Aircraft Neural Network Corruption Detection With Test-Input Validation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Aircraft control systems employing machine learning models, such as neural networks, are susceptible to corruption due to electromagnetic radiation, leading to potential erroneous aircraft control outputs.

Innovation Solution

An aircraft control system comprising a test module that monitors the operating status of the aircraft control module, writes test inputs, reads outputs, and selectively removes authority if outputs do not conform to expected values, with a backup system to replace the aircraft control module in case of corruption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If machine learning models are used in aircraft control systems, then development time is reduced, but reliability deteriorates due to susceptibility to corruption from electromagnetic radiation

Engineering Contradiction:
Improvedevelopment timeVSAvoidreliability
Core Design Contradiction:
Loss of timeVSReliability

Solution Approach 1:

The patent implements preliminary testing of machine learning models before deployment in aircraft control systems. The test module performs corruption tests by introducing electromagnetic radiation interference during the testing phase, allowing the system to identify and correct vulnerabilities before actual operation, thus maintaining both reduced development time and improved reliability

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent establishes a feedback mechanism where the test module continuously monitors the machine learning model's performance and compares it against expected outcomes. When corruption is detected through feedback from test results, the system can alert operators or automatically switch to backup models, maintaining reliability while preserving the efficiency benefits of machine learning

Inventive Principle:
Principle #23Feedback

2Extent of automation

If machine learning models are deployed in aircraft control, then automation is increased, but safety deteriorates due to potential corruption leading to erroneous control outputs

Engineering Contradiction:
ImproveautomationVSAvoidsafety
Core Design Contradiction:
Extent of automationVSReliability

Solution Approach 1:

The patent introduces a test module as an intermediary between the machine learning model and the aircraft control system. This intermediary continuously validates the model's outputs by comparing them against expected results and can intercept erroneous control signals caused by corruption, thereby maintaining high automation while ensuring safety through an intermediate validation layer

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent prepares backup machine learning models in advance that can immediately replace corrupted primary models. This beforehand cushioning ensures that if corruption occurs during automated operation, the system can switch to a pre-prepared backup model without interrupting flight control, thus maintaining both automation and safety

Inventive Principle:
Principle #11Beforehand cushioning (Prior cushioning)

Data Source

PatentEP4043980B1System for preventing corruption of neural network insafety-critical systems on aircraft
Publication Date: 2025.06.11 AIRBUS OPERATIONS LTD
  • EP4043980B1 patent drawingFigure 1
  • EP4043980B1 patent drawingFigure 2
  • EP4043980B1 patent drawingFigure 3

AI summary

Disclosed is an aircraft control system (100) comprising an aircraft control module (110) and a test module (120). The aircraft control module (110) comprises a memory storing a machine learning model (130) adapted, based on received live aircraft operating inputs (102a-102f), to generate live aircraft control outputs (104a-104b), to perform aircraft control. The test module (120) comprises a memory storing input data representing test aircraft operating inputs (102a'-102f') and respective output data representing expected aircraft control outputs (114a, 114b) that are expected to be generated by the aircraft control module in response to an input of the respective test aircraft operating inputs (102a'-102f'). The test module (120) is adapted, while the aircraft control module (110) is not performing aircraft control, to write test aircraft operating inputs (102a'-102f') to the aircraft control module, read test aircraft control outputs (104a'-104b') of the aircraft control module (110) generated based on the respective test aircraft operating inputs (102a'-102f'), and selectively remove the authority of the aircraft control module (110) to perform aircraft control if the test aircraft control outputs (104a'-104b') do not conform to the respective expected aircraft control outputs (114a-114b).