Method for detecting anomalies in system of aircraft

Abstract

The invention relates to a method for detecting anomalies in a system (110) of an aircraft (100), the method comprising:-obtaining a current sequence of measurements of one or more physical quantities (PT1, PT2) of the system (110) during a time period during which the system (110) is operating; -providing, by the encoder-decoder (126), a current reconstructed sequence on the basis of the current sequence of measurements; and-comparing the reconstructed current sequence with the current sequence of measurements to obtain a current sequence of anomalies. The method further comprises:-calculating a current distribution function of the current anomalous sequence; calculating an area separating the current distribution function from the reference distribution function; and-comparing the area with a predetermined threshold value.

Description

Technical Field

[0001] The invention relates to a method for detecting anomalies in a system of an aircraft, a corresponding computer program and a monitoring device. Background Art

[0002] It is known to use a method for detecting anomalies in aircraft equipment, the method comprising:

[0003] - obtaining a sequence of measured values ​​of one or more physical quantities of the device during a period of time when the device is running, the sequence of measured values ​​being referred to as the current sequence of measured values;

[0004] a neural network of encoder / decoder type provides a reconstructed sequence based on the current sequence of measurements, which is called the reconstructed current sequence; and

[0005] - Compare the reconstructed current sequence with the current measurement value sequence to obtain an anomaly sequence, which is referred to as the current anomaly sequence.

[0006] More specifically, as is well known, the encoder / decoder 126 includes an encoder portion and a decoder portion. The encoder portion is configured to perform compression of information provided as input to the encoder portion to provide intermediate information that is incomplete due to compression. The decoder portion is configured to perform decompression, i.e., reconstruction of the input information from the incomplete intermediate information.

[0007] Thus, it is known that the method is used for aircraft turbines by comparing each anomaly score in an anomaly sequence with a predetermined detection threshold. If the anomaly score is above the detection threshold, an anomaly is detected. The detection threshold is defined by maximizing the F score for a plurality of marked measurement value sequences, some of which are normal and others of which are abnormal.

[0008] However, the inventors have discovered that this known approach does not work well with some aircraft equipment, such as valves used to deliver hot air to the passages of an air inlet lip de-icing system.

[0009] Therefore, it may be desirable to provide a method for detecting anomalies that avoids at least some of the problems and constraints described above. Summary of the invention

[0010] A method for detecting anomalies in an aircraft system is proposed, the method comprising:

[0011] - obtaining a sequence of measured values ​​of one or more physical quantities of the system during a period of time when the system is running, the sequence of measured values ​​being referred to as the current sequence of measured values;

[0012] - a reconstructed sequence is provided by the encoder / decoder based on the current sequence of measured values, which is called the reconstructed current sequence; and

[0013] - Compare the reconstructed current sequence with the current measurement value sequence to obtain an anomaly sequence, which is referred to as the current anomaly sequence.

[0014] Features:

[0015] - Calculate the distribution function of the current anomaly sequence, which is called the current distribution function;

[0016] - calculating the area that separates the current distribution function from the reference distribution function; and

[0017] - Comparing the area to a predetermined threshold.

[0018] This is because the encoder / decoder is generally not able to reconstruct large changes of short duration in the sequence of measured values, which is particularly due to the loss of information caused by the compression of information on the part of the encoder. In this case, the reconstructed current sequence is very different from the current sequence of measured values ​​during this short period of significant change, resulting in high anomalies in the abnormal sequence, exceeding the detection threshold when the system is operating normally. However, the measurements of many aircraft systems can show significant short-term changes, which do not represent an abnormal or slightly abnormal signal. For example, in the valves of the de-icing system, the pressure measurement value usually exceeds the expected nominal value immediately at the beginning of pressure build-up. However, the area separating the current distribution function from the reference distribution function is not very sensitive to short-duration changes (even large short-duration changes), precisely because the changes are of short duration. In this way, the present invention makes it possible to avoid false anomaly detection, which may occur by using a detection threshold.

[0019] The invention also comprises one or more of the following advantageous features, employed in any technically possible combination:

[0020] Advantageously, the system is configured to operate in a plurality of operating configurations, and the method further comprises:

[0021] - selecting, from the encoders / decoders respectively associated with operating configurations, an encoder / decoder associated with the operating configuration in which the device is during the current sequence of measurements.

[0022] Also advantageously, the encoder / decoder is a pre-trained learning system for reconstructing a sequence of measurements acquired during normal operation of the system.

[0023] Also advantageously, the encoder / decoder comprises an encoder neural network and a decoder neural network.

[0024] Also advantageously, the neural networks are two recurrent neural networks with long short-term memory.

[0025] Also advantageously, the system is a system for de-icing an air inlet lip, the de-icing system being configured to collect hot air from the turbine and comprising a channel for delivering the hot air to the air inlet lip and at least one valve on the delivery channel.

[0026] Advantageously still, the de-icing system comprises two valves connected in series on the delivery channel.

[0027] A computer program is also proposed, which can be downloaded from a communication network and / or recorded on a computer-readable medium, characterized in that when the computer program is executed on a computer, the computer program comprises instructions for executing the steps of the method according to the invention.

[0028] Also proposed is a device for monitoring an aircraft system, the device comprising:

[0029] - means for obtaining a sequence of measured values ​​of one or more physical quantities of the device during a period of time when the device is in operation, the sequence of measured values ​​being referred to as the current sequence of measured values;

[0030] - means for providing a reconstructed sequence from a current sequence of measured values ​​using an encoder / decoder, the reconstructed sequence being called reconstructed current sequence; and

[0031] - a module for comparing the reconstructed current sequence with the current sequence of measured values ​​to obtain an anomaly sequence, which is called the current anomaly sequence;

[0032] Features:

[0033] - a module for calculating the distribution function of the current anomaly sequence, which distribution function is called the current distribution function;

[0034] - means for calculating the area separating the current distribution function from the reference distribution function; and

[0035] - means for comparing the area to a predetermined threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The invention will be better understood from the following description, given by way of example only and with reference to the accompanying drawings, in which:

[0037] - Figure 1 is a schematic functional view of an example of an aircraft embodying the present invention,

[0038] - Figure 2 is Figure 1A functional diagram of an example of an encoder / decoder used in an aircraft is shown.

[0039] - Figure 3 is a functional view of an example of an apparatus for monitoring a de-icing device of an air inlet lip of an aircraft,

[0040] - Figure 4 is a block diagram of an example of a method for monitoring de-icing equipment,

[0041] - Figure 5 is a block diagram of a method for configuring a monitoring device,

[0042] - Figure 6 is a graph showing a plurality of distribution functions calculated by a monitoring device and a reference distribution function,

[0043] - Figure 7 is a graph showing an example of a current measurement value sequence in case of normal behavior and a reconstructed current sequence and an associated current abnormal sequence;

[0044] - Figure 8 is a graph showing an example of a current measurement value sequence and a reconstructed current sequence and an associated current anomaly sequence in case of abnormal behavior;

[0045] - Figure 8 It is shown separately Figure 7 and Figure 8 Plot of the reference distribution function and the two distribution functions. DETAILED DESCRIPTION

[0046] Reference Figure 1 , an example of an aircraft 1OO in which the present invention is implemented will now be described.

[0047] The aircraft 100 firstly includes a turbine 102 .

[0048] The aircraft 100 further comprises a nacelle 104 surrounding the turbine 102. In particular, the nacelle 104 has an air inlet lip 106 defining an air inlet 108 for the turbine 102.

[0049] The aircraft 100 also includes a system 110 for deicing the air inlet lip 106. The deicing system 110 is particularly configured to take hot air from the turbine 102 and includes a channel 112 for delivering the hot air to the air inlet lip 106. The deicing system 110 includes, for example, a first valve 114 and a second valve 116 connected in series with each other on the delivery channel 112. This redundancy ensures the correct operation of the deicing system 110 in the event of a failure of one of the two valves 114, 116. Each valve 114, 116 is configured to be selectively in an open state and a closed state, in which the valve 114, 116 is configured to regulate the downstream pressure of the hot air (at the outlet of the valve in question) to a predetermined regulated pressure, and in a closed state, the valve 114, 116 is configured to prevent the hot air from circulating.

[0050] The aircraft 100 also includes a device 118 for controlling the valves 114 , 116 .

[0051] Thus, the de-icing system 110 is configured to operate in a plurality of operating configurations, depending on the state in which the control device 118 causes the valves 114, 116 to be placed. In a first operating configuration, referred to as "open," both valves 114, 116 are open. In a second operating configuration, referred to as "closed," the valve 114 is closed and the second valve 116 is open. In a third operating configuration, referred to as "autotest," the valves 114, 116 are tested at the beginning of a flight of the aircraft 100 when the pressure of the hot air upstream of the valves 114, 116 is low.

[0052] The aircraft 100 further comprises a first sensor 120 for a pressure PT1 downstream of the first valve 114 and a second sensor 122 for a pressure PT2 downstream of the second valve 116 .

[0053] The aircraft 100 also comprises a device 124 for monitoring the valves 114 , 116 on the basis of the downstream pressures PT1 , PT2 .

[0054] In particular, the monitoring device 124 comprises at least one neural network 126 of the encoder / decoder type. Preferably, an encoder / decoder 126 is provided for each configuration of valves 114,116.

[0055] Each encoder / decoder 126 includes an encoder portion and a decoder portion. The encoder portion is configured to perform compression of information provided as input to the encoder portion to provide intermediate information that is incomplete due to compression. The decoder portion is configured to perform decompression, that is, to reconstruct the input information from the incomplete intermediate information. In the described example, the input information consists of a time series of measured values ​​of the downstream pressures PT1 and PT2. The decoder portion is configured to provide a reconstructed time series. The encoder / decoder 126 is configured to operate using a sequence of measured values ​​acquired during the assumed normal operation of the valves 114 and 116 (i.e., in the absence of anomalies). This means that the encoder / decoder is configured so that when a sequence of measured values ​​is acquired during the assumed normal operation of the valves 114 and 116, the reconstructed sequence is as close as possible to the sequence of measured values.

[0056] The encoder / decoder 126 is, for example, a learning system and is in this case pre-trained on the basis of a plurality of measurement value sequences (referred to as measurement value training sequences), all of which are acquired during an assumed normal operation of the valves 114 , 116 .

[0057] To obtain these actuation sequences, for example, measurement value sequences obtained during the operation of the valves 114, 116 can be used, for which the failure times of at least one of the valves 114, 116 or the maintenance times of the valves 114, 116 are known. Therefore, it can be assumed that the valves 114, 116 were operating normally and fully before these times. In this way, the measurement values ​​before these moments of a predetermined duration (e.g., one month) can be considered to be obtained during the normal operation of the valves 114, 116 and are used to obtain the training sequences.

[0058] Will refer to Figure 2 An example of the encoder / decoder 126 is described in more detail.

[0059] The monitoring device 124 comprises, for example, a computer system including a data processing unit 128 (e.g., a microprocessor) and a main memory 130 (e.g., a random access memory) accessible by the processing unit 128. The computer system also comprises, for example, a network interface and / or a computer-readable medium, such as a local medium (e.g., a local hard disk 132), or a remote medium (e.g., a remote hard disk accessible via a network interface via a communication network), or a removable medium (e.g., a Universal Serial Bus (USB) key, or a Compact Disc (CD) or a Digital Versatile Disc (DVD)) that can be read by an appropriate computer system driver (e.g., a USB port or a CD and / or DVD disk drive). A computer program 134 containing instructions for the processing unit 128 is stored on the medium 132 and / or can be downloaded via a network interface. The computer program 134 is intended, for example, to be loaded into the main memory 130 so that the processing unit 128 can execute the instructions of the computer program. In order to more easily describe the computer program 132, the instructions will be described as being organized into software modules in the following. However, this presentation does not predetermine the form of the computer program 132 , which may be in any form. In particular, the encoder / decoder 126 may be implemented in the form of a software module of the computer program 132 .

[0060] Alternatively, all or some of these modules may be implemented in the form of hardware modules, that is, in the form of electronic circuits (eg, micro-wires), without including a computer program.

[0061] Reference Figure 2 , an example of the encoder / decoder 126 will now be described.

[0062] The encoder / decoder 126 is configured to receive a measurement sequence combining P measurements x(1) ... x(p) ... x(P) obtained at corresponding measurement times marked by an index p ranging from 1 to P. The measurements x(p) may be one-dimensional (a single physical quantity measured at each measurement time) or multi-dimensional (multiple physical quantities measured at each measurement time). In the example described, each measurement x(p) combines the measurement of the downstream pressure PT1 and the measurement of the downstream pressure PT2 at the measurement time in question. Thus, the measurement sequence X may be written as:

[0063] [Mathematical formula 1]

[0064]

[0065] The encoder / decoder 126 is configured to provide a reconstructed sequence X' from the measurement value sequence X. Thus, in the described example, the reconstructed sequence X' may be written as:

[0066] [Mathematical formula 2]

[0067]

[0068] The encoder / decoder 126 includes, for example, an encoder neural network 202 and a decoder neural network 204 that implement an encoder portion and a decoder portion, respectively.

[0069] The neural networks 202 and 204 are, for example, two recurrent neural networks for long short-term memory (LSTM).

[0070] In this case, the encoder neural network 202 may include P consecutive units 202-1...202-p...202-P, each of the P consecutive units being configured to calculate an internal value h from a corresponding one of the measured values ​​x(1)...x(p)...x(P), an internal value of the previous unit, and a parameter of the unit under consideration. E (1)……h E (p)……h E (P). Each unit 202-p has an architecture such as that described in https: / / fr.wikipedia.org / wiki / Réseau_de_neurones_récurrents, so that each internal value h E (p) is obtained, for example, by the following equation:

[0071] [Mathematical formula 3]

[0072]

[0073] in, represents the Hadamard matrix product, σ g represents the Sigmoid function, σ h represents the hyperbolic tangent function, fE(t) represents the activation state of the encoder forget gate, i E (t) represents the activation state of the encoder input gate, o E (t) represents the activation state of the encoder output gate, Indicates the activation status of the encoder input unit, and c E (t) represents the internal state of the encoder unit.

[0074] The decoder neural network 204 similarly includes P consecutive units 204-1...204-p...204-P. Each unit is configured by parameters and by the internal value h of the next unit. D '(p+1) and the reconstructed value x'(p+1) (except for the last unit 204-P) to calculate the internal value h D (1)……h D (p)……h D (P) and reconstructed values ​​x'(1) ... x'(p) ... x'(p). Each unit 204-p has an architecture such as that described in https: / / fr.wikipedia.org / wiki / Réseau_de_neurones_récurrents, so that each internal value h E (p) is obtained, for example, by the following equation:

[0075] [Mathematical formula 3]

[0076]

[0077] Where ° represents the Hadamard matrix product, σ g represents the Sigmoid function, σ h represents the hyperbolic tangent function, fx(t) represents the activation state of the decoder forget gate, i D (t) represents the activation state of the decoder input gate, o D (t) represents the activation state of the decoder output gate, Indicates the activation state of the decoder input unit, c D (t) represents the internal state of the decoder unit, and W, U and b represent the weight matrix and bias parameters of the decoder.

[0078] The internal value h of the last unit 204-P of the decoder neural network 204 D (P) is taken to be equal to the internal value h of the last unit 202-P of the encoder neural network 202 E (P). Therefore, the internal value h E (P) = h D (P) forms the compressed intermediate information of the encoder / decoder 126.

[0079] Reference Figure 3 , an example of the computer program 132 will now be described.

[0080] The computer program 132 comprises a module 302 configured to obtain a sequence of measured values ​​of one or more physical quantities of the de-icing system 110 (referred to as the current sequence of measured values ​​X) during a period of time in which the de-icing system 110 is operated in one of the operating configurations of the de-icing system. In this way, the value of the or each physical quantity is obtained for each of a plurality of consecutive measurement times p. More specifically, in the example shown, the physical quantity is the downstream pressure PT1, PT2 provided by the sensors 120, 122.

[0081] Computer program 132 also includes module 304 configured to select encoder / decoder 126 associated with the operating configuration that de-icing system 110 was in during measurement time p (referred to as the current operating configuration). Module 304 is configured to receive an indication of the operating configuration from control device 118, for example.

[0082] The computer program 132 further comprises a module 306 configured to use the encoder / decoder 126 selected by the module 304 to obtain a reconstructed sequence from the current sequence of measured values ​​X (referred to as reconstructed current sequence X′).

[0083] The computer program 132 further comprises a module 308 configured to compare the current reconstructed sequence X′ with the current measured value sequence X to obtain an anomaly sequence (referred to as current anomaly sequence A) representing anomalies in the reconstructed current sequence X′ relative to the current measured value sequence X.

[0084] For example, the module 308 is configured to first calculate, at each measurement time p, an error e(p) between a measurement value x(p) of the current measurement value sequence X at the measurement time p and a reconstructed value x'(p) of the reconstructed current sequence X' at the measurement time p, for example, by using the following mathematical formula to calculate the error:

[0085] [Mathematical formula 3]

[0086] e(p)=|x(p)-x′(p)|

[0087] Among them, |......| is the absolute value function.

[0088] The module 308 is also configured to calculate an anomaly a(p) (also referred to as an "anomaly score") at each measurement time p, for example from the error e(p) at the measurement time p. The anomalies a(p) calculated in this way form an anomaly sequence A.

[0089] For example, it is assumed that the error e(p) follows a predetermined probability distribution characterized by one or more parameters. These parameters include, for example, a mean value μ and / or a standard deviation ∑. The probability distribution is, for example, a normal distribution. In this way, the anomaly a(p) can be defined using one or more parameters of the probability law to be far from what is expected by the probability law.

[0090] For example, particularly in the case of a normal distribution, the anomaly a(p) can be calculated from the mean μ and standard deviation Σ of the probability distribution by the following mathematical formula:

[0091] [Formula 4]

[0092] a(p)=(e(p)-μ) T ∑- 1 (e(p)-μ)

[0093] The computer program 132 further comprises a module 310 configured to calculate a cumulative distribution function F of the anomalies a(p). A cumulative distribution function is also referred to as a distribution function.

[0094] To this end, the module 310 is preferably configured to calculate a predetermined number N (for example one thousand, in order to have a sufficiently accurate approximation) of equally spaced outlier values ​​a1 . . . a between zero and the maximum outlier a(p) n …a N Then, module 310 is configured to calculate an abnormal value a less than or equal to a for multiple values ​​of a, for example, by the following formula: n Quantity:

[0095] [Formula 5]

[0096]

[0097] Among them, when a n When it is in the interval [-∞, a], 1 [-∞，a] (a n ) is one, otherwise it is zero.

[0098] The computer program 132 also includes a module 312 configured to calculate the difference between the cumulative distribution function F and the reference cumulative distribution function F that would be expected in the absence of the anomaly. s The area of ​​separation will be referred to later. Figure 5 Describe how to obtain the reference cumulative distribution function F s Preferably, a reference distribution function F is provided for each running configuration. s , and compared with the distribution function associated with the current running configuration.

[0099] Preferably, module 312 is further configured to normalize the calculated area A. For example, the normalization includes dividing the area A by the reference distribution function F s The area under the x-axis (i.e., the area between the x-axis and the reference distribution function F s the area between).

[0100] The computer program 132 also includes a module 314 configured to detect anomalies based on the comparison.

[0101] More specifically, the calculated area is compared to a predetermined threshold, and if the area is greater than the threshold, an abnormality is detected. If the area is below the threshold, no abnormality is detected.

[0102] Multiple thresholds may also be provided to distinguish between low-level anomalies (area greater than a minimum threshold) and high-level anomalies (area greater than a maximum threshold).

[0103] For example, the detected anomaly may be examined by an engineer, who may then send a report to an airline using the aircraft to warn the airline, which may then perform maintenance operations on valves 114 and 116 to repair or replace the valves.

[0104] Reference Figure 4 , an example of a method 400 for monitoring the de-icing system 110 will now be described.

[0105] In step 402 , module 302 obtains a current sequence of measurements X acquired while de-icing system 110 is operating in one of the de-icing system's operating configurations, referred to as the current operating configuration.

[0106] In step 404, module 304 selects the encoder / decoder 126 associated with the current operating configuration.

[0107] In step 406 , the module 306 uses the selected encoder / decoder 126 to obtain from the current sequence of measured values ​​X a reconstructed current sequence X′.

[0108] In step 408 , the module 308 compares the reconstructed current sequence X′ with the current measurement value sequence X to obtain the current abnormal sequence A.

[0109] In step 410 , the module 310 calculates the cumulative distribution function F of the current abnormal sequence A.

[0110] In step 412, module 312 calculates the cumulative distribution function F and the reference cumulative distribution function F s Separated area A.

[0111] In step 414 , the module 314 detects anomalies by comparing the calculated area to a predetermined threshold.

[0112] Reference Figure 5 , a method 500 for configuring the monitoring device 124 will now be described.

[0113] For example, the following steps are performed for each operating configuration of de-icing system 110 .

[0114] During step 502, a plurality of measurement sequences (referred to as training measurement sequences X) are obtained while de-icing system 110 is operating normally. * ). For example, each training sequence X * Contains P measurements.

[0115] In step 504, the encoder / decoder 126 performs the following operations according to the training sequence X. * Training is performed to provide training sequences X * Reconstruct the sequence X' as close as possible * .

[0116] During step 506, a plurality of measurement sequences (referred to as initialization measurement sequence X°) are obtained while de-icing system 110 is operating normally. Preferably, initialization sequence X° is different from training sequence X°. * The initialization sequences X° each comprise, for example, P measurements.

[0117] In step 508, the slave encoder / decoder 126 is used to provide a reconstructed sequence X°' from the initialization sequence X°.

[0118] During a step 510 , for each initialization sequence X°, an error sequence E° between the initialization sequence X° and the corresponding reconstructed sequence X°′ is calculated.

[0119] For each measurement time instant p, each error sequence E° includes the error e°(p) between the measured value X°(p) of the initialization sequence X° at the measurement time instant p and the reconstructed value X°'(p) of the reconstructed sequence X°' at the measurement time instant p, obtained, for example, according to the above equation [Formula 3].

[0120] In step 512, one or more parameters of the probability law that the error sequences E° should follow are calculated based on the error sequences E°. For example, it can be assumed that the error sequences E° follow a normal distribution characterized by a mean value μ and a standard deviation ∑. Therefore, the mean value μ and / or the standard deviation ∑ can be calculated based on the error sequences E°. This can be done using Maximum Likelihood Estimation (MLE).

[0121] In step 514, for each error sequence E°, an abnormal sequence A° is calculated from the mean value μ and the standard deviation Σ according to the above formula [Formula 4], for example.

[0122] In step 516, for each abnormal sequence A°, the distribution function (referred to as the healthy distribution function F°) of the considered abnormal sequence A° is calculated, for example, according to the above formula [Mathematical Formula 5].

[0123] In step 518, the reference distribution function F s Calculated according to the distribution function F° (eg by averaging the distribution function F°).

[0124] Reference Figure 6 , showing the reference distribution function F s An example of and multiple examples of distribution functions F1, F2, F3.

[0125] The distribution function F1 corresponds to the normal operation of the valves 114, 116 and is therefore very close to the reference distribution function F s , so that the distribution function F1 and the reference distribution function F s The separated area is very small.

[0126] On the other hand, the distribution function F3 corresponds to the abnormal operation of the valves 114, 116 and is therefore different from the reference distribution function F s Very different, so that the distribution function F3 and the reference distribution function F s The separated area is very large.

[0127] The distribution function F2 corresponds to normal operation of the valves 114, 116 but with small errors (i.e. for small values ​​of a). Due to the environment of the aircraft (which can sometimes be very noisy), such small errors may occur inconspicuously but frequently. By considering the area between the curves, the operation of the valves 114, 116 can be correctly classified as normal. For example, this may not be the case if the criterion is the classical criterion of the maximum vertical difference between the curves. In the latter case, since the error is small, the larger vertical distance (given by Figure 6 The double arrows in the figure indicate that the energy consumption (in ) will be very high.

[0128] Reference Figure 7, shows an example of a current measurement value sequence X for normal behavior of a valve 114 or 116 (time on the x-axis and measurement values ​​on the ordinate), as well as a reconstructed current sequence X' and an associated current abnormal sequence A. It can be seen that the current measurement value sequence X includes an initial peak value PX, which corresponds to, for example, a pressure rise after the valve 114 or 116 is opened. However, since the encoder / decoder 126 cannot reconstruct the peak value PX, the reconstructed current sequence X' does not include the corresponding peak value, so that the current abnormal sequence A includes a peak value PA at the time of the peak value PX.

[0129] In the case where a low detection threshold is used (e.g. 110), all time series that exhibit the behavior shown will be considered abnormal, even though the system (valve 114 or 116) is operating normally. In contrast, if a high threshold is chosen (e.g. 160), abnormal behavior will not be allowed to be detected. To illustrate this point, Figure 8 A current measurement value sequence X~ (time on the X-axis and measurement values ​​on the ordinate) with an initial peak value PX~ for an abnormal behavior of a valve 114 or 116, as well as a reconstructed current sequence X'~ and an associated current abnormal sequence A~ are shown. Because the behavior is abnormal, the reconstructed current sequence X'~ deviates slightly from the current measurement value sequence X'~. This slight deviation is sometimes referred to as a "small drift". Therefore, after the initial peak value PA~, the anomalies in the abnormal sequence are slightly greater than zero. In order to detect a slight drift, a relatively low threshold must be used, which will result in a false detection of anomalies due to the initial peak value PA~. In contrast, a high threshold value above the initial peak value PA~ will not allow the detection of a small drift.

[0130] Therefore, in case the detection threshold is directly applied to the current abnormal sequence, due to the presence of the initial peak, either abnormal behavior or healthy behavior will be inferred from almost all the current measurement value sequences, depending on the selected detection threshold. Both situations are unsatisfactory.

[0131] Reference Fig. 9 , shows the reference distribution function FS, where Figure 7 The distribution function F(X) of the current measurement sequence X in Figure 8 The distribution function F(X~) of the current measurement value sequence X~ in . It can be seen that it is similar to the reference distribution function F s In contrast, the presence of the initial peak PX does not change the distribution function F(X) much, resulting in a very small difference in area. On the other hand, the distribution function F(X') is similar to the reference distribution function F s are very different, resulting in large differences in area. In this way, it is possible to Figure 7 The normal behavior and Figure 8This will make it impossible to use the detection threshold on the current abnormal sequence A, A~.

[0132] In summary, it should be noted that the present invention is not limited to the embodiments described above. In fact, it should be understood by those skilled in the art that various modifications can be made to the embodiments described above based on the teachings just disclosed.

[0133] In the above detailed description of the present invention, the terms used should not be construed to limit the present invention to the embodiments disclosed in this specification, but should be construed to include all equivalent means that can be anticipated by those skilled in the art by applying their general knowledge to the implementation methods of the just disclosed teachings.

Claims

1. A method (400) for detecting an anomaly in a system (110) of an aircraft (100), the method comprising: - obtaining (402) a sequence of measured values ​​of one or more physical quantities (PT1, PT2) of the system (110) during a period of operation of the system (110), the sequence of measured values ​​being referred to as the current sequence of measured values ​​(X); - providing (406) a reconstructed sequence based on the current sequence of measured values ​​(X) by the encoder / decoder (126), said reconstructed sequence being referred to as reconstructed current sequence (X'); as well as - comparing (408) the reconstructed current sequence (X') with the current measurement value sequence (X) to obtain an anomaly sequence, which is referred to as a current anomaly sequence (A); Its characteristics are: - calculating (410) a distribution function of said current anomaly sequence (A), said distribution function being referred to as current distribution function (F); - Calculate (412) the current distribution function (F) and the reference distribution function (F s ) area separated; and - Comparing the area with a predetermined threshold (414).

2. The method (400) of claim 1, wherein: The system (110) is configured to operate in a plurality of operating configurations, and the method further comprises: - selecting (404) from the encoders / decoders (126) respectively associated with the operating configurations an encoder / decoder associated with the operating configuration in which the device is during the current sequence of measurement values ​​(X).

3. The method (400) according to claim 1 or 2, wherein: The encoder / decoder (126) is a pre-trained learning system for reconstructing a sequence of measurements acquired during normal operation of the system (110).

4. The method (400) of claim 3, wherein: The encoder / decoder (126) includes an encoder neural network (202) and a decoder neural network (204).

5. The method (400) of claim 4, wherein: The neural networks (202, 204) are two recurrent neural networks with long short-term memory.

6. The method (400) according to any one of claims 1 to 5, wherein: The system (110) is a system for deicing an air inlet lip (106), the deicing system being configured to collect hot air from a turbine (102) and comprising a passage (112) for delivering the hot air to the air inlet lip (106) and at least one valve (114, 116) on the delivery passage (112).

7. The method (400) of claim 6, wherein: The de-icing system (110) includes two valves (114, 116) connected in series on the delivery channel (112).

8. A computer program (132) capable of being downloaded from a communication network and / or recorded on a computer-readable medium, characterized in that When the computer program (132) is executed on a computer, the computer program comprises instructions for performing the steps of the method (400) according to any one of claims 1 to 7.

9. A device (124) for monitoring a system (110) of an aircraft (100), comprising: - a module (302) for obtaining a sequence of measured values ​​of one or more physical quantities (PT1, PT2) of the device during a period of operation of the device, said sequence of measured values ​​being referred to as the current sequence of measured values ​​(X); - a module (306) for providing a reconstructed sequence from said current sequence of measured values ​​(X) using an encoder / decoder (126), said reconstructed sequence being called reconstructed current sequence (X'); as well as - a module (308) for comparing the reconstructed current sequence (X') with the current sequence of measured values ​​(X) to obtain an anomaly sequence, said anomaly sequence being referred to as current anomaly sequence (A); Its characteristics are: - a module (310) for calculating a distribution function of said current anomaly sequence (A), said distribution function being called current distribution function (F); - used to calculate the current distribution function (F) and the reference distribution function (F s ) separate area modules (312); and - means (314) for comparing said area with a predetermined threshold.

Images