Methods for backing up flight data and generating an abnormal event detection module used to trigger said backup.

Abstract

Methods for backing up flight data and generating an abnormal event detection module used to trigger said backup. This method, a module (MDEA) for detecting an abnormal event during a flight, comprises the following steps: - obtaining (E10) data from healthy flights; - associating (E30), for subsets (EIi) of reference event indicators produced during healthy flights, thresholds (THi) allowing the determination of partial opinions on said reference event from the values ​​of these indicators; - configuring said module with said thresholds (THi) and with a decision model (LD) allowing the determination of a definitive opinion on an event of interest from partial opinions determined for this event of interest; - said module being configured to generate a distress signal based on said definitive opinion determined for an event produced during a flight. Fig. 9

Description

Title of the invention: Methods for saving flight data and generating an abnormal event detection module used to trigger said saving. Background of the invention

[0001] The present invention is in the field of aeronautics and more specifically in the field of regulatory flight recorders, otherwise called "on-board recorders" or "black boxes".

[0002] It sometimes happens that after a plane crash: • the black boxes cannot be recovered; and / or • that the costs of searching for the black boxes are very high, and / or • that one or more of the black boxes are damaged.

[0003] Consequently, civil aviation regulations have recently evolved, and the ICAO (International Civil Aviation Organization) has published various documents, in particular document 10054: "Manual on Locating Aircraft in Distress and Retrieving Data from Flight Recorders." This document describes, among other things, solutions for rapidly retrieving data from black boxes. One solution in document 10054 involves retrieving flight recorder data in the event of a potential distress situation, before the flight becomes fatal.

[0004] This document is based on another EUROCAE document ED-237: "Minimum aviation System performance specification for criteria to detect in-flight aircraft distress events to trigger transmission of flight information" which recommends a list of criteria to be implemented to trigger entry into distress mode.

[0005] This ED-273 document is itself based on work carried out under the supervision of the BEA (Bureau of Enquiry and Analysis for Civil Aviation Safety) on a set of sixty-eight flight accidents. By analyzing the data collected during these flights, the BEA developed a series of thirteen criteria based on an assessment of what constitutes an emergency situation. The approach is binary in that a condition is either true or false. If a condition is true, an emergency situation is considered to have begun. Otherwise, if all the conditions are false, then the flight is considered normal. These thirteen criteria are:

[0006] - Excessive bank angle (in English "excessive bank");

[0007] - Excessive pitch

[0008] - Stall

[0009] - Low calibrated speed (in English, "low CAS")

[0010] - Excessive vertical speed

[0011] - Overspeed

[0012] - Unusual load factors

[0013] - Excessive roll command

[0014] - Excessive use of the rudder rudder»)

[0015] - Impact Warning and Alarm System Warning (in English " Terrain awareness and warning System TAWS warning)

[0016] - Altitude too low (in English "too low altitude")

[0017] - Traffic alert and collision avoidance system alert (in English “Traffic Alert and Collision Avoidance System)

[0018] - Cabin altitude warning.

[0019] An analysis of all sixty-eight accidental flights confirms that all flights with a fatal outcome are detected by applying these criteria, which validates the definition of these thirteen criteria.

[0020] However, using these criteria, the detection of the distress situation occurs only a very short time, on average about thirty seconds, before the accident. This thirty-second period is unfortunately insufficient to allow the transmission of the data contained in a black box via a satellite communication solution. The inventors have estimated that it would take approximately four minutes to transfer the equivalent of twenty minutes of recording from a black box via satellite. Object and summary of the invention

[0021] The invention therefore aims at a solution for anticipating the detection of an accident of an aircraft to come, without thereby generating excessively untimely false alarms and for transmitting, before the occurrence of the accident, a copy of all or a significant amount of the data contained in one or more black boxes of the aircraft to an external receiver.

[0022] The invention aims primarily to identify new symptomatic or warning criteria of a distress situation, i.e. not identified on the sixty-eight accidental flights and to initiate an emergency communication (transmission of data to a receiver external to the aircraft).

[0023] In the remainder of this document, the following terms will be used: - "Accidental flight" means a flight during which an accident or major incident occurred, preventing its maintenance data from being downloaded; and - "Safe flight" or simply "safe flight", a flight during which no major accident or incident occurred, so that the aircraft was able to land, and its maintenance data could be downloaded.

[0024] It is recalled that prior art studies, in particular those of the BEA, relate to the analysis of sixty-eight crashed flights.

[0025] The invention, on the contrary, proposes to analyze the data collected during flights that were healthy or considered healthy and which were unloaded into maintenance computers at the end of these flights.

[0026] Thus, and according to a first aspect, the invention relates to a method of generating a module for detecting an abnormal event occurring during a flight in an aircraft.

[0027] In the remainder of this document, a "reference event" will be defined as an event that occurred during a healthy flight at time t and that can be represented by a set of indicators at that time t. Reference events are used to generate the module for detecting an abnormal event.

[0028] Once the detection module has been generated, it can be used to determine whether an event, called an "event of interest" to distinguish it from the reference events used to generate the module, is a normal or atypical event.

[0029] A preferred use of the detection module is to determine, when used in an aircraft in flight, whether an event of interest that occurs during flight is normal or atypical so as to trigger an emergency communication, i.e. the sending of flight data to an external receiver.

[0030] But the detection module can also be used on the ground, to detect whether an event of interest occurring during a healthy or accidental flight is detected as normal or atypical so as to retrain and / or validate the detection module.

[0031] Hereafter we will call: - "indicator" the type of an indicator (for example "pitch angle") and; - "value of the indicator", a value of this indicator at a time t, for example 20°.

[0032] In accordance with the invention, subsets of the complete set of indicators representing a reference event will be used. Those skilled in the art understand that such an incomplete subset provides only a partial / incomplete / imperfect representation of the reference event.

[0033] Thus, the method for generating an abnormal event detection module according to the invention comprises: - a step of obtaining flight data acquired during flights considered to be safe; - for at least a subset of indicators selected from a set of indicators obtained from this data and representative of a reference event that occurred during a moment of said safe flight, a step of associating, with this subset of indicators, a threshold allowing the determination of a partial opinion on the normal or atypical nature of this reference event based solely on the values ​​of the indicators in said subset; and - a configuration step for the detection module with: (i) the indicators and the threshold for each subset of indicators, and with (ii) a decision model enabling the determination of a definitive opinion on the normal or atypical nature of an event of interest from partial opinions determined for that event of interest, a partial opinion being determined for each subset of indicators from (i) the values ​​of those indicators obtained from data representative of the event of interest and (ii) the threshold associated with that subset; - said abnormal event detection module being configured to, when implemented in an aircraft in flight, generate or not generate a distress signal depending on said final opinion determined for an event of interest occurring during the flight.

[0034] In one embodiment of the invention, the abnormal event detection module is a computer program configured to be executed by a processor of the aircraft.

[0035] The generation method, which is the subject of the invention, uses data from flights that were safe or considered safe, namely flights that ended without a major accident occurring.

[0036] Once generated, this detection module is capable of determining a definitive opinion as to whether an event of interest occurring during a healthy or accidental flight is a normal or atypical event.

[0037] According to the invention, this final opinion is obtained from partial opinions and a decision model. This decision model can, for example, be based on a set of detection rules, or on a model obtained by learning, for example using an algorithm of the Isolation Forest or LOF (Local Outlier Factor) type.

[0038] Partial opinions on the normal or atypical nature of an event are described as partial because they are obtained by considering only a subset of the set of indicators that completely represents that event.

[0039] In one embodiment, the decision model is configured to take into account a plurality of partial opinions determined for a plurality of subsets of indicators during a time window in order to determine the final opinion on the normal or atypical nature of a reference event.

[0040] The invention also relates to a module for detecting an abnormal event during a flight in an aircraft, this module being characterized in that it was generated by a generation process as mentioned above.

[0041] The invention also relates to a method for saving flight data acquired during a flight of an aircraft, said aircraft comprising a detection module abnormal event as mentioned above, this process involving: - a data collection step during the flight and storage of said collected data in an aircraft memory; - a step of copying at least part of the data collected in an aircraft black box; - a step of providing said data collected as input to said detection module; and - a step of triggering the transmission of at least part of the collected data to an external receiver on and only on detection of a distress signal generated by said detection module.

[0042] Correspondingly, the invention relates to a device for saving flight data acquired during a flight of an aircraft, said aircraft comprising a module for detecting an abnormal event as mentioned above, this device comprising: - a module for collecting data during flight and storing said collected data in a memory; - a module for copying at least part of the said data collected in a black box of the aircraft; - a module for supplying the said data collected as input to the detection module; and - a triggering module for the transmission of at least part of the collected data to an external receiver on and only on detection of a distress signal generated by said detection module.

[0043] The invention also relates to an aircraft comprising a safeguarding device as mentioned above.

[0044] The backup device can for example be integrated into a Flight-Data Acquisition Unit which receives various flight data, in analog or digital form from a number of sensors and avionics systems, and which routes at least some of this data to the black box.

[0045] Thus, and in general, the invention considers that among the flight data considered healthy, atypical events occur which would have warranted the triggering of an emergency communication of flight data because these atypical events could potentially have led to an accident of the aircraft, for example in slightly different circumstances, or by a different reaction of the aircraft pilots.

[0046] The invention therefore proposes to enable the detection of these atypical events by analysis of data acquired and recorded during healthy flights.

[0047] The new triggering conditions identified by the invention from data collected during healthy flights make it possible to anticipate more accident situations and to detect them earlier.

[0048] This healthy flight data is, for example, data from an Aircraft Condition Monitoring System (ACMS). It should be noted that an aircraft condition monitoring system is a predictive maintenance tool consisting of a high-capacity flight data acquisition unit and associated sensors that sample, monitor, and record flight information and parameters from the aircraft's main systems and components.

[0049] Very advantageously, the ACMS system data used to define the new triggering criteria, and collected on healthy flights, are much more voluminous and much richer than the black box data from the sixty-eight accidental flights analyzed by the BEA to identify the thirteen criteria mentioned above.

[0050] The ACMS system data used to determine the criteria by the invention are much more extensive because they are collected from a very large number of flights considered safe. They are also available at a higher frequency (several Hertz) than black box data.

[0051] The data used by the invention may contain data of the type recorded in black boxes but also other types of data.

[0052] By way of example, the ACMS system data on A330-300 aircraft comprises more than 695 parameters, a very large number of which are not recorded in the black boxes. By way of example, the invention can take into account all the criteria defined by the BEA and other flight data, for example at least one piece of data from: - notifications produced by measurement systems (e.g., indicator lights) “Master Caution” and “Master Warning”); - overspeed data under conditions of landing gear extended, or flaps in landing position (in English Overspeed Train / Overspeed Flaps) ); - fire indicators other than engine fire indicators; - vibration signals; - one or more specific words spoken by a pilot or co-pilot, for example MAYDAY; - an inability of the pilot or co-pilot to fly the aircraft; - a critical situation in the cockpit, for example the presence of smoke; - an obstacle detected by an outdoor camera;

[0053] Data transmission to the external receiver can, for example, be done by radio transmission via satellite in particular or via GSM network.

[0054] The data sent can be both data collected before the detection of the generated distress signal and representative of the detection of an abnormal event and data collected after this detection.

[0055] In particular, the present invention can implement a method for backing up operating data of an aircraft comprising the steps of collecting operating data and recording it as it is taken on board the aircraft, detecting that the aircraft is likely to have an accident, and from the detection, transmitting to the external receiver, on the one hand, the data collected as soon as they are collected and, on the other hand, the data stored in a reverse chronological order of the order of recording.

[0056] This backup can be implemented using the backup procedure described in document FR2967647 (Al).

[0057] In one embodiment, the method for generating a module for detecting an abnormal event during a flight in an aircraft comprises, for a plurality of subsets of indicators representative of reference events that occurred during said healthy flights: - a step of calculating a normality score for each of said reference events from the values ​​of the indicators of said subset for that event; - the threshold associated with a subset of indicators being determined from a type of distribution of said normality scores calculated for that subset.

[0058] In one embodiment, the threshold is determined according to the shape of a tail of the distribution.

[0059] In one embodiment, the normality scores calculated for a subset of indicators are obtained by a learned algorithm associated with that subset.

[0060] In a particular embodiment, among the aforementioned indicators, a distinction is made between so-called first-level indicators and so-called second-level indicators.

[0061] First-level indicators can, for example, be represented as vectors, each component of which represents a flight data point or a meteorological data point at a time t.

[0062] Second-level indicators can be obtained by aggregating first-level indicators, for example, using a function of the average, minimum, or maximum type, over a given time period, for example, during the last ten seconds. In this embodiment, the use of second-level indicators makes it possible to determine a partial opinion based on indicator values ​​considered during a time window corresponding to this aggregation period.

[0063] The decision model used in the invention can be of a different nature. It can Specifically, this involves: - a model based on a combination of logical conditions (for example if condition 1 and (condition 2 or condition 3) ...); or - a model based on state machines.

[0064] In one example, an event is considered atypical (partial opinion) if the normality score of that event exceeds said threshold.

[0065] In a particular embodiment, the different stages of the data generation or data saving process are determined by computer program instructions or are implemented by a silicon chip which includes transistors adapted to constitute model gates of a programmable or non-programmable wired model.

[0066] Consequently, the invention also relates to a computer program on an information medium, this program being capable of being implemented in a controller computer, this program comprising instructions adapted to the implementation of the steps of a data generation and / or saving process as described above.

[0067] This program may use any programming language, and be in the form of source code, object code, or code intermediate between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0068] The invention also relates to a computer-readable information carrier containing instructions for a computer program as mentioned above. The information carrier can be any entity or device capable of storing the program. For example, the carrier can include a storage means, such as a ROM, non-volatile flash memory, or a magnetic recording means, such as a hard drive. Alternatively, the information carrier can be a transmissible medium, such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The program according to the invention can, in particular, be uploaded to a network such as the Internet. Alternatively, the information carrier can 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 process in question. Brief description of the drawings

[0069] Other features and advantages of the present invention will become apparent from the description below, with reference to the accompanying drawings which illustrate non-limiting examples of embodiments. In the figures:

[0070] [Fig. 1] Fig. 1 represents the correlation for the values ​​of two indicators (normal and longitudinal acceleration of the aircraft) for a set of events obtained from healthy flight data;

[0071] [Fig.2] The [Fig.2] represents a first example of thresholds on these two indicators for the events of the [Fig. 1];

[0072] [Fig.3] The [Fig.3] represents a second example of thresholds on these two indicators for the events of the [Fig.1];

[0073] [Fig.4] The [Fig.4] represents the result of a classification of the indicators of the events of the [Fig.1] by a mathematical tool or by an operator;

[0074] [Fig.5] The [Fig.5] represents a third example of a threshold for the indicators of the events of the [Fig.1];

[0075] [Fig.6] The [Fig.6] represents an aircraft conforming to a particular embodiment of the invention;

[0076] [Fig.7] Fig.7 represents the main steps of a data backup process according to one embodiment of the invention;

[0077] [Fig.8] Fig.8 illustrates an example of a decision model that can be used in the invention;

[0078] [Fig.9] Fig.9 represents the main steps of a method for generating a module for detecting an abnormal event according to an embodiment of the invention;

[0079] [Fig. 10] The [Fig. 10] represents a first example of a distribution of normality scores;

[0080] [Fig. 11] The [Fig. 11] represents a second example of normality score distribution;

[0081] [Fig. 12] The [Fig. 12] represents a third example of normality score distribution;

[0082] [Fig. 13] Figure 13 represents four normality score distributions. Description of different embodiments

[0083] We will now describe a data backup method and a method for generating an abnormal event detection module according to a particular embodiment of the invention.

[0084] In general, and as described below in detail, the data backup method uses an abnormal event detection module according to the invention to trigger, during a flight, the transmission of flight data if the module detects an abnormal event during the flight.

[0085] This detection module uses thresholds stored in the detection module by the process of generating this module from indicator values ​​associated with reference events that occurred during healthy flights.

[0086] It is recalled that a "reference event" is an event that has produced during a healthy flight at time t and which can be represented by a (complete) set of indicators at that time t, the reference events being used to generate a module for detecting an abnormal event within the meaning of the invention.

[0087] It is also recalled that in one embodiment, the invention proposes to use first-level indicators which can, for example, be represented as vectors, each component of which represents a raw flight data or a meteorological data at a time t, and second-level indicators which can be obtained from first-level indicators; these may include, in particular, the time elapsed since a previous event, such as the start-up of the autopilot, or aggregated data, for example, an average value of an indicator over the last 10 seconds.

[0088] Figures 1 to 5 illustrate different ways of determining a threshold to discriminate between normal and atypical events.

[0089] In these figures, by way of example, we consider an event composed of two indicators, namely the normal acceleration nz and the longitudinal acceleration ny of an aircraft.

[0090] Fig. 1 presents the correlation of these indicators in each event, each point corresponds to an event, potentially from several flights.

[0091] Figure 2 shows: - by horizontal and vertical lines, a threshold for each of the indicators, approximately 1.31 for normal acceleration nz and 0.08 for longitudinal acceleration ny; and - a hatched area in which are located the points where at least one value among the normal acceleration nz or the longitudinal acceleration ny is greater than the corresponding threshold.

[0092] Considering each point of the hatched area as an event announcing a future accident would lead to a large number of false positives.

[0093] In [Fig.3], a hatched area is shown in which are located the points representing the events whose normal acceleration nz and longitudinal acceleration ny of the aircraft are simultaneously greater than their respective thresholds.

[0094] Considering each point in this hatched area as an event foreshadowing an impending accident can considerably reduce the number of false positives compared to the example in [Fig. 2]. On the other hand, events foreshadowing an event might not be detected.

[0095] Figure 4 shows the result of a classification of the events of the [Fig. 1] by a mathematical tool, or by manual selection. This tool produces: - an unhatched area corresponding to nominal points; - gridded areas corresponding to classes of atypical points.

[0096] From this classification, it is possible to define a combination of these pa to measure and compare this combination against a threshold represented by an oblique line.

[0097] For example, [Fig.5] illustrates an example in which a weighted sum of normal and longitudinal acceleration indicators is compared to a threshold, with points located in the hatched area triggering, when detected, an urgent transmission of at least part of the data collected during the flight.

[0098] With reference to Figures 6 and 7, we will now describe the main steps of a data backup method according to an embodiment of the invention. This method is implemented in an AER aircraft in flight. It uses an MDEA module for detecting an abnormal event according to the invention. Figure 9, which will be described later, shows a method according to the invention for generating this detection module.

[0099] During the flight of the aircraft, flight data dv and meteorological data dm are acquired during an acquisition step F10.

[0100] This flight data allows us to determine, for a given time t of the flight, indicators of an event occurring at that time t. This data is recorded in a MEM memory of the aircraft and part of this data is copied into a black box referenced FDR.

[0101] In a particular embodiment, among the aforementioned event indicators, a distinction is made between so-called first-level event indicators and so-called second-level event indicators.

[0102] First-level indicators can, for example, be represented as vectors, each component of which represents a flight data dv or a meteorological data dm at a time t.

[0103] Flight data constituting first-level indicators within the meaning of the invention may, for example, be selected from the following data: - the aircraft's roll angle, - the aircraft's pitch angle, - the aircraft's vertical speed, - the aircraft's ground speed, - an aircraft load factor, - the aircraft's altitude, - the aircraft's deviation from the glide path of an Instrument Landing System (ILS), - the aircraft's flap deflection, - the aircraft's slat deflection, and - the angle of the wind relative to the aircraft's heading.

[0104] Wind speed is an example of meteorological data.

[0105] An indicator considered at a time t (for example the roll angle), may in an example include values ​​of this indicator at the last N times t, t-1, ..., tNl. N may for example be chosen equal to 3, two consecutive times ti, t-(i+l) being for example separated by 1 second.

[0106] In the embodiment described here, second-level indicators can be obtained (i) by aggregating first-level indicators or (ii) by aggregating categorical parameters. It should be noted that in statistics, a categorical parameter takes the value of categories, as opposed to quantitative variables.

[0107] For example, a second-level indicator could be: (i) the sum, minimum value, or maximum value of first-level indicators over a period of time, for example the maximum value of the pitch angle during the last 5 seconds, a value obtained by principal component analysis of pairs of first-level indicators, for example from the pitch angle and the vertical speed; (ii) the time elapsed since the last TCAS alert, the number of TCAS alerts over a period of time, for example since the last 10 minutes, the total time of autopilot engagement since for example the last 5 minutes

[0108] An EV(t) event of the aircraft flight at time t is defined by the values ​​of all indicators at time t.

[0109] In the embodiment described here, during a step F15, subsets of indicators selected from this complete set of indicators are formed.

[0110] For example, if we have a set of 10 indicators (first or second level), we can create five subsets of indicators, such as: - the first subset can consist of indicators 1 to 4, - the second subset can consist of indicators 6 to 10, - the third subset can be made up of indicator 5, - the fourth subset can be made up of indicators 1, 5 and 9, - the fifth subset can be made up of indicators 3, 7 and 9.

[0111] During a step F20 of the data backup process, the MDEA module for detecting an abnormal event determines for an EV event occurring at a time t, and for each subset of indicators EI;, a partial opinion ap; on the normal (N) or atypical (A) nature of this event from the values ​​of the indicators of this event for this subset.

[0112] For example, for a given event, - in view of the values ​​of the indicators of the first subset at time T, the partial opinion determined by API for the first subset of indicators may be that this event is normal; - in view of the values ​​of the indicators of the second subset at time T, the opinion partial ap2 determined for the second subset of indicators may be that this event is atypical; - in view of the values ​​of the indicators of the third subset at time T, the partial opinion determined for the third subset of indicators may be that this event is normal; - in view of the values ​​of the indicators of the fourth subset at time T, the partial opinion determined for the fourth subset of indicators may be that this event is normal; and - in view of the values ​​of the indicators of the fifth subset at time T, the partial opinion determined for the fifth subset of indicators may be that this event is atypical.

[0113] For this purpose, the MDEA abnormal event detection module calculates, for each subset of indicators EI;, a normality score sn; based on the values ​​of only the indicators of the subset El at time t, using an AAi algorithm associated with this subset Eh, and compares this normality score sn; with a threshold TH; determined for this subset of indicators EI;.

[0114] For example, the more atypical the algorithm considers the event, the higher the normality score assigned to that event. As already mentioned and as detailed below, each AA algorithm was trained on reference events from safe flights to define a normality score for the subsets of indicators EI, as well as the threshold TH of that subset.

[0115] During a step F30, the MDEA abnormal event detection module determines, using a decision model, from the partial opinions ap, a final opinion AD on the normal nature N or atypical A of the event EV at time t.

[0116] If (and only if) the MDEA abnormal event detection module determines that the EV event at time t is an atypical event A, this module emits a distress signal SD.

[0117] Figure 8 illustrates an LD decision model that can be used by an MDEA abnormal event detection module according to the invention to generate an SD distress signal.

[0118] In this figure, five subsets of indicators Eli to EI5 associated with five algorithms AAi to AA5 are shown. Five indicators I6 to I10 are also shown. The indicators of the five subsets of indicators Eli to EI5 and the indicators I6 to ho are part of a set of indicators which represent an event of interest EV occurring during a time t of the flight of an aircraft.

[0119] By way of example, an indicator of the EI4 assembly is a level two indicator that represents a type of flight phase (e.g. takeoff, climb, cruise, descent, approach to landing...) obtained by a AEF state machine.

[0120] Each algorithm AA; determines a partial opinion ap; on the normal or atypical nature of the event of interest from the associated subset EI;.

[0121] In this example, the indicators I6 to Ro are indicators defined by the BEA.

[0122] The LD decision model is represented by the thick-lined frame.

[0123] In this example, an SD distress signal is generated if and only if at least One of the following five conditions is met: condition 1 / the partial API opinion determined for the first subset of indicators Eli and the partial API opinion determined for the second subset of indicators EI2 are both of the opinion that the EV event is atypical; condition 2 / the partial opinion ap2 determined for the second subset of indicators EI2 and the partial opinion ap3 determined for the third subset of indicators EI3 are both of the opinion that the EV event is atypical; condition 3 / a counter, which counts the number of consecutive times a categorical indicator of the fourth subset EI4 takes a determined modality, has exceeded a value X in a time window of one minute; Condition 4 / a value determined from the indicators of the fifth sub- The EI5 set of indicators is within a range defined by a filter; condition 5 / at least one risk of distress RD corresponding to a criterion defined by an indicator I6 to Ro of the BEA is detected.

[0124] In one embodiment of the invention, the abnormal event detection module MDEA can generate a distress signal SD if the following combination of indicator values ​​is detected: - (roll greater than 35° and pilot action to restore the pitch attitude and increase in roll) or (roll greater than 45° (BEA condition)) - (time interval between two stalls less than one minute and low thrust power of the engines) or (vertical speed greater than -9000 feet per minute (BEA condition)) - (dive angle greater than -11° and vertical speed greater than -7000 feet per minute) or (sharp angle greater than -20° (BEA condition)) - (obstacle detection by video analysis and detection of a vertical or horizontal avoidance command) - detection of pilot or co-pilot inability to fly based on indicators representing a video feed acquired by a camera facing the pilot, data measured by a sensor integrated into the (co)pilot's seat, words spoken in the cockpit... - (smoke detection in the cockpit) and (loss of image sharpness in a video stream acquired by a camera in the cockpit).

[0125] In the embodiment described here, if the event detection module an abnormal MDEA generates an SD distress signal, an aircraft COM communication module triggers the transmission (step F40) of flight data dv and meteorological data contained in MEM memory to an external receiver.

[0126] With reference to Figures 9 and 10, the main steps of a method for generating an abnormal event detection module are now described.

[0127] This process consists in particular of: - select from a set of indicators representative of an event of at least one safe flight, the indicators of the subsets of indicators EI; These subsets of indicators are those which will be used by the data backup process in step F15 already described; - teach each AA algorithm, associated with a subset of EI indicators, to generate a normality score based on the values ​​associated with these indicators at a given time; - calculate the TH threshold; of each subset of Eh indicators; and - define the LD decision model mentioned in step F30, for example illustrated in [Fig.8].

[0128] One or more TH thresholds; may be determined by a flight safety expert.

[0129] During an E10 step, flight data dv is obtained, and possibly some meteorological data acquired during healthy flights, for example from maintenance data from previous flights.

[0130] The indicators of the EI subsets; are selected during a step E15.

[0131] During a step E20, we learn from a plurality of EVk events from healthy flights, and for each subset of indicators El,, the algorithms AA i to generate a normality score sni>k for each EVk event.

[0132] Thus, for each subset of indicators EI;, with the dedicated learned algorithm, we obtain a distribution of these normality scores.

[0133] In one embodiment, the TH threshold; associated with a subset of Eh indicators is determined from the shape of the distribution of normality scores.

[0134] Thus, in this embodiment, during a step E30, for each subset of indicators EI;, a type TD; of the distribution of normality scores calculated for this subset is determined.

[0135] Figures 10 to 12 represent typical cases of normality score distributions that can be obtained by these algorithms. A particular embodiment of the invention proposes to automatically determine the threshold from the shape of these distributions.

[0136] Fig. 10 represents a first type of TDi distribution in which the number of occurrences of normality scores drops sharply from a normality score, For example, 2 in this figure. For this type of TDI distribution, the threshold can be chosen slightly higher than this value by taking a greater or lesser safety margin. Two thresholds are shown according to two options, 1 and 2, corresponding to two safety margins.

[0137] Figure 11 represents a second type TD2 distribution that is more continuous than that of Figure 10. In this example, events with a normality score around 6 have a much lower probability (the ordinate scale is logarithmic) than events with a normality score around 1.

[0138] For this type of TD2 distribution, we can define the value of the threshold beyond which we will consider an atypical event, by setting an acceptable rate of false positives, for example 1 per 1000. This choice is illustrated by option 3 in [Fig.1 1].

[0139] Figure 12 represents a third type TD3 two-component (i.e., two-contribution) distribution in which the second component appears to correspond to extreme normality score values ​​outside the main distribution. For this type of TD3 distribution, the threshold can be chosen between these extreme values ​​and the maximum value of the main distribution. This choice is illustrated by option 4 in Figure 12.

[0140] Figures 10 to 12 are only examples presented to illustrate that normality score distributions can be of very different types.

[0141] The invention proposes to automatically detect the type of distribution. For this purpose, a method of characterizing distribution tails can be used which produces an index called a "tail index" which characterizes the shape of the distribution tail.

[0142] Different methods can be applied to define this index, for example: - [1] the “Hill esthnator” method described in the document BM Hill, “A simple general approach to inference about the tail of a distribution”, Annals of Statistics vol.3, 1975, pp. 1163-1174.); or - [2] the method described in the document “J. Boonradsamee, W. Bodhisuwan, and U. Jaroengeratikun, “A New Selecting k Method of Hill's Estimator”, Thai Journal of Mathematics, 2021, pp. 153-163”

[0143] By way of illustration, a distribution tail characterization algorithm was applied to the four distributions in [Fig. 13]: - For the uniform distribution of [Fig.10](a), the tail index obtained is equal to 160 - For the beta distribution of [Fig.10](b), the tail index obtained is equal to -4.6 - For the normal distribution of [Fig. 10](c), the tail index obtained is equal to 0.31 - For the exponential distribution of [Fig.10](d), the tail index obtained is equal to 0.28

[0144] Thus, in one embodiment of the invention, at step E30, for each distribution of normality scores obtained for a subset of indicators, a tail characterization method of this distribution is applied to automatically determine the type of this distribution, and automatically set the threshold TH; associated with this subset of indicators according to this type, for example in accordance with options 1 to 4 defined previously.

[0145] Typically, for a distribution with a very high tail index, for example greater than 100, the threshold slightly above the highest normality score can be chosen in accordance with options 1 and 2.

[0146] Similarly, for a distribution whose tail index is positive but relatively low, for example between 0 and 0.5, the threshold can be chosen according to a predetermined number of false positives in accordance with option 3.

[0147] In a particular embodiment, the LD decision model can be determined by a flight safety expert.

[0148] During a step E40, for each event EVk, and for each subset of indicators Eh, a partial opinion apijk is determined on the normal nature N or atypical A of this event as a function of the values ​​of the indicators of this subset.

[0149] In the embodiment described here, this step involves comparing the normality score sni>ka associated with the EI values; calculated by the AA algorithm; at step E20 with the threshold TH; determined at step E30.

[0150] Then, during a step E50, the LD decision model is implemented to determine whether the EVk event is normal or atypical based on the partial opinions api>k.

[0151] In a manner known to a person skilled in the art of learning, an expert can use different training and validation datasets to determine the hyperparameters of AA algorithms;

[0152] The learned algorithms AA;, the thresholds TH; and the decision model LD are stored (step E60) in a memory of the MDEA abnormal event detection module so that they can be used during the flight of an aircraft to determine: - partial opinions ap; on the normal or atypical nature of an event during the flight using learned algorithms and thresholds; - the normal or typical nature of an event based on partial opinions and decision model and; - trigger a distress signal if (and only if) an event is considered atypical.

Claims

Demands

1. A method for generating a module (MDEA) for detecting an abnormal event occurring during flight in an aircraft, this method being implemented by a computer and comprising: - a step (E10) of obtaining flight data acquired during flights considered to be safe; - for at least one subset of indicators selected from a set of indicators obtained from said data and representative of a reference event that occurred during a moment of said safe flight, a step (E30) of associating, with this subset (Eb) of indicators, a threshold (TH;) allowing for the determination of a partial opinion on the normal or atypical nature of this reference event based on the values ​​of the indicators of said subset (EI;); - a step of configuring said detection module (MDEA) with: (i) the indicators and the threshold (TH;) of each subset of indicators, and with (ii) a decision model (DL) enabling the determination of a definitive opinion on the normal or atypical nature of an event of interest from partial opinions determined for that event of interest, a partial opinion being determined for each subset of indicators from (i) the values ​​of those indicators obtained from data representative of the event of interest and (ii) the threshold associated with that subset; - said abnormal event detection module being configured to, when implemented in an aircraft in flight, generate or not generate a distress signal according to said final opinion determined for an event of interest occurring during the flight.

2. A method for generating according to claim 1 comprising, for a plurality of subsets (Eh), indicators representative of reference events that occurred during said safe flights: - a step (E20) of calculating a normality score (sni>k) for each (EVk) of said reference events from the values ​​of the indicators of said subset (EI;) for this event (EVk); - the threshold (TH;) associated (E30) with a subset (EI;) of indicators being determined from a type of distribution of said normality scores calculated for this subset (Ef).

3. Generation method according to claim 2 wherein said threshold (TH;) is determined (E30) as a function of the shape of a tail of said distribution of said normality scores (sni>k).

4. A generation method according to claim 2 or 3 wherein said normality scores (sni>k) calculated (E20) for a subset of indicators (EI;) are obtained by a learning algorithm (AA;) associated with this subset.

5. A generation method according to any one of claims 1 to 4, wherein the decision model (LD) is configured to take into account a plurality of partial opinions determined for a plurality of subsets (Eli) of indicators during a time window in order to determine the final opinion on whether a reference event is normal or atypical.

6. A module for detecting an abnormal event during a flight in an aircraft, this module being characterized in that it was generated by a method according to any one of claims 1 to 5.

7. A method for saving flight data acquired during a flight of an aircraft, said aircraft (AER) comprising a module (MDEA) for detecting an abnormal event according to claim 6, said method comprising: - a step of collecting data (dv, dm) during the flight and storing said collected data in a memory (MEM) of the aircraft; - a step of copying at least a part of said collected data into a black box (FDR) of the aircraft; - a step of providing said collected data as input to said detection module (MDEA); and - a step of triggering the transmission of at least a part of the collected data to an external receiver on and only on detection of a distress signal (SD) generated by said module.

8. A device for saving flight data acquired during a flight of an aircraft, said aircraft comprising a module for detecting an abnormal event according to claim 6, said device comprising: - a module for collecting data during flight and storing said collected data in a memory (MEM) of the aircraft; - a module for copying at least a portion of said collected data into a black box (FDR) of the aircraft; and - a module for providing said collected data as input to said detection module; and - a triggering module for the transmission of at least part of the collected data to an external receiver on and only on detection of a distress signal (SD) generated by said abnormal event detection module.

9. Aircraft comprising a data backup device according to claim 8 or a processor configured to implement an abnormal event detection module according to claim 6.

10. Computer program (PG) comprising instructions for carrying out the steps of the generation process according to any one of claims 1 to 5 and / or instructions for carrying out the steps of the saving process according to claim 7 when said program is executed by a computer.

Images