Machine Learning in Avionics

Abstract

The document relates to a system and a method for managing a flight of an aircraft, comprising the steps of receiving data (200) from a record of a flight of the aircraft; the data comprising data from sensors and / or data from onboard avionics; determining the state of the aircraft at point N (220) based on the received data (200); determining the state of the aircraft at point N+1 (240) based on the state of the aircraft at point N (220) by applying a model learned by means of machine learning (292). The development describes the use of flight parameters SEP, FF and N1; offline and / or online unsupervised machine learning according to various algorithms and neural networks. Software aspects are described.

Description

Technical Field

[0001] The present invention relates generally to the field of avionics. In particular, the present invention relates to a method and system for predicting the future state of an aircraft. Background Art

[0002] Methods known in the prior art (eg WO 2017042166 or US9290262) are typically based on a data set that models the performance of an aircraft.

[0003] Various methods for modeling aircraft performance are known. Tools such as "BADA" (abbreviation for "Base of Aircraft Data") or the "Safety-Line" of EUCASS ("European Conference on Aerospace Sciences") have limitations. The BADA model is limited with respect to thrust and drag. The EUCASS model is only applicable to aircraft propelled by turbojets (i.e., N1 driven).

[0004] Aeronautical technical problems usually involve many different parameters and, therefore, currently available optimizations converge rarely, poorly or not at all. This lack of convergence (or convergence towards local minima) is often explicitly observed when the process being modeled is large-scale (and follows different models depending on the perspective).

[0005] Methods based on models incorporating the physical equations of the aircraft generally depend on the quality of the model and on the knowledge of the actual behavior of the aircraft. In practice, these models are generally not as robust in the face of variability in the actual behavior of a given aircraft as in the face of variability in the actual behavior of an "average" (modeled) aircraft.

[0006] There is a need in the industry for improved methods and systems for optimizing all or some of the operations of an aircraft. Summary of the invention

[0007] The document relates to a system and method for managing a flight of an aircraft, comprising the steps of receiving data (200) from a record of a flight of an aircraft; the data comprising data from sensors and / or data from onboard avionics; determining the state of the aircraft at point N (220) based on the received data (200); determining the state of the aircraft at point N+1 (240) based on the state of the aircraft at point N (220) by applying a model learned by means of machine learning (292). Developments describe the use of flight parameters SEP, FF and N1; offline and / or online unsupervised machine learning based on various algorithms and neural networks. Software aspects are described. A method for managing a flight of an aircraft is described, comprising the steps of receiving data from a record of a flight of an aircraft; the data comprising data from sensors and / or data from onboard avionics; determining the state of the aircraft at point N based on the received data; determining the state of the aircraft at point N+1 based on the state of the aircraft at point N by applying a model learned by machine learning (292). In this embodiment, learning is performed end-to-end, that is, learning includes a PERFDB (performance calculation) step and a TRAJ / PRED (trajectory calculation) step; output data is directly determined by learning performed on input data.

[0008] In one embodiment, the step of determining the state of the aircraft at point N+1 based on the state of the aircraft at point N includes the following steps: determining flight parameters SEP, FF and N1 based on the state of the aircraft at point N by applying a model learned by machine learning; and determining the state of the aircraft at point N+1 based on the values ​​of the flight parameters SEP, FF and N1 by means of trajectory calculation, wherein the SEP value represents the energy available for the aircraft to climb, the FF value represents the change in fuel weight, and the N1 value represents the first stage rotation speed of the engine that affects fuel consumption.

[0009] In one embodiment, machine learning is unsupervised. Unsupervised learning aims to find underlying structures based on unlabeled (or unmarked) data. Category numbers and definitions are not given a priori. This type of learning includes, for example, deep learning techniques. Advantages associated with this type of learning include the use of large amounts of computing power on accumulated big data, no need for human control, and the discovery of trends, patterns, or relationships that may not necessarily be understandable to humans but may be efficient.

[0010] In one embodiment, machine learning is supervised. Advantageously, certain attributes of the data may be known, e.g., SEP, N1, or FF. The desired output is known, and the record category or class is known (label, tag). Supervised learning allows for human intervention, and may therefore produce effective models, e.g., models that converge faster. In contrast, human presuppositions may limit possibilities (there is no such limitation in the unsupervised case).

[0011] In one embodiment, the machine learning is performed offline. The records may be records of past flights (data mining approach). This embodiment is advantageous because it allows for the reuse of existing data (of which there is a lot and which is currently underutilized).

[0012] In one embodiment, machine learning is performed online. In one embodiment, machine learning can be performed incrementally or online. Based on an averaged general model (aircraft type or series), a specific aircraft can be characterized and gradually improved (by serial number or tail number) as it conducts its own flights. When the model is known, learning can continue through the data stream (to improve the existing model without starting from scratch). Offline machine learning learns based on the complete data set, while online learning can continue learning on the aircraft (transfer learning) without having to re-ingest the starting data.

[0013] It should be noted that the machine learning implemented in the method according to the invention may comprise two types of learning: offline learning (which allows, for example, parameterizing a generic aircraft model) and online learning (which then allows parameterizing a model unique to each specific aircraft). (However, offline learning may also be used to specify a specific aircraft.) It is also possible to use only one type of learning (one airline may be only interested in a class of aircraft, while another airline may want to know the specific characteristics of a given aircraft, for example, for fine optimization of fuel consumption).

[0014] In one embodiment, machine learning includes one or more algorithms selected from the group consisting of: support vector machines; classifiers; neural networks; decision trees and / or steps from statistical methods such as Gaussian mixture modeling, logistic regression, linear discriminant analysis, and / or genetic algorithms.

[0015] A computer program product is described, the computer program comprising code instructions for performing one or more of the steps of the method when the program is executed on a computer.

[0016] A system for implementing one or more of the steps of the method is described, the system comprising one or more avionics systems, such as a flight management system FMS and / or an electronic flight bag EFB.

[0017] In one embodiment, the system further comprises one or more neural networks selected from the group consisting of: an artificial neural network; a non-recurrent artificial neural network; a recurrent neural network; a feed-forward neural network; a convolutional neural network; and / or a generative adversarial neural network.

[0018] Advantageously, the method according to the invention allows prediction of aircraft performance independently of the models provided by the manufacturer.

[0019] Advantageously, the method according to the invention allows learning to continue without time constraints (eg online learning, in particular enhanced by using data streams from recordings of commercial flights).

[0020] Advantageously, the method according to the invention can be implemented in an onboard trajectory prediction and / or trajectory calculation system and in particular in an electronic flight bag (EFB). The invention can be implemented in a computer such as an FMS (Flight Management System) or in a system set interconnecting an FMS with one or more EFBs.

[0021] Potential applications of the invention relate to calculating trajectories, assisting aircraft manufacturers in establishing aircraft performance, optimizing airline flight operations, flight simulation, assisting in mission management, assisting in the piloting of an aircraft, adjusting avionics systems in a broader sense, or predictive maintenance by modeling changes in the performance of an aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Various aspects and advantages of the present invention will become apparent with reference to the following drawings, in support of the description of a preferred but non-limiting mode of implementation of the invention:

[0023] Figure 1 Some of the technical objectives pursued by the present invention are shown;

[0024] Figure 2 The relationship between the performance model and the predictive trajectory calculation is shown;

[0025] Figure 3 shows the coupling between machine learning and integration within trajectory prediction;

[0026] Figure 4 Some aspects of the invention according to a particular processing chain are shown;

[0027] Figure 5 Various learning modes according to various embodiments of the present invention are shown. DETAILED DESCRIPTION

[0028] Various types of machine learning are possible. Machine learning is a field in computer science that uses statistical techniques to give computer systems the ability to "learn" using data (e.g., to gradually improve performance on a particular task) without being explicitly programmed for that purpose.

[0029] Machine learning is useful for detecting and identifying trends, patterns, or relationships. Often, it is easier to collect data (e.g., data from a video or board game) than to explicitly write a program that manages the game in question. Furthermore, neural networks (either hardware embodiments or software emulations of machine learning) can be repurposed to process new data. Machine learning can be performed on particularly large amounts of data, i.e., by using as much data as possible (e.g., stability, convergence, weak signals, etc.). New data can be added continuously, and learning can be improved.

[0030] Various learning algorithms may be used in conjunction with features according to the invention. The method may include one or more algorithms from the group consisting of: support vector machine (SVM); "boosting" (classifier); neural network (in unsupervised learning); decision tree ("random forest"); statistical methods, e.g., Gaussian mixture modeling, logistic regression, linear discriminant analysis, and genetic algorithms.

[0031] Machine learning tasks are usually classified into two broad categories, depending on whether there is a "signal" or learning input or an "information feedback" or usable output.

[0032] The expression "supervised learning" refers to the situation where a computer is presented with exemplary inputs and exemplary outputs (whether real or expected). The learning then consists in identifying the linking rules that match the inputs to the outputs (these rules may or may not be understandable to humans).

[0033] The expression “semi-supervised learning” refers to situations where the computer receives only an incomplete dataset: for example, some output data is missing.

[0034] The expression "reinforcement learning" consists in learning, based on experiments, what actions to take in order to optimize a quantitative reward over time. Through iterative experiments, a decision-making behavior (called a strategy or policy, which is a function of relating the actions to be performed to the current state) is determined to be optimal if it maximizes the sum of rewards over time.

[0035] The expression "unsupervised learning" (also called "deep learning") refers to the situation where there are no labels (no explanations, descriptions, etc.), leaving it to the learning algorithm to find one or more structures between the input and output. Unsupervised learning can be a goal in itself (discovering hidden structure in data) or a means to achieve a goal (through feature learning).

[0036] In computer science, an "online algorithm" is an algorithm that receives its input as a stream of data, as opposed to all at once, and must make decisions on the fly. In the context of machine learning, the term "incremental learning algorithm" may be used.

[0037] Since the incremental learning algorithm does not have access to all the data, the incremental learning algorithm must make choices that may ultimately be a posteriori non-optimal. A competitive analysis can be performed by comparing the performance of an incremental learning algorithm and an equivalent algorithm that has all the data available to it on the same data. Online algorithms specifically include, for example, the following algorithms: k-server, BALANCE2, balanced-slack, double cover, equipoise, handicap, harmonic, random-slack, tight-span, tree, and work function algorithms. Online algorithms are related to probabilistic and approximate algorithms.

[0038] Depending on the embodiment, the human contribution in the machine learning step can vary. In some embodiments, machine learning is applied to the machine learning itself (reflexive). In fact, the entire learning process can be automated, especially by using multiple models and comparing the results produced by these models. In most cases, humans will be involved in machine learning ("humans in the loop"). Developers and managers are responsible for maintaining the data set: data ingestion, data cleaning, model discovery, etc. In some cases, human intervention is not required, and learning will be fully automated once the data has become available.

[0039] Machine learning used in conjunction with features of the present invention generally benefits from having access to large amounts of data. The expression "big data" refers to the collection and analysis of data on a large scale. The concept is associated with technical characteristics including: volume (e.g., large data sets, even if they are redundant); variety (e.g., using many different sources); velocity (e.g., data is "new" or constantly updated in a changing or dynamic environment); exhibiting a certain degree of authenticity (e.g., weak signals that are drowned out by noise are not removed and may therefore be detected or amplified); and ultimately representing a certain value (e.g., usefulness from a technical and / or professional (i.e., business) perspective).

[0040] In one embodiment, an "on-policy" learning method can be used. An on-policy method is an iterative method that alternates between a policy evaluation phase and a policy improvement phase. An on-policy method determines the choice of the next action (control) based on the current estimate of the value (or quality) function in the current state; after observing the new current state and the received reinforcement signal, the model already used will be updated. A classic example of this type of method is the SARSA algorithm.

[0041] In one embodiment, an "off-policy" learning method may be used. An off-policy method is insensitive to the way an action is chosen at a given time, instead it is only used to observe control policies with a sufficient degree of exploration. Therefore, an off-policy method is free to observe different control policies (which may be suboptimal). The classic example of an off-policy algorithm is the Q-learning algorithm.

[0042] Figure 1 Some of the technical objectives pursued by the present invention are shown.

[0043] For optimization, it is common practice to use models 101 provided by the manufacturer. These models are usually generic, i.e. theoretical, static and data-poor. These models relate to "average" or "model" aircraft, which are difficult to handle or ultimately not very relevant in certain contexts. In other words, there is a need for improved aircraft models 102 that are "real" (i.e. personalized (different for different aircraft)), dynamic and based on a large amount of data (which is available anyway).

[0044] More specifically, it is advantageous to be able to specify the performance of a specific aircraft in an individualized manner, in particular for optimization (e.g. with respect to fuel consumption). Depending on maintenance events performed on the aircraft or depending on the mission (e.g. load distribution, etc.), the performance data may vary for the same aircraft (which has different loads, dynamic aspects).

[0045] Being able to assess the instantaneous performance of an aircraft in real time, either on board the aircraft or via remote computing, offers significant advantages to airlines that must manage fleets of aircraft.

[0046] The embodiments of the present invention described below at least partially satisfy the needs stated above.

[0047] Figure 2 The relationship between the performance model and the predictive trajectory calculation is shown.

[0048] The trajectory includes multiple points, including point N and point N+1 (trajectory points, matching points on the flight plan, waypoints).

[0049] The performance model 210 (PERFDB) determines the aircraft state; the performance model 210 comprises tables 211 and one or more performance calculators 212. The tables 211 may be estimated (the parameters from the parameterized model are estimated based on actual flight).

[0050] At input, data coming from sensors or from avionics 200 are manipulated (eg the state of the aircraft at point N). At output, parameters SEP, N1 and FF are determined (directly or indirectly) at said point N (220).

[0051] These output parameters 220 are then used in a trajectory calculation model 230 (integrator and propagator 231 , etc.) which predicts the future (or next) state 240 of the aircraft based on the current (or previous) state.

[0052] Aircraft status

[0053] In practice, the state of the aircraft at any given time can be characterized (approximately but satisfactorily) by three parameters: parameters SEP, FF and N1. These parameters are data that can be measured by the aircraft sensors and are therefore accessible in the flight records.

[0054] The abbreviation SEP for "Specific Excess Power" refers to the energy available for an aircraft to climb, i.e., the aircraft's ability to climb divided by its weight (which is not a constant). SEP is not directly measured, but rather is calculated based on measurable data (e.g., altitude, speed, gravity constant, etc.).

[0055] The abbreviation FF for "Fuel Flow" refers to the change in fuel weight.

[0056] Formula 1

[0057]

[0058] The abbreviation N1 refers to the first stage rotation speed of the engine, which is the factor that has the greatest influence on fuel consumption. The available power is closely related to the speed N1.

[0059] The parameters SEP, FF and N1 are closely related. In particular, the engine thrust mode and vertical guidance are the determining factors of the parameters SEP, FF and N1.

[0060] A model-based approach may consist in modeling the interdependencies between SEP, FF and N1, i.e. by formulating a system of equations involving these parameters (e.g. by modeling aerodynamics and / or engine thrust modes and / or vertical guidance). Thus, a particularly efficient (e.g. convergent and fast) optimization may be obtained.

[0061] Model-free methods based on machine learning.

[0062] According to one embodiment of the invention, an advantageous (model-free) alternative is to apply a machine learning method. No prior knowledge is required. In other words, no model is pre-set, whether aerodynamic model or engine model or other model: machine learning matches data sets at input and output, which are real data (i.e. data directly measured or indirectly determined).

[0063] Various machine learning methods may be applied at various levels: between 200 and 220 on the one hand (machine learning 291 ), and between 200 and 240 on the other hand (machine learning 292 ).

[0064] Figure 3 Shown is the coupling between machine learning and integration in trajectory prediction.

[0065] The integral calculation 330 for predicting the trajectory of the aircraft involves only the parameters 311 predicted at point N+1, which are obtained by learning 300 based on point N. The predicted data 311 and the measured data 312 are used to continue training 3100 the model 300.

[0066] The performance model 210 (PERFDB) determines the aircraft state; the performance model 210 comprises tables 211 and one or more performance calculators 212. The tables 211 may be estimated (the parameters from the parameterized model are estimated based on actual flight).

[0067] At the input, data from sensors and / or from avionics are manipulated. At the output, the parameters SEP, N1 and FF 220 are determined (directly or indirectly).

[0068] In one embodiment of the invention, these output parameters are used in a trajectory calculation model 230 (integrator and propagator 231, etc.) which predicts the future (or next) state of the aircraft based on the current (or previous) state.

[0069] The flight parameters at point N are received (measured and / or calculated) and then submitted to the learning model, which:

[0070] - The data has been batch processed (unsupervised learning on a collection of data); or

[0071] - Streaming the data (incremental or online learning, see below).

[0072] In one embodiment, input data (input) and output data (output) are received and / or provided. Machine learning is performed on the output data, and then learning establishes a "link" between the input and the output.

[0073] In one ("differential") embodiment, the learned model is potentially modified by learning the differences between predicted data and data actually measured in flight. The predicted data and / or measured data are manipulated by the flight computer.

[0074] Figure 4 Some aspects of the invention are shown according to a specific processing chain.

[0075] Measurements from sensors and / or calculations and / or other observations are collected in step 410 and possibly filtered and formatted in step 420; a "reduced" (i.e., reference) corpus is defined in step 430 (via feedback from human expertise and / or machine filtering), and an aircraft state table 440 comprising a plurality of aircraft states is determined. By considering a particular aircraft state, the learning model 300 predicts the aircraft state at the next point N+1 based on the data at the previous point N. The learned model 450 is gradually and / or iteratively improved and possibly validated or modified 460 by an operator, and may optionally be the subject of reports and statistics 470 (for certification bodies, traffic control agencies, manufacturers, equipment manufacturers, etc.).

[0076] Figure 5 Various learning modes according to various embodiments of the present invention are shown.

[0077] In one embodiment, for a given aircraft 500, for which it is desired to have a better understanding of its properties (e.g., in order to determine its flight behavior), the method includes a first step 512 of modeling the link between (i) the state of the aircraft and (ii) the parameters (SEP, FF, N1). To this end, the model is trained by machine learning on a large number of flight records related to the type of aircraft in question. In a second step, the model is on the aircraft (i.e., on-board implementation) and is improved so that it is improved, enhanced, or improved by serial number, i.e., it is dedicated to or specific to the aircraft in question (each aircraft is unique and slightly different from other aircraft of the same aircraft type or category). To this end, the (generic) model is made specific by performing machine learning on flight data specific to the aircraft in question. The data can be (past) flight records of the aircraft (offline 521), or received in real time or by streaming (online 522).

[0078] This last embodiment is an "end-to-end" model: a learning model 300 (e.g., a neural network) is trained with the aircraft state as input and SEP, FF, and N1 as output. Then, an (existing) integrator is used with the output from this model to predict the future state of the aircraft.

[0079] The knowledge learned for one aircraft 520 can be reproduced at the scale of a fleet of aircraft 530. The generic model can then be matched to an average of aircraft of the same type.

[0080] In one embodiment, for a given aircraft 500, for which it is desired to have a better understanding of its properties (in order to determine its flight behavior), the method includes a first step 511 of modeling the link between (i) the state of the aircraft at point N and (ii) the state of the aircraft at point N+1. This step includes the step of using trajectory prediction / integration 230. To this end, the model is trained by machine learning on a large number of flight records related to the type of aircraft in question. In a second step, the model is on the aircraft (i.e., onboard implementation) and is improved so that it is improved, enhanced, or improved by serial number, i.e., it is dedicated to or specific to the aircraft in question (each aircraft is unique and slightly different from other aircraft of the same aircraft type or category). To this end, the (generic) model is made specific by performing machine learning on flight data specific to the aircraft in question. The data can be (past) flight records of the aircraft (offline 521), or received in real time or by streaming (online 522).

[0081] The present invention can be implemented based on hardware and / or software elements. The present invention can be used as a computer program product on a computer readable medium.

[0082] Machine learning may correspond to a hardware architecture that can be emulated or simulated by a computer (e.g., CPU-GPU), but occasionally it cannot be emulated or simulated by a computer (there may be dedicated circuits for learning).

[0083] Depending on the embodiment, the method according to the invention may be implemented on or by one or more neural networks. The neural network according to the invention may include one or more neural networks selected from the following neural networks: a) artificial neural networks; b) non-cyclic artificial neural networks, such as multi-layer perceptrons as opposed to recurrent neural networks; c) feed-forward neural networks; d) Hopfield neural networks (discrete-time recurrent neural network models whose connection matrix is ​​symmetric and zeros are located on the diagonal, and whose dynamics are asynchronous, with one neuron updated per unit time); e) recurrent neural networks (composed of interconnected units that interact nonlinearly and have at least one cycle in their structure); f) convolutional neural networks ("CNN" or "ConvNet", which is a type of feed-forward non-cyclic artificial neural network constructed by stacking multi-layer perceptrons); or g) generative adversarial neural networks (GANs, which are a class of unsupervised learning algorithms).

[0084] In one embodiment, the learning calculations are performed off-line on a ground computer.

[0085] Advantageously, if the computer is located on board an aircraft and has access to flight data, the model can be trained specifically to model the specific aircraft it is located on with greater accuracy using augmentation methods.

[0086] Advantageously, if the computer is located on board the aircraft and connected to the parameter recorder, the flight data can be used in a real-time architecture to improve the understanding of the aircraft performance in real time.

[0087] In a variant embodiment, one or more steps of the method according to the invention are implemented in the form of a computer program hosted on an EFB (Electronic Flight Bag).

[0088] In a variant embodiment, one or more steps of the method may be implemented in an FMS (Flight Management System) computer or in an FM function of a flight computer.

Claims

1. A method for managing a flight of an aircraft, The following steps are involved: receiving data (200) from a recording of said flight of the aircraft; said data comprising data from sensors and / or data from onboard avionics equipment; Determining the state of the aircraft at point N (220) based on the received data (200); Determining flight parameters SEP, FF and N1 (220) based on the aircraft state (200) at point N by applying the model learned by the first machine learning (291); Determining the state of the aircraft (240) at point N+1 based on the values ​​of flight parameters SEP, FF and N1 by means of trajectory calculation (230) by applying the model learned by means of the second machine learning (292), Among them, the SEP value represents the energy available for the aircraft to climb, the FF value represents the change in fuel weight, and the N1 value represents the first stage rotation speed of the engine which affects fuel consumption.

2. According to the method of claim 1, the first machine learning (291) and the second machine learning (292) are unsupervised.

3. According to the method of claim 1, the first machine learning and the second machine learning are supervised.

4. According to the method according to any one of claims 1-3, the first machine learning and the second machine learning are performed offline.

5. According to the method according to any one of claims 1-3, the first machine learning and the second machine learning are performed online.

6. According to the method described in any one of claims 1-3, the first machine learning and the second machine learning include one or more algorithms selected from the following algorithms: support vector machine; classifier; neural network; decision tree and / or steps from statistical methods including any one of Gaussian mixture model, logistic regression, linear discriminant analysis and / or genetic algorithm.

7. A computer program product, the computer program comprising code instructions for executing the steps of the method according to any one of claims 1 to 6 when the computer program is executed on a computer.

8. A system for implementing the steps of the method according to any one of claims 1 to 6, the system comprising one or more avionics systems.

9. The system of claim 8, further comprising one or more neural networks selected from the group consisting of: an artificial neural network; a non-recurrent artificial neural network; a recurrent neural network; a feed-forward neural network; a convolutional neural network; and / or a generative adversarial neural network. 10 . The system of claim 8 , wherein the one or more avionics systems include a flight management system (FMS) and / or an electronic flight bag (EFB).

Images