• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    The present invention relates to a method for detecting icing conditions for an aircraft comprising a measurement of the parameters (320) of embedded systems, with the exception of probes dedicated to the detection of ice, a transformation (330) of the measurements of the symptomatic parameters. to obtain P -uplets of values of explanatory variables of icing conditions, a classification (370) of said measurements by classifiers on the basis of the P -uplets of values thus obtained, said classifiers having been previously trained in a supervised manner, each classifier providing a prediction of membership in a class of icing conditions, the predictions of the different classifiers being consolidated (380) to provide a consolidated prediction of icing conditions (390).
    公开号:FR3077059A1
    申请号:FR1850553
    申请日:2018-01-24
    公开日:2019-07-26
    发明作者:Alice Calmels；Jean PASQUIE
    申请人:Airbus Operations SAS；
    IPC主号:
    专利说明:

    METHOD FOR DETECTION OF ICING CONDITIONS FOR AN AIRCRAFT BY SUPERVISED AUTOMATIC LEARNING
    DESCRIPTION
    TECHNICAL AREA
    The present invention relates generally to the forecasting of weather conditions for an aircraft. It also concerns the field of supervised learning (supervised! Machine learning).
    PRIOR STATE OF THE ART
    The occurrence of icing conditions in flight poses a risk to aircraft. In fact, when an aircraft encounters such conditions, it is exposed to an accumulation of frost on its surfaces which can lead to a significant increase in the weight of the aircraft, a loss of lift, problems with actuating the control surfaces, communication and antenna malfunctions, anemometric probe measurement errors, engine thrust losses, these various malfunctions which can ultimately lead to a loss of control of the aircraft.
    To overcome these malfunctions, aircraft authorized to fly in icing conditions are equipped with ad hoc protection systems, in particular heating systems, integrated into the elements to be protected (airfoil, probes, engine air intakes, etc.) preventing training or accumulation of frost.
    The activation of these protection systems is generally based on the pilot's judgment after he has visually identified the presence of icing conditions. This identification being necessarily imperfect, it is generally resorted to mechanical or optical detection systems to help the pilot in his judgment. Thus, it is common to equip an aircraft with probes (or sensors) mounted on the skin of the aircraft and to use the measurements obtained to diagnose the presence of frost. However, these detection systems generally do not automatically activate the protection systems. An assessment of the measures by the pilot remains necessary taking into account the flight phase, the criticality of the functions performed by the elements affected by the frost and the associated safety margins, in order to avoid any untimely triggering of the protection systems.
    Current systems for detecting icing conditions have several drawbacks.
    First of all, these systems are installed on the skin of the fuselage or on an aircraft surface, which, on the one hand, requires drilling the fuselage / the surface in question, providing mechanical reinforcements near the hole , deploy electrical wiring and install additional acquisition systems in electrical cabinets. In addition, the sensors / probes often protrude from the skin of the fuselage and consequently generate induced drag, which affects the performance of the aircraft.
    Then, the current systems for detecting icing conditions have relatively limited performance in the sense that they are only capable of meeting certain very limited conditions for the formation of frost. They are generally ineffective when it comes to detecting the formation of large drops of water or large crystals.
    An object of the present invention is to provide a method of detecting icing conditions for an aircraft, which at least partially remedies the above drawbacks, in particular which does not require additional drilling and wiring operations, neither increases nor the weight of the aircraft or its aerodynamic drag, and allows both to understand a wide range of icing conditions and to provide a more precise diagnosis than in the prior art.
    STATEMENT OF THE INVENTION
    The present invention relates to a method for detecting icing conditions for an aircraft, comprising the following steps:
    measurement of parameters of aircraft systems with the exception of external ice detection probes, said systems being incapable of degraded operation in the presence of ice and said parameters being symptomatic of the presence of ice on the aircraft;
    transformation of the measurements of these parameters into P -uplets of values of variables explaining the icing conditions;
    classification of the measurements by at least one classifier previously trained in a supervised manner, said classifier providing a prediction of belonging to a class of icing conditions.
    The system parameters of the aircraft are advantageously chosen from the parameter lists of ATA21, ATA27, ATA28, ATA30, ATA32, ATA34, ATA36, ATA70 to ATA79.
    The parameters can be selected from temperatures, heating circuit currents, pressures, differences between actuator controls and returns, kinematic, altimetric, barometric and anemometric parameters.
    The transformation of parameter measurements into values of explanatory variables includes for example the calculation of an average, a median, a standard deviation, a variance, a Fourier transform, a low-pass or pass filtering -up, a decomposition in wavelets, a calculation of spectral density.
    The classification step is advantageously carried out by a plurality of classifiers, the respective predictions of these classifiers being consolidated, the result of the consolidation giving a prediction of the presence / absence of frost or not, or else a degree of severity of the conditions of icing.
    The classifiers preferably use classification models chosen from a decision tree classification model, a linear discriminant analysis classification model, a quadratic discriminant analysis classification model, a forest decision tree classification model, a classification model using a bagging of decision trees, a classification model using logistic regression, a classification model using the k nearest neighbors method, a classification model using a boosting of weak classifiers.
    The parameter measurements can in particular be transmitted by the device to a ground station, said ground station carrying out said transformation and classification steps then returning the result of consolidation to the aircraft.
    The present invention also relates to a supervised training method of the method for predicting icing conditions defined above, this supervised training method comprising the following steps:
    measuring said aircraft system parameters during a flight in icing conditions;
    transformation of the measurements of these parameters into P -uplets of values of variables explaining the icing conditions;
    detection of the presence of icing conditions during said flight by means of dedicated probes located on the aircraft;
    allocation of icing conditions classes to said measurements from the conditions detected in the previous step;
    training of a plurality of classifiers from the explanatory variables and the corresponding assigned classes;
    comparison of the prediction performance of said classifiers by cross validation on all of the measurements;
    selection of at least one classifier from among the best performing classifiers to predict icing conditions.
    Preferably, the detection of the presence of icing conditions during said flight also uses meteorological sources external to said aircraft.
    The predictive performance of a classifier is estimated, for example, from the mean absolute value of the prediction error or the quadratic mean value of the prediction error or the mean prediction success rate over all measures.
    The classification models are advantageously chosen from a decision tree classification model, a classification model with linear discriminant analysis, a classification model with quadratic discriminant analysis, a forest tree classification model, a classification model using a bagging of decision trees, a classification model using logistic regression, a classification model using the k nearest neighbor method, a classification model using a boosting of weak classifiers.
    BRIEF DESCRIPTION OF THE DRAWINGS
    Other characteristics and advantages of the invention will appear on reading a preferred embodiment of the invention, made with reference to the attached figures among which:
    Fig. 1 schematically represents a supervised learning method for training a method for detecting icing conditions according to an embodiment of the invention;
    Fig. 2 illustrates the performance of a plurality of classification models after supervised learning according to FIG. 1;
    Fig. 3 schematically represents a method for detecting icing conditions according to an embodiment of the invention, after supervised learning according to FIG. 1;
    Figs. 4A-4D represent examples of detection of icing conditions by different classifiers.
    DETAILED PRESENTATION OF PARTICULAR EMBODIMENTS
    The idea behind the invention is to use the available aircraft data, without the development and installation of specific external probes / sensors and therefore without implantation of probes / sensors on the skin of the aircraft to detect the presence of conditions. icing and estimate, if necessary, their degree of severity. By specific sensors is meant here sensors whose measurements are exclusively intended for the detection of the presence of frost (for example an ice crystal detector). By available airplane data is meant data from airplane systems whose functioning is not degraded by the formation of frost, in other words the reliable data in such a situation (for example the data coming from an anemometric probe, risking be blocked by frost, are not considered available data).
    When an aircraft encounters icing conditions, certain elements of systems, such as sensors or steering actuators have a characteristic response, symptomatic of the presence of frost. These symptoms can vary in number, intensity and frequency depending on the type of icing conditions encountered (ice crystals or supercooled water for example) and their degree of severity (concentration of ice crystals). Consequently, it is possible to select certain symptomatic parameters (of icing conditions) from the available data of the aircraft systems and to deduce therefrom explanatory parameters (also called explanatory variables or “features”) of the presence of icing conditions. and where appropriate their severity. These explanatory variables are used as input to one or more classifiers to determine whether or not there is frost (binary classification) or to apprehend the degree of severity of icing conditions (classification with K modalities where K is the number of classes). The classification models are trained by supervised learning from available data acquired by means of specific instrumentation during flights carried out in icing conditions, as explained below.
    Fig. 1 schematically represents a supervised learning method for a system for detecting icing conditions according to an embodiment of the invention.
    Prior to supervised learning, available data obtained during test flights of an aircraft in icing conditions are collected at 110. These data are measurements of parameters of aircraft systems which are incapable of operating in degraded mode in the presence of frost. In addition, the test aircraft is equipped during these flights with sensors dedicated to the direct detection of ice, known as test sensors.
    Consequently, aircraft parameters are available on the one hand, and on the other hand, a diagnosis of the presence or absence of frost, possibly supplemented by a measurement of the water content (crystals and supercooled water).
    Then, among the parameters measured, a plurality M of parameters symptomatic of the presence of frost is selected at 120. These symptomatic parameters will advantageously be chosen from those listed in the chapters ATA 21 (air conditioning and pressurization), ATA 27 (flight controls ), ATA 28 (fuel), ATA 30 (frost and rain protection), ATA 32 (landing gear), ATA 34 (navigation), ATA 36 (pneumatic), ATA 70 to ATA 79 (engines, controls and engine indications) , canals
    FADEC for Full Authority Digital Engine Control). The symptomatic parameters selected are typically temperatures, heating circuit currents, pressures, differences between actuator controls and returns, kinematic parameters (accelerations, rotation rate, speed), altimetric (altitude), barometric (pressure barometric) and anemometric (apparent wind speed).
    For example, it is possible to choose in chapter ATA 21 the parameters related to the temperatures at different points in the cabin, to the temperatures of the ducts of the pressurization systems and to the defrosting controls, in chapter ATA 27 the related parameters acceleration, clinometry, surface control (elevator and / or ailerons and / or spoiler), in chapter ATA 28 the parameters linked to the fuel temperatures in each compartment, in chapter ATA 30 the parameters linked to energy consumption and power supply of the de-icing and wind protection components, in the ATA chapter 32 the parameters linked to the landing gear temperatures, in the ATA 34 the parameters linked to the attitude of the device (pitch, roll, yaw), dynamic measurements (acceleration, rotation rate), anemobarometric measurements, in chapter ATA 36 the parameters related to cooling system, and in chapter ΑΤΑ 7X the parameters related to vibrations and / or motor regulation and control information.
    According to the invention, the symptomatic parameters are not limited to the ATA chapters listed above and can come from other chapters such as chapter ATA 42 (Integrated Modular Avionics).
    The measurements of symptomatic parameters acquired at a given instant form a sample. All the samples acquired during a measurement campaign are noted S.
    The measured symptomatic parameters are then transformed into explanatory variables of the icing conditions (operation called "features extraction") in 130. The transformation making it possible to pass from symptomatic parameters to the explanatory variables can notably include a calculation of an average, a median, a standard deviation, a variance, a Fourier transform, a low-pass or high-pass filtering, a wavelet decomposition, a calculation of spectral density, etc.
    The purpose of this operation is to remove or reduce the non-explanatory information in the measurements and to prevent overfitting on the symptomatic parameters measured during the tests. In the following, the explanatory variables are denoted X 1 , ..., To each sample of S is thus associated a P -uplet of values of explanatory variables.
    The icing conditions, for their part, can be simply represented by a variable to be explained (or target variable) Y. This target variable is binary, when it is simply a question of predicting the presence or absence of frost, or K modalities, when it is a question of predicting the degree of severity of icing conditions (K then being the number of degrees of severity).
    The measurements from the test sensors are acquired in parallel in 140. These test sensors are, for example, Lyman-alpha hygrometric sensors capable of giving the total water content or TWC (Total Water Content), regardless of the nature of its phase ( liquid or vapor).
    The measurements of the test sensors can optionally be supplemented at 150 with contextual meteorological information from external sources. From the measurements of the test sensors and, where appropriate from this contextual meteorological information, the presence of frost or not is determined in 160. Thus, a class (class labeling) can be assigned to each P -uplet of values of variables explanatory (and therefore for each sample of S) in 130. The classification can be simply binary (absence or presence of frost) or with K modalities, for example the following classification with 4 modalities:
    • absence of icing condition (TWC <0.5) • weak icing condition (0.5 <TWC <1.5) • moderate icing condition (1.5 <TWC <3) • severe icing condition (TWC> 3)
    Other classes may be envisaged by those skilled in the art without departing from the scope of the present invention. For example, a classification may be provided making a distinction between the rate of crystals and the rate of supercooled water.
    From the P-tuples of explanatory variables and the classes assigned to them, several classification models, F V ..., F Q can be trained on the set S, as indicated in 170. A classification model is a function F associating with any P-tuple (x 1 , x 2 , ..., x F ) of values of explanatory variables a prediction y of the value of the variable to be explained Y. More precisely, training a classifier consists in defining, in the space of explanatory variables, the domains associated with each of the possible modalities of the variable to be explained.
    Different classification models can be envisaged, some examples of which are provided below:
    First, we can use a decision tree type model such as CART (Classification And Regression Tree). A classification by decision tree is carried out by partitioning by dichotomy the space of the explanatory variables according to a tree structure, a class being associated with each sheet of the decision tree. This classification model is trained on a part T of the set S (training data set) and tested on the remaining part V of this set (validation data set).
    Alternatively, the same classification model (for example, a decision tree) can be trained on subsets T X , T 2 , ..., T N of S obtained by subsampling S randomly. The N classifiers resulting from this training can be aggregated using an aggregation function (for example majority vote). This technique is known in supervised learning under the name of bagging (or boodstrap agregating).
    Alternatively, a Random Forest Classifier type classification model can also be used. According to this approach, elementary classifiers with decision trees are trained on subsets of S, each classifier using only part of the variables to be explained. The elementary classifiers thus obtained are then aggregated by a majority vote decision, the predicted class being that collecting the majority of the votes of the elementary classifiers.
    Alternatively, a so-called boosting technique combining predictions from several weak classifiers can also be used. A classifier is said to be weak if its prediction error rate is slightly better than that of a purely random prediction (random guessing). By combining the successive predictions of these weak classifiers, a classifier with a low error rate (high accuracy level) can be obtained. Weak classifiers can for example be decision tree classifiers. There are different types of boosting depending on whether the weak classifier is trained on the samples corresponding to the highest prediction errors of the previous weak classifier (Adaboost) or on the quadratic errors of prediction of this classifier (Gradient Boosting).
    The classification model can also be based on a linear discriminant analysis or LDA (Linear Discriminant Analysis) or even a quadratic discriminant analysis or QDA (Quadratic Discriminant Analysis). The linear discriminant analysis assumes that the covariance matrix of the explanatory variables is identical for the different classes. The decision boundaries in the variable space are then hyperplanes. When the covariance matrices in the different classes are not identical, the decision function has a quadratic form (QDA): it is possible to reduce to the previous case by placing oneself in a larger dimension representing not only the variables explanatory themselves but also quadratic variables (two by two products and square explanatory variables).
    Alternatively again, the classification model could use a classification according to the method of k nearest neighbors or £ -NN (k Nearest Neighbors). In this method, to predict the class associated with a given P-tuplet Ω of values of explanatory variables, we proceed to the search for the P-tuples, o) p ..., (i) t closest to Ω, having been obtained during learning. The class of the P-tuplet is then predicted to be the preponderant class (majority vote) among the classes respectively associated with o) p ..., o t .
    Finally, a classification model based on a logistic regression (multinomial logistic regression in the case of a variable to be explained with K modalities) could alternatively be used. According to this approach, the posterior probabilities of the different classes, knowing a F-tuplet of values of explanatory variables, are modeled by means of linear functions. The coefficients of these functions can be determined as those maximizing the logarithmic likelihood on the training set, the search for the maximum being carried out iteratively on the values of the coefficients.
    Of course, still other types of classifiers could be envisaged, such as, for example, state vector machines.
    A description of the different classification models mentioned above can be found in the work by T. Hastie et al. entitled "The elements of statistical learning", 2 nd edition, 2017, published by Springer.
    The performances of the Q classifiers, F x , ..., F Qt corresponding to the different classification models, can then be compared using cross validation, as indicated in 180. According to this approach, the set S of samples is partitioned into subsets (or batches) S r , γ = ι, ..., Γ, each classifier F q being trained on a batch of samples S r and its performance evaluated on sub / = 1 ,. ., Γ γψλ remaining set S À , this for Ag Γ]. The performance of a classifier can be evaluated in terms of the average of the absolute value or the prediction error or mean square error on each subset, the best performing classifier leading to the lowest average error. . Alternatively, the performance of a classifier can be assessed in terms of average classification success rate for each subset.
    The most efficient classifier (s) at the end of the supervised learning phase can then be selected, as indicated in 190.
    The classifiers thus selected will then be used in the phase of predicting icing conditions as explained below.
    Fig. 2 represents the performance of a plurality of classification models at the end of supervised learning according to FIG. 1.
    The different types of classification models are indicated on the abscissa, namely:
    • TREE is a decision tree classification model;
    • QDA is a classification model using a quadratic discriminant analysis;
    • LDA is a classification model using a linear discriminant analysis;
    • RF is a classification model using a forest of decision trees;
    • BAGG is a classification model using a bagging of decision trees;
    • LR is a classification model using logistic regression;
    • KNN is a classification model using the k nearest neighbor method;
    • BOOST is a classification model using a boosting of weak classifiers.
    In the figure, the success rate (accuracy) of each classifier has been indicated on the ordinate. More precisely, the distribution of the success rate has been shown for each classifier using a boxplot. The distribution of the success rate is relative to the different partitions of all the samples used in the cross validation. The whiskers correspond to the minimum value and to the maximum value of the success rate, the lower and upper ends of a box correspond respectively to the lower quartile and to the upper quartile, the horizontal bar inside the box corresponds to the value median.
    Advantageously, a classifier having a high success rate, with a high median value is chosen, for example, the LR model and / or the BOOST model.
    Fig. 3 schematically represents a method for detecting icing conditions according to an embodiment of the invention, after supervised learning according to FIG. 1.
    The detection method is implemented during an operational flight of an aircraft, generally of the same type as that used for test flights except that this time it does not include test sensors (ice detector) capable of directly indicating the presence or absence of frost.
    In step 310, the data available from the aircraft systems are collected at regular intervals, these systems being incapable of degraded operation in the presence of frost. In other words, the modification of behavior or the change of state of these systems in the presence of frost makes it possible to confirm their presence without altering flight safety.
    In step 320, the measurements of the symptomatic parameters of the presence of frost are extracted and, if necessary stored in a memory. These symptomatic parameters are in principle the same as those chosen for the learning method. In other words, these symptomatic parameters will have been chosen from those listed in the chapters ATA21, ATA27, ATA28, ATA30, ATA32, ATA34, ATA36, ATA70 to ATA79. However, if the classifiers chosen at the end of the learning period do not use certain symptomatic parameters, these can be omitted in this acquisition phase.
    In step 330, the measurements of the symptomatic parameters into values of explanatory variables are transformed, as explained previously in relation to step 130 of FIG. 1.
    The transformation of the symptomatic parameters is carried out by a calculation module, such as a processor configured for this purpose.
    In step 370, each classifier, trained on the test flight data and selected in step 190 of supervised learning, predicts the class of icing conditions associated with the values of explanatory variables obtained in the previous step. The classification is of the same type as that trained during the learning phase. It can be binary or log 2 K -aire depending on whether a prediction of the presence / absence of frost or whether a prediction of the degree of severity of the icing conditions is desired.
    In any event, when several classifiers have been selected, their respective predictions are consolidated in step 380, for example according to a majority voting procedure. When the number of classifiers selected is even, it may be agreed that one of them has a casting vote. According to a variant, regressors can be used in place of the classifiers to each estimate a (continuous) degree of severity, and perform an average between them before possible discretization.
    Depending on the result of consolidation, it is determined in 390 whether there is frost or not (binary classification) or the degree of severity of the icing conditions (multinomial classification).
    The classification, consolidation and prediction steps are performed by one or more calculation modules. These calculation modules can be hardware modules or software modules, for example software modules executed by the aforementioned processor. Where appropriate, the classifiers can be implemented in separate processors operating in parallel, consolidation and prediction being performed by a programmable combinational logic circuit, such as an FPGA. Those skilled in the art can envisage different modes of implementing these steps without departing from the scope of the present invention.
    The method for predicting icing conditions can be performed entirely on board the aircraft, in on-board equipment such as FWC (Flight Warning Computer) or EFB (Electronic Flight Bag), after the classifiers have been trained on the ground (or in the test aircraft). Alternatively, the symptomatic parameters can be transmitted to the ground for remote monitoring of the ARTHM (Airbus Real Time Health Monitoring) type with return of the prediction result to the aircraft. In all cases, the predicted or estimated icing condition may be displayed on a cockpit screen and possibly generate an alarm. The pilot will then have the option of activating the ice protection systems. Alternatively, the icing condition may automatically trigger ice protection systems.
    By way of example, a method for detecting icing conditions is described below.
    The symptomatic parameters were chosen in ATA27, ATA34 and ATA7X, namely the redundant FCPC flight control parameters (Flight Control Primary Computer) * _FCPC1_COM; * _FCPC2_COM; * _FCPC3_COM, the redundant kinematic parameters ADIRU (Air Data Inertial Reference Unit) ADIRU _ * _ 1, ADIRU _ * _ 2, ADIRU _ * _ 3 and the A-B FADEC control channels of the two motors, namely POLOCAL_ [1; 2] A; POLOCAL_ [1; 2] B; T12LOCAL_ [1; 2] A; T12LOCAL_ [1; 2] B.
    The explanatory variables are obtained by taking:
    the median value of * _FCPC1_COM, * _FCPC2_COM, * _FCPC3_COM then by calculating the min, max, mode and median value of the value obtained on a sliding window of 10s;
    the median value of ADIRU _ * _ 1, ADIRU _ * _ 2, ADIRU _ * _ 3 then by calculating the min, max, mode and median value of the value obtained on a sliding window of 10s;
    the maximum value of POLOCAL_ [1; 2] A and POLOCAL_ [1; 2] B, then by calculating the min, max, mode and median value of the value obtained on a sliding window of 10s;
    the maximum value of T12LOCAL_ [1; 2] A; T12LOCAL_ [1; 2] B, then by calculating the min, max, mode and median of the value obtained on a sliding window of 10s.
    The classification was at 4 degrees of severity of icing conditions as indicated above.
    However, instead of using a single multinomial classifier, 4 binary classifiers were used for each of the TWC intervals. The 4 classifiers are based on independent Gradient Boosting classification models. Table I shows the performance of the classifiers in terms of success rate (accuracy) and precision (sensitivity):
    τ (%) σ (%) model # 1 96.80 97.16 model # 2 96.07 97.96 model # 3 97.32 82.08 model # 4 97.36 88.33
    Table I
    The success rate τ is defined as the ratio between the sum of the number of predictions that are actually positive (7P) and the number of predictions that are actually negative (77V) over all of the positive and negative predictions (wrong or not) :
    TP + TN ~ TP + TN + FP + FN where FP (resp. FN) is the number of false positive (false negative) predictions.
    Precision is the fraction of positive predictions that are actually positive:
    TP σ = ---- TP + FP
    It is therefore possible to correctly classify icing conditions based on symptomatic parameters without adding any specific probe (frost sensor) and with a success rate of around 80%.
    Figs 4A to 4D represent the results of classification of icing conditions by classifiers based on the four classification models mentioned above.
    The first classification model in Fig.
    4A predicts the condition (variable to explain binary) 7WC <0.5, the second classification model in Fig.
    4B, predicts the condition 0.5 <7WC <1.5, the third classification model in Fig.
    4C predicts the condition 1.5 <TWC <3 and finally the fourth classification model in Fig.
    4D predicts the TWC condition> 3. For each of these figures, the real icing condition has been represented in the upper part and, in the lower part, the icing condition predicted by the corresponding classifier. Those skilled in the art will note a very good correlation between the predicted conditions and the real conditions regardless of the severity of the icing conditions.
    权利要求:
    Claims (11)
    [1" id="c-fr-0001]
    1. Method for detecting icing conditions for an aircraft, characterized in that it comprises the following steps:
    measuring (320) parameters of aircraft systems except for external ice detection probes, said systems being insusceptible to degraded operation in the presence of ice and said parameters being symptomatic of the presence of ice on the aircraft;
    transformation (330) of the measurements of these parameters into F-tuples of values of variables explaining the icing conditions;
    classification (370) of the measurements by at least one classifier previously trained in a supervised manner, said classifier providing a prediction of belonging to a class of icing conditions.
    [2" id="c-fr-0002]
    2. Method for detecting icing conditions according to claim 1, characterized in that the aircraft system parameters are chosen from the parameter lists of ATA21, ATA27, ATA28, ATA30, ATA32, ATA34, ATA36, ATA70 to ATA79.
    [3" id="c-fr-0003]
    3. Method for detecting icing conditions according to claim 1, characterized in that the parameters are selected from temperatures, heating circuit currents, pressures, deviations between actuator commands and returns, kinematic parameters, altimetric, barometric and anemometric.
    [4" id="c-fr-0004]
    4. Method for detecting icing conditions according to any one of the preceding claims, characterized in that the transformation of the parameter measurements into values of explanatory variables comprises the calculation of an average, a median, a deviation -type, of a variance, a Fourier transform, a low-pass or high-pass filtering, a decomposition in wavelets, a calculation of spectral density.
    [5" id="c-fr-0005]
    5. Method for detecting icing conditions according to any one of the preceding claims, characterized in that the classification step (370) is carried out by a plurality of classifiers, the respective predictions of these classifiers being consolidated (380), the result of consolidation giving a prediction (390) of the presence / absence of frost or not, or a degree of severity of the icing conditions.
    [6" id="c-fr-0006]
    6. Method for detecting icing conditions according to claim 5, characterized in that the classifiers use classification models chosen from a decision tree classification model, a classification model with linear discriminant analysis, a classification model with quadratic discriminant analysis, a forest classification model of decision trees, a classification model using a bagging of decision trees, a classification model using logistic regression, a classification model using the k nearest neighbor method, a classification model using a boosting of weak classifiers.
    [7" id="c-fr-0007]
    7. Method for detecting icing conditions according to claim 5 or 6, characterized in that the parameter measurements are transmitted by the apparatus to a ground station, said ground station carrying out said transformation and classification steps then returning the result of consolidation to the aircraft.
    [8" id="c-fr-0008]
    8. supervised training method of a method for predicting icing conditions according to claim 1, characterized by the following steps:
    measuring (120) said aircraft system parameters during a flight in icing conditions;
    transformation (130) of the measurements of these parameters into P-tuples of values of variables explaining the icing conditions;
    detecting (140) the presence of icing conditions during said flight by means of dedicated probes located on the aircraft;
    assigning (160) icing condition classes to said measurements from the conditions detected in the previous step;
    training (170) of a plurality of classifiers from the explanatory variables and the corresponding assigned classes;
    comparison (180) of the predictive performance of said classifiers by cross validation on all of the measurements;
    selecting (190) at least one classifier from among the best performing classifiers to predict icing conditions.
    [9" id="c-fr-0009]
    9. Supervised training method according to claim 8, characterized in that the detection of the presence of icing conditions during said flight also uses meteorological sources external to said aircraft.
    [10" id="c-fr-0010]
    10. Supervised training method according to claim 8 or 9, characterized in that the predictive performance of a classifier is estimated from the mean absolute value of the prediction error or from the mean squared value of the prediction error or the average prediction success rate on all measures.
    [11" id="c-fr-0011]
    11. Supervised training method according to claim 10, characterized in that the classifiers use classification models chosen from a classification model with decision tree, a classification model with linear discriminant analysis, a classification model with discriminant analysis quadratic, a forest tree classification model of decision trees, a classification model using a bagging of decision trees, a classification model using logistic regression, a classification model using the k nearest neighbor method, a classification using a boosting of weak classifiers.
    类似技术:
    公开号 | 公开日 | 专利标题
    EP3517442B1|2020-04-22|Method for detecting freezing conditions for an aircraft by supervised automatic learning
    CN104508624B|2017-06-30|The method and system of aircraft data is asked and retrieved during aircraft flight
    FR2970358A1|2012-07-13|PROGNOSTIC OF DURATION BEFORE MAINTENANCE BY FUSION BETWEEN MODELING AND SIMULATION, FOR ELECTRONIC EQUIPMENTS ON BOARD IN AN AIRCRAFT
    CA2826608A1|2012-08-23|Monitoring of an aircraft engine for anticipating maintenance operations
    WO2012049396A1|2012-04-19|System for monitoring an engine test bed
    US10311202B2|2019-06-04|Probabilistic load and damage modeling for fatigue life management
    EP3039497B1|2018-08-29|Monitoring of an aircraft engine to anticipate maintenance operations
    FR2991486A1|2013-12-06|METHOD AND DEVICE FOR ASSISTANCE IN FOLLOWING MISSION OF AN AIRCRAFT
    EP2966526B1|2017-01-11|A method and a system for merging health indicators of a device
    FR3004422A1|2014-10-17|METHOD FOR PREDICTING ANOMALY IN A SUCTION AIR CIRCUIT
    FR2905778A1|2008-03-14|METHOD FOR VERIFYING RELEVANCE OF A MASS VALUE OF AN AIRCRAFT
    US20130246860A1|2013-09-19|System monitoring
    US8521341B2|2013-08-27|Methods and systems for fault determination for aircraft
    FR3046268A1|2017-06-30|AIRCRAFT FLIGHT DATA OPERATION SYSTEM
    US20160011073A1|2016-01-14|Load Estimation System for Aerodynamic Structures
    EP3572332B1|2020-10-14|Apparatus for detecting anomalous aircraft behavior at takeoff
    Li et al.2018|Operational anomaly detection in flight data using a multivariate gaussian mixture model
    Laurence III et al.2017|Wind tunnel results for a distributed flush airdata system
    Lv et al.2018|A novel method of overrun risk measurement and assessment using large scale QAR data
    KR101193116B1|2012-10-19|Method of recognizing flight pattern of rotor craft, apparatus for recognizing flight pattern implementing the same and recording medium
    EP3546365A1|2019-10-02|Detection of icing conditions for an aircraft by analysis of electrical power consumption
    EP3923104A1|2021-12-15|Method and system for controlling a damage level of at least one part of an aircraft, associated aircraft
    US20190193866A1|2019-06-27|Method and assistance system for detecting a degradation of light performance
    Mavris et al.2010|Investigation of a health monitoring methodology for future natural laminar flow transport aircraft
    Wang et al.2019|Turbofan Engine Baseline Model Extraction Based on FDR Data
    同族专利:
    公开号 | 公开日
    EP3517442B1|2020-04-22|
    EP3517442A1|2019-07-31|
    CN110070098A|2019-07-30|
    FR3077059B1|2020-01-31|
    US20190225346A1|2019-07-25|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    EP2979980A1|2014-07-29|2016-02-03|Airbus Helicopters|Method and device for detecting when an aircraft flies in icing conditions|
    EP3269944A1|2016-07-11|2018-01-17|Rolls-Royce plc|A method of operating a gas turbine engine|
    GB2590671A|2019-12-23|2021-07-07|Airbus Operations Ltd|Control system|
    CN111291505B|2020-05-08|2020-10-09|中国空气动力研究与发展中心低速空气动力研究所|Wing-type icing shape prediction method and device based on depth confidence network|
    CN112046761B|2020-08-04|2021-10-15|中国空气动力研究与发展中心计算空气动力研究所|Airplane icing on-line detection method based on statistical test and filtering|
    法律状态:
    2019-01-24| PLFP| Fee payment|Year of fee payment: 2 |
    2019-07-26| PLSC| Search report ready|Effective date: 20190726 |
    2020-01-21| PLFP| Fee payment|Year of fee payment: 3 |
    2021-10-08| ST| Notification of lapse|Effective date: 20210905 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1850553A|FR3077059B1|2018-01-24|2018-01-24|METHOD FOR DETECTING ICING CONDITIONS FOR AN AIRCRAFT BY SUPERVISED AUTOMATIC LEARNING|
    FR1850553|2018-01-24|FR1850553A| FR3077059B1|2018-01-24|2018-01-24|METHOD FOR DETECTING ICING CONDITIONS FOR AN AIRCRAFT BY SUPERVISED AUTOMATIC LEARNING|
    CN201910042002.XA| CN110070098A|2018-01-24|2019-01-17|Method for learning the frost situation come explorer vehicle automatically by supervised|
    EP19152850.4A| EP3517442B1|2018-01-24|2019-01-21|Method for detecting freezing conditions for an aircraft by supervised automatic learning|
    US16/255,106| US20190225346A1|2018-01-24|2019-01-23|Method for detecting freezing conditions for an aircraft by supervised automatic learning|
    [返回顶部]