FR3109436A1 - Method and electronic device for determining at least one recommended action for an aircraft, associated computer program and electronic display system - Google Patents

Abstract

Method and electronic device for determining at least one recommended action for an aircraft, associated computer program and electronic display system This method for determining at least one recommended action for an aircraft is implemented by a device determination electronics and comprises the following steps: - acquisition of a set of image(s) (24) of geo-referenced zone(s) relating to the aircraft, - reception of a group (28) of contextual data(s) relating to the aircraft, - calculation of a list (32) of action(s) recommended for the aircraft, via an algorithm of artificial intelligence and from the set of image(s) (24) of geo-referenced zone(s) and the group (28) of contextual data(s), the algorithm of artificial intelligence comprising as input variables, on the one hand, the set of image(s) (24) of geo-referenced zone(s), and on the other hand, the group (28) of contextual data(s), and as an output variable e the list (32) of recommended action(s). Figure for the abstract: Figure 2

Description

Method and electronic device for determining at least one recommended action for an aircraft, associated computer program and electronic display system

The present invention relates to a method for determining at least one action recommended for an aircraft, the method being implemented by an electronic determination device.

The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement such a determination method.

The invention also relates to an electronic device for determining at least one action recommended for an aircraft, and an electronic display system comprising such a determination device.

The invention relates to the field of decision support systems for a user of an aircraft, such as an aircraft pilot, in particular for the real-time management of complex problems, such as an emergency landing .

As the aeronautical context becomes more complex (density of air traffic, ambiguous situations, speed of technological means, number of actors in a theater of operation, etc.), a problem-solving support system is desirable for operators.

Current decision support systems are essentially based on a detailed analysis of human business know-how and the capture of its knowledge, so that the system can manage the largest possible number of cases already seen by humans.

However, these decision support systems are not very robust when faced with new contexts or situations.

The object of the invention is then to propose a method, and an associated electronic device, for determining at least one action recommended for an aircraft, which is more robust and effective vis-à-vis new contexts or situations.

To this end, the subject of the invention is a method for determining at least one recommended action for an aircraft, the method being implemented by an electronic determination device and comprising the following steps:

- acquisition of a set of image(s) of geo-referenced zone(s) relating to the aircraft,

- reception of a group of contextual data(s) relating to the aircraft,

- calculation of a list of action(s) recommended for the aircraft, via an artificial intelligence algorithm and from the set of image(s) of geo-referenced zone(s) s) and the group of contextual data(s), the artificial intelligence algorithm comprising as input variables, on the one hand, the set of image(s) of geo-zone(s) referenced, and on the other hand, the group of contextual data(s), and as an output variable the list of recommended action(s).

Thus, with the method according to the invention, the calculation of the list of recommended action(s) from the artificial intelligence algorithm and the set of image(s) of zone(s) geo-referenced(s) and the group of contextual data(s) allows greater robustness and greater efficiency vis-à-vis new contexts or situations. The artificial intelligence algorithm then adapts to new contexts or situations, by processing as input variables the set of image(s) of geo-referenced zone(s) and the group of data contextual(s), and delivering as output the list of recommended action(s) from these input variables.

Preferably, the artificial intelligence algorithm is trained via reinforcement learning, for example via a Q-learning algorithm with a neural estimator (DQN) or even an actor-critic type policy gradient algorithm, such as a DDPG algorithm, a PPO algorithm or an SAC algorithm.

When the reinforcement learning is an actor-critic type policy gradient algorithm, the artificial intelligence algorithm preferably further comprises at least one neural network and the reinforcement learning comprising at least one other neural network .

According to other advantageous aspects of the invention, the determination method comprises one or more of the following characteristics, taken in isolation or in all technically possible combinations:

the method further comprises a step of displaying, on a display screen, the list of action(s) recommended for the aircraft;

- the set of image(s) of geo-referenced area(s) comprises at least one image from among the group consisting of: an image of an area of interest, such as an image containing one or several geo-referenced meteorological information; an image of a threat area, such as an image containing one or more geo-referenced obstacle information; an image of an area of uncertainty containing geo-referenced information indicating, for a given position, the probability that an object may be present at this position;

- the group of contextual data(s) comprises at least one data item from the group consisting of: a quantity of resource(s) available, such as a quantity of fuel, a number of equipment(s) ) carried away; a number of identified entity(ies) present in a respective geo-referenced area; and a number of unknown entity(ies) present in a respective geo-referenced area;

- the list of recommended action(s) comprises at least one action from the group consisting of: a heading instruction, a speed instruction, a position instruction, an altitude instruction, use of a respective sensor, a use of a respective equipment, a selection of a mode/parameter of use of a sensor or equipment;

- the artificial intelligence algorithm includes at least one neural network;

the artificial intelligence algorithm preferably comprising a network of neurons headed by a first sub-network of neurons having as input variable the set of image(s) of geo-referenced zone(s) ( s), a second sub-network of neurons having as input variable the group of contextual data(s), and at the bottom, a third sub-network of neurons connected at the output of the first and second sub-networks of neurons and having as output variable the list of recommended action(s);

the first sub-network of neurons preferably comprising layers of convolutions, the second and third sub-networks of neurons each being a multilayer perceptron comprising dense layers;

- the artificial intelligence algorithm is obtained via reinforcement learning;

reinforcement learning preferably being selected from the group consisting of: a Q-learning algorithm with a neural estimator; an actor-critic type policy gradient algorithm, such as a DDPG algorithm, a PPO algorithm or an SAC algorithm;

when the reinforcement learning is an actor-critic type policy gradient algorithm, the artificial intelligence algorithm preferably further comprising at least one neural network called actor(s) and the reinforcement learning comprising at least one other so-called critical neural network(s);

- reinforcement learning involves solving an optimization problem of at least one reward function,

reinforcement learning preferably implementing a plurality of reward functions with respective learning of the artificial intelligence algorithm for each reward function,

when the artificial intelligence algorithm comprises at least one neural network, the reinforcement learning then preferably also comprises obtaining a set of network weights for each reward function;

- the method further comprises a step of selecting a reward function from among the plurality of reward functions, for example from an interaction of a user;

- the calculation step includes several successive iterations of the artificial intelligence algorithm, in order to have an anticipated implementation of the artificial intelligence algorithm at subsequent successive instants;

- the method also includes an evaluation step, via a game theory algorithm, of one or more preponderant parameters in the calculation of the list of recommended action(s).

The invention also relates to a computer program comprising software instructions which, when they are executed by a computer, implement a determination method as defined above.

The invention also relates to an electronic device for determining at least one recommended action for an aircraft, the device comprising:

- an acquisition module configured to acquire a set of image(s) of geo-referenced zone(s) relating to the aircraft,

- a reception module configured to receive a group of contextual data(s) relating to the aircraft,

- a calculation module configured to calculate a list of recommended action(s) for the aircraft, via an artificial intelligence algorithm and from the set of zone(s) image(s) geo-referenced data(s) and of the group of contextual data(s), the artificial intelligence algorithm comprising as input variables, on the one hand, the set of image(s) of zone( s) geo-referenced, and on the other hand, the group of contextual data(s), and as an output variable the list of recommended action(s).

The invention also relates to an electronic display system for an aircraft, the system comprising:

- a display screen,

- an electronic device for determining at least one action recommended for the aircraft, the determination device being as defined above, and

- a display module, on the display screen, of a list of action(s) recommended for the aircraft.

These characteristics and advantages of the invention will appear more clearly on reading the following description, given solely by way of non-limiting example, and made with reference to the appended drawings, in which:

FIG. 1 is a schematic representation of an aircraft comprising an electronic display system according to the invention, the system comprising an electronic device for determining at least one action recommended for the aircraft, connected to a display screen , avionics systems and a database,

FIG. 2 is a schematic representation of an artificial intelligence algorithm implemented by the electronic determination device of FIG. 1, in order to determine at least one recommended action for the aircraft;

FIG. 3 is a flowchart of a method, according to the invention, for determining at least one action recommended for the aircraft, the method being implemented by the determination device of FIG. 1; And

FIG. 4 is a flowchart representing several successive iterations of the artificial intelligence algorithm during the calculation of the list of actions recommended for the aircraft, this for anticipated implementation of the artificial intelligence algorithm at instants successive later ones.

In FIG. 1, an aircraft 10 comprises several avionic systems 12, a database 14, such as a cartographic database, and an electronic display system 16 connected to the avionic system 12 and to the database 14, the display system 16 comprising a display screen 18 and an electronic device 20 for determining at least one recommended action for the aircraft 10.

The aircraft 10 is for example an airplane, as represented in the example of FIG. 2. Alternatively, the aircraft 10 is a helicopter, a drone that can be controlled remotely by a pilot, or else an autonomous aircraft without an operator. Those skilled in the art will observe that in the case where the aircraft 10 is an autonomous aircraft without an operator, it preferably does not include an on-board display system inside the aircraft, the display system then being placed on the ground.

The avionic systems 12 are known per se and are capable of transmitting to the display system 16, and in particular to the electronic determination device 20, various avionic data, for example so-called "aircraft" data, such as position, speed , the acceleration, the orientation, the heading or even the altitude of the aircraft 10, and/or so-called “navigation” data, such as a flight plan.

The avionic systems 12 are also able to receive instructions, or even commands, from the determination device 20. These avionic systems 12 able to receive instructions, comprise for example at least one system from among:

a flight control system, also denoted FCS ( Flight Control System ) or FBW ( Fly By Wire ), to act on a set of control surfaces and actuators of the aircraft. In the case of a fixed-wing aircraft, the control surfaces are, for example, ailerons, the elevator or the rudder. In the case of a rotary-wing aircraft, the control surfaces are, for example, collective pitch, cyclic pitch or tail rotor pitch;

an engine control system, also denoted ECU ( Engine Control Unit ) for varying the energy delivered by an engine of the aircraft, such as a jet engine, a turboprop or even a turbine;

- at least one guidance system, such as an autopilot device, also denoted AFCS ( Auto-Flight Control System ), also called autopilot and denoted PA or AP ( Automatic Pilot ), such as an aircraft flight management system, also denoted FMS ( Flight Management System ). In addition, the guidance system is an auto-thrust device, also called auto-throttle.

The database 14 is optional, typically a navigation database, and is known per se. The navigation database is also called NAVDB (from the English NAVigation Data Base ), and includes in particular data relating to prohibited flight spaces or zones, data relating to landing strips on which the aircraft 10 is likely to land, these data being typically a position of a landing runway threshold, an orientation of the landing runway, a runway length, an altitude or a decision point, etc.

In the example of Figure 1, the database 14 is a database external to the determination device 20. Alternatively, not shown, the database 14 is a database internal to the determination device 20.

The display system 16 is, for example, a head-down display system and/or a head-up display system, also called HUD ( Head-Up Display ). The head-down display system is, for example, a navigation data display system (English Navigation Display ).

When display system 16 is a head-down display system, display screen 18 is a head-down display screen, and when display system 16 is a head-up display system, the display screen 18 is a head-down display system. The display screen 18 is a head-up display screen. As a variant, the display system 16 is a remote display system, in particular a display system external to the aircraft 10, such as a display system in a ground station, or even the remote control or the vision glasses of a drone operator, in particular when the aircraft 10 is an autonomous aircraft without an operator.

The electronic determination device 20 is configured to determine at least one action recommended for the aircraft 10. The determination device 20 then forms a decision aid device, in particular for an operator, such as the pilot of the aircraft 10 This decision support device is particularly useful for the real-time management of complex problem(s), and aims in particular to help the pilot improve the management of critical or complex situation(s). (s), such as an emergency landing.

The electronic determination device 20 comprises a module 22 for acquiring a set of image(s) 24 of geo-referenced zone(s) relating to the aircraft 10, a module 26 for receiving a group of contextual data(s) 28 relating to the aircraft 10, and a module 30 for calculating a list of recommended action(s) 32 for the aircraft 10 .

As an optional addition, the determination device 20 comprises a module 34 for displaying the list 32 of recommended action(s) on the display screen 18 and/or a module 36 for transmitting the list 32 of recommended action(s) to at least one corresponding avionic system 12 .

As a further optional addition or as a variant, the determination device 20 comprises a module 38 for evaluating the predominant parameter(s) in the calculation of the list 32 of recommended action(s).

As a further optional addition or as a variant, the determination device 20 comprises a module 40 for selecting a reward function from among a plurality of reward functions, the reward function being associated with an artificial intelligence algorithm implemented by the calculation module 30 for obtaining the list 32 of recommended action(s).

In the example of Figure 1, the electronic determination device 20 comprises an information processing unit 50 formed for example of a memory 52 and a processor 54 associated with the memory 52.

In the example of Figure 1, the acquisition module 22, the reception module 26 and the calculation module 30, as well as optionally the display module 34, the transmission module 36, the module of evaluation 38 and the selection module 40, are each produced in the form of software, or a software brick, executable by the processor 54. The memory 52 of the electronic determination device 20 is then capable of storing software for acquisition of the set of image(s) 24 of geo-referenced zone(s) relating to the aircraft, software for receiving the group 28 of contextual data(s) and software for calculating the list 32 of recommended action(s) for the aircraft. As an optional addition, the memory 52 of the electronic determination device 20 is capable of storing software for displaying the list 32 of recommended action(s) on the display screen 18, software for transmitting the list 32 of recommended action(s) to at least one respective avionics system 12, software for evaluating the predominant parameter(s) in the calculation of the list 32 of recommended action(s) (s) and software for selecting the reward function from the plurality of reward functions. The processor 54 is then capable of executing each of the software programs among the acquisition software, the reception software and the calculation software, as well as, as an optional complement, the display software, the transmission software, the evaluation software and/or selection software.

When in a variant, not shown, the database 14 is an internal database of the determination device 20, it is typically able to be stored in a memory of the determination device 20, such as the memory 52.

In a variant not shown, the acquisition module 22, the reception module 26, and the calculation module 30, as well as in optional addition the display module 34, the transmission module 36, the evaluation module 38 and the selection module 40, are each made in the form of a programmable logic component, such as an FPGA ( Field Programmable Gate Array ), or even in the form of a dedicated integrated circuit, such as a ASIC (from the English Application Specific Integrated Circuit ).

When the electronic determination device 20 is produced in the form of one or more software, that is to say in the form of a computer program, it is also capable of being recorded on a medium, not shown, readable by computer. The computer-readable medium is, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. By way of example, the readable medium is an optical disc, a magneto-optical disc, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card. On the readable medium is then stored a computer program comprising software instructions.

The acquisition module 22 is configured to acquire the set of image(s) 24 of geo-referenced zone(s) relating to the aircraft 10, for example from the database 14. The set of image(s) 32 is typically constructed prior to the implementation of the determination device 20.

Each image 24 is for example a geo-referenced map, or even an aeronautical cartography, and is typically constructed from raw data, such as data from a navigation database, and/or from virtual representations of a context surrounding the aircraft 10. The image or images 24 are for example constructed by level of importance according to a mission to be carried out by the aircraft 10, or even by aeronautical function. A first image 24 thus represents for example the important aeronautical zone(s), in particular those where a final state or a desired objective is located. A second image 24 then represents an area of intermediate importance, and a third image 24 represents an area of low importance. In addition, other images 24 typically represent a geographical area to be avoided, determined for example from meteorological information, or even a no-fly zone.

In addition, during the preliminary construction of these images 24, business data from high-level behavioral analysis models, such as doctrines, business uses, etc., are also taken into account to enrich these images 24 in order to to obtain geo-referenced graphic maps containing business criteria. In addition or as a variant, these graphic maps are also built from physical and behavioral data of a simulation level, such as data relating to the position of the aircraft, meteorological information, etc.

As a further optional addition, these images 24, or maps, are interactive, and can then be modified by a user, typically via a man-machine interface, such as a graphical interface.

As a variant, the image(s) 24 are images acquired via one or more sensors, not shown.

The set of image(s) 24 of geo-referenced area(s) then comprises at least one image 24 from among the group consisting of: an image of an area of interest, such as an image containing one or more geo-referenced meteorological information; an image of a threat area, such as an image containing one or more geo-referenced obstacle(s) information; an image of an area of uncertainty containing geo-referenced information indicating, for a given position, the probability that an object may be present at this position.

The set of images 24 comprises for example M images 24. The images 24 are for example two-dimensional images, and are represented either in color, for example on three RGB channels, or in gray level. The images 24 are for example images of rectangular shape and of dimension N₁ x N₂, where N₁, NOT₂represent respective numbers of pixels, or even square in which case N₁is of course equal to N₂.

The reception module 26 is configured to receive the group 28 of contextual data(s) relating to the aircraft 10, for example from the database 14.

The group 28 of contextual data(s) comprises for example at least one data item from the group consisting of: a quantity of available resource(s), such as a quantity of fuel, a number of equipment (s) carried away; a number of identified entity(ies) present in a respective geo-referenced area; and a number of unknown entity(ies) present in a respective geo-referenced area.

The group 28 of contextual data(s) is for example in the form of a vector of dimension P ₁ . This vector is then used to concatenate all the contextual data in the form of a single representation.

The calculation module 30 is configured to calculate the list 32 of action(s) recommended via an artificial intelligence algorithm from the set of image(s) 24 of geo-zone(s). referenced(s) and of the group 28 of contextual data(s), the artificial intelligence algorithm comprising as input variables, on the one hand, said set of image(s) 24, and on the other hand, the group 28 of contextual data(s), and as an output variable the list 32 of recommended action(s).

At the output of the learning algorithm, the list 32 of recommended action(s) is for example in the form of a vector of size P ₂ concatenating a set of recommended action(s) for the aircraft 10, that is to say a set of action(s), such as instructions or commands, recommended for the aircraft 10.

By recommended action is typically meant an instruction or a command that it is advisable to follow or perform for the aircraft 10, for example via a corresponding avionic system 12 .

The list 32 of recommended action(s) then comprises for example at least one action chosen from the group consisting of: a heading setpoint, a speed setpoint, a position setpoint, an altitude setpoint, a use of a respective sensor, use of a respective equipment, selection of a mode/parameter of use of a sensor or equipment.

The artificial intelligence algorithm is configured to receive as input M 24 images of dimensions N₁ x N₂and a vector of dimension P₁representing the group 28 of contextual data(s), and outputting a vector of dimension P₂representing the list 32 of recommended action(s), where M, N₁, NOT₂, P₁and P₂are each an integer greater than or equal to 1.

The artificial intelligence algorithm includes for example at least one neural network 60, visible in Figure 2.

In the example of FIG. 2, the artificial intelligence algorithm is the neural network 60 headed by a first sub-network of neurons 62 whose input variable is the set of image(s) 24 of geo-referenced zone(s), a second sub-network of neurons 64 having as input variable the group 28 of contextual data(s), and at the bottom, a third sub-network of neurons 66 connected at the output of the first 62 and second 64 sub-networks of neurons and having as output variable the list 32 of recommended action(s). The first sub-network of neurons 62 typically comprises convolution layers, and is also denoted CNN (from the English Convolutional Neural Network ). The second 64 and third 66 neural sub-networks are each typically a multilayer perceptron, also denoted MLP ( Multi Layer Perceptron ) comprising dense layers.

The artificial intelligence algorithm is for example obtained via learning from reference input variables, that is to say from a part of image(s) of geo-zone(s). referenced reference(s), and on the other hand contextual reference data(s), and with a list of recommended reference action(s) as output variables.

The artificial intelligence algorithm is preferably obtained via reinforcement learning, such as deep reinforcement learning, making it possible to learn a policy of action(s) for a given control problem and having the objective of maximizing a return function G.

The feedback function G of such reinforcement learning verifies for example the following equation:

where G _t represents the feedback function G at time t,

R _t represents the reward at time t, and

γ represents a factor, with a value between 0 and 1 and penalizing future rewards.

An optimal action policy then makes it possible to maximize a state-value function, such as the state -value function V, and the optimal action policy, denoted π _* then verify for example the following equation:

where the state value function V verifies for example the following equation:

Alternatively or in addition, the optimal action policy π_* maximizes an action value functionTo ction-value function)Q, for example according to the following equation:

where the action value function Q verifies for example the following equation:

where S _t represents the observation state of the system at time t,

s represents a value taken by the observation state in a state space,

A _t represents the action that the system must take at time t; And

A represents the value of this action in an action space.

In other words, the reinforcement learning of the artificial intelligence algorithm involves solving an optimization problem for at least one reward function, denoted R.

Reinforcement learning is for example chosen from the group consisting of : a Q-learning algorithm with a neural estimator, also called DQN ( Deep Q ear n ing ); an Actor-Critic type policy gradient algorithm.

The actor-critic type policy gradient algorithm typically involves a neural network to learn which actions are defined by the optimal policy, corresponding to the actor, while judging the chosen policy, corresponding to the critics.

The actor-critical type policy gradient algorithm is for example the DDPG ( Deep Determining Policy Gradient ) algorithm, the PPO ( Proximal Policy Optimization ) algorithm, or the SAC ( Soft Actor Critic ) algorithm.

When reinforcement learning is an actor-critic type policy gradient algorithm, the artificial intelligence algorithm then typically comprises at least one neural network 60 called actor(s), and reinforcement learning comprises at minus one other neural network, not shown, called critical(s).

In the case of an actor-critical type algorithm, the function of value(s), such as the state value function V or the action value function Q is then learned by the critical entity, such as that the critical neural network, and the actor entity, such as the actor neural network, learns the policy, i.e. the set of actions to be taken in a given state, in order to go into the direction suggested by the critical entity.

Thus, and in other words, in actor-critic type deep reinforcement learning, a first neural network learns the value function, either the state value function V or the action value function Q, and a second neural network learns the policy π.

Those skilled in the art will also understand that the so-called actor neural network and the so-called critical neural network preferably have similar architectures, with the difference that the so-called critical neural network takes as input the action in addition to the current state, and that their outputs are of a different nature. The output of the critical neural network is an estimate of the state value function or the action value function for a given transition. This estimate is for example formed from the state at time t, the action taken at time t, the reward obtained by this action, and the state obtained at the following time t+1. The actor neural network output corresponds to the list of actions to be taken at time t.

Those skilled in the art will then observe that the neural network 60 represented in FIG. 2 is the actor neural network, the critical neural network not being represented.

As an optional complement, reinforcement learning preferably implements a plurality of reward functions with respective training of the artificial intelligence algorithm for each reward function. The reward function is typically decided according to the objective sought, and is for example linked to the type of images 24 and/or contextual data of the group 28 supplied as input variables of the artificial intelligence algorithm, and the type of objective set for the mission, in particular the type of recommended action expected at the output of the artificial intelligence algorithm.

When the AI algorithm includes at least one 60 neural network, then reinforcement learning preferably includes obtaining a set of network weights for each reward function.

For example, if the mission corresponding to the set objective is an emergency landing mission, then the policy should penalize passage through turbulent weather zones, and on the contrary reward the choice of an area of nearest landing point to the access point. By then typically taking a first metric corresponding to the proximity to the meteorological zones to be avoided, a second metric corresponding to the distance to the selected runway and a third metric corresponding to the duration of availability of the runway, the reward function R will verify the following equation:

where R _t represents the reward function at time instant t.

represents a parameterized aggregation function, linear or not,

F ₁ represents a utility function of the first metric corresponding to the proximity to the meteorological zones to be avoided,

F ₂ represents a utility function of the second metric corresponding to the proximity to the landing strip;

F ₃ represents a utility function of the third metric corresponding to the duration of availability of the track.

The display module 34 is configured to display the list 32 of recommended action(s) on the display screen 18. The display module 34 is for example configured to display this list 32 of action (s) recommended in the form of symbol(s) or icon(s) superimposed on one or more images 24 of the set of image(s) of geo-referenced zone(s) relating to the aircraft 10, in order to obtain a resulting image 70.

In the example of FIG. 2, the recommended action is a position instruction, represented in the form of an arrow on the resulting image 70, the object position of said instruction corresponding to the end of the arrow.

As a variant or in addition, the transmission module 36 is configured to transmit the list 32 of action(s) recommended(s) to at least one avionic system 12 respectively. The transmission module 36 then makes it possible to directly transmit one or more instructions or commands, included in said list 32 of recommended action(s), to at least one corresponding avionic system 12, in order to act on the behavior of the aircraft 10.

As a further optional addition, the evaluation module 38 is configured to evaluate, via a game theory algorithm, one or more preponderant parameters in the calculation of the list 32 of recommended action(s). The game theory algorithm is then, for example, adapted to receive as input the set of image(s) 24 of geo-referenced zone(s) relating to the aircraft 10, from a hand, and on the other hand, the list 32 of recommended action(s), then to deliver as output one or more preponderant parameters in the calculation performed by the artificial intelligence algorithm.

The preponderant parameter(s) supplied by the evaluation module 38 are for example one or more images 24, that is to say or several graphics cards, which have been particularly taken into account in the calculation carried out by the algorithm of 'artificial intelligence.

The evaluation module 28 then makes it possible to provide the user with the causal explanation of the proposals of the artificial intelligence algorithm, in particular of the proposals of the neural network 60, and then makes it possible to confront the vision of the problem from the point view of the artificial intelligence algorithm with the user's business vision and the criteria that the latter would have considered to be preponderant in view of the current situation of the aircraft 10.

The game theory algorithm implemented by the evaluation module 38 is for example the algorithm based on Shapley's value making it possible to give the contribution of each player to an overall score. More generally, when several attributes contribute to the same overall score or result, the Shapley value makes it possible to calculate the contribution of each attribute to this overall score. The calculation of this value corresponds to the average of the contributions obtained thanks to this attribute, by taking all the possible combinations of attributes.

The Shapley value is for example calculated according to the following equation:

where X ₁ , …, X _p represent the different attributes.

Val is the global score function.

Those skilled in the art will then understand that the attributes correspond to the various parameters influencing the choice of the artificial intelligence algorithm, such as the neural network 60, these influencing parameters being in particular the images 24 or graphics cards, of the set provided as input variable of the artificial intelligence algorithm. The value function Val then corresponds to the probability that the artificial intelligence algorithm, such as the neural network 60, chooses the action selected as output.

That is, the game theory algorithm, such as the Shapley value-based algorithm, allows - for an action decision given by the artificial intelligence algorithm, i.e. for example by the so-called actor neural network - to find the contribution of each parameter, for example of each image 24 of the set supplied as input to the artificial intelligence algorithm, and thus to select the one having the greatest impact.

According to this optional addition, the image 24 having had the greatest impact on the list 32 of recommended action(s) provided by the artificial intelligence algorithm is additionally displayed on the display screen 18 by the display module 34.

Those skilled in the art will then understand that, in the example of FIG. 2, the resulting image 70, displayed by the display module 24, corresponds to the image 24 having had the greatest impact on which is superimposed the list 32 of actions recommended by the calculation module 30.

As a further optional addition, the selection module 40 is configured to select a reward function from among the plurality of reward functions R, for example based on a user interaction. The selection thus made then influences the learning of the artificial intelligence algorithm implemented by the calculation module 30.

According to this optional complement, the selection module 40 is indeed configured to then select the reward function R which will be used for learning by reinforcement of the artificial intelligence algorithm, as described above.

The person skilled in the art will then understand that this selection ultimately has an influence on the list 32 of recommended action(s) which will be calculated by the calculation module 30.

As a further optional addition, the calculation module 30 is configured to carry out several successive iterations of the artificial intelligence algorithm, in order to have an anticipated implementation of the artificial intelligence algorithm at successive subsequent instants, and then calculate different successive lists 32 of recommended action(s), these successive lists 32 being associated with successive subsequent instants.

According to this optional complement, the calculation module 30 is then configured to, starting from a situation, that is to say from a state, at a time instant t, calculate a list 32 of recommended actions at a first subsequent instant, denoted t+1, and then give a situation or state of the aircraft at this subsequent instant t+1. The list 32 of action(s) recommended at this later instant t+1 is in particular likely to be displayed in the display module 34 on the display screen 18. Iteratively and starting from the calculated situation of the aircraft at this subsequent instant t+1, the calculation module 30 is then configured to calculate during a new iteration a list 32 of recommended action(s) at a second subsequent instant, denoted t+2, in order to obtain the situation of the aircraft at this second subsequent instant t+2, and so on, as represented in figure 4.

The operation of the electronic display system 16 according to the invention, and in particular of the electronic determination device 20, will now be described with reference to FIG. 3 representing a flowchart of the method according to the invention for determining at least a recommended action for the aircraft 10, as well as an optional addition with regard to FIG. 4 representing a flowchart of the calculation of the list 32 of recommended actions when this calculation comprises several successive iterations of the artificial intelligence algorithm.

During an initial step 100, the determination device 20 acquires, via its acquisition module 22, the set of image(s) 24 of geo-referenced zone(s) relating to the aircraft 10, for example from the database 14. These images 24 stored in the database 14 have typically been constructed beforehand via a respective man-machine interface, such as a graphical interface.

These images 24 are then typically discretized spatial data on the area of interest, such as meteorological data, threat areas, areas of interest, the direction in which a threat is located, trajectories of past obstacles or future, areas of uncertainty in which no information is available, or even representations of the links between entities connected on the same network by a mesh.

During a next step 110, the determination device 20 receives, via its reception module 26, the group 28 of contextual data(s) relating to the aircraft 10, this group 24 of data( s) contextual(s) being for example in the form of a vector of dimension P ₁ . This contextual data is then typically non-spatial data, for example the number of resources available, such as a quantity of fuel, or a number of equipment carried, the number of identified entities present on the scene, that is i.e. in a respective geo-referenced area, or even the number of unknown entities present on the scene.

The determination device 20 then calculates, during the next step 120 and via its calculation module 30, the list 32 of recommended action(s) for the aircraft 10 via the artificial intelligence algorithm, such as the neural network 60. The list 32 of recommended action(s) is then typically in the form of a respective vector of size P ₂ concatenating a set of recommended actions for the aircraft 10. These actions recommended are for example a heading, a speed, a position to be reached, an altitude, the choice of equipment, the selection of a sensor, the selection of a mode/parameter for use of a sensor or equipment.

As described previously, the artificial intelligence algorithm implemented by the calculation module 30 results from learning, preferably from reinforcement learning. This reinforcement learning is for example a DQN algorithm, i.e. a Q-learning algorithm with a neural estimator, or an actor-critic type policy gradient algorithm.

At the end of the calculation step 120, the determination device 20 displays, during step 130 and via its display module 34, the list 32 of recommended action(s) on the display screen 18 and/or transmits, via its transmission module 36 and during step 140, the list 32 of action(s) recommended(s) to at least one avionic system 12 respectively.

As an optional complement, during the optional step 150, the determination device 20 evaluates, via its evaluation module 38, implementing the game theory algorithm, the predominant parameter(s) in the action calculation 32 (s) recommended. This or these preponderant parameters are for example the image or images 24 which have had the greatest impact on the calculation, carried out by the artificial intelligence algorithm, of the list 32 of recommended action(s). The game theory algorithm is for example based on the Shapley value.

During this evaluation step 150, the predominant parameter or parameters are optionally displayed on the display screen 18.

As represented in FIG. 3, this evaluation step is carried out, on the one hand, from the list 32 of recommended actions calculated by the calculation module 30 during step 120, and on the other hand, from the set of images 24 acquired during step 100 by the acquisition module 22 and supplied as input to the artificial intelligence algorithm.

As a further optional addition, during an optional step 160, the determination device 20 selects, via its selection module 40 and typically on the basis of user interaction, a reward function from among the plurality of reward functions. R used for artificial intelligence algorithm learning. The user interaction leading to this selection results for example from the list 32 of recommended action(s) displayed during step 130 and/or from the preponderant parameters evaluated during step 150 and allowing the user to better understand the parameters having influenced the calculation of the list 32 of recommended action(s).

This selection of reward function during step 160 will then have an influence on a future calculation of list(s) 32 of recommended action(s), this calculation of list(s) 32 of action(s) ) predicted(s) then being performed using the learned set of network weights with the selected reward function.

As a further optional addition, the calculation step 120 comprises several successive iterations of the artificial intelligence algorithm, in order to have an anticipated implementation of the artificial intelligence algorithm at successive subsequent instants, as shown in Figure 4.

In the example of FIG. 4, starting from a situation, or a state, of the aircraft at the time instant t represented by a sub-step 200, the calculation module 30 requests during a sub - step 210 and with the artificial intelligence algorithm, the list 32 of action(s) recommended at the first subsequent time instant t+1, which makes it possible to succeed during a sub-step 220 to a situation, or state, of the aircraft 10 at the first subsequent time instant t+1. The list 32 of action(s) recommended at the first subsequent time instant t+1 and materializing the situation of the aircraft 10 at this instant is also displayed on the display screen 18 during a sub -step 230, and so on. The calculation module 30 is in particular capable of continuing this iteration starting from the situation, or state, of the aircraft at the first subsequent time instant t+1 in order to calculate a new list 32 of recommended action(s) at a new subsequent time instant, such as the second subsequent time instant t+2.

Thus, the display system 16 according to the invention, and in particular the determination device 20, as well as the associated determination method, make it possible to provide the user, such as the pilot of the aircraft 10, with a significant decision-making support, in particular for the real-time management of complex problem(s). Indeed, the determination device 20 makes it possible to easily and efficiently calculate the list 32 of recommended action(s), that is to say the list 32 of action(s) that is recommended for the user to perform with the aircraft 10, or that it is recommended to be performed by the avionic system or systems 12 to which said list 32 of action(s) has been transmitted, in order to manage the corresponding complex problem . The action or actions to be taken for the aircraft 10 are typically a heading to follow, a speed to respect, a position to reach, an altitude instruction, the choice of equipment, the selection of a sensor or even the selection of a mode/use parameter of a sensor or equipment.

In addition, through the evaluation module 38, the determination device 20 makes it possible to implement two-way communication between the user and the determination device 20, in particular between the user and the artificial intelligence algorithm. implemented by the determination device 20, which amounts to putting the user in the loop. This two-way communication between the user and the artificial intelligence algorithm also comes from the possible selection of reinforcement learning, and in particular from the reward function used for said learning, this via the selection module 40 and on data that is entirely visual and interpretable by the user. In addition, the images 24 are preferentially interactive images for the immediate taking into account of modifications desired by the user.

The evaluation via the game theory algorithm of the recommended actions, proposed by the determination device 20, also provides the user with an explanation of the proposals of the artificial intelligence algorithm, which allows him to validate or not said artificial intelligence algorithm proposals.

In addition, the determination device 20 makes it possible to calculate with greater robustness and greater efficiency the list 32 of action(s) recommended with respect to new contexts or situations, by implementation of the artificial intelligence algorithm which adapts to such new contexts or situations by processing as input variables the set of images 24 of geo-referenced zone(s) and the group 28 of contextual data and delivering as output the list 32 of action(s) recommended from these input variables, the artificial intelligence algorithm having been previously trained on a wide variety of images 24 and contextual data corresponding to various contexts and situations.

Furthermore, when the artificial intelligence algorithm is trained via reinforcement learning, such as a DQN algorithm or else an actor-critic type policy gradient algorithm, the calculation of the list 32 of action(s ) recommended is even more robust and efficient.

It can thus be seen that the determination device 20 according to the invention and the associated determination method make it possible to determine one or more recommended actions for the aircraft 10 in a more robust and more efficient manner with respect to new contexts. or situations.

Claims (10)

Method for determining at least one recommended action for an aircraft (10), the method being implemented by an electronic determination device (20) and comprising the following steps:
- acquisition (100) of a set of image(s) (24) of geo-referenced zone(s) relating to the aircraft (10),
- reception (110) of a group (28) of contextual data(s) relating to the aircraft (10),
- calculation (120) of a list (32) of recommended action(s) for the aircraft (10), via an artificial intelligence algorithm and from the set of image(s) (24) of geo-referenced zone(s) and of the group (28) of contextual data(s), the artificial intelligence algorithm comprising as input variables, on the one hand, the set of image(s) (24) of geo-referenced zone(s), and on the other hand, the group (28) of contextual data(s), and as an output variable the list (32) of recommended action(s). Method according to claim 1, in which the method further comprises a step (130) of displaying, on a display screen (18), the list (32) of action(s) recommended for the aircraft (10). Method according to claim 1 or 2, in which the set of image(s) (24) of geo-referenced zone(s) comprises at least one image (24) among the group consisting of: an image of an area of interest, such as an image containing one or more geo-referenced meteorological information; an image of a threat area, such as an image containing one or more geo-referenced obstacle information; an image of an area of uncertainty containing geo-referenced information indicating, for a given position, the probability that an object may be present at this position. Method according to any one of the preceding claims, in which the group (28) of contextual data(s) comprises at least one data item from among the group consisting of: a quantity of resource(s) available, such a quantity of fuel, a number of equipment(s) carried; a number of identified entity(ies) present in a respective geo-referenced area; and a number of unknown entity(ies) present in a respective geo-referenced area. Method according to any one of the preceding claims, in which the list (32) of recommended action(s) comprises at least one action from the group consisting of: a heading instruction, a speed instruction, an of position, an altitude instruction, use of a respective sensor, use of respective equipment, selection of a mode/parameter for use of a sensor or equipment. A method according to any preceding claim, wherein the artificial intelligence algorithm includes at least one neural network (60);
the artificial intelligence algorithm preferably comprising a neural network (60) headed by a first neural sub-network (62) having the set of image(s) (24) as input variable of geo-referenced zone(s), a second sub-network of neurons (64) having as input variable the group (28) of contextual data(s), and at the tail end, a third sub-network neural network (66) connected at the output of the first (62) and second (64) neural sub-networks and having as output variable the list (32) of recommended action(s);
the first neural subnetwork (62) preferably having layers of convolutions, the second (64) and third (66) neural subnetworks each being a multilayer perceptron (MLP) having dense layers. A method according to any preceding claim, wherein the artificial intelligence algorithm is obtained via reinforcement learning;
the reinforcement learning being preferably selected from the group consisting of: a Q-learning algorithm with a neural estimator (DQN); an actor-critic type policy gradient algorithm, such as a DDPG algorithm, a PPO algorithm or an SAC algorithm;
when the reinforcement learning is an actor-critical type policy gradient algorithm, the artificial intelligence algorithm preferably further comprising at least one neural network (60) called actor(s) and the reinforcement learning comprising at least one other so-called critical neural network. A computer program comprising software instructions which, when executed by a computer, implement a method according to any preceding claim. Electronic device (20) for determining at least one recommended action for an aircraft (10), the device (20) comprising:
- an acquisition module (22) configured to acquire a set of image(s) (24) of geo-referenced zone(s) relating to the aircraft (10),
- a reception module (26) configured to receive a group (28) of contextual data(s) relating to the aircraft (10),
- a calculation module (30) configured to calculate a list (32) of recommended action(s) for the aircraft (10), via an artificial intelligence algorithm and from the set of image(s) (24) of geo-referenced zone(s) and of the group (28) of contextual data(s), the artificial intelligence algorithm comprising as input variables, on the one hand, the set of image(s) (24) of geo-referenced zone(s), and on the other hand, the group (28) of contextual data(s), and output variable the list (32) of recommended action(s). An electronic display system (16) for an aircraft (10), the system (16) comprising:
- a display screen (18), and
- an electronic device (20) for determining at least one recommended action for the aircraft (10),
characterized in that the device (20) is according to the preceding claim, and further comprises a display module (34) configured to display, on the display screen (18), a list (32) of actions (s) recommended for the aircraft (10).