Enhanced security for landing gear control algorithms

Abstract

A trained machine learning algorithm emulates a control architecture for an embedded control system (e.g., a landing gear control system for an aircraft), which may provide additional security by obfuscating details of the control architecture. An original control architecture for the embedded control system is obtained, wherein the original control architecture receives input signals and outputs one or more output signals that control one or more functions of the embedded control system. A trained machine learning algorithm maps a set of input values of the input signals to a set of output values for the output signal(s). The trained machine learning algorithm emulates the input-output relationship of the original control architecture of the embedded control system. The trained machine learning algorithm comprises one or more matrices of weights by which a vector of inputs corresponding to the set of input values is multiplied to generate the output signal(s).

Description

[0001] ENHANCED SECURITY FOR LANDING GEAR CONTROL ALGORITHMS

[0002] FIELD OF DISCLOSURE

[0003] The present disclosure relates to security of computer code for control systems, such as landing gear control systems for aircraft.

[0004] BACKGROUND

[0005] Suppliers of aircraft subsystems devote considerable time and money to developing computer code for the controllers of those subsystems and to protecting related confidential information and trade secrets. For stand-alone or federated subsystems, it may be possible to preserve some trade secrets by not sharing underlying computer code. However, in the aviation industry, there is an increasing drive towards integration of aircraft subsystems, which makes it more difficult to preserve trade secrets. As an example, integrated modular avionics systems (IMA) in modem aircraft can help to reduce the weight (and therefore increase fuel efficiency) of the aircraft, as the code to run multiple subsystems (e.g., flight control, landing gear control, aileron control, etc.) can be run on a single controller rather than multiple controllers. In many IMA architectures, suppliers provide the code required to run their systems to an airframe rather than a federated control box. By providing code to an airframer, this poses a risk that is not present when providing a federated control box, namely, the risk of exposure of trade secrets. More generally, even if such code is not required to be shared, trade secrets may still be exposed if such code is inadvertently released or deliberately stolen, especially if it is in human readable form.

[0006] SUMMARY

[0007] The present disclosure provides examples of enhanced security tools and techniques for protection of control algorithms, such as landing gear control algorithms. In particular, disclosed examples involve obfuscation of deterministic control algorithms using neural networks. The obfuscation process results in control logic that is nearly impossible to reverse-engineer while still emulating to a very high degree of accuracy the same behavior and reliability as the original algorithm, thereby allowing the control logic to be more freely shared and even to be used in place of the original control algorithm in the system itself.

[0008] In accordance with an aspect of the present disclosure, a method of generating a trained machine learning algorithm to emulate a control architecture for an embedded control system comprises obtaining an original control architecture for the embedded control system, wherein the original control architecture receives input signals and outputs one or more output signals that control one or more functions of the embedded control system; and training a machine learning algorithm that maps a set of input values of the input signals to a set of output values for the one or more output signals, wherein the trained machine learning algorithm emulates the input-output relationship of the original control architecture of the embedded control system, and wherein the trained machine learning algorithm comprises one or more activation functions and one or more matrices of weights by which a vector of inputs corresponding to the set of input values is multiplied to generate the one or more output signals.

[0009] In any of the embodiments of the present disclosure, the embedded control system may include a vehicle control system. Alternatively, the embedded control system may include a control system for some other equipment or device, such as a factory control system.

[0010] In any of the embodiments of the present disclosure, the embedded control system may include an aircraft control system. Alternatively, the embedded control system may include a control system for some other vehicle, such as a train control system, an automobile control system, etc.

[0011] In any of the embodiments of the present disclosure, the embedded control system may include a landing gear control system. Alternatively, the embedded control system may include a control system for a different component or subsystem of an aircraft.

[0012] In any of the embodiments of the present disclosure, the trained machine learning algorithm may include a neural network, such as a multilayer perceptron (MLP) comprising fully connected neurons with a nonlinear activation function, organized in at least three layers, including an input layer corresponding to the set of input values, an output layer corresponding to the set of output values, and one or more hidden layers.

[0013] In any of the embodiments of the present disclosure, the trained machine learning algorithm may be installed in the embedded control system as a replacement for the original control architecture.

[0014] In any of the embodiments of the present disclosure, the original control architecture may include a state machine and / or a feedback controller.

[0015] In any of the embodiments of the present disclosure, the input signals may include one or more user control signals (e.g., one or more of a brake application signal, a landing gear extension / retraction signal, or a steering control signal) and one or more sensor acquisition signals.

[0016] In any of the embodiments of the present disclosure, the sensor acquisition signal(s) may include one or more of a wheel speed signal, a tire sensor signal (e.g., tire temperature signal, tire wear signal, tire grip signal), a steering sensor signal (e.g., a rotary position sensor signal), or a brake sensor signal (e.g., brake temperature signal, brake wear signal).

[0017] In any of the embodiments of the present disclosure, the one or more output signals may include an actuator signal (e.g., brake actuator signal, steering actuator signal, alarm / notification actuator signal to indicate dangerous conditions, required maintenance conditions, informational conditions, etc., with a visual or audible alarm or notification).

[0018] In accordance with another aspect of the present disclosure, a computer-readable storage medium (e.g., anon-transitory storage medium) has stored thereon computing logic that, when executed by a processor, causes the processor to perform one or more operations of an embedded control system, wherein the computing logic comprises a trained machine learning algorithm generated according to any of the disclosed methods.

[0019] In accordance with another aspect of the present disclosure, a control system for an aircraft comprises at least one computer-readable storage medium having stored therein control logic that, when executed, causes the control system to perform operations to control one or more functions of the aircraft, wherein the control logic comprises a trained machine learning algorithm generated according to any of the disclosed methods.

[0020] In accordance with another aspect of the present disclosure, an aircraft comprises a control system including at least one computer-readable storage medium having stored therein control logic that, when executed, causes the control system to perform operations to control one or more functions of the aircraft, wherein the control logic comprises a trained machine learning algorithm generated according to any of the disclosed methods.

[0021] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This summary is not intended to identify key features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0022] DESCRIPTION OF THE DRAWINGS

[0023] The foregoing aspects and many of the attendant advantages of the claimed subject matter will become more readily appreciated as the same become better understood by reference to the following detailed description, when taken in conjunction with the accompanying drawings, wherein:

[0024] FIGURE 1A depicts one example of an aircraft, such as a passenger or cargo aircraft, shown in bottom view, in which technologies and / or methodologies of the present disclosure may be employed;

[0025] FIGURE IB is a functional block diagram of a landing gear system in which technologies and / or methodologies of the present disclosure may be employed;

[0026] FIGURE 2 is a block diagram of an embedded control system and method for generating a trained machine learning algorithm to emulate a control architecture for the embedded control system, in accordance with aspects of the present disclosure; and FIGURE 3 is a graph illustrating performance of a trained machine learning algorithm that gives equivalent performance to a control architecture that it is modeling, in accordance with aspects of the present disclosure.

[0027] DETAILED DESCRIPTION

[0028] The detailed description set forth above in connection with the appended drawings, where like numerals reference like elements, are intended as a description of various embodiments of the present disclosure and are not intended to represent the only embodiments. Each embodiment described in this disclosure is provided merely as an example or illustration and should not be construed as preferred or advantageous over other embodiments. The illustrative examples provided herein are not intended to be exhaustive or to limit the disclosure to the precise forms disclosed.

[0029] As will be described in more detail below, the present disclosure provides examples of enhanced security tools and techniques for protection of control algorithms, such as for landing gear control algorithms. In particular, disclosed examples involve obfuscation of deterministic control algorithms using neural networks. In an illustrative scenario, an original algorithm is developed using human-readable code, which is effective for the purpose of the algorithm but is vulnerable to unauthorized disclosure, copying, or reverse engineering. In described embodiments, a neural -network based obfuscation process is applied to the original algorithm, resulting in control logic that is nearly impossible to reverse-engineer while still exhibiting to a very high degree the same behavior and reliability as the original algorithm. The control logic can then be more freely shared and, because it operates in a highly similar manner, can even be used in place of the original control algorithm in the system itself.

[0030] Although embodiments of the present disclosure may be described with reference to landing gear control systems for aircraft, one skilled in the relevant art will appreciate that the disclosed embodiments are illustrative in nature and therefore should not be construed as limited to such an application. It should therefore be apparent that the disclosed technologies and methodologies have wide application, and therefore may be suitable for use with many types of vehicles, including automobiles, buses, trains, heavy equipment, and the like, or control systems for other types of equipment. Accordingly, the following descriptions and illustrations herein should not limit the scope of the claimed subject matter.

[0031] FIGURE 1A depicts one example of an aircraft 100, such as a passenger or cargo aircraft, shown in bottom view, in which technologies and / or methodologies of the present disclosure may be employed. The aircraft 100 can include a landing gear system 101 comprising a nose landing gear system 110, a left main landing gear system 120a, and a right main landing gear system 120b. As used herein, left and right directions are in reference to the perspective of the pilot seated in the cockpit facing the standard forward direction of travel. The landing gear systems 110, 120a, and 120b can include various components configured to support and control movement of the aircraft 100 on the ground surface, e.g., wheels, tires, brakes, shock absorbers, hydraulics, sensors, controllers, etc.; these components are not all shown for the sake of clarity. Functions that may be controlled by landing gear systems include braking functions, extension / retraction functions, and steering functions.

[0032] The landing gear system 101 is shown as an example of a control system in which described embodiments may be implemented. It should be appreciated that the aircraft 100 and the landing gear system 101 illustrated in FIGURE 1A should not be considered limiting on the present disclosure, and the landing gear systems can be arranged in various other configurations with fewer or additional components as desired. In addition, the particular location of the landing gear systems, the quantity of wheels and tires, and the other aspects of the aircraft 100 illustrated in FIGURE 1A should not be considered limiting on the present disclosure, as the components may be positioned at various locations.

[0033] FIGURE IB is a functional block diagram of the landing gear system 101 in accordance with aspects of the present disclosure, showing a configuration of the system 101 including the nose landing gear system 110 and the left and right main landing gear systems 120a and 120b. The system 101 includes a master brake controller 102 that is configured to manage braking actuation timing and force based on various input device signals from the aircraft 100. Examples of input device signals sent to and received by the master brake controller 102 are a mechanical brake input device 130 (e.g., hydraulic, pneumatic, mechanical, etc. — from a pilot, personnel manipulating an emergency brake lever, or other manually actuated system), an electronic brake input device 132 (e.g., a signal from an aircraft control system, a potentiometer, a position sensor, an autopilot system, or the like), and other input signals sent to the master brake controller 102. Each of these signals can be interpreted by the master brake controller 102 to actuate the brakes (e.g., an eddy current brake, a friction brake, or a combination thereof) in accordance with a control scheme of the aircraft 100. The master brake controller 102 can additionally be coupled to various other systems and sensors to provide feedback based on, e.g., wheel speed, brake force requirements, tire grip, and other similar aspects.

[0034] The nose landing gear system 110 typically includes two wheels and tires and may, in some circumstances, include brakes. The master brake controller 102 can be coupled to a left brake controller associated with the left main landing gear system 120a and configured to control one or more brakes (e.g., friction brakes) arranged in the left main landing gear system 120a; and a right brake controller associated with the right main landing gear system 120b and configured to control one or more brakes arranged in the right main landing gear system 120b. The main landing gear systems 120a and 120b can include any number of wheels and tires that can individually include a brake. In other embodiments, any number of brakes can be included in the main landing gear systems 120a and 120b.

[0035] While brake controllers are disclosed as examples of control systems in which described embodiments may be implemented, it should be understood that described embodiments may be implemented in a wide variety of control systems. FIGURE 2 is a block diagram of a control system 200, e.g., a landing gear control system such as a master brake controller 102, or some other control system, according to aspects of the present disclosure. The control system 200 includes control architecture 205. In this example, the control architecture is characterized by the combination of a state machine 210 and a feedback controller 220, such as a proportional-integral-derivative (PID) controller. Alternatively, other control architectures may be accommodated, such as control architectures with multiple state machines, multiple controllers, or different types of controllers.

[0036] The control architecture 205 receives input signals and generates one or more output signals. In the example shown in FIGURE 2, the input signals include one or more user control signals 230 and one or more sensor acquisition signals 240, and the output signal(s) include actuator commands 250. The relationship between the input signals and the output signal(s) (or input-output relationship) is a deterministic relationship that is dependent on functional characteristics and state of the control architecture (e.g., state machine 210 and the feedback controller 220).

[0037] In an illustrative control system 200 (e.g., a landing gear control system), the user control signals 230 may include, e.g., landing gear extension / retraction signals, brake application control signals, and steering control signals, and the sensor acquisition signals 240 may include, e.g., wheel speed signals, tire pressure signals, brake condition signals. In the example shown in FIGURE 2, these input signals are provided as input to feedback controller 220 along with the system state from state machine 210. In some embodiments (e.g., for PID controllers), the inputs may further include the derivative and integral of the state. The output is the actuator commands 250 (e.g., applying or releasing brakes in a landing gear control system) that the control system 200 applies in response to the set of inputs.

[0038] In a training phase, a training module may be used to generate matrices of weights that are used to map inputs to outputs corresponding to an original control algorithm (e.g., control architecture 205). The training module may be used to repeatedly train a neural network to obtain values for the matrices that provide the correct output results for corresponding inputs, such that the correct outputs are provided consistently in a repeatable manner.

[0039] In the example shown in FIGURE 2, a machine learning algorithm is trained in training module 260 based on the input-output relationship of the control architecture 205. This relationship (e.g., sets of particular input values corresponding to particular output values) is provided as input to the training module 260, which trains a machine learning algorithm to emulate this relationship and thereby emulate the function of the control architecture 205. This training results in a trained machine learning algorithm 270. In some embodiments, the trained machine learning algorithm 270 contains at least one matrix of weights and one or more activation functions that are very difficult to reverse-engineer into a distinct controller architecture, but deterministically replicates the behavior of the algorithm in control architecture 205. The machine learning algorithm may include other functional blocks as well, such as one or more batch normalization blocks that may be used to improve training of the algorithm.

[0040] In some embodiments, such as in aircraft that may be subject to regulation by an aviation authority that forbids “on-the-fly” training, all training is performed prior to the aircraft’s entry into service. In such embodiments, no training occurs during operation of the aircraft; once the trained machine learning algorithm is uploaded to the control box of the aircraft, the trained machine learning algorithm (e.g., the matrix of weights) does not change, and no parameters are tuned dynamically.

[0041] As shown in FIGURE 2, the same inputs that were provided to control architecture 205 (e.g., user control signals 230, sensor acquisition signals 240) and are now provided to the trained machine learning algorithm 270 produce the same outputs or substantially the same outputs (e.g., actuator commands 250). The trained machine learning algorithm 270 may then be provided to another party as a “black box” solution accurately representing the behavior of the control architecture 205. In an illustrative scenario, the trained machine learning algorithm 270 is provided to an airframer to be incorporated into an integrated modular avionics system (IMA) for accurate and reliable representation of the behavior of the control architecture 205 without revealing the underlying algorithmic details of the control architecture 205.

[0042] In some embodiments, the machine learning algorithm used to model the control architecture 205 is a deterministic multilayer perceptron (MLP) comprising an input layer, one or more hidden layers, and an output layer. Although the MLP is very difficult to reverse engineer, its deterministic nature means that like the control architecture 205, it will give the same response (output) to the same set of inputs every time, which allows it to accurately model the behavior of the control architecture 205.

[0043] In some embodiments, the computation for one layer of an MLP (which can be repeated for additional layers) is as follows:

[0044] Oi = o(( o Wo + bo) where Oo is the inputs, O i is the output, IL i s a weight matrix, b is a bias vector (which may be 0 if no bias vector is desired), and o is an activation function, such as a sigmoid or ReLU (rectified linear unit) function.

[0045] In use, a vector of inputs corresponding to a set of input values for the control system (e.g., values for user control signals 230 such as extension / retraction signals, brake application control signals, steering control signals; values for sensor acquisition signals 240 such as wheel speed signals, tire pressure signals, and brake condition signals) can be multiplied by one or more matrices of weights to generate correct output results (e.g., actuator commands 250 for applying or releasing brakes) for the corresponding inputs.

[0046] In this way, the trained machine learning algorithm 270 may be submitted for testing or provided to an airframer to be incorporated into an IMA system for accurate and reliable representation of the behavior of the control architecture 205 without revealing the underlying algorithmic details of the control architecture 205. Once testing has confirmed its reliability within a desired tolerance or error rate, the trained machine learning algorithm 270 may even be substituted for the original control architecture 205 and installed as control logic in control system 200 (e.g., a landing gear control system for an aircraft) that operates as a substitute or replacement for the original control architecture 205.

[0047] FIGURE 3 is a graph indicating that a trained machine learning algorithm (in this case, the trained neural network / MLP) gives equivalent performance as a PID controller that it is modeling in terms of plant response (system response). Mathematically, the MLP is response in computed as:

[0048] C =0o(z-7) where C is the control output from the MLP, I is the vector of inputs, cr is the activation function (e.g., sigmoid function) applied element- wise, and Woand W are matrices of weights that are determined by the training process. In some embodiments, a similar process may be applied to other scenarios besides PID controllers, such as state logic or anti -windup controls.

[0049] Using this method, the controller and the details of the control architecture are modeled in the weight matrices. In the example shown in FIGURE 3, a simple mechanism (e.g., applying a force to a mass to move the mass along a surface from an initial point 0, where the mass is connected to the initial point by a spring / damper that restricts the mass’ movement) is emulated by the trained machine learning algorithm. This simple mechanism is depicted in FIGURE 3 for ease of graphical illustration. However, a large number of different control architectures can be represented in this manner, provided the size of the matrix or matrices (e.g., Woand Wj) is / are sufficiently large, making the exact control architecture challenging to reverse engineer for any party with access to the control logic.

[0050] Alternatively, other machine learning approaches may be used, e.g., transformers, recurrent neural networks, or other deterministic neural network approaches using weight matrices to model system behavior. Requirements can be set and tested on allowable deviations between the trained machine learning algorithm and the original architecture before submitting the trained machine learning algorithm for testing or installing it in an aircraft system controller or other equipment. In such embodiments, training is repeated until the trained machine learning algorithm replicates the original control architecture within the required tolerance.

[0051] While landing gear control systems are disclosed as examples of control systems in which described embodiments may be implemented, it should be understood that described embodiments may be implemented in a wide variety of control systems, which may include controllers, control circuits, control units, etc.

[0052] In some embodiments, controllers, control circuits, control units, etc., comprise hardware circuity (e.g., analog and / or digital circuits, etc.) for carrying out functionality described herein. In other embodiments, the control circuit or the various control units include, for example, a processor and memory. The memory may include computer readable storage media in read-only memory (ROM), random-access memory (RAM), and keep-alive memory (KAM), for example. The KAM may be used to store various operating variables or program instructions while the processor is powered down. The program instructions, when executed by the processor, is configured to carry out the functionality of the control circuit described herein. In that regard, the processor and / or memory storing the program instructions forms a special purpose controller or control circuit specifically configured to carry out the methodologies and technologies described herein. The computer-readable storage media may be implemented using any of a number of known memory devices such as PROMs (programmable read-only memory), EPROMs (electrically PROM), EEPROMs (electrically erasable PROM), flash memory, or any other electric, magnetic, optical, or combination memory devices capable of storing data, instructions, programs, modules, etc.

[0053] As used herein, the term "processor" is not limited to integrated circuits referred to in the art as a computer, but broadly refers to a microcontroller, a microcomputer, a microprocessor, a programmable logic controller, an application specific integrated circuit, other programmable circuits, combinations of the above, among others. Therefore, as used herein, the term "processor" can be used to generally describe these aforementioned components, and can be either hardware or software, or combinations thereof, that implement logic for carrying out various aspects of the present disclosure. Similarly, the terms "module" and "unit" can include logic that may be implemented in either hardware or software, or combinations thereof.

[0054] A collection of exemplary embodiments, including at least some explicitly enumerated as “ECs” (Example Combinations), providing additional description of a variety of embodiment types in accordance with the concepts described herein are provided below. These examples are not meant to be mutually exclusive, exhaustive, or restrictive; and the claimed subject matter is not limited to these example embodiments but rather encompasses all possible modifications and variations within the scope of the issued claims and their equivalents.

[0055] EC A. A method of generating a trained machine learning algorithm to emulate a control architecture for an embedded control system, the method comprising: obtaining an original control architecture for the embedded control system, wherein the original control architecture receives input signals and outputs one or more output signals that control one or more functions of the embedded control system; training a machine learning algorithm that maps a set of input values of the input signals to a set of output values for the one or more output signals, wherein the trained machine learning algorithm emulates the input-output relationship of the original control architecture of the embedded control system, and wherein the trained machine learning algorithm comprises one or more matrices of weights by which a vector of inputs corresponding to the set of input values is multiplied to generate the one or more output signals. EC B. The method of EC A, wherein the embedded control system comprises a vehicle control system.

[0056] EC C. The method of EC B, wherein the vehicle control system comprises an aircraft control system.

[0057] EC D. The method of EC C, wherein the aircraft control system comprises a landing gear control system.

[0058] EC E. The method of any of EC A to EC D, wherein the trained machine learning algorithm comprises a neural network.

[0059] EC F. The method of EC E wherein the neural network comprises a multilayer perceptron (MLP) including an input layer corresponding to the set of input values, an output layer corresponding to the set of output values, and one or more hidden layers.

[0060] EC G. The method of any of EC A to EC F, further comprising installing the trained machine learning algorithm in the embedded control system as a replacement for the original control architecture.

[0061] EC H. The method of any of EC A to EC G, wherein the original control architecture comprises a state machine and a feedback controller.

[0062] EC I. The method of any of EC A to EC H, wherein the input signals comprise one or more user control signals and one or more sensor acquisition signals.

[0063] EC J. The method of EC I, wherein the one or more user control signals comprise one or more of a brake application signal, a landing gear extension / retraction signal, or a steering control signal.

[0064] EC K. The method of EC I or EC J, wherein the one or more sensor acquisition signals comprise one or more of a wheel speed signal, a tire sensor signal (e.g., a tire temperature sensor signal, a tire wear sensor signal, a tire pressure sensor signal, a tire grip sensor signal), a steering sensor signal (e.g., a rotary position sensor signal), or a brake sensor signal (e.g., a brake temperature sensor signal, a brake wear sensor signal).

[0065] EC L. The method of any of EC A to EC K, wherein the one or more output signals include an actuator signal.

[0066] EC M. A computer-readable storage medium having stored thereon computing logic that, when executed by a processor, causes the processor to perform one or more operations of an embedded control system, wherein the computing logic comprises a trained machine learning algorithm generated according to the method of any of EC A to EC L.

[0067] EC N. A control system for an aircraft, comprising at least one computer-readable storage medium having stored therein control logic that, when executed, causes the control system to perform operations to control one or more functions of the aircraft, wherein the control logic comprises a trained machine learning algorithm generated according to the method of any of EC A to EC L.

[0068] EC O. An aircraft comprising the control system of EC N.

[0069] In the foregoing description, specific details are set forth to provide a thorough understanding of exemplary embodiments of the present disclosure. It will be apparent to one skilled in the art, however, that the embodiments disclosed herein may be practiced without embodying all of the specific details. In some instances, well-known process steps have not been described in detail in order not to unnecessarily obscure various aspects of the present disclosure. Further, it will be appreciated that embodiments of the present disclosure may employ any combination of features described herein. The present application may reference quantities and numbers. Unless specifically stated, such quantities and numbers are not to be considered restrictive, but exemplary of the possible quantities or numbers associated with the present application. Also in this regard, the present application may use the term “plurality” to reference a quantity or number. In this regard, the term “plurality” is meant to be any number that is more than one, for example, two, three, four, five, etc. The terms “about,” “approximately,” “near,” etc. , mean plus or minus 10% of the stated value. F or the purposes of the present disclosure, the phrase “at least one of A and B” is equivalent to “A and / or B” or vice versa, namely “A” alone, “B” alone or “A and B.” Similarly, the phrase “at least one of A, B, and C,” for example, means (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C), including all further possible permutations when greater than three elements are listed.

[0070] It should be noted that for purposes of this disclosure, terminology such as “right,” “left,” “front,” “rear,” etc., should be construed as descriptive and not limiting the scope of the claimed subject matter. Further, the use of “including,” “comprising,” or “having” and variations thereof herein is meant to encompass the items listed thereafter and equivalents thereof as well as additional items. Unless limited otherwise, the terms “connected,” “coupled,” and “mounted” and variations thereof herein are used broadly and encompass direct and indirect connections, couplings, and mountings.

[0071] Throughout this specification, terms of art may be used. These terms are to take on their ordinary meaning in the art from which they come, unless specifically defined herein or the context of their use would clearly suggest otherwise.

[0072] The principles, representative embodiments, and modes of operation of the present disclosure have been described in the foregoing description. However, aspects of the present disclosure, which are intended to be protected, are not to be construed as limited to the particular embodiments disclosed. Further, the embodiments described herein are to be regarded as illustrative rather than restrictive. It will be appreciated that variations and changes may be made by others, and equivalents employed, without departing from the present disclosure.

Claims

CLAIMSThe embodiments of the invention in which an exclusive property or privilege is claimed are defined as follows:

1. A method of generating a trained machine learning algorithm to emulate a control architecture for an embedded control system, the method comprising: obtaining an original control architecture for the embedded control system, wherein the original control architecture receives input signals and outputs one or more output signals that control one or more functions of the embedded control system; training a machine learning algorithm that maps a set of input values of the input signals to a set of output values for the one or more output signals, wherein the trained machine learning algorithm emulates the input-output relationship of the original control architecture of the embedded control system, and wherein the trained machine learning algorithm comprises an artificial neural network that includes an activation function and one or more matrices of weights by which a vector of inputs corresponding to the set of input values is multiplied to generate the one or more output signals.

2. The method of Claim 1, wherein the embedded control system comprises a vehicle control system.

3. The method of Claim 2, wherein the vehicle control system comprises an aircraft control system.

4. The method of Claim 3, wherein the aircraft control system comprises a landing gear control system.

5. The method of Claim 1, wherein the artificial neural network comprises a multilayer perceptron (MLP) including the activation function and connected neurons organized in a plurality of layers including an input layer corresponding to the set of inputvalues, an output layer corresponding to the set of output values, and one or more hidden layers.

6. The method of any of Claims 1 to 5, wherein the artificial neural network further includes a batch normalization layer.

7. The method of any of Claims 1 to 6 further comprising installing the trained machine learning algorithm in the embedded control system as a replacement for the original control architecture.

8. The method of any of Claims 1 to 7, wherein the original control architecture comprises a state machine and / or a feedback controller.

9. The method of any of Claims 1 to 8, wherein the input signals comprise one or more user control signals and one or more sensor acquisition signals.

10. The method of Claim 9, wherein the one or more user control signals comprise one or more of a brake application signal, a landing gear extension / retraction signal, or a steering control signal.

11. The method of Claim 9 or Claim 10, wherein the one or more sensor acquisition signals comprise one or more of a wheel speed signal, a tire sensor signal, a steering sensor signal, or a brake sensor signal.

12. The method of any of Claims 1 to 11 , wherein the one or more output signals include an actuator signal.

13. A computer-readable storage medium having stored thereon computing logic that, when executed by a processor, causes the processor to perform one or more operations of an embedded control system, wherein the computing logic comprises atrained machine learning algorithm generated according to the method of any of Claims 1 to 12.

14. A control system for an aircraft, comprising at least one computer-readable storage medium having stored therein control logic that, when executed, causes the control system to perform operations to control one or more functions of the aircraft, wherein the control logic comprises a trained machine learning algorithm generated according to the method of any of Claims 1 to 12.

15. An aircraft comprising the control system of Claim 14.

Images