quantile qT calculated for a probability level r amounts to solving the unconstrained optimization problem according to which g = argmin\g (?) |, with gr ( 0 = 1¾ 0 - ^ |=| k^L^-P (r' )dx ■TI'

[0102] As a reminder, the minimum argument function (argmin) of a function represents the value of the variable for which the value of the function concerned reaches its minimum, the value for which the function is minimized being unique.

[0103] It is first necessary to estimate the parameters of the quantile distribution. First, a parametric modeling of the health indicator CI conditioned on one or more context parameters is carried out by a mixture of Gamma distributions. We then use a method for determining the uncertainties on the parameters of the mixture of Gamma distributions by a WLB method called "Weighted Likelihood Boostrapp". At each sampling according to the WLB method, a quantile value ^r is estimated. This sampling step is carried out P times in order to obtain a sample of quantile value ÿT, called

[0104] Now, the quantity q ~ logN(fl*, 2^) is considered as a random variable modeled by a log Normal distribution of parameters:

[0105] We can then calculate an average: » 1 ( $ 1 'and — t yNh n 'and = 10g | ] Pt “ Nh a variance: = ^(^ + 1)^2^ = 7^^,(^ '

[0106] [Fig.4] illustrates a histogram 41 of the probability density of a health indicator CI, according to the value x of this health indicator CI. The curves 42,43,44 represent the variations of the Gamma components identified by the mixture model according to the value x of this health indicator CL. The histogram 45 corresponds to the quantile sample estimated and modeled by a Log Normal distribution 46.

[0107] Furthermore, to enable the construction of a regression model via decision trees, the method performs a recursive division of the context parameter space in order to divide it into several ranges for which the threshold relating to the CI health indicator is statistically constant, namely that, for a range of values ​​of a given context parameter, two distinct thresholds resulting from a further division of the range of values ​​of the context parameter into two sub-ranges have for example a statistical difference less than a predetermined value. This predetermined value is for example equal to the critical value of the Welch statistic. Such a division criterion makes it possible to reduce the dissimilarity of the quantile distributions on the context parameter space for example.

[0108] The method according to the invention can, for example, make it possible to express the conditional variability of the health indicators CI from the variability of the quantile distributions, the dissimilarity of the quantile distributions on the space of the context parameters being estimated from a similarity statistic or Welch statistic Sj(k) between two quantile distributions characterized by their respective means / C] and / ½ as well as by their respective variances 2^.and £t2 according to the relation: cr • n,,

[0109] The indices j and k respectively designate the index j of the context parameter Çj, j varying from 1 to p, and the iteration indicating the index k of the course on this context parameter Çj, Nb being the dimension of the quantile sample Q . The index k of the course is an iterator indexed on the vector defined by the context parameter Çj. The means of the samples Qt of quantiles and as well as the variances and E-[k) are estimated from the health indicator CI conditioned on the context parameter Çj at the course index k.

[0110] A maximum value of the Welch statistic can be determined over the entire context parameter space and allow the value of the context parameter and the context parameter to be discretized to be determined. A discretization coordinate ik* can then be chosen equal to the maximum value of the Welch statistic Welch following the context parameter Çj and the traversal index k, such that: Q = argmax. k [ Sj ( k ) ] ' with the function argmax(S), giving as result the point(s) for which the function S reaches its maximum value.

[0111] This discretization coordinate rk' makes it possible to separate a range of values ​​of a Z context parameter into two sub-ranges. The recursive division process on the context parameter space can stop when, for example, at least one of the following two conditions is verified, allowing us to conclude that the threshold relating to the health indicator is statistically constant over these two sub-ranges.

[0112] For a first condition, a minimum number of values ​​required to estimate a quantile distribution is reached.

[0113] For a second condition, a hypothesis Ho, according to which two means of quantile distributions are no longer statistically different, cannot be rejected. The hypothesis Ho can be written: — ZZ2. This hypothesis Ho cannot in particular be rejected if the value of the Welch statistic Sj(k ) is less than the critical value estimated from the inverse distribution of the Student law following a level

[0114]

[0115] of probability a. A low value of the parameter a favors the acceptance of the hypothesis Ho. This estimated critical value is for example equal to fyj, with v, the number of degrees of freedom such that: ( / 7-1).(27+2+ )2 • y = _—2— L —_ This number v of degrees of freedom conditions the critical value of the Welch statistic and depends on the variance of the quantile sample as well as the dimension of

[0116]

[0117]

[0118]

[0119]

[0120] The construction and use of a decision tree 300 are based on the notion of Boolean choice allowing to take a direction in the space of context parameters as illustrated in [Fig.5] with an example of a single decision tree 300. The Boolean choices concern only the root node 310 and the decision nodes 320. The root node 310 is the first node of the decision tree 300 from which the decision nodes 320 derive. For a pair of context parameters or a single context parameter, if the condition analyzed at a node 310,320 is true, the left branch is followed. Otherwise, the right branch must be followed. This rule is repeated until arriving at a leaf node 330 containing the parameters of the quantile distributions. Regardless of the number of decision trees 300, the estimations of the root node 310, decision nodes 320 and leaf nodes 330 follow the following procedure from the dataset D={x,Ç}. The example shown in [Fig.5] concerns a single 300 decision tree and a space of two context parameters, Z={Çi,Ç2}. A minimum number Nm of values ​​x of the health indicator CI conditioned on the context parameters Ç as well as a maximum number Ny of gamma distributions for the mixture model must be fixed. At the first iteration, k=l, the data block containing the values ​​x of the health indicator CI as a function of the two context parameters Ç, C2 is split into two data blocks Bi,B2 at the coordinate ^=1 while respecting the condition of a minimum number Nm of values ​​x of the health indicator CI conditioned on the first context parameter as shown in [Fig.6].

[0121] A mixture model of gamma distributions is estimated for each block of data B i,B2, according to the relation / rr y „ RU The means and P \, J jL^n—y*-H'P \ A Pn / / C, as well as the variances Br, and are estimated at a probability level r. The sta Welch's tistic Sj(k) is also calculated.

[0122] Then, the next iteration, k=2, is performed, the context parameter Çi being in-

[0123]

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130]

[0131] cremented (^-2). The process is thus carried out again for the first context parameter Çi by successively splitting the data blocks for each iteration until at least one of the two previously mentioned conditions is verified. Then the process is also performed for the second context parameter Ç2. A maximum value of the Welch statistic can be determined over the entire context parameter space, with the discretization coordinate rC being defined equal to this maximum value 5* of the Welch statistic as represented on [Fig.7]. The values ​​x of the health indicator CI are then separated, according to a two-dimensional discretization of the health indicator CI for each of the two context parameters Çi£2 into two subsets designated for example "data sets" with the value and the context parameter determined previously. Figure 8 represents such a two-dimensional discretization as a function of the context parameters and Ç, according to the process of building a decision tree. The two left and right blocks are separated at the coordinate A decision node 320 of the decision tree 300 is then created, or even a node root 310 if it is the first node of tree 300. The entire process is then repeated for the creation of each decision node 320 of the decision tree 300. For each branch, this process ends at the creation of a leaf node 330, if the value of the Welch statistic is less than the critical value of the Welch statistic (f 1 - yp estimated or if the minimum number of values ​​for the health indicator CI is reached in one of the blocks Bi, B2. The procedure for constructing a single decision tree 300 is thus complete when each branch ends with a leaf node 330. An example using an aggregation of Nt decision trees C, • ■ •, În, is shown in [Fig.9], the procedure for constructing such a set of Nt decision trees being similar. Several decision trees can then be estimated. For each decision tree, a random draw of the data formed by a health indicator CI and a set of context parameters ç j is carried out. Nt corresponds to the number of random draws to produce a set of decision trees T = (^, t2...tNJ. The different decision trees thus constructed can then be used during the operational phase 200 by averaging the parameters of the quantile distributions of each decision tree traversed for a value x of the health indicator CI and the associated context parameter values ​​t.

[0132] Using the decision tree(s) 300, the function / (^) can then be estimated following the creation of a correlation between the quantile distributions as a function of the context parameters Ç. Once the function f( Ç ) has been estimated, it is then possible to predict, for a set of context parameters Ç or a single context parameter Ç, the parameters of the associated quantile distributions.

[0133] For a new set of context parameters Ç or a single context parameter Ç, each decision tree t has a leaf node l(Çbt).

[0134] The predicted quantile distribution on the set of Nt decision trees for a set of context parameters Ç is then written: = <) = LogN ( ^(0 ) '

[0135] A mean and variance of the quantile distribution can be estimated at the leaf node l(Çbt) respecting the condition l <ep( f^)}.

[0136] The average can be written: t)'

[0137] In particular, the value of the average of the quantile associated with the leaf node l(Çbt) of the tree t as a function of the context parameter(s) This is then (t, t) =

[0138] The variance £*. contains the average of the variances of the Nt decision trees and the variance of the estimate of the average over all the generated decision trees such that: ku ) o+3^(^- '))”

[0139] In particular, the value of the variance of the quantile associated with the leaf node l(Çbt) of the tree t as a function of the context parameter(s) Ç is written: <( M ï = ONE f

[0140] In this way, a variant threshold model H(t, Ç, y) as represented in [Fig.10], can be defined for a health indicator x as a function of two context parameters with a probability level r of false alarms equal to 0.9999 and with three different confidence rates, also designated uncertainty levels y, equal respectively to 0.05 (5%), 0.55 (55%) and 0.95 (95%).

[0141] Following the completion of the initial phase 100 and the establishment of the threshold variation model, the operational monitoring phase 200 can be carried out at each use of the vehicle 1 involving the operation of the mechanical system 10, as long as a significant maintenance operation likely to call into question the reliability of the threshold variation model of a mechanical system 10 of the vehicle 1 is not carried out. The operational monitoring phase 200 comprises the following steps.

[0142] During a measurement step 220, successive vibration operational values ​​and context operational values ​​originating respectively from the vibration sensor(s) 20 and the context sensor(s) 25 are measured and transmitted to the memory 6, 6' to be stored there and / or possibly transmitted directly to the computer 50.

[0143] This measurement step 220 is similar to the measurement step 120 of the initial vibration values ​​and the successive initial context values ​​originating respectively from the vibration sensor(s) 20 and the context sensor(s) 25.

[0144] Then, during a determination step 230, several operational health values ​​of each CI health indicator are determined in the usual manner as a function of the vibration operational values, using the calculator 50. This determination step 230 is similar to the determination step 130 of several initial health values ​​of each CI health indicator, and each operational health value of a CI health indicator is associated with one of the context operational values ​​of one or more context parameters.

[0145] As for the determination step 130, the determination step 230 of several operational health values ​​of each health indicator CI can be carried out by the calculation unit 51 on board the vehicle 1, during the use of the vehicle 1 for example, or by the calculation unit 55 of the station 40 after the vehicle 1 has stopped. Each operational value of a health indicator CI can then be stored either in the memory 6,6' present in the vehicle 1, or in the memory 56 of the station 40.

[0146] Then during a determination step 250, a threshold specific to the operational health value of each health indicator CI is determined using the threshold variation model and the associated context operational value(s).

[0147] Figure 11 represents values ​​x of a health indicator CI as a function of time t, as well as curves 31, 32, 33 of variation of the threshold relating to this health indicator CI with respectively an uncertainty level y equal to 0.05, 0.55 and 0.95 and with a probability level r of false alarms equal to 0.9999. For comparison, an algebraically constant threshold 35 estimated for example from an inverse cumulative distribution of a Gaussian distribution with a probability level r=0.9999 is also represented.

[0148] It can thus be seen that the threshold variation curves 31, 32, 33 can be higher or lower than the algebraically constant threshold 35, depending on the time intervals. For example, over a first time interval 37, the algebraically constant threshold 35 is lower than the threshold variation curves 31, 32, 33 and therefore more penalizing in the sense of potentially triggering false alarms.

[0149] Finally, during a triggering step 260, an alert is triggered for indicate to an operator or driver of vehicle 1 a risk of the presence of a fault on the mechanical system 10 if the operational health value of each health indicator CI is greater than the determined threshold.

[0150] In this way, the detection of a fault on the mechanical system 10 can be anticipated in a reliable, efficient and robust manner from the first signs of its appearance, the presence of this fault being able in fact to be detected through the vibrations of this mechanical system 10.

[0151] For example, the use of one of the threshold variation curves 31, 32, 33 shown in FIG. 11 thus advantageously avoids the generation of a false alarm for a value Xi of the health indicator CI. The use of one of the threshold variation curves 31, 32, 33 also avoids the non-detection of a potential fault for a value x2 of the health indicator CI.

[0152] Naturally, the present invention is subject to numerous variations as to its implementation. Although several embodiments have been described, it is understood that it is not conceivable to exhaustively identify all possible modes. It is of course conceivable to replace a means described by an equivalent means without departing from the scope of the present invention and the claims.

Claims

Claims

1. Method for monitoring the health of a mechanical system (10) equipping a vehicle (1), said mechanical system (10) comprising at least one moving member (15) and at least one vibration sensor (20) emitting a vibration signal, said mechanical system (10) or said vehicle (1) comprising at least one context sensor (25) emitting a context signal relating to at least one context parameter, said at least one context parameter being chosen from one or more functional parameters of said mechanical system (10), one or more navigation parameters of said vehicle (1) and one or more atmospheric parameters, said method being composed of an initial phase of defining a threshold variation model as a function of said at least one context parameter followed by a phase of operational monitoring of the mechanical system, said initial phase (100) comprising the following steps: - measurements (120) of successive initial vibration values ​​from said at least one vibration sensor (20) and of successive initial context values ​​from said at least one context sensor (25), - determination (130), using a calculator (5, 55), of several initial health values ​​of at least one health indicator CI relating to said mechanical system (10) as a function of said initial vibration values, each initial health value being associated with one of said initial context values ​​of said at least one context parameter, - definition (150) for each CI health indicator, of a threshold variation model, said threshold variation model being defined as a function of said initial health values ​​of said CI health indicator and said initial context values ​​of said at least one context parameter, by partitioning a domain formed by said initial context values ​​of said at least one context parameter into several ranges of values ​​for which said threshold relating to said CI health indicator is statistically constant over each range of values, each range being associated with several of said initial context values, and by determining values ​​of said threshold with a method using regressions on parameters of quantile distributions conditioned on context parameters and domain decompositions of said at least one context parameter using at least one decision tree, said operational monitoring phase (200) comprising the following steps during operation of the mechanical system (10): - measurements (220) of successive vibration operational values ​​from said at least one vibration sensor (20), and successive context operational values ​​from said at least one context sensor (25), - determination (230), with said calculator (5, 55) of an operational health value of at least one CI health indicator as a function of said vibration operational values, said operational health value of at least one CI health indicator being associated with one of said context operational values ​​relating to one or more context parameters, - determination (250) of said threshold specific to said operational health value of said at least one CI health indicator using said threshold variation model and said at least one associated context operational value,- triggering (260) an alert signaling a risk of the presence of a fault on said mechanical system (10) if said operational health value of said at least one health indicator CI is greater than said determined threshold.,

2. The method of claim 1, wherein said definition (150) of said threshold variation model is performed by aggregating at least two decision trees.

3. A method according to any one of claims 1 to 2, wherein said defining (150) of said threshold variation model is performed with a predetermined minimum false alarm rate and a predetermined confidence rate in said minimum false alarm rate.

4. A method according to any one of claims 1 to 3, wherein said initial health values ​​of said at least one in- CI health indicator is calculated from a decomposition into Fourier coefficients of said successive initial vibratory values ​​modeled with an assumption of cyclostationarity of order one and two of said vibratory signal, which follows a Chi-square law, generalized by a GAMMA distribution.

5. Method according to any one of claims 1 to 4, for which said initial phase (100) comprises an analysis step (140) for analyzing a sensitivity of said at least one health indicator to said context parameters and for selecting influential context parameters.

6. Method according to any one of claims 1 to 4, for which said initial phase (100) comprises a preselection step (110) for selecting context parameters taken into account by said method from a multitude of context parameters measured by said at least one context sensor (25).

7. Method according to any one of claims 1 to 6, for which said threshold relating to said at least one CI health indicator is considered to be statistically constant over a range of contiguous values ​​of a context parameter if, for any division of said range of values ​​into two sub-ranges of contiguous values, said thresholds estimated respectively over said two sub-ranges of values ​​are considered identical according to a statistical test.

8. Computer program comprising instructions which, when said program is executed on a computer, lead to implementing the method according to any one of claims 1 to 7.

9. Monitoring device (9) configured to monitor a mechanical system (10) equipping a vehicle (1), said mechanical system (10) comprising at least one moving member (15), said monitoring device (9) comprising at least one vibration sensor (20) emitting a vibration signal, at least one context sensor (25) emitting at least one context signal relating to at least one context parameter and a computer (50), characterized in that said monitoring device (9) is configured for implementing the method according to any one of claims 1 to 7.

10. Monitoring device (9) according to claim 9, wherein said computer (50) comprises a calculation unit (51) and a memory (6,6') on board said vehicle (1).

11. Monitoring device (9) according to any one of claims 9 to 10, for which said calculator (50) comprises a calculation unit (56) and a memory (56) separate from said vehicle (1).

12. Mechanical system (10) comprising at least one moving member (15), characterized in that said mechanical system (10) comprises a monitoring device (9) according to any one of claims 9 to 11.

13. Aircraft (1) comprising a mechanical system (10) and a monitoring device (9) according to any one of claims 9 to 11.

Images