Method and device for detecting malfunctions in an aircraft

Abstract

Method for detecting malfunctions of an aircraft component associated with at least one status indicator, said method comprising, for at least one acquired time sequence (Seq1, Seq2) comprising acquired values ​​of said at least one indicator, the steps of: - determining (E20) an approximate time sequence (Seq1*, Seq2*) of approximate values ​​of said acquired values, by an approximation module (AE1, AE2); - determining (E30) a score (A1, A2) of overall abnormality for said acquired time sequence (Seq1, Seq2) from differences between said acquired (Seq1, Seq2) and approximated values; and - detection (E40) of a malfunction of said based on a comparison of said overall abnormality score with a first threshold, said approximation module being configured to minimize approximation errors between time sequences of reference values ​​and time sequences of approximated values.Figure for the abridged version: Fig 1.

Description

Title of the invention: Method and device for detecting an aircraft operating anomaly Prior art

[0001] The invention lies in the general field of methods for monitoring a system.

[0002] The invention is more particularly situated in the context of detecting an operating anomaly of one or more components of an aircraft. The invention finds a particularly advantageous, although in no way limiting, application in the case of an aircraft comprising turbomachine type engines.

[0003] An aircraft has components whose operation is critical to the safety of the aircraft, and whose monitoring is necessary. For example, the oil circulation system is vital for the proper functioning of an engine since it allows its lubrication and thermal regulation. Failures of such a system can lead to engine shutdowns in flight.

[0004] In order to prevent such failures, methods of the prior art make it possible to detect operating anomalies. An anomaly is an abnormal (operating) state of one or more components which may result in a reduction in capacity or the loss of capacity of these components to perform their functions.

[0005] Methods of the prior art make it possible to detect pre-identified anomalies from snapshots of measurements (snapshot type data) taken only during certain flight phases.

[0006] However, operating anomalies not covered by existing algorithms may occur without being detected. Statement of the invention

[0007] The present invention aims to remedy all or part of the drawbacks of the prior art, in particular those set out above, by proposing a solution making it possible to effectively monitor a component of an aircraft.

[0008] To this end, according to a first aspect, there is proposed a method for detecting an operating anomaly with respect to reference values, of at least one component of an aircraft associated with at least one status indicator (the component is associated with the status indicator), said method comprising, for at least one acquired time sequence (for example, the method comprises the acquisition of the acquired time sequence) comprising successive acquired values ​​of said at least an indicator at different times during operation of said at least one component, steps of: - determination of an approximate time sequence comprising approximate values ​​of said values ​​acquired at said times, the approximation being carried out by an approximation module; - determining an overall abnormality score for said acquired time sequence from differences between said acquired values ​​and said approximated values; and - detecting an operating anomaly of said component based on a comparison of the overall abnormality score with a first threshold, said approximation module being previously configured to minimize approximation errors between time sequences of reference values ​​and time sequences of values ​​approximated from said reference values, said reference values ​​being representative of normal operations, so that approximation errors between said acquired values ​​and said approximated values ​​are greater if the acquired values ​​are representative of abnormal operation.

[0009] Correlatively, according to a second aspect of the invention, there is proposed a device for detecting an operating anomaly with respect to reference values, of at least one component of an aircraft associated with at least one state indicator, said device comprising, for at least one acquired time sequence comprising successive acquired values ​​of said at least one indicator at different times during operation of said at least one component: - a module for determining an approximate time sequence comprising approximate values ​​of said values ​​acquired at said times, the approximation being carried out by an approximation module; - a module for determining an overall abnormality score for said acquired time sequence from differences between said acquired values ​​and said approximate values; and - a module for detecting an operating anomaly of said component based on a comparison of said overall abnormality score with a first threshold, said approximation module being previously configured to minimize approximation errors between time sequences of reference values ​​and time sequences of values ​​approximated from said reference values, said reference values ​​being representative of normal operations, so that approximation errors between said acquired values ​​and said approximated values ​​are greater if the acquired values ​​are representative of abnormal operation.

[0010] Generally speaking, the invention allows the monitoring of a component of an aircraft. In particular, the invention allows the detection of an operating anomaly of a component. Component is understood to mean an isolated element (for example a sensor) or a functional set of elements (for example an oil circulation system for lubricating a turbine and regulating its temperature, an electronic aircraft control system, a fuel circulation system, or a turbine). The method can be applied to several components.

[0011] The method can for example be implemented “on the ground” by analyzing the state indicators measured during the different phases of a flight, including in particular the taxi, takeoff, flight and landing phases.

[0012] This method thus makes it possible to detect abnormal operation of a component which occurred during a flight of the aircraft, and to prevent potential malfunctions during future flights.

[0013] This method is based on status indicators associated with the component. By status indicator associated with the component, we mean a quantity dependent on the operation of the component. For example, if this component is an oil circulation system of the aircraft engine, status indicators may be the oil temperature, the oil pressure, or the rotational speed of a turbine of the engine, because this speed depends on the lubrication of the turbine by the oil system.

[0014] Values ​​taken by these indicators over time form a time sequence. Thus, an acquired time sequence means values ​​of state indicators measured successively over the flight time by specific sensors and recorded, or values ​​calculated from such measured values. Each value is associated with an instant, for example the instant corresponding to the measurement of this value. For example, a time sequence may contain values ​​of the oil temperature and pressure measured every second during the flight of the aircraft. In this example, each element of this sequence corresponds to a vector comprising a temperature value and a pressure value measured at a given instant.

[0015] Thus, a time sequence acquired for a component is representative of the operation of the component over a range of flight times.

[0016] An approximated time sequence is a reproduction by a so-called approximation module of a time sequence from partial information on the time sequence. The date associated with each element of the approximated sequence corresponds to the date of an element of the acquired sequence.

[0017] For example, an approximation module performs a regression, from values ​​of a sequence corresponding to a time range, of values ​​of the acquired sequence corresponding to a date later than this time range.

[0018] In another example, the approximation module is an auto-encoder which performs a projection of the acquired sequence into a mathematical space, then produces an approximation of this same acquired sequence from this projection.

[0019] An approximate sequence may have differences with the acquired time sequence.

[0020] In the present application, an approximation error is a quantity representative of the differences between the acquired values ​​and the approximated values ​​corresponding respectively to the same dates. For example, such an error is the sum of the differences in absolute value between the elements of the approximated sequence and the elements of the acquired sequence. In another example, such an error is the sum of the square of the differences.

[0021] The approximation module, which determines the approximated time sequence, is previously (i.e. before implementing the detection method) configured to reproduce reference sequences with little approximation error. These reference sequences comprise state indicator values ​​of the same type as the state indicators mentioned above.

[0022] These reference sequences comprise state indicator values ​​acquired during a plurality of aircraft flights.

[0023] The use of the approximation module thus configured advantageously makes it possible to detect abnormal operation of the component relative to the operations of which the reference time sequences are representative. Indeed, the approximation module is configured to provide better approximations of a plurality of reference time sequences than for a sequence representative of abnormal operation of the component. Thus, the approximation module makes it possible to discriminate between abnormal operations and therefore to recognize an anomaly.

[0024] In one embodiment, the flights from which the reference sequences were acquired are flights during which no malfunction occurred and which were followed by a significant number of flights without malfunction. By malfunction, here is meant an event causing a loss of functionality of a component or which may affect the safety of the aircraft (for example: an engine shutdown in flight, an oil leak, an overheating of the oil).

[0025] In one embodiment of the method, a said approximation module is obtained by a machine learning algorithm from said reference time sequences.

[0026] Advantageously, this embodiment makes it possible to detect abnormal operation based solely on past observations (the reference sequences). This embodiment thus has the advantage of not requiring a priori knowledge linked to the physics and / or parameters of the aircraft. In particular, the use of a machine learning algorithm does not require the establishment of a physical model of the aircraft or any of its components.

[0027] According to one embodiment, a said approximation module is an artificial neural network. By "artificial neural network", reference is made here to a parameterized function comprising one or more layers of artificial neurons connected to each other.

[0028] An artificial neural network, for example a convolutional neural network, accurately learns the normal operation of the component from reference sequences. In particular, a neural network can integrate the information contained in a large volume of data (i.e. a large number of reference sequences). In the present case, such information is, for example, particular temporal patterns of variations in the values ​​of state indicators. A neural network can thus learn a wide variety of operations and thus contribute to improving the reliability of anomaly detection.

[0029] The artificial neural network can be implemented on a computer-type device.

[0030] According to one embodiment, a said approximation module is an artificial neural network comprising an encoder and a decoder, - a said encoder determining a compressed sequence from a said time sequence provided as input to the neural network, said compressed time sequence being smaller than said time sequence provided as input, - a said decoder performing a reconstruction from the compressed sequence to obtain at the output of the neural network a said approximate time sequence.

[0031] Such a neural network is also called an auto-encoder network. This is trained to reconstruct reference sequences. By compressed sequence, we mean here a projection of a sequence into a mathematical space, such that the projection is of reduced size compared to the sequence provided as input. For the auto-encoder to perform a good approximation (i.e.with a low approximation error) of a sequence from a compression of this sequence, it is necessary that this compression includes sufficient information characterizing this sequence. Thus, an auto-encoder configured in accordance with the invention extracts from a sequence the information characteristic of normal operations. If the sequence at the input of the auto-encoder corresponds to an abnormal operation, then its compression will not include all the information necessary to make a good approximation.

[0032] Furthermore, an autoencoder network is capable of recognizing complex temporal features that are difficult to detect with an analytical model, and can therefore distinguish more precisely between abnormal and normal operation.

[0033] When learning to reproduce reference sequences, an autoencoder learns to recognize recurring patterns and tends to ignore the rarest patterns. Thus, the use of an autoencoder is particularly advantageous for anomaly detection from flight data since a large majority of aircraft flights do not exhibit malfunctions. Thus, when learning from reference sequences acquired for a plurality of past flights, the autoencoder learns to recognize patterns representative of flights without malfunctions.

[0034] According to one embodiment, the anomaly detection method is implemented for at least a first and a second acquired time sequence, said second acquired time sequence being determined by sub-sampling of said first acquired time sequence, and in which a said approximate time sequence from a said acquired sequence is determined by a said approximation module configured from reference time sequences of the same sampling frequency as this acquired sequence.

[0035] In this embodiment, two acquired time sequences corresponding to the same time range of operation of the aircraft component are used for the detection of an anomaly, but these sequences capture variations on different time scales. Indeed, these two sequences correspond to different sampling frequencies.

[0036] Certain anomalies are detectable only on variations over short times (visible to the extent that the sampling frequency of the analyzed sequence is quite high), while other anomalies correspond to variations over longer times (therefore detectable more easily for a lower sampling frequency). Consequently, this embodiment advantageously allows the detection of a greater number of anomalies.

[0037] In this embodiment, the first and second acquired time sequences are respectively approximated by two distinct approximation modules, and an overall abnormality score is determined for each of the first and second acquired sequences.

[0038] According to one embodiment of the method, each value of said second time sequence is an average over an interval of values ​​of said first time sequence, two consecutive values ​​of said second sequence being determined respectively from two contiguous intervals of said first sequence.

[0039] In this embodiment, the subsampling applied to the first sequence is also a low-pass filter which makes it possible to limit the maximum frequency of the signals of the second sequence. This advantageously makes it possible to detect additional anomalies, in particular anomalies detectable on low-frequency signals.

[0040] According to one embodiment of the method, the step of determining a global abnormality score comprises, for each of the one or more time sequences, sub-steps of: - determination of a sequence of instantaneous abnormality scores, a said instantaneous abnormality score being obtained, for a said instant, from the difference between the acquired value and the approximate value of this instant; - said overall abnormality score being obtained from a sum of at least some said instantaneous abnormality scores.

[0041] By instantaneous abnormality score, we mean here a quantity representative of a difference between an acquired value and an approximate value.

[0042] In one embodiment, a said instantaneous abnormality score is a Mahalanobis distance determined from a said difference and the mean and variance parameters. In another embodiment, a said instantaneous abnormality score is a difference divided by a mean difference.

[0043] According to one embodiment of the method, said overall abnormality score is obtained from the sum of a subset of said instantaneous abnormality scores greater than a second threshold.

[0044] The inventors have observed that the use of a second threshold in accordance with this embodiment advantageously makes it possible to reduce the number of false positives in the anomaly detection, and therefore to obtain more reliable detection.

[0045] According to one embodiment of the method, the calculation of said overall abnormality score only takes into account the instantaneous abnormality score of an instant if this instant is in a consecutive series of instants for which all the instantaneous abnormality scores are greater than said second threshold, a said consecutive series comprising a number of instants greater than a fixed number.

[0046] For example, if this fixed number is equal to 3, only the instantaneous abnormality scores belonging to intervals of at least three scores higher than the second threshold are counted in the sum which determines the overall abnormality score.

[0047] It should be emphasized that in embodiments in which several acquired sequences are approximated respectively by different approximation modules, different fixed numbers can be used to determine the overall abnormality scores for each of the acquired sequences.

[0048] In this embodiment, only the indicator values ​​which deviate from normal values ​​over a sufficiently large time page are taken into account in determining the overall abnormality score, and therefore the anomaly detection. Thus, This embodiment advantageously makes it possible to reduce the number of false positives in anomaly detection and therefore to have more reliable detection.

[0049] According to one embodiment of the method: - said sum is a sum of a subset of said instantaneous abnormality scores greater than a second threshold; and - the calculation of said overall abnormality score only takes into account the instantaneous abnormality score of an instant if this instant is in a consecutive series of instants for which all the instantaneous abnormality scores are greater than said second threshold, a said consecutive series comprising a number of instants greater than a fixed number.

[0050] According to one embodiment of the method, said at least one sum comprises instantaneous abnormality scores in a time interval starting after a start phase of operation of said one or more components and ending before an end phase of operation of said one or more components.

[0051] This embodiment makes it possible to limit false positives in anomaly detection because the start of operation and end of operation phases can be subject to large variations in state indicator values. This embodiment therefore advantageously makes it possible to obtain more reliable detection.

[0052] In one embodiment of the method, said at least one status indicator belongs to at least one of the following categories: - oil level; - oil temperature; - oil pressure; - oil filter pressure drop; - rotation speed of a turbine of said aircraft; - ambient pressure; - room temperature; - fuel flow; - position of a valve controlling the return of fuel to the tank.

[0053] Preferably, this embodiment makes it possible in particular to reliably monitor the operation of an oil circulation system in an engine or a turbomachine of an aircraft.

[0054] According to one embodiment of the method: - said at least one acquired time sequence comprises acquired values ​​of a plurality of state indicators; and - said at least one approximate time sequence comprises approximate values ​​of said acquired values ​​of a part of the state indicators of said plurality of indicators, the approximation being carried out from the acquired values ​​of said plurality of indicators.

[0055] In this embodiment, the approximation module takes as input the acquired values ​​of a plurality of status indicators (for example the indicators of the categories mentioned above) and determines the approximations of only a part of these acquired values, in particular the acquired values ​​of only a part of the plurality of indicators (for example only the status indicators of the first four categories mentioned above: the oil level, the oil temperature, the oil pressure, and the pressure drop of the oil filter). The overall abnormality score is then determined from differences between the part of the acquired values ​​and the corresponding approximated values. The rest of the indicators then serve as contextual information used by the approximation modules.

[0056] This embodiment makes it possible to improve the reliability of the approximation module (and consequently the reliability of anomaly detection) since the latter uses indicator values ​​that can be correlated with the values ​​of state indicators to be approximated. For example, the rotation speed of a turbine can be influenced by the operation of the oil circulation system (therefore is correlated with the values ​​of oil level, oil temperature, and oil pressure).

[0057] According to one embodiment, the method comprises a step of identifying the states causing the anomaly.

[0058] For example, the identification is performed by determining the times at which the instantaneous abnormality scores are greater than the second threshold. The values ​​of the state indicators at these times can be compared to average values ​​or thresholds to determine which indicators have abnormal values ​​and are therefore responsible for the anomaly.

[0059] According to one aspect of the invention, there is provided a computer program of instructions for implementing the steps of a method as described above, when the computer program is executed by a processor or a computer.

[0060] The computer program may be formed of one or more sub-parts stored in the same memory or in separate memories. The program may use any programming language, and be in the form of source code, object code, or intermediate code between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0061] According to one aspect of the invention, there is provided a computer-readable information medium comprising a computer program as described above.

[0062] The information carrier may be any entity or device capable of storing the program. For example, the carrier may comprise a storage means, such as non-volatile memory or ROM, for example a CD-ROM or a ROM microelectronic circuit, or a magnetic recording means, for example a floppy disk or a hard disk. On the other hand, the storage medium may be a transmissible medium such as an electrical or optical signal, which may be conveyed via an electrical or optical cable, by radio or by a telecommunications network or by a computer network or by other means. The program according to the invention may in particular be downloaded onto a computer network. Alternatively, the information medium may be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the method in question. Brief description of the drawings

[0063] Other features and advantages of the present invention will emerge from the description provided below of embodiments of the invention. These embodiments are given by way of illustrative example and are not of any limiting nature. The description provided below is illustrated by the attached drawings: 1. [Fig.l] represents in the form of a flowchart the steps of the detection method according to one embodiment. 2. [Fig.2] represents the functional architecture of a detection device according to one embodiment 3. [Fig.3A] represents sub-steps of the acquisition step of the detection method according to one embodiment. [Fig.3B] represents a first and a second time sequence according to one embodiment. 4. [Fig.4] represents sub-steps of the step of determining an overall abnormality score according to one embodiment. 5. [Fig.5A] and [Fig.5B] represent a first and a second auto-encoder neural network used to implement the step of determining a first and a second approximate time sequence, according to one embodiment. 6. [Fig.6A] represents a sequence of instantaneous abnormality scores determined with a detection method according to one embodiment. [Fig.6B] represents a sequence of values ​​of a status indicator. 7. [Fig.7] represents the hardware architecture of an anomaly detection device according to one embodiment. Description of the embodiments

[0064] The present invention relates to a method and a device for detecting anomalies in one or more components of an aircraft. In the following description, the invention is used to detect operating anomalies of a system of oil circulation of a turbomachine of an aircraft. This particular embodiment is used as an example and is in no way limiting. In particular, the invention can also be applied to an electronic control system of the aircraft, a fuel circulation system, a turbine, a set of sensors, or any other functional component of an aircraft.

[0065] [Fig.l] represents in the form of a flowchart the steps of the anomaly detection method according to one embodiment.

[0066] This embodiment comprises a step E10 during which time sequences Seq1 and Seq2 of state indicators of the oil circulation system are acquired during a flight. In this embodiment, each element of a time sequence comprises values ​​of state indicators of the system. Each of these values ​​is representative of a state of the system at a given instant or over a given time range.

[0067] A status indicator is for example an oil level, an oil temperature, an oil pressure, a rotational speed of an engine turbine, a position of a valve, or any other indicator impacted by the operation of the oil system.

[0068] In the example described here, each element of an acquired time sequence comprises the values ​​of 10 status indicators: - an oil temperature value, - an oil level value, - an oil pressure value, - an oil filter pressure drop, - a rotation speed of a turbine in a low pressure regime, - a rotation speed of a turbine in a high pressure regime, - the ambient pressure, - the ambient temperature, - the fuel flow, - the position of a valve controlling the return of fuel to the fuel tank.

[0069] The first four aforementioned status indicators directly characterize the oil circulation system while the following are indicators indirectly correlated with the operation of the oil circulation system.

[0070] Filter pressure drop refers to a drop in oil pressure at the filter, for example due to friction between the oil and the filter.

[0071] The other six indicators do not directly characterize the oil circulation system but their values ​​are influenced by the operation of the oil system. Thus, the evolution of these indicators over time constitutes information relevant to the state of the oil system. Subsequently, these six indicators are called contextual.

[0072] In this example, the engine (for example a turbojet) comprises two turbines operating in two different pressure regimes: low and high pressure.

[0073] Ambient pressure and temperature refers to the pressure and temperature outside the aircraft.

[0074] The invention also covers embodiments in which only one state indicator is used and each element of a time sequence has only one state indicator value.

[0075] The status indicators come for example from measurements by sensors connected to a system for saving the measurements made by these sensors over time. These indicators are thus measured and then recorded successively during the operation of the oil system, in particular during the flight of the aircraft.

[0076] [Fig.3A] represents an embodiment in which the acquisition step E10 comprises two sub-steps Eli and El2. During the sub-step Eli, values ​​of state indicators forming a first time sequence Seql are acquired.

[0077] For example, these values ​​correspond to measurements made at regular time intervals during operation of the oil system. For example, these measurements were made every second. Thus, in this example, the first time sequence Seql corresponds to a sampling frequency of 1 Hertz.

[0078] During sub-step E12, a second time sequence Seq2 is determined from the first sequence Seql by sub-sampling.

[0079] [Fig.3B] represents an example of values ​​acquired over a time range (of 400 seconds) forming a time sequence Seq1, and an example of a time sequence Seq2 determined by sub-sampling of the sequence Seq2. In this example, only one value of a status indicator (in particular the oil level) is represented per time step.

[0080] In this example, the sampling frequency of the sequence Seql is 1Hz (one value per second) while the sampling frequency of the sequence Seq2 is 10mHz (one value every 100 seconds).

[0081] In one embodiment, the subsampling consists of: - decompose the first sequence into contiguous intervals of K elements each, and - determine the average of the elements of each interval, the second sequence being the sequence of averages thus obtained.

[0082] In this embodiment, the subsampling makes it possible both to divide the sampling frequency of the first sequence by K, and to apply a low-pass filter which divides the frequency of the signals of the first sequence by K. For information purposes, in the example described here, the number K is chosen to be equal to 100. In this case, the sequence Seq2 corresponds to a sampling frequency of 10 mHz

[0083] In a simplified example for illustrating the embodiment mentioned above, the first acquired time sequence comprises four elements, each element comprising two state indicator values: (3; 4), (2; 3), (4; 5) and (4; 6). In this example, K is equal to 2, so the second sequence is determined by taking the average of the elements (3; 4) and (2; 3) and the average of the elements (4; 5) and (4; 6). The second sequence therefore includes the two elements (2,5; 3,5) and (4; 5,5).

[0084] The invention is not limited to the sub-sampling method described above. For example, in another embodiment, the sub-sampling consists of selecting an element of the first sequence every K elements. In this case, the sampling sequence of the second sequence is equal to the sampling frequency of the first sequence divided by K.

[0085] During a step E20, approximation modules AE1 and AE2 perform approximations of the respective sequences Seql and Seq2, and thus determine approximate sequences Seql* and Seq2*.

[0086] In the embodiment described here, the modules AE1 and AE2 are auto-encoder neural networks.

[0087] In this embodiment, the module AE1 has been trained to approximate a first group of reference sequences corresponding to the same sampling frequency as that of the sequence Seq1 (1Hz for example), and the module AE2 has been trained to approximate a second group of reference sequences corresponding to the same sampling frequency as the sequence Seq2 (10 mHz for example).

[0088] The reference sequences of the second group are for example determined, from the sequences of the first group, by the same sub-sampling method which makes it possible to obtain Seq2 from Seql.

[0089] [Fig.5A] illustrates an example of the architecture of the auto-encoder network AE1.

[0090] This network includes an encoder performing the compression of the Seql sequence provided as input. It consists of five ConvlD convolution layers preceded (preceded in the input-to-output direction) by BatchNorm batch normalization layers. Except for the OutConv layers presented below, the ConvlD convolution layers all include a nonlinear activation layer at the output of the convolution. The ConvlD convolution layers are also followed by MaxPool max pooling layers, except for the last convolution layer. These pooling layers reduce the sequence length provided as input to the network and thus allow its compression. The five ConvlD convolution layers of the encoder respectively include filters of dimensions 20, 40, 60, 80 and 100. The dimension of a filter corresponds to the dimension of the elements at the output of the convolution layer. For example, the first convolution layer takes as input a sequence of elements each comprising 10 values, and provides as output a sequence of elements each comprising 20 values, these values ​​corresponding to the projection of elements of the input sequence into a mathematical space of dimension 20.

[0091] This encoder is followed by a decoder which determines the Seql* sequence from the encoder output. This decoder has four composite layers, each composite layer consisting of a ConvlDTranspose deconvolution layer followed by a BatchNorm layer, a ConvlD layer followed by another BatchNorm layer. The decoder's ConvlDTranspose deconvolution layers have filters of dimensions 100, 80, 60 and 40 respectively. The decoder's ConvlD convolution layers have filters of dimensions 80, 60, 40 and 20 respectively. The last OutConv layer of the decoder is a convolution layer with a filter of dimension 10. The OutConv layer outputs the Seql* sequence.

[0092] [Fig.5B] illustrates an example of the architecture of the AE2 autoencoder network.

[0093] The AE2 encoder comprises three multiple convolution layers ConvlD and three dilated convolution layers ConvlDdilated, each followed by a BatchNorm packet normalization layer. Each of the multiple dilated convolution layers ConvlDdilated is a stack of several dilated convolution layers with filters of the same dimension, followed by BatchNorm layers. The three multiple dilated convolution layers of the encoder each comprise 5, 4 and 2 stacked layers, respectively. Each dilated convolution layer has a dilation step size of 2. These dilated convolution layers are used to detect anomalies occurring over a long period of time.

[0094] Except for the last one, the ConvlD convolution layers are followed by a MaxPool pooling layer. The six convolution layers of the encoder have filters of dimensions 64, 64, 96, 96, 128 and 128 respectively.

[0095] The AE2 decoder has two deconvolution layers, two multiple dilated convolution layers, and one convolution layer, each followed by a BatchNorm layer. The encoder's deconvolution layers have filters of dimension 128 and 96, respectively, while its convolution layers have filters of dimension 128 and 96, respectively. The encoder's two multiple dilated convolution layers each have 4 and 5 layers, respectively.

[0096] The last OutConv layer of the decoder is a convolution layer with a filter of dimension 10. The OutConv layer provides the sequence Seq2* as output.

[0097] The invention is not limited to the auto-encoder network architectures as illustrated in Figures 5A and 5B. In other embodiments, auto-encoder networks may have a number of layers and parameters different from the networks described above. In embodiments, the auto-encoder networks comprise layers of recurrent networks, layers of encoder and / or decoder type transformers, or any other architecture allowing the analysis of time sequences.

[0098] Furthermore, the invention does not limit the approximation modules to auto-encoder networks. In other embodiments, the approximation modules perform regressions as mentioned previously. These regressions can be produced by machine learning models such as artificial neural networks. In embodiments, the approximation modules perform regressions as mentioned by applying an analytical model of the evolution of the states of the monitored components. Such an analytical model is for example a Kalman filter.

[0099] The AE1 module takes as input the sequence Seql. In the example described here, the encoder determines a compressed sequence by passing the sequence Seql through its successive layers. The compressed sequence then passes through the successive layers of the decoder of AE1 which provides as output the approximate sequence Seql*. The process is similar for determining the approximate sequence Seq2* of Seq2 with the AE2 module.

[0100] The autoencoder AE1 was trained from a set of reference sequences. These reference sequences were acquired during flights of a set of aircraft. In the example described here, each element of a reference sequence contains values ​​of the same 10 state indicators mentioned previously and contained in the sequence Seql.

[0101] In one embodiment, the reference sequences come from flights deemed “healthy”, i.e. without malfunction during the flight, and which were followed by at least a significant number (for example 10) of flights without malfunction.

[0102] During its training, the autoencoder is configured to minimize approximation errors between the reference sequences and sequences that the autoencoder determines. Thus, the training of AE1 comprises a plurality of steps, each comprising: - the determination of a sequence from a reference sequence; - the calculation of an approximation error between the sequence thus determined and the corresponding reference sequence; - modification of the AE1 autoencoder parameters to minimize the approximation error thus calculated

[0103] Modifying the parameters of the AE1 autoencoder to minimize approximation errors can be done with a gradient descent algorithm. The gradient descent algorithm aims to correct errors according to the importance of the contribution of each parameter to it.

[0104] The approximation error is for example a mean squared error between the values ​​of the sequence determined by AE1 and the values ​​of the corresponding reference sequence.

[0105] The reference sequences used for training the autoencoder AE2 are determined in the same way as the sequence Seq2 is determined from the Seql. Thus, each reference sequence used for training AE2 is determined by subsampling a reference sequence used for training AE1.

[0106] The training of AE2 is similar to the training of AE1.

[0107] During a step E30, overall abnormality scores A1 and A2 are determined for the respective sequences Seq1 and Seq2.

[0108] In embodiments, the modules AE1 and AE2 determine approximate values ​​of only a portion of the state indicators. For example, the input time sequences of the modules AE1 and AE2 contain the acquired values ​​of the ten state indicators mentioned above (the indicators directly characterizing the oil system and the contextual indicators), and the approximate values ​​correspond to the four state indicators directly characterizing the oil circulation system, in particular: - the oil temperature, - the oil level, - the oil pressure, and - the pressure drop of the oil filter.

[0109] Furthermore, in this same example, the overall abnormality scores are determined from differences between the acquired values ​​corresponding to these four state indicators and the corresponding approximate values. Thus, in this example, the overall abnormality scores A1 and A2 do not take into account contextual indicator values.

[0110] [Fig.4] represents sub-steps of step E30 according to one embodiment.

[0111] During a sub-step E31, for each element of an instant t of the sequence Seql, a difference rlt is calculated between this element and the corresponding approximated element. In a simplified example used for illustration purposes, the sequence Seql has two elements, each with three state indicator values: (3; 4; 1) and (2; 3; 2). In this example, the sequence Seql* approximated by AE1 has the two elements (3,1; 4,2; 0,8) and (2,5; 2,7; 1,9). In this case, the differences between the sequence Seql and the sequence Seql* are the two elements (-0.1; -0.2; 0.2) and (-0.5; 0.3; 0.1).

[0112] In the same way, the differences r2t between the sequence Seq2 and the sequence Seq2* approximated by Seq2 are calculated.

[0113] During a sub-step E32, from each difference rlt (r2t) an instantaneous abnormality score alt (a2t) is calculated. For example, such an instantaneous abnormality score is a Mahalanobis distance. In this case, an instantaneous abnormality score alt or a2t is calculated by the following formulas:

[0114] [Math.l] alt= (rVnjV^rViiJ

[0115] [Math.2] a2t=

[0116] Where is a mean difference and is the inverse of a covariance matrix If a difference rlt has a single value, 'Li, has a single value, and ^1. also has a single value. If a difference is a vector with multiple values, 1¼ is a vector with as many values. The parameter can be determined by calculating, with the trained AE1 autoencoder, an average of differences calculated from reference sequences and sequences approximated by AE1 from these reference sequences. Similarly, the covariance matrix can be determined from the covariances between reference values ​​and values ​​approximated by AE1.

[0117] Similarly, the parameters M2 and °2 are mean and covariance parameters determined from the trained AE2 autoencoder.

[0118] The invention is not limited to the choice of a Mahalanobis distance. For example, in another embodiment, an instantaneous abnormality score corresponding to a time t is an average of the absolute values ​​of the difference values ​​rlt or r2t.

[0119] During a sub-step E33 the overall abnormality scores A1 and A2 are determined, from the sequences of instantaneous abnormality scores alt and a2t.

[0120] In the embodiment illustrated in Figure 4, a global abnormality score Al(A2) is a sum a2t) over instantaneous abnormality scores alt(a2t).

[0121] In one embodiment, an abnormality score Al (A2) is the sum of the instantaneous abnormality scores alt (a2t) which are greater than a fixed threshold S2. This threshold S2 can be determined as a quantile of a set of instantaneous abnormality scores calculated from a set of reference sequences with the module AE1 (AE2). For example, this threshold S2 can be equal to the 99th quantile of this set of instantaneous abnormality scores.

[0122] It should be emphasized that in embodiments, different threshold values ​​S2 and S2' may be chosen for the alt scores determined from the auto-encoder AE1 and for the a2t scores determined from the auto-encoder AE2.

[0123] In one embodiment, only intervals of more than N instantaneous abnormality scores above the threshold S2 are taken into account in the sum which determines the overall abnormality score Al (A2), where N is a fixed number.

[0124] In embodiments, only the intervals of more than N instantaneous abnormality scores above the threshold S2 are taken into account in the sum which determines the overall abnormality score A1, and only the intervals of more than N' instantaneous abnormality scores above the threshold S2' are taken into account in the sum which determines the overall abnormality score A2, N and N' being numbers fixed independently for each auto-encoder AE1 and AE2. For example, N can be chosen equal to 1 in order to detect point anomalies, and N' equal to 2 to be able to detect non-point anomalies.

[0125] During a step E40, each global abnormality score A1 and A2 is compared with a threshold SI. This threshold is for example equal to the 98th quantile of a set of global abnormality scores determined from a set of reference sequences with the modules AE1 and AE2.

[0126] In embodiments, the global abnormality scores A1 and A2 are respectively compared with distinct thresholds SI and SI'. For example, SI is equal to the 98th quantile of a set of global abnormality scores determined from a set of reference sequences with the module AE1, and SI' is equal to the 98th quantile of a set of global abnormality scores determined from a set of reference sequences with the module AE2.

[0127] In one embodiment, if the abnormality score A1 is greater than the threshold SI, or if the score A2 is greater than the threshold SI', then the anomaly is detected.

[0128] It should be noted that the thresholds SI and S2 as well as the number N (as well as SI', S2' and N'), mentioned above, are parameters which can be chosen so as to optimize the anomaly detection on a set of reference sequences.

[0129] For example, these reference sequences may include sequences corresponding to flights without detected malfunctions, and flights on which malfunctions have been detected, for example with conventional methods. In this case, the parameters SI, S2 and N may be chosen so as to jointly minimize the number of false positives and the number of false negatives in the anomaly detection from these reference sequences. In this example, a false positive corresponds to the detection of an anomaly (an overall abnormality score is greater than the SI threshold) on a reference sequence while this sequence corresponds to a flight without malfunction. Conversely, a false negative corresponds to the non-detection of anomaly (the overall abnormality score is lower than the SI threshold) on a reference sequence corresponding to a flight presenting a malfunction.

[0130] During a step E50, the status indicators causing an anomaly which was detected during the step E40 are identified.

[0131] For example, during this step, the instants at which the instantaneous abnormality scores are greater than the threshold S2 are identified. For these instants thus identified, the state indicators for which the differences rlt and r2t between the acquired values ​​and the approximated values ​​are the greatest are then identified. Thus, the state indicator(s) whose approximated values ​​deviate the most from the acquired values ​​represent the state responsible for the anomaly.

[0132] Figures 6A and 6B illustrate an example where abnormal variations in the oil level are the cause of the anomaly.

[0133] [Fig.6A] represents instantaneous abnormality scores over time, as well as the threshold S2 represented by a horizontal line. Between time steps 2200 and 3200, the instantaneous abnormality scores are higher than the threshold S2.

[0134] [Fig.6B] represents the differences over time between the acquired oil level values ​​and the corresponding approximate values. Between time steps 2200 and 3200, the approximate oil level values ​​deviate significantly from the acquired values, which indicates that the oil level is responsible for the anomaly, in other words, that the evolution of the oil level during this flight is abnormal.

[0135] [Fig.2] represents the functional architecture of an anomaly detection DA device configured to implement the steps of the anomaly detection method represented in [Fig.l]. The DA device comprises: - a module M10 for implementing step E10 of acquiring the sequences of state indicator values ​​Seql and Seq2; - a module M20 for implementing step E20 of determining the sequences of approximate values ​​Seql* and Seq2*, this module comprising the approximation modules AE1 and AE2 described previously; - an M30 module to implement step E30 of the global abnormality scores Al and A2 for the sequences Seql and Seq2; - a module M40 for implementing step E40 of detecting an operating anomaly; and - an M50 module to implement step E50 of identifying the states causing the anomaly.

[0136] [Fig.7] represents the hardware architecture of a DA anomaly detection device according to a particular embodiment of the invention.

[0137] In the embodiment described here, the device DA has a hardware architecture of a computer. It comprises in particular a processor D1, a read-only memory D2, a random access memory D3, a rewritable non-volatile memory D4 and communication means D5.

[0138] The read-only memory D2 of the device DA constitutes a recording medium in accordance with the invention, readable by the processor D1 and on which is recorded a computer program PGI in accordance with the invention, this program comprising instructions for the execution of the steps of an anomaly detection method according to the invention described previously with reference to [Fig.l].

[0139] The PGI computer program defines functional modules of the DA device shown in [Fig.2].

Claims

1. Claims Method for detecting an operating anomaly with respect to reference values, of at least one component of an aircraft associated with at least one status indicator, said method comprising, for at least two acquired time sequences (Seql, Seq2), each comprising successive acquired values ​​of said at least one indicator at different times during operation of said at least one component, steps of: - determination (E20), for each of the acquired time sequences, of an approximate time sequence (Seql*, Seq2*) comprising approximate values ​​of said values ​​acquired at said instants for this acquired time sequence, the approximation being carried out by an approximation module (AE1, AE2); - determination (E30), for each of said acquired time sequences (Seql, Seq2), of a score (A1, A2) of overall abnormality from differences between said acquired values ​​(Seql, Seq2) and said approximate values ​​(Seql*, Seq2*) for this acquired time sequence; and - detection (E40), for each of said acquired time sequences, of an operating anomaly of said component as a function of a comparison of said overall abnormality score determined for this sequence with a first threshold, said approximation module (AE1, AE2) being previously configured to minimize approximation errors between time sequences of reference values ​​and time sequences of values ​​approximated from said reference values, said reference values ​​being representative of normal operations, so that approximation errors between said acquired values ​​(Seql, Seq2) and said approximated values ​​(Seql*, Seq2*) are greater if the acquired values ​​are representative of abnormal operation, said second acquired time sequence (Seq2) being determined by sub-sampling said first acquired time sequence (Seql),a said approximated time sequence (Seql*, Seq2*) from a said acquired sequence (Seql, Seq2) being determined by a said configured approximation module (AE1, AE2), from reference time sequences of the same sampling frequency as this acquired sequence (Seql, Seq2)

2. A method according to claim 1 wherein each value of said second time sequence is an average over an interval of values ​​of said first time sequence, two consecutive values ​​of said second sequence being determined respectively from two contiguous intervals of said first sequence.

3. Method according to one of claims 1 to 2 in which a said approximation module (AE1, AE2) is obtained by a machine learning algorithm from said reference time sequences.

4. Method according to claim 3 in which a said approximation module (AE1, AE2) is an artificial neural network comprising an encoder and a decoder, - a said encoder determining a compressed sequence from a said time sequence provided as input to the neural network, said compressed time sequence being of a size smaller than said time sequence provided as input, - a said decoder performing a reconstruction from the compressed sequence to obtain at the output of the neural network a said approximate time sequence.

5. Method according to one of claims 1 to 4 in which the step of determining (E30) a global abnormality score (A1, A2) comprises, for each of the one or more time sequences, sub-steps of: - determining (E31) a sequence of instantaneous abnormality scores, a said instantaneous abnormality score (alt, a2t) being obtained (E32), for a said instant, from the difference (rlt, r2t) between the acquired value (Seql, Seq2) and the approximate value (Seql*, Seq2*) of this instant; - said global abnormality score being obtained (E33) from a sum of at least some of said instantaneous abnormality scores.

6. Method according to claim 5 in which: - said sum is a sum of a subset of said instantaneous abnormality scores greater than a second threshold (S2); and - the calculation of said overall abnormality score only takes into account the instantaneous abnormality score of an instant if this instant is in a consecutive series of instants for which all the instantaneous abnormality scores are greater than said second threshold, a said consecutive series comprising a number of instants greater than a fixed number.

7. Method according to one of claims 5 or 6 in which said sum comprises instantaneous abnormality scores in a time interval starting after a start phase of operation of said one or more components and ending before an end phase of operation of said one or more components.

8. Method according to one of claims 1 to 7 wherein said component is an oil circulation system and said at least one status indicator belongs to at least one of the following categories: - oil level; - oil temperature; - oil pressure; - oil filter pressure drop; - rotation speed of a turbine of said aircraft; - ambient pressure; - ambient temperature; - fuel flow rate; - position of a valve controlling the return of fuel to the tank.

9. Method according to one of claims 1 to 8 comprising a step of identifying (E50) the states causing the anomaly.

10. Device (DA) for detecting an operating anomaly with respect to reference values, of at least one component of an aircraft associated with at least one status indicator, said device comprising, for at least two acquired time sequences (Seql, Seq2), each comprising successive acquired values ​​of said at least one indicator at different times during operation of said at least one component: - a determination module (M20), for each of said acquired time sequences, of an approximate time sequence (Seql*, Seq2*) comprising approximate values ​​of said acquired values ​​at said times for this acquired time sequence, the approximation being carried out by an approximation module (AE1, AE2); - a determination module (M30) for each of said acquired time sequences (Seql, Seq2), of a global abnormality score (A1, A2) from differences between said acquired values ​​(Seql, Seq2) and said approximate values ​​(Seql*, Seq2*) for this acquired time sequence; and - a detection module (M40), for each of said acquired time sequences, of an operating anomaly of said component as a function of a comparison of said global abnormality score determined for this sequence, with a first threshold, said approximation module (AE1, AE2) being previously configured to minimize approximation errors between time sequences of reference values ​​and time sequences of values ​​approximated from said reference values, said reference values ​​being representative of normal operations, so that approximation errors between said acquired values ​​(Seql,Seq2) and said approximate values ​​(Seql*, Seq2*) are greater if the acquired values ​​are representative of abnormal operation, said second acquired time sequence (Seq2) being determined by sub-sampling said first acquired time sequence (Seql), a said approximate time sequence (Seql*, Seq2*) from a said acquired sequence (Seql, Seq2) being determined by a said approximation module (AE1, AE2) configured from reference time sequences of the same sampling frequency as this acquired sequence (Seql, Seq2).,

11. Computer program (PGI) comprising instructions for implementing the steps of a method according to any one of claims 1 to 9, when said computer program is executed by a processor.

12. Recording medium (D2) readable by a computer and on which a computer program (PGI) according to claim 11 is recorded.

Images