Method of training a neural network to control an aircraft system
A method of training a neural network to control an aircraft system. The method includes obtaining an operational mode data space representing a set of operational modes for the aircraft system, wherein each operational mode represents a configuration of the aircraft system where performance of an aircraft component is impaired according to an aircraft system model. Each operational mode includes probability data indicating a respective probability of the operational mode occurring, according to the model. The method includes generating a reduced operational mode data space including operational modes having a probability greater than a theoretical operational probability threshold, and generating a training operational mode data space including operational modes within the reduced operational mode data space that have a probability less than a real-world operational probability threshold. The method includes training the neural network using operational modes within the training operational mode data space.
1 . A method of training a neural network to control an aircraft braking system having a plurality of aircraft braking system components, the method comprising:
obtaining an operational mode data space representing a set of operational modes for the aircraft braking system, wherein each operational mode represents a configuration of the aircraft braking system where performance of one or more of the plurality of aircraft braking system components is impaired and comprises probability data indicating a respective positive and non-zero probability of the operational mode occurring, according to a model of the aircraft braking system, wherein the plurality of aircraft braking system components comprise one or more of a power supply selector valve and an actuator;
generating a reduced operational mode data space limited to the operational modes having a probability greater than a theoretical operational probability threshold and excluding the operational modes having a probability less than the theoretical operational probability;
generating a training operational mode data space limited to the operational modes within the reduced operational mode data space that have a probability less than a real world operational probability threshold and excluding the operational modes within the reduced operational mode data space having a probability greater than the real-world operational probability threshold; and
training the neural network using operational modes within the training operational mode data space, to control reconfiguration of the braking system by controlling one or more of the power supply selector valve and the actuator.
2 . The method according to claim 1 , wherein the method comprises stratifying the training operational mode data space to generate a plurality of sub-spaces based on the obtained probability data, selecting a first set of operational modes from the plurality of sub-spaces, selecting a second set of operational modes from the plurality of sub-spaces, the second set of operational modes different from the first set of operational modes, training the neural network using the first set of operational modes, and testing the neural network using the second set of operational modes.
3 . The method according to claim 2 , wherein each operational mode of the first set of operational modes is different from each operational mode of the second set of operational modes.
4 . The method according to claim 2 , wherein selecting the first and second sets of operational modes comprises randomly selecting the first and second sets of operational modes from the plurality of sub-spaces.
5 . The method according to claim 2 , wherein selecting the first and second sets of operational modes comprises selecting operational modes from the plurality of sub-spaces such that each of the first and second sets of operational modes has the same distribution of probabilities as the training operational mode data space.
6 . The method according to claim 2 , wherein the method further comprises stratifying the plurality of sub-spaces based on a structure of the aircraft braking system to generate a plurality of further sub-spaces, selecting the first set of operational modes from the plurality of further sub-spaces and selecting the second set of operational modes from the plurality of further sub-spaces.
7 . The method according to claim 1 , wherein the method further comprises generating an operational mode data space using operational modes present in the reduced operational mode data space but absent from the training operational mode data space, and verifying operation of the neural network using operational modes within the operational mode data space.
8 . The method according to claim 1 , wherein the theoretical operational probability threshold is set such that operational modes present in the operational mode data space but absent from the reduced operational mode data space comprise a cumulative probability less than an acceptable safety level for the aircraft braking system.
9 . The method according to claim 1 , wherein a size of the operational mode data space is determined based on a maximum acceptable probability of a determined critical operational mode according to the model of the aircraft braking system, and an operational mode having a highest probability according to the model of the aircraft braking system.
10 . The method according to claim 1 , wherein the neural network comprises a plurality of nodes, and the method comprises monitoring outputs from the plurality of nodes during training of the neural network to determine activity data representing activity of the plurality of nodes, identifying a node or combination of nodes having activity levels below an activity level threshold using the activity data, and modifying the operational modes used to train the neural network based on the identified node or combination of nodes.
11 . The method according to claim 10 , wherein modifying the operational modes used to train the neural network based on the identified node or combination of nodes comprises utilising a greater number of operational modes from within the training operational mode data space.
12 . The method according to claim 10 , wherein modifying the operational modes used to train the neural network based on the identified node or combination of nodes comprises adding operational modes to the training operational mode data space from the reduced operational mode data space.
13 . The method according to claim 10 , wherein modifying the operational modes used to train the neural network based on the identified node or combination of nodes comprises adding operational modes to the training operational mode data space from the operational mode data space.
14 . The method according to claim 1 , wherein the model of the aircraft braking system comprises a directed graph comprising one or more vertices, and each vertex represents operation of at least one of the aircraft braking system components of the aircraft braking system.
15 . The method according to claim 1 , wherein the aircraft braking system comprises a braking system, and each operational mode represents a respective set of possible braking component impairments according to a model of the aircraft braking system.
16 . The method according to claim 1 , wherein operational modes present in the operational mode data space but absent from the reduced operational mode data space define a discarded operational mode data space, and the discarded operational mode data space comprises a cumulative probability less than an acceptable safety level for the aircraft braking system.
17 . A method of training a neural network to control an aircraft braking system having a plurality of aircraft braking system components, the method comprising:
obtaining an aircraft braking system operational mode data space representing a set of operational modes for the aircraft braking system, wherein each operational mode represents a configuration of the aircraft braking system where performance of one or more of the plurality of aircraft braking system components is impaired, according to a model of the aircraft braking system, wherein the plurality of aircraft braking system components comprise one or more of a power supply selector valve and an actuator;
obtaining probability data indicative of probabilities of the operational modes occurring, wherein each of the probabilities is positive and greater than a zero probability;
comparing obtained probability data for each operational mode to a theoretical operational probability threshold, wherein each of the probabilities is greater than a zero probability;
removing operational modes having a probability less than or equal to the theoretical operational probability threshold from the aircraft braking system operational mode data space to leave behind a reduced aircraft braking system operational mode data space, wherein the reduced aircraft braking system operational mode data space is limited to the operational modes from the operational mode data space having a probability greater than the theoretical operational probability threshold and excludes the operational modes in the operational mode data space having a probability equal or less than the theoretical operational probability threshold;
comparing obtained probability data for operational modes within the reduced aircraft braking system operational mode data space to a real-world operational probability threshold;
removing operational modes having a probability greater than or equal to the real-world operational probability threshold from the reduced aircraft braking system operational mode data space to leave behind a training aircraft braking system operational mode data space, wherein the operational modes remaining in the training aircraft braking system operational mode data space is limited to the operational modes in the reduced aircraft braking system operational mode data space having a probability less than the real-world operational probability threshold; and
training the neural network using operational modes within the training aircraft braking system operational mode data space, to control reconfiguration of the braking system by controlling one or more of the power supply selector valve and the actuator.
18 . A data carrier comprising machine-readable instructions for the operation of a processor of a computer system to perform the method according to claim 1 .
19 . A computer system for training a neural network to control the aircraft braking system, the computer system comprising a processor and a data carrier according to claim 18 .
20 . An aircraft braking system comprising a neural network trained according to the method of claim 1 .
21 . The method of claim 1 , wherein:
the generating of the reduced operational mode includes selecting for the reduced operational mode data space the operational modes from the operational mode data space having a probability greater than the theoretical operational probability threshold and excluding for use in the reduced operational mode data space the operational modes in the operational mode data space having a probability equal or less than the theoretical operational probability threshold; and
the generating of the training operational mode data space selects for the training operational mode data space the operational modes in the reduced operational mode data space having a probability less than the real-world operational probability threshold and excludes for use in the training operational mode data space operational modes the operational modes in the reduced operational mode data space having a probability equal to or greater than the real-world operational probability threshold.