阅读说明：本技术 用于估计飞行器空中数据的基于模型和飞行信息的组合训练的神经网络系统 (Neural network system for estimating combined training of aircraft aerial data based on model and flight information ) 是由 布鲁诺·贾维尔·卡瓦略 古斯塔沃·乔斯·扎姆布雷诺 莱安德罗·费尔南德斯·贝加莫 朱利亚诺·德 于 2017-10-11 设计创作，主要内容包括：使用被训练以独立于任何来自其值基于气流压力测量的空中数据传感器的信号的神经网络来估计飞行器空中数据。(Aircraft air data is estimated using a neural network trained to be independent of any signal from an air data sensor whose value is based on airflow pressure measurements.)

1. A method of generating estimated over-the-air data, comprising:

on-board the aircraft, using sensors to sense parameters other than air data based on airflow pressure measurements; and

operating a trained neural network on the aircraft to estimate airborne data in response to the sensed parameter other than the airborne data based on airflow pressure measurements.

2. The method of claim 1, further comprising: training the neural network on sensor data other than airborne data.

3. The method of claim 1, further comprising: filtering the aerial data estimated by the neural network by a low pass filter, a Kalman filter, or a complementary filter.

4. The method of claim 1, further comprising: preprocessing the sensed parameter prior to applying the sensed parameter as input to the neural network.

5. The method of claim 1, wherein the sensed parameter comprises a group consisting essentially of W, GS, H, Gamma, Thrust, Theta, Nx, Ny, Nz, and delta temperature or TAT.

6. The method of claim 1, wherein the sensed parameters include some or all of H-Stab, Elev, CG, Ice, Phi, TAT, Q, P, and R.

7. The method of claim 1, further comprising: the neural network airspeed estimate is combined with at least one inertial measurement using a kalman filter or a complementary filter.

8. The method of claim 1, wherein the parameter sensed is independent of an airflow pressure measurement.

9. The method of claim 1, further comprising: detecting a failure of the airborne data and in response to the detected failure, using a value relating to the airborne data captured prior to the failure and either maintaining the value or updating it with a sensor parameter other than the airborne data and inputting the updated value to the neural network.

10. A system for generating estimated over-the-air data, comprising:

at least one sensor on-board an aircraft configured to sense a parameter other than airborne data; and

a processor on at least one aircraft operatively coupled to receive the sensed parameter, the processor operating a trained neural network to estimate aerial data in response to the sensed parameter.

11. The system of claim 11, wherein the processor is configured to perform low pass filtering, kalman filtering, or complementary filtering on the estimated aerial data prior to providing the estimated aerial data to a flight computer and/or display.

12. The system of claim 11, wherein the processor is configured to detect a failure of the over-the-air data and to update the over-the-air data-related values captured prior to the over-the-air data failure using the sensed parameter other than the over-the-air data.

13. The system of claim 11, wherein the estimated airborne data comprises airspeed.

14. The system of claim 11, wherein the estimated aerial data comprises a mach number.

15. The system of claim 11, wherein the estimated aerial data comprises an angle of attack.

16. The system of claim 11, wherein the estimated aerial data comprises a sideslip angle.

17. The system of claim 11, wherein the estimated airborne data comprises dynamic pressure.

18. A method of training a neural network, comprising:

inputting training data including model data and/or flight test data to a neural network, the training data including at least some parameters independent of the airborne data; and

the neural network is trained such that, in the event of an airborne data failure, the neural network is able to estimate the airborne data based solely on parameters that are independent of the airborne data.

19. The method of claim 19, wherein the neural network is trained to estimate aircraft air data independently of any signals from air data sensors whose values are based on airflow pressure measurements.

Technical Field

The technology herein relates to machine learning and neural networks, and more particularly to methods and apparatus for estimating aircraft air data by means of a neural network trained independently of any signal from a conventional air data sensor whose values are based on airflow pressure measurements.

Background

Airspeed, the speed of an aircraft relative to the mass of air surrounding it, is a critical signal in all aircraft operations. It is used by both pilots and onboard systems to influence their operational decisions or control system actions.

In fly-by-wire aircraft, in particular, the need for availability and integrity of this airspeed parameter has increased. In view of the highly integrated nature of the functions, airspeed is critical to ensure safety of these aircraft or at least to provide sufficient processing quality in flight. The incorrect airspeed values may result in unique fault conditions that previous system architectures would not normally experience, thereby resulting in unacceptable stability and control performance possibilities.

For example, as a result of unreliable airspeed events, flight crew may become confused and unable to diagnose a stall condition. Such accidents have occurred at least once in the aviation history. On day 1 of 6 months in 2009, airbus a330-203 operated by french air (AF 447 flight) takes off from riyork heat, inland, to paris. Temporary inconsistencies were observed between the measured airspeeds, possibly as a result of the pitot tube probe becoming clogged with ice crystals. This inconsistency causes the autopilot to disengage and makes the aircraft flight crew unclear of airspeed. The crew cannot then diagnose the aerodynamic stall condition caused by low speed and head-up attitude. Thus, the flight crew presents a lack of input that could otherwise make recovery possible.

Drawings

The following detailed description of exemplary, non-limiting illustrative embodiments will be read in conjunction with the following drawings, in which:

FIG. 1 shows an example, non-limiting development process of a neural network, divided into a pre-flight development phase consisting of training and embedding the network and a second part of the in-flight operation where the estimation is performed in real time.

FIG. 2 provides more detail regarding an example, non-limiting operation of a neural network onboard a flight computer. Fig. 2, if associated with fig. 1, would represent a more detailed view of items (18), (15), (16), (19) and (10). The pre-processing step (28) and the filter (24) are examples of such detailed information.

Fig. 3 goes further into detail regarding an exemplary non-limiting input to a neural network, which is divided into mandatory (31) and alternative (32). They represent both (17) and (18) in fig. 1, and (21) and (22) in fig. 2. Further details given in fig. 3 relate to the example of the filter (35) and the possibility of over-the-air data output (36).

Detailed Description

After a legal aviation accident, certification authorities in different countries have expanded their attention to the impact of erroneous airborne data in aircraft. Some countries now require that the consequences of erroneous data from two sensors of the same type do not result in catastrophic failure. For more than two sensors of the same type, the consequences of erroneous data should be minimized where feasible. It is mentioned that the use of independent information sources, analytical redundancy or model-based fault detection and isolation techniques are contemplated. To avoid these types of situations, redundant sensors and heating devices may be used to reduce certain environmental effects.

In this regard, it is helpful to link the ideas of the same type of sensors mentioned by the institution with the methods used by conventional sensors, such as pitot tubes, to measure airborne data. Pitot tubes cannot directly measure airspeed. Instead, pitot tube sensors measure total and static pressures. The difference between these two measurements is sensed and used to calculate airborne data, including aircraft airspeed and dynamic pressure.

Unfortunately, in many cases, the pressure line of a pitot tube sensor may become blocked by an unexpected object. These obstructions may include, but are not limited to: ice, rain, hail, insects, volcanic ash, or sand. Pitot tubes may also face other threats that may cause erroneous over-the-air data calculations. Since they are external and elevated sensors, bird strikes may affect them.

Large jet aircraft have multiple pitot tube sensors, but a blocked pitot tube may eliminate the advantage of redundant sensors in providing multiple sensed parameters for comparison and analysis. Furthermore, if all of these sensors are of the same type, they may all be blocked at the same time (common causal event). This would result in inaccurate identification of airspeed indications and other parameters based on pressure measurements. The fault isolation logic may not be able to detect simultaneous faults. This is the case where the agency requires avoidance by adding a different type of sensor.

In many other cases, different sensors (i.e., using different types of sensors) are also necessary or helpful in order to mitigate common causes or common mode failures. Maintenance errors or development errors can also be mitigated by using different types of sensors.

For example, angle of attack (attack) blade sensors may be used as the different sensors. It is an airborne data sensor for measuring the angle of attack of an aircraft, where the blades are attached to a freely rotatable shaft. Once the angle of attack is known, the resultant airspeed may be calculated from the angle of attack value in combination with some other parameters such as weight and altitude (this method is used by patent US2010/0100260a 1).

Some of these sensors operate based on the doppler effect (changes to the frequency or wavelength of the waves or other periodic events at which the observer moves relative to its source).

However, sensors that introduce new technology often require special attention-i.e., the development of new algorithms for voting or otherwise processing their signals. Of course, new sensors may fail in new unpredictable ways. System logic should be prepared to handle this failure mode. For example, an angle of attack blade may become stuck for some reason. In addition, bearing in mind that increasing the number of sensors means increasing the weight of the aircraft and increasing the associated system complexity.

To avoid all these problems, some model-based approaches are being considered. For example, to estimate airspeed Using only the global positioning system and inertial navigation system, An extended Kalman Filter may be designed Using the 6DOF equations of motion (Yong-gonjng Park and Chan Gook Park have their paper name "WindVelocity Estimation with airspeed Sensor Using Kalman Filter Without airspeed Sensor in colored Measurement Noise" on the 30 th Congress of the International Council of the Aero scientific of 2016). However, this approach relies on implementing a good and reliable aircraft model in the kalman filter. Since modern transonic aircraft may have complex and highly non-linear aerodynamic and thrust models, this approach may be difficult to implement and adjust, or become too complex to be practical.

New method

The technology herein provides a reliable and advantageous method to compute independent airborne data sources that addresses the above-mentioned problem, i.e., it is designed to be completely independent of traditional airborne data sensors, and it uses other sensors already present in modern aircraft systems. Such a system would mitigate common causal patterns of failure without increasing the weight, cost, and complexity of the aircraft system.

The example non-limiting techniques herein estimate the airborne data using a low-cost software-based neural network and a correlation algorithm.

Definition of "data over the air

For the purposes of this disclosure, from the term "over-the-air data", any element in the following list of options can be understood:

angle of attack: an angle formed between a reference in the aircraft body and an airspeed vector in a plane perpendicular to the wing;

side slip angle: an angle formed between a reference in the aircraft body and an airspeed vector in a plane containing the wing;

space velocity: the speed of the aircraft relative to its surrounding air mass;

mach number: the ratio between the magnitude of the airspeed and the speed of sound.

Dynamic pressure: kinetic energy per unit volume of fluid moving around the aircraft.

The techniques disclosed herein are not limited to always estimating the entire air data set mentioned above. Depending on the particular implementation or design goals, it may be implemented to estimate some of the parameters, or only one of them. Also, if desired, more than one neural network may be implemented to estimate the airborne data signals.

Neural network method

Neural networks are computing systems inspired by biological neural networks that make up the animal brain. Such systems learn (i.e., are trained to step up performance) to perform tasks or estimate parameters by considering examples. Neural networks consist of a collection of gains and biases (neurons), which are passed through some non-linear activation function. These neurons are organized into layers that connect inputs to desired outputs.

Training neural networks

FIG. 1 illustrates an example non-limiting development process of a neural network, divided into a pre-flight development phase consisting of training and embedding the network, and a second part that performs estimated in-flight operations in real time.

The neural network of the exemplary non-limiting embodiments herein is trained (13) on a desktop or other computer through data obtained from aircraft model simulations (11) and real flight results (12). Training involves collecting a series of input-output conjugate samples (model or flight test data) and using an optimization algorithm to adjust the neural network bias and gain to provide the best possible estimate.

As shown in FIG. 1, in an exemplary non-limiting embodiment, examples for training a neural network may be obtained from two different data sources: aircraft representation model (11) or real flight test data (12).

1. A model (11) of the aircraft. During the development of an aircraft, many analyses use aircraft models. It may include best knowledge obtained from the wind tunnel, computational fluid dynamics, previous experience, mathematical models, or flight test data. In general, obtaining data using a model may be cheaper and faster than using real aircraft flight data.

2. Real flight test data (12). Signals recorded along the flight during the aircraft development phase (12) may also be used as an example for training the neural network.

Regardless of how the entire training data set is constructed, i.e., from which source they are derived or how much data is derived from each source, it is important that they must cover the flight envelope of the aircraft.

The process of training the network (13) shown in fig. 1 consists of using an optimization algorithm to adjust the neural network bias and gain to provide the best nonlinear input-output relationship between the input variables and the desired output. It is contemplated that the training step (13) of the process comprises the following iterative sequence: training the neural network, verifying the obtained relationships, analyzing the accuracy of the estimates, and if necessary rearranging the topology of the network, restarting the loop, or freezing the obtained relationships as a final result once the training is sufficiently complete.

After the neural network has been trained by the method shown in fig. 1, the gain, bias and topology of the neural network are frozen and, as a result, are ready to be embedded in the aircraft (14). The frozen relationship is stored in non-transitory storage in any on-board computer (27), which on-board computer (27) is able to perform the same kind of estimation in real time over the air. Note that the neural networks (23) (34) should not remain trained at this stage (14).

There are an countless number of possibilities for use as storage devices. Some options include:

1. a special purpose computer.

2. The neural network is included as part of the flight control software and embedded in the flight control computer.

3. Built into flight management computer

4. Avionics software and embedded in any avionic computer.

On-board operation of trained neural networks

Once the desired performance is achieved, the neural network is frozen and implemented in any on-board computer (27) on the aircraft to provide real-time air data estimation as shown in fig. 1 and 2.

Fig. 1 shows that the embedded neural network receives inputs (18) and data measured by aircraft sensors (17) (block 15), processes the data (block 16) and computes outputs (block 19). Additional systems may be used to process the calculated output (block 10).

When embedded in an aircraft (14) as shown in the lower part of fig. 1, the neural network receives inputs from aircraft sensors (15). These inputs represent the same list of signals previously used in training the neural network algorithms (11) (12), but now they are either read in real time by the aircraft sensors (17) or are informed as input parameters by the pilot (18). By inputting these signals on a pre-trained neural network (16), the on-board computer calculates the desired output in real time.

FIG. 2 illustrates the overall on-board neural network system and provides more detail regarding an exemplary, non-limiting operation of the on-board neural network of the flight computer. Fig. 2, if associated with fig. 1, will represent a more detailed view of items (18), (15) (16), (19) and (10) of fig. 1. The pre-processing step (28) and the filter (24) are examples of such detailed information.

As shown in fig. 2, the previously trained neural network (23) consumes real-time data from the aircraft sensors (21), sometimes after pre-processing (28) it, and computes an estimated aerial data output. Some low pass filter, kalman filter or complementary filter may be used to filter the aerial data pair output of the neural network estimate (24). Depending on the particular implementation, the output (19) of the neural network may be filtered (24) (35) to reduce the noise behavior of the signal. The bandwidth (24) (35) of the filter may be adjusted to suit the particular implementation and type of filter used. The test may indicate that a particular type of filter is recommended in each case. Exemplary non-limiting embodiments cover options including, for example, a low pass filter, a kalman filter, and a complementary filter (24) (35). For example, by combining a neural network airspeed estimate with the aircraft longitudinal acceleration measured in a Kalman or complementary filter, the neural network estimating airspeed may reasonably eliminate the noise behavior of its signal. In some embodiments, the function of such filtering may be to "fuse" the estimated aerial data with other input data to provide a more accurate, reliable, and/or robust output.

The filtered estimated aerial data may be used by some on-board computer, such as a flight control computer (25), or may be used to display synthetic aerial data (such as airspeed) to the pilot (26). Some of the logic (10) that may consume the output of the neural network (unfiltered or filtered, as previously described) is flight control system logic (25) (e.g., control law or general design error monitoring) and flight crew indications (26) (e.g., indications of estimated airspeed of the pilot via an avionics display screen).

Sensor input to neural networks

The method disclosed herein does not require a list of additional sensors specifically installed on the aircraft to provide the estimate. An exemplary non-limiting embodiment uses the currently available sensors of modern aircraft systems to supply the necessary parameters during flight to estimate the desired synthetic airborne data parameters (see FIG. 3). In an exemplary, non-limiting embodiment, the input list is carefully selected to form inputs only by sensors of a different type than those typically used to measure airborne data (i.e., only by sensors that do not depend on the measurement of total and static pressures, which have pressure lines, such as pitot tubes, that may be blocked by unexpected objects).

Fig. 3 goes further into detail regarding an exemplary non-limiting input to a neural network, which is divided into mandatory (31) and alternative (32). They represent both (17) and (18) in fig. 1, and (21) and (22) in fig. 2. Further details given in fig. 3 relate to the example of the filter (35) and the possibility of over-the-air data output (36).

As shown in fig. 3, in one exemplary non-limiting embodiment, the neural network is based on the following list of "mandatory" inputs (31) (although please see the following description, "mandatory" does not necessarily mean "strictly necessary" in all embodiments and in all contexts):

weight (W): estimated aircraft weight. In some embodiments, once the weight estimate may depend on the availability of the over-the-air data, the value received from the Flight Management System (FMS) represents a surrogate in some implementations. Another option is to implement some logic to freeze the current estimated weight stored in memory before some aircraft systems detect an air data failure, and then use some estimated fuel flow to digitally integrate the current weight over time.

Ground Speed (GS): typically measured by an Inertial Reference Unit (IRU) of the aircraft (e.g., accelerometer and/or gyroscope sensors) or a Global Positioning System (GPS), but may be supplied by other systems in some particular implementations. Corresponding to the speed of the aircraft relative to the ground, considered as an inertial reference;

geometric height (H): typically measured by the Global Positioning System (GPS) of the aircraft, but may be supplied by other systems in some particular embodiments. Corresponding to altitude;

trajectory angle (Gamma): typically measured by the Inertial Reference Unit (IRU) of the aircraft (see above), but may be supplied by other systems in some particular embodiments. Since the inertial velocity of an aircraft is a vector tangential to its trajectory, Gamma is the angle formed between this vector and the plane of the earth;

for example, fan speed (N1), or Thrust lever angle (T L A), or Engine Pressure Ratio (EPR), or others may be used.

Posture angle (Theta): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. Corresponding to one of the euler angles measured between the central axis of the fuselage and the plane of the earth;

longitudinal acceleration load factor (Nx): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. Corresponding to the resultant force of the external forces acting on the longitudinal axis of the aircraft normalized by the actual weight of the aircraft. Instead of Nx, any parameter having some physical relationship with the magnitude of the aircraft longitudinal acceleration may be used.

Vertical acceleration load factor (Nz): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. Corresponding to the resultant force of the external forces acting on the vertical axis of the aircraft normalized by the actual weight of the aircraft. Instead of Nz, any parameter having some physical relationship to the magnitude of the vertical acceleration of the aircraft may be used.

Lateral acceleration load factor (Ny): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. Corresponding to the resultant force of the external forces acting on the transverse axis of the aircraft normalized by the actual weight of the aircraft. Instead of Ny, any parameter having some physical relationship to the magnitude of the aircraft lateral acceleration may be used.

Delta temperature (delta temperature) from standard atmosphere: calculation corresponding to deviation of the outside air temperature from the standard atmosphere model expectation such as ISA (international standard atmosphere ICAO model), but is not limited thereto. The calculation uses the measured Total Air Temperature (TAT), mach number and pressure altitude. Since this algorithm depends on the parameters measured by the aircraft airborne data system, special attention is required to mitigate common causes that may lead to loss of the airborne data system. Then, using some logic similar to the weight signal, the signal may need to be pre-processed. The last trustworthiness value computed before the over-the-air data was lost may be retained. Another option is that if the Total Air Temperature (TAT) sensor and the airborne data sensor do not have a common mode failure, such as the presence of ice crystals or a design error, the delta temperature (TAT) may be replaced with the Total Air Temperature (TAT) as an input to the estimation logic.

A second input list of the neural network is presented below. This list (32) contains signals that may improve the accuracy of the estimation in some specific embodiments and/or signals that may represent alternatives if the previously mentioned mandatory signals are not available. Thus, the term "mandatory" as used above does not mean that the example embodiment cannot function or be designed without a specific signal.

Horizontal stabilizer position (H-Stab): typically supplied by a Horizontal Stabilizer Control Unit (HSCU), but may be supplied by other systems in some particular implementations. Corresponding to a measure of the deflection of the aircraft horizontal stabilizer relative to a predetermined reference in the aircraft.

Lifter position (Elev): typically supplied by sensors belonging to the primary flight control system, but may be supplied by other systems in some particular embodiments. Corresponding to a measurement of the deflection of the aircraft's lifters with respect to a predetermined reference in the aircraft.

Center of Gravity (CG): an estimated aircraft center of gravity. In some embodiments, once the CG estimate may depend on air data availability, the values received from the Flight Management System (FMS) represent alternatives in some implementations. Another option is to retain the last trustworthy calculated value of the CG and then from this point to integrate the fuel consumption to estimate the CG stroke when some aircraft systems detect an air data failure.

Ice detection flag (Ice): typically supplied by an Air Management System (AMS), but may be supplied by other systems in some particular embodiments. Corresponding to the boolean flag, indicating the presence of ice;

tilt angle (Phi): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. Corresponding to one of the euler angles measured between the wing axis and the earth plane;

total Air Temperature (TAT): usually measured by the aerial data system of the aircraft, which means that special measures must be taken to mitigate the loss of TAT, but may in some particular embodiments be supplied by other systems as well. Corresponding to the air temperature measured at the point of stagnation of the air flow (possibly within the pitot tube);

pitch rate (Q): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. An angular rate of change corresponding to an axis formed by the aircraft about the wing;

roll rate (P): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. An angular rate of change corresponding to the aircraft about an axis formed by the fuselage;

yaw rate (R): typically measured by the aircraft's Inertial Reference Unit (IRU), but may be supplied by other systems in some particular embodiments. An angular rate of change corresponding to the aircraft about an axis perpendicular to both the wing and fuselage axes;

the input list described herein includes a subset selected for training the neural network (11) (12). These subsets may vary according to the desired neural network output (36) and also based on how many neural networks (23) (34) are to be used for estimation.

Example non-limiting sensor Signal preprocessing

There are some cases where the signal used for estimation is pre-processed (28) (33) before being applied to the neural network (23) (34). For example, in some applications, the estimated weight and the estimated center of gravity may be a function of the airborne data. Then, an option is desirable in view of the fact that the neural network may need to function in case of complete loss of data in the air, and in view of the fact that the weight is one of its mandatory inputs (31). One possible solution is to use the weight or center of gravity as communicated by the Flight Management System (FMS). If these values already contain information about the fuel consumption, i.e. if the weight values decrease over time by an amount that is indicative of the fuel consumption, and if the center of gravity moves accordingly, they may be a good alternative. However, in the case where the fuel consumption cannot be properly notified by these signals, another possibility is to retain the last reliable calculated value and then to integrate the fuel consumption from this point on, or alternatively to use the amount of fuel in the fuel tank. In both cases, since the values from the Flight Management System (FMS) may assume significant holes in the estimation (especially weight), some form of verification of these signals may be necessary, as they typically depend on manual input given by the pilot while the aircraft is on the ground. Verification will prevent erroneous neural network estimates due to human error.

Another pre-processing (28) (33) related to weight and center of gravity estimation is related to some interlocks that may freeze the estimation if the aircraft has assumed an abnormal attitude. They may be formed by thresholds defining flight envelopes beyond which the estimated values will return large errors. For example, interlocking may be used with Ny, Nz, Phi, Theta, Alpha, and even with a flag indicating the presence of ice.

While the invention has been described in connection with what is presently considered to be the most practical and preferred embodiment, it is to be understood that the invention is not to be limited to the disclosed embodiment, but on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.