Taking Off with Machine Learning for Flight Modeling

• Published in: IISE Annual Conference. Proceedings pp. 1 - 6
• Main Authors: Walker, Joel R, Claudio, David
• Format: Journal Article
• Language: English
• Published: Norcross Institute of Industrial and Systems Engineers (IISE) 01.01.2023
• Subjects: 20th century; Aerodynamics; Air transportation; Air travel; Aircraft; Aircraft reliability; Airline industry; Artificial intelligence; Aviation; Certification; Commercial aircraft; Design; Flight envelopes; Flight simulators; Flight tests; Identification; Machine learning; Methods; Modelling; Motion simulators; Nonlinear dynamics; Simulator fidelity; System identification
• DOI: 10.21872/2023IISE_3170

Flight test for aircraft certification is a fundamental method to ensure safe aircraft and air travel worldwide. Flight test data is collected through a limited number of discrete subspace points in the flight envelope to verify and validate preselected linear model coefficients chosen prior to the test program. The finalized model is presented as evidence that the platform meets the certification requirements throughout the expected flight envelope. It is also utilized to develop high-fidelity full-motion simulators, the primary training method for business aviation and airline industries. System identification utilizing local linear approximations is still the dominant flight test approach exacerbating the development of a global model when trying to capture typical non-linear flight dynamics. This makes it appealing to leverage the robust and highly evolving machine learning techniques successfully used to find physical and physics-based insights. This article summarizes the historic developments of flight modeling and proposes coupling machine learning methodologies with airworthiness model validation to reduce the uncertainties of current approaches while providing dynamic and real-time model development with little programmatic overhead. Machine learning allows to actively capture flight dynamics independently of the historical preselected model validation and reduces risks to the flight test crew. The limited number of studies using machine learning for flight dynamics modeling are discussed, showing their strong potential and specific scope limitations. The constrained overlap of these two topics is also discussed with comments and discussion toward potential areas of exploration.