METHOD AND SYSTEM FOR ICING DETECTION ON AN AIRCRAFT

Abstract

A method for detecting the presence of ice on an aircraft, comprising: obtaining (201) real-time measurements (provided by sensors) of a set of flight parameters including a glide slope, angle of attack, Mach number, rotational speed of a low-pressure assembly of the aircraft's propulsion engine, fuel flow rate, aircraft mass, position of an adjustable horizontal stabilizer, static temperature, and static pressure; feeding (202) the obtained measurements into an artificial intelligence model, trained and configured to calculate an ice presence indicator based on measurements of the flight parameter set; and providing (203) the ice presence indicator calculated by the model. This new indirect method is effective because it relies on an artificial intelligence model using numerous flight parameters. Figure to be published with the abbreviation: Fig. 2

Description

Title of the invention: METHOD AND SYSTEM FOR DETECTING ICING ON AN AIRCRAFT technical field

[0001] The present invention relates to a method and system for detecting ice on an aircraft. It is particularly applicable to the automatic and real-time detection of aircraft icing, especially the accretion of ice on the aircraft wings. PRIOR TECHNOLOGY

[0002] The formation of ice on an aircraft, particularly on its wings, can significantly affect flight qualities and performance, to the point of disrupting aircraft control. Indeed, an accumulation (or accretion) of ice on the aircraft surfaces (especially on the wings) can lead to a significant increase in aircraft weight, a loss of lift, control surface actuation problems, communication and antenna malfunctions, airspeed sensor measurement errors, and engine thrust losses; these various malfunctions can disrupt aircraft control.

[0003] To mitigate these malfunctions, aircraft authorized to fly in icing conditions are equipped with protection systems (also called "anti-icing systems"), including heating systems integrated into the components to be protected (wings, probes, engine air intakes, etc.) to prevent the formation or accumulation of ice. Activation of these protection systems generally relies on the pilot's judgment after visually identifying the presence of icing conditions. Since this identification is necessarily imperfect, one or more detection systems are generally used to assist the pilot in their judgment.

[0004] Thus, three types of methods are commonly known to be implemented by these detection systems to detect the presence of ice on an aircraft: - direct methods, based on the addition of specific sensors on the wings and fuselage of the aircraft. By specific sensors, we mean sensors whose measurements are exclusively intended for the detection of the presence of ice; - Indirect methods, based on the use of non-specific sensors already present on board to estimate the aircraft's dynamics or performance. These are estimation methods (using, for example, a Kalman filter) of a specific quantity in real time to be compared with a tabulated reference value to verify the presence or absence of ice; and - atmospheric methods, based on the use of radar and / or a dedicated camera to classify clouds and calculate the probability that they will generate frost formation.

[0005] A common drawback of both direct and atmospheric methods is that they require the addition of various sensors directly onto the aircraft. This involves drilling through the wing and fuselage, providing mechanical reinforcements near the hole, deploying electrical wiring, and installing additional data acquisition systems in electrical cabinets. The sensors also represent an additional cost and weight. Furthermore, the sensors often protrude from the wing and fuselage skin and consequently generate induced drag, which affects aircraft performance.

[0006] One advantage of indirect methods is that they do not require the addition of various sensors directly onto the aircraft's wings and fuselage. Known indirect methods offer interesting performance, but this could be improved. For example, even the best-performing known indirect methods are highly sensitive to turbulence and sensor bias.

[0007] There is therefore a need to provide an indirect method that is more efficient than known indirect methods for detecting, automatically and in real time, the presence of frost on an aircraft, in particular on the wings. Description of the invention

[0008] A method for detecting the presence of frost on an aircraft is proposed, the method being implemented by a detection system comprising electronic circuitry, the method comprising: - to obtain real-time measurements of a set of flight parameters of the aircraft, the measurements being provided by sensors included in the aircraft, the set of flight parameters including: a slope, denoted yair; an angle of attack, denoted “; a Mach number, denoted M; a rotational speed of a low-pressure assembly of the aircraft's propulsion engine, denoted NI; a fuel flow rate, denoted F_flow; an aircraft mass, denoted m; a position of an adjustable horizontal stabilizer, denoted POSTHS; a static temperature, denoted Ts; and a static pressure, denoted Ps; - to feed the measurements obtained into a machine learning-based model, such as artificial intelligence, trained and configured to calculate an indicator of the presence of ice on the aircraft based on measurements of the aircraft's flight parameter set; and - provide the frost presence indicator calculated by the machine learning-based model.

[0009] Thus, the proposed solution is a new indirect method for detecting ice on an aircraft. It allows for the automatic and real-time detection of the presence of ice on an aircraft, particularly on the wings. It is more effective than known indirect methods because it relies on a machine learning model, a type of artificial intelligence, and uses numerous aircraft flight parameters.

[0010] According to a particular embodiment, the aircraft flight parameter set further includes at least one information redundancy parameter belonging to the group comprising: a true speed, denoted Vc; and a total pressure, denoted Pt.

[0011] According to a particular embodiment, the aircraft flight parameter set further includes at least one context parameter belonging to the group comprising: a roll angle, denoted cp; a pitch angle, denoted 0; a yaw angle, denoted W; a roll rate, denoted pl; a pitch rate, denoted ql; a yaw rate, denoted rl; a sideslip angle, denoted [3; a pressure altitude, denoted Zp; a vertical speed, denoted Vz; and a center of gravity information, denoted CG.

[0012] According to a particular embodiment, the machine learning-based model is a deep neural network.

[0013] According to a particular embodiment, the deep neural network is a convolutional neural network comprising at least one convolution layer and at least one subsampling layer.

[0014] A computer program product is also proposed, comprising instructions leading to the execution, by a processor, of the process mentioned above according to any one of its embodiments, when said instructions are executed by the processor.

[0015] A storage medium is also proposed, storing such instructions.

[0016] A system for detecting the presence of frost on an aircraft is also proposed, the system comprising electronic circuitry configured to implement: - to obtain real-time measurements of a set of aircraft flight parameters, the measurements being provided by sensors included in the aircraft, the set of flight parameters comprising: a glide slope, denoted yair; an angle of attack, denoted a; a Mach number, denoted M; a rotational speed of a low-pressure assembly of the aircraft's propulsion engine, denoted NI; a fuel flow rate, denoted F_flow; an aircraft mass, denoted m; a position of an adjustable horizontal stabilizer, denoted POSTHS; a static temperature, denoted Ts; and a static pressure, denoted Ps; - to inject the measurements obtained into a machine learning-based model, of the artificial intelligence type, trained and configured for calculate an indicator of the presence of ice on the aircraft based on measurements of the aircraft's flight parameter set; and - provide the frost presence indicator calculated by the machine learning-based model.

[0017] An aircraft comprising an aircraft as mentioned above is also proposed. Brief description of the drawings

[0018] The features of the invention mentioned above, as well as others, will become clearer upon reading the following description of at least one exemplary embodiment, said description being made in relation to the accompanying drawings, among which:

[0019] [Fig-1] schematically illustrates, in side view, an aircraft equipped with a system of detection of the presence of frost;

[0020] [Fig.2] schematically illustrates an example of an algorithm for detecting the presence of frost, in an embodiment;

[0021] [Fig.3] schematically illustrates a set of forces exerted on the aircraft;

[0022] [Fig.4] schematically illustrates the model used by the detection system of the presence of frost in [Fig. 1], in one embodiment; and

[0023] [Fig.5] schematically illustrates an example of the hardware architecture of the system of detection of the presence of frost in [Fig.l].

[0024] DETAILED DESCRIPTION OF IMPROVEMENTS

[0025] Figure 1 schematically illustrates, in side view, an aircraft 100. The aircraft 100 includes a system 101 for detecting the presence of ice on the aircraft. System 101 is implemented as electronic circuitry and is typically integrated into the aircraft's avionics. For example, system 101 is hosted in a flight control computer onboard the aircraft (e.g., a Flight Control and Guidance System (FCGS)).

[0026] In this case, the ice detection process performed by system 101 is therefore entirely carried out on board the aircraft. As detailed below, this process uses, during the production phase (also called the "use phase"), a machine learning-based model to calculate an ice presence indicator. This model is pre-estimated during a learning phase (also called the "training phase").

[0027] In one variant (not illustrated), the ice detection system comprises a first part, installed on board the aircraft, and a second part, installed on the ground. For example, the first onboard part receives measurements of the aircraft's flight parameter set, and transmits these measurements to the second part on the ground; The second part on the ground injects these measurements into the model and transmits to the aircraft the ice presence indicator calculated by the model.

[0028] In the embodiment of [Fig. 1], as in the aforementioned variant, the ice presence indicator calculated by the model can be displayed on a screen in the aircraft cockpit and may optionally generate an alarm (i.e., trigger an alert). The aircraft pilot then has the option of activating the anti-icing systems. Alternatively, the ice presence indicator automatically activates the anti-icing systems.

[0029] System 101 implements a method (algorithm) for detecting the presence of frost. This method is schematically illustrated in [Fig.2], in a particular embodiment.

[0030] In a step 201, the system 101 obtains real-time measurements of a set of flight parameters of the aircraft. The measurements are provided by sensors included in the aircraft. The set of flight parameters includes: a glide slope, denoted yair (in °); an angle of attack, denoted a (in °); a Mach number, denoted M (unitless); a rotational speed of a low-pressure assembly of the aircraft's propulsion engine, denoted NI (in %); a fuel flow rate, denoted F_flow (in kg / s); an aircraft mass, denoted m (in kg); a position of an adjustable horizontal stabilizer (or THS for "Trimmable Horizontal Stabilizer"), denoted POSTHS (in °); a static temperature, denoted Ts (in K); and a static pressure, denoted Ps (in Pa).

[0031] An explanation of the choice of this set of flight parameters of the aircraft will be found below in relation to [Fig.3].

[0032] In a first variant, the aircraft flight parameter set further includes at least one information redundancy parameter belonging to the group comprising: a true speed, denoted Vc (in ft / s); and a total pressure, denoted Pt (in Pa).

[0033] In a second embodiment, the aircraft flight parameter set further includes at least one context parameter belonging to the group comprising: a roll angle, denoted q> (in °); a pitch angle, denoted 0 (in °); a yaw angle, denoted W (in °); a roll rate, denoted pl (in 7s); a pitch rate, denoted ql (in 7s); a yaw rate, denoted rl (in 7s); a sideslip angle, denoted [3 (in °); a pressure altitude, denoted Zp (in ft); a vertical speed, denoted Vz (in ft / s); and a center of gravity information, denoted CG (in %).

[0034] In a third variant, the aircraft flight parameter set further includes both at least one of the information redundancy parameters of the first variant and at least one of the context parameters of the second variant.

[0035] In a particular case of this third variant, the aircraft's flight parameter set further includes all the information redundancy parameters of the first variant and all the context parameters of the second variant. The flight parameter set then comprises NP = 21 flight parameters.

[0036] In step 202, the system 101 feeds the obtained measurements into a machine learning-based model of the artificial intelligence type. It is trained and configured to calculate an indicator of the presence of ice on the aircraft based on measurements of the aircraft's flight parameter set.

[0037] The artificial intelligence type model is, for example, chosen from: - a neural network; - a Bayesian network; and - deep learning.

[0038] In a particular implementation, the model is a deep neural network that performs deep learning. In a particular embodiment, the deep neural network is a convolutional neural network (or "CNN" for "Convolutional Neural Network") comprising at least one convolutional layer and at least one subsampling layer (also called a "pooling layer").

[0039] The general principle is as follows: when an aircraft equipped with standard sensors (altitude-pressure, Mach number, temperature, NI engine speed, etc.) flies through a cloud that triggers ice formation on the wings, the newly formed ice will affect the aircraft's performance (additional drag, modification of lift if the angle of attack is sufficiently high). The model used in step 202 is capable of detecting this change in the aircraft's current state (reflected by the temporal evolution of the set of flight parameters observed in real time) and estimating the associated ice level. It therefore warns the pilot that ice is forming on the wings, who can then choose whether or not to activate the anti-icing system if a de-icing threshold is reached.

[0040] In a step 203, the system 101 provides the frost presence indicator calculated by the machine learning-based model.

[0041] Thus, the proposed solution consists of an Artificial Intelligence (AI) approach using a machine learning-based model, which enables the efficient and robust detection of ice in real time, possibly specifically for certain flight phases (e.g., the cruise phase or final approach), using measurements already present on board the aircraft (no need to add new sensors or specific equipment). The proposed solution makes it possible, in real time, to guarantee the detection of a given level of ice accretion on the wings.

[0042] Such an AI algorithm can be based on a supervised learning phase, made possible by real flight data recorded on aircraft (aircraft in company and / or test aircraft with artificial ice forms), possibly supplemented by simulated flight data obtained with simulators.

[0043] The proposed solution can easily be adapted to an aircraft model and / or an aircraft configuration. This simply requires training the algorithm on data associated with the desired aircraft model, adding configuration information in addition to the aforementioned flight parameters. Similarly, the developed method can be just as easily adapted to a particular flight phase if sufficient training data associated with that flight phase is available.

[0044] Typically, the detection of frost or the absence of frost is characterized as a binary classification problem (0: no frost; 1: presence of frost). However, we also seek to characterize several intermediate frost levels (multi-class classification). For example, we distinguish five frost classes, each defined by a particular value of a Kice accretion level (which constitutes the frost presence indicator discussed above) between 0 and 1: Kice = 0: no frost on the sail; Kice = 0.25: low frost accretion on the sail; Kice = 0.5: average frost accretion on the sail; Kice = 0.75: strong frost accretion on the sail; and Kice = 1: sail completely frosted.

[0045] Depending on the class of frost detected, and by comparison with a predetermined activation threshold (also called "de-icing threshold"), it is possible to issue an alert to the pilot and / or activate the protection system (anti-icing or "anti-ice" systems) at the appropriate time in order to optimize the energy performance of the aircraft (compromise between the loss of performance due to the drag caused by frost and the loss of performance caused by the anti-icing system).

[0046] Fig. 3 schematically illustrates a set of forces exerted on the aircraft 100 and explaining the choice of the set of flight parameters used by the model (based on machine learning) to calculate the ice presence indicator.

[0047] The forces exerted on the aircraft 100 are: the thrust T (for "Thrust" in English), the drag D (for "Drag" in English), the weight W (for "Weight" in English) and the lift L (for "Lift" in English).

[0048] The flight parameters used by the model were chosen based on the aircraft's flight dynamics. In fact, under normal flight conditions (at an angle of attack lower than aprot, the maximum angle of attack defined by the flight controls), icing acts exclusively on the aircraft's longitudinal balance, via the addition of drag. The aircraft's longitudinal balance is therefore expressed as follows:

[0049] T*cos(a) -DW*sin(yair) = ax*m

[0050] With:

[0051] T = / 1(^1, F_flow,M, P^

[0052] D = q*S*Cx + hDiMg [00531 ax, = ^=df(M,Tfdx

[0054] Where A Dicing is the term of drag due to frost, Cx is the drag coefficient, is the dynamic pressure and the reference area.

[0055] By isolating the term AD^g (which is the one we are trying to determine, in order to go back to the level of frost on the sail), we obtain the following equation:

[0056] Dicing = T*cos (a) - »^g*sin (yair) + - q*S*Cx

[0057] By explicitly stating the dependencies of the different terms (the dynamic pressure 9 is 0.5*l,4*M2*Ps, the drag coefficient Cx depends on a, M, Ps, POSTHS and the aircraft configuration Confy), we obtain:

[0058] = f pjpcttfa)M. P,.Coiif, POSTHS}

[0059] With fl the actual thrust generated by the aircraft and / 21a the actual drag excluding that due to icing. These functions are unknown in flight but depend on parameters that are constantly measured.

[0060] Based on the results of this equation, a first set of relevant input parameters for the model was defined: yair, a, M, NI, F_flow, m, POSTHS, Ts and Ps. As mentioned above, the flight parameter set may also include one or more information redundancy parameters (from among Vc and Pt) and / or one or more context parameters (from among q>, 0, W, pl, ql, rl, [3, Zp, Vz and CG).

[0061] During the model training phase (for example, in supervised mode), labeled data is fed into the model; that is, values ​​of the aforementioned flight parameters at flight times for which the wing icing level is known. The labeled data constitutes a basis for training, validation, and testing the model. The labeled data consists of real flight data, possibly supplemented by simulated flight data obtained with a simulator.

[0062] For example, to obtain more training data, a simulator is used to model the aircraft's behavior when subjected to ice accretion. Several simulation scenarios are used, for instance, to represent as comprehensively as possible the maneuvers and situations encountered during cruise. Each scenario consists of one or more maneuvers commanded to the aircraft's autopilot. Examples of scenarios: level flight at a point of Given a flight scenario, the scenario involves a change of speed in level flight with simulated wind gusts and intermittent turbulence, a constant climb with simulated wind gusts and intermittent turbulence, a constant descent with simulated wind gusts and intermittent turbulence, a constant vertical speed climb, a constant vertical speed descent, a 180° left turn with simulated wind gusts and intermittent turbulence, and a 180° right turn with simulated wind gusts and intermittent turbulence. For each scenario, the recorded parameters are sampled at a frequency of 4 Hz over time. This offers a good compromise between computation time and accurate tracking of the temporal evolution of the recorded signals. Each scenario is run at, for example, 10,000 different flight points, each flight point distinguished by the value of one of the flight parameters or by the level of icing.Next, the data is formatted, depending on the model developed.

[0063] In one alternative, if turbulence can be robustly detected on board the aircraft, it may be advantageous to train the model solely on flight data (real or simulated) without turbulence. Once on board, it will suffice to disable icing detection as soon as turbulence is detected to disregard false alarms in this scenario.

[0064] Figure 4 schematically illustrates the model used by the detection system of the presence of frost in [Fig.1], in one embodiment. The model here is a deep neural network, and more specifically a convolutional neural network (CNN) 400 comprising (in order): an input layer 401, a first convolution layer 402, a first subsampling layer (also called a "pooling layer") 403, a second convolution layer 404, a second subsampling layer 405, a flattening layer 406, a first linear layer 407 and a second linear layer 408.

[0065] Each layer plays a specific role: - Convolution layers 402 and 404, for example, use a "ReLu" type activation function (for "Rectified Linear Unit") and transform the current state of the aircraft provided as input (in the form of signals which are measurements of the set of NP flight parameters discussed above) into NP explanatory parameters (also called "explanatory variables" or "features") which encode the information contained in the time evolution of the signals. In other words, each of the N flight parameters constitutes one of the NP explanatory parameters; - The 403 and 405 pooling layers allow us to reduce the dimensionality of the neural network (in order to make it easier embeddable) and to prevent to some extent overfitting, that is, overlearning of the neural network on the training set, which is responsible for a lack of reliability of predictions on unknown data; - Linear layers 407 and 408, for example, use a "ReLu" type activation function and utilize the aircraft's current state encoding to predict the output ice class (five distinct ice levels to discriminate); and - the second linear layer 408 (output layer) uses for example the "softmax" activation function, at the output of the neural network, which makes it possible to obtain a probability of belonging to each class and facilitates the interpretation of the results.

[0066] One advantage of the 400 convolutional neural network is that it allows for consideration of the temporal correlation between different measurement points. In fact, the pitch angle at t = 20 s is strongly correlated with that measured at t = 20.25 s and much less so with that at t = 1 s. To improve the accuracy of frost level classification, a convolutional neural network architecture capable of accounting for this temporal evolution is therefore used.

[0067] Model 400 in [Fig. 4] is suitable for onboard deployment because it does not have too many layers or excessively large layers. Depending on whether performance or portability is prioritized, it is possible to choose a different neural network architecture. For example, adding new convolutional layers and another linear output layer makes it even more reliable in detecting the frost level, at the cost of a model that is more complex to run on an onboard computer. Conversely, removing a convolutional layer and a subsampling (pooling) layer from the presented architecture reduces performance but makes the solution more easily portable.

[0068] In one embodiment, each row of the database used to train the 400 model is a 120x21 matrix (21 explanatory parameters (features), each corresponding to a signal of 120 time points of measurement corresponding to a duration (time window) of 30s with a sampling at 4 Hz).

[0069] In the case where the architecture is a neural network, a Min-Max normalization is performed on each value of the different signals. For each point pk of each signal i associated with an explanatory parameter (feature) j:

[0070] p -min(y, Vf VA:) P^ — max(j, VÊ VÂyminiJ, Vi, Vk] 1 ]

[0071] Training is then performed on a portion of the rows of the training dataset, keeping the remainder as a validation and test set. The cost function

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] The function selected for training is, for example, the "Cross Entropy Loss" function. This is a classic loss function in a multi-class classification problem. For a batch of N data points during training, the formula is as follows: With : - A is the number of samples in the batch; - C is the number of classes (here 5); - Jy = 1 if the data is associated with class j, 0 otherwise; and - Pÿ the predicted probability that the data 1 belongs to class j. It is also possible to use a variant of the "Cross Entropy Loss" cost function when training the model. For example, if log-probabilities are used, for a data point n from a set of N data points during training, the formula for this cost function is as follows: , i J exp(xw-) | Z»- With : - C is the number of classes (here 5); - a manual weight to be assigned to each class in the case of data inhomogeneous; - xni the predicted probability that the data n is of the class; and - = 1 if the data n is indeed of class c, 0 otherwise. Once the model is trained and validated, it can be used onboard to detect the level of frost on the sails in real time. Note that the neural network in the embodiment presented above is considered "fixed" because its coefficients remain constant once calculated offline. There is no online adaptation (i.e., no updating of the coefficients), so the solution remains deterministic. In the case where, for the training phase, real flight data is supplemented with simulated flight data, it is possible to improve the realism of the simulations. Indeed, if flight points are generated by randomly selecting the parameters that characterize the model (altitude, initial speed, heading, etc.), and if each parameter is correctly selected between two bounds corresponding to operational limits, some of the ultimately generated flight points may be located outside the flight envelope, or even be physically impossible. For example, it is possible to generate a flight point where the aircraft has a low mass and is moving slowly at high altitude; this is at the stall limit, outside the flight envelope, and the resulting data is useless. for model training. To address this issue, charts available in the technical documentation of a typical aircraft were digitally coded to account for the interdependencies between certain parameters associated with flight points (e.g., center of gravity, mass, altitude-pressure, Mach number). By using double interpolations and numerically calculating, for example using a Support Vector Machine (SVM) algorithm, the boundary equation between valid and invalid parameter pairs (e.g., mass / Mach), it is possible to generate only valid flight points for model simulations. This offers three main advantages: because anomalous behavior is eliminated during simulations, all simulations performed are valid and can be used to build the training and validation dataset; The calculation on the simulator is faster because the aircraft balancing phase is very quick since no selected flight point is associated with extreme conditions; more data is obtained at the end of each simulation series since no flight point is rejected by the simulator.

[0080] It is also possible to improve the realism of the signals obtained by simulations by adding a bias to the various sensors after the fact, in order to reflect the actual biases of the sensors on the aircraft (which differ depending on the aircraft). A random bias is thus drawn within the limits defined by the technical specifications of the different sensors. For example, here are some typical biases: pressure altitude of 100 ft, Mach number of 0.02, static temperature of 2°C, NI of 2%, mass of 5 T, etc.

[0081] It is also possible to add measurement noise to the recorded signals. This brings the signals used during training closer to the real signals so that the good performance of the model on the simulated data is more easily transferable to the real data.

[0082] It is also possible to improve the robustness of the model by increasing the variability of the simulated data. To do this, the time windows, for example of 30 seconds (containing 120 time points of measurement obtained with a sampling at 4 Hz), are not successive but are placed randomly along the simulation (and without overlap), which makes it possible to take a smaller number of them.

[0083] Figure 5 schematically illustrates an example of the hardware architecture of the frost detection system 101, which comprises, connected by a communication bus 510: a processor or CPU (Central Processing Unit) 501; a RAM (Random Access Memory) 502; a ROM (Read Only Memory) 503, for example, Flash memory; a data storage device, such as a HDD (Hard Disk Drive), or a storage media reader, such as an SD (Secure Digital) card reader 504; at least one interface communication 505 allowing the ice detection system 101 to interact with one or more pieces of equipment on aircraft 10, and more particularly with avionics equipment on aircraft 100.

[0084] The processor 501 is capable of executing instructions loaded into RAM 502 from ROM 303, external memory (not shown), a storage medium such as an SD card, or a communication network (not shown). When the frost detection system 101 is powered on, the processor 501 is able to read instructions from RAM 302 and execute them. These instructions form a computer program causing the processor 501 to implement the behaviors, steps, and algorithm described herein.

[0085] All or part of the behaviors, steps, and algorithm described herein can thus be implemented in software form by executing a set of instructions by a programmable machine, such as a DSP (Digital Signal Processor) or a microcontroller, or be implemented in hardware form by a dedicated machine or component (chip) or a dedicated set of components (chipset), such as an FPGA (Field-Programmable Gate Array) or an ASIC (Application-Specific Integrated Circuit). Generally speaking, the 101 frost detection system comprises electronic circuitry arranged and configured to implement the behaviors, steps, and algorithms described herein.

Claims

Demands

1. A method for detecting the presence of ice on an aircraft (100), the method being implemented by a detection system (101) comprising electronic circuitry, the method comprising: - obtaining (201) real-time measurements of a set of flight parameters of the aircraft, the measurements being provided by sensors included in the aircraft, the set of flight parameters comprising: a glide slope, denoted yair; an angle of attack, denoted a; a Mach number, denoted M; a rotational speed of a low-pressure assembly of the aircraft's propulsion engine, denoted NI; a fuel flow rate, denoted F_flow; a mass of the aircraft, denoted m; a position of an adjustable horizontal stabilizer, denoted POSTHS; a static temperature, denoted Ts; and a static pressure, denoted Ps;- inject (202) the measurements obtained into a machine learning-based model, of the artificial intelligence type, trained and configured to calculate an indicator of the presence of ice on the aircraft as a function of measurements of the aircraft's flight parameter set; and - provide (203) the indicator of the presence of ice calculated by the machine learning-based model.

2. A method according to claim 1, wherein the aircraft flight parameter set further comprises at least one information redundancy parameter belonging to the group comprising: a true speed, denoted Vc; and a total pressure, denoted Pt.

3. A method according to any one of claims 1 and 2, wherein the aircraft flight parameter set further comprises at least one context parameter belonging to the group comprising: a roll angle, denoted q>; a pitch angle, denoted 0; a yaw angle, denoted W; a roll rate, denoted pl; a pitch rate, denoted ql; a yaw rate, denoted rl; a sideslip angle, denoted [3; a pressure altitude, denoted Zp; a vertical rate, denoted Vz; and a center of gravity information, denoted CG.

4. A method according to any one of claims 1 to 3, wherein the machine learning-based model is a deep neural network (400).

5. A method according to claim 4, wherein the deep neural network (400) is a convolutional neural network comprising at least one convolutional layer (402, 404) and at least one subsampling layer (403, 405).

6. Product computer program, comprising instructions causing a processor (501) to execute the method according to any one of claims 1 to 5, when said instructions are executed by the processor.

7. Storage medium (503), storing a computer program comprising instructions causing a processor (501) to execute the method according to any one of claims 1 to 5, when said instructions are read and executed by the processor.

8. System (101) for detecting the presence of ice on an aircraft (100), the system comprising electronic circuitry configured to implement: - obtaining (201) real-time measurements of a set of flight parameters of the aircraft, the measurements being provided by sensors included in the aircraft, the set of flight parameters comprising: a glide slope, denoted see; an angle of attack, denoted (l); a Mach number, denoted M; a rotational speed of a low-pressure assembly of the aircraft's propulsion engine, denoted NI; a fuel flow rate, denoted F_flow; a mass of the aircraft, denoted m; a position of an adjustable horizontal stabilizer, denoted POSTHS; a static temperature, denoted Ts; and a static pressure, denoted Ps;- inject (202) the measurements obtained into a machine learning-based model, of the artificial intelligence type, trained and configured to calculate an indicator of the presence of ice on the aircraft as a function of measurements of the aircraft's flight parameter set; and - provide (203) the indicator of the presence of ice calculated by the machine learning-based model.

9. Aircraft (100) comprising a system (101) for detecting the presence of frost according to claim 8.