Aircraft dynamics simulation using a novel physics-based learning method

Yang Yu, Houpu Yao, Yongming Liu

Research output: Contribution to journalArticle

Abstract

The fast and accurate prediction of flight trajectories is crucial for the real-time prognostics of the air transportation system. However, the computation costs of simulating aircraft dynamics can be expensive or even prohibitive especially for a large number of aircrafts in the airspace system. This study presents a novel physics-based learning method as a model order reduction (MOR) method for the simulation of aircraft dynamics. The idea of physics-based learning method is to integrate the underlying physics of aircraft dynamical systems into machine learning models to reduce training costs and enhance simulation performances. A recently proposed recurrent neural network (RNN) known as the deep residual RNN (DR-RNN) is used as a tool of physics-based learning. The application of the physics-based learning method is demonstrated on simulating the dynamics of a Boeing 747-100 aircraft. The results show that the DR-RNN can accurately predict aircraft responses and shows excellent extrapolation performances. Furthermore, a purely data-driven approach using the long short-term memory (LSTM) network is also used for the simulation. The comparison demonstrates that incorporating the physics of aircraft dynamics into the learning model can significantly improve prediction performances and effectively reduce training costs compared with using purely data-driven methods. Finally, it is found that the physics-based learning method exhibits superior computation efficiency compared with a classical numerical method since the physics-based learning method can use large time step sizes that violate the numerical stability condition while being explicit in time.

Original languageEnglish (US)
Pages (from-to)254-264
Number of pages11
JournalAerospace Science and Technology
Volume87
DOIs
StatePublished - Apr 1 2019
Externally publishedYes

Fingerprint

Physics
Aircraft
Computer simulation
Recurrent neural networks
Costs
Convergence of numerical methods
Extrapolation
Learning systems
Numerical methods
Dynamical systems
Trajectories

Keywords

  • Data-driven learning
  • Flight mechanics
  • Model order reduction
  • Physics-based learning
  • Recurrent neural network

ASJC Scopus subject areas

  • Aerospace Engineering

Cite this

Aircraft dynamics simulation using a novel physics-based learning method. / Yu, Yang; Yao, Houpu; Liu, Yongming.

In: Aerospace Science and Technology, Vol. 87, 01.04.2019, p. 254-264.

Research output: Contribution to journalArticle

@article{c0b9f899f5ca4861bde67e4843aae57d,
title = "Aircraft dynamics simulation using a novel physics-based learning method",
abstract = "The fast and accurate prediction of flight trajectories is crucial for the real-time prognostics of the air transportation system. However, the computation costs of simulating aircraft dynamics can be expensive or even prohibitive especially for a large number of aircrafts in the airspace system. This study presents a novel physics-based learning method as a model order reduction (MOR) method for the simulation of aircraft dynamics. The idea of physics-based learning method is to integrate the underlying physics of aircraft dynamical systems into machine learning models to reduce training costs and enhance simulation performances. A recently proposed recurrent neural network (RNN) known as the deep residual RNN (DR-RNN) is used as a tool of physics-based learning. The application of the physics-based learning method is demonstrated on simulating the dynamics of a Boeing 747-100 aircraft. The results show that the DR-RNN can accurately predict aircraft responses and shows excellent extrapolation performances. Furthermore, a purely data-driven approach using the long short-term memory (LSTM) network is also used for the simulation. The comparison demonstrates that incorporating the physics of aircraft dynamics into the learning model can significantly improve prediction performances and effectively reduce training costs compared with using purely data-driven methods. Finally, it is found that the physics-based learning method exhibits superior computation efficiency compared with a classical numerical method since the physics-based learning method can use large time step sizes that violate the numerical stability condition while being explicit in time.",
keywords = "Data-driven learning, Flight mechanics, Model order reduction, Physics-based learning, Recurrent neural network",
author = "Yang Yu and Houpu Yao and Yongming Liu",
year = "2019",
month = "4",
day = "1",
doi = "10.1016/j.ast.2019.02.021",
language = "English (US)",
volume = "87",
pages = "254--264",
journal = "Aerospace Science and Technology",
issn = "1270-9638",
publisher = "Elsevier Masson SAS",

}

TY - JOUR

T1 - Aircraft dynamics simulation using a novel physics-based learning method

AU - Yu, Yang

AU - Yao, Houpu

AU - Liu, Yongming

PY - 2019/4/1

Y1 - 2019/4/1

N2 - The fast and accurate prediction of flight trajectories is crucial for the real-time prognostics of the air transportation system. However, the computation costs of simulating aircraft dynamics can be expensive or even prohibitive especially for a large number of aircrafts in the airspace system. This study presents a novel physics-based learning method as a model order reduction (MOR) method for the simulation of aircraft dynamics. The idea of physics-based learning method is to integrate the underlying physics of aircraft dynamical systems into machine learning models to reduce training costs and enhance simulation performances. A recently proposed recurrent neural network (RNN) known as the deep residual RNN (DR-RNN) is used as a tool of physics-based learning. The application of the physics-based learning method is demonstrated on simulating the dynamics of a Boeing 747-100 aircraft. The results show that the DR-RNN can accurately predict aircraft responses and shows excellent extrapolation performances. Furthermore, a purely data-driven approach using the long short-term memory (LSTM) network is also used for the simulation. The comparison demonstrates that incorporating the physics of aircraft dynamics into the learning model can significantly improve prediction performances and effectively reduce training costs compared with using purely data-driven methods. Finally, it is found that the physics-based learning method exhibits superior computation efficiency compared with a classical numerical method since the physics-based learning method can use large time step sizes that violate the numerical stability condition while being explicit in time.

AB - The fast and accurate prediction of flight trajectories is crucial for the real-time prognostics of the air transportation system. However, the computation costs of simulating aircraft dynamics can be expensive or even prohibitive especially for a large number of aircrafts in the airspace system. This study presents a novel physics-based learning method as a model order reduction (MOR) method for the simulation of aircraft dynamics. The idea of physics-based learning method is to integrate the underlying physics of aircraft dynamical systems into machine learning models to reduce training costs and enhance simulation performances. A recently proposed recurrent neural network (RNN) known as the deep residual RNN (DR-RNN) is used as a tool of physics-based learning. The application of the physics-based learning method is demonstrated on simulating the dynamics of a Boeing 747-100 aircraft. The results show that the DR-RNN can accurately predict aircraft responses and shows excellent extrapolation performances. Furthermore, a purely data-driven approach using the long short-term memory (LSTM) network is also used for the simulation. The comparison demonstrates that incorporating the physics of aircraft dynamics into the learning model can significantly improve prediction performances and effectively reduce training costs compared with using purely data-driven methods. Finally, it is found that the physics-based learning method exhibits superior computation efficiency compared with a classical numerical method since the physics-based learning method can use large time step sizes that violate the numerical stability condition while being explicit in time.

KW - Data-driven learning

KW - Flight mechanics

KW - Model order reduction

KW - Physics-based learning

KW - Recurrent neural network

UR - http://www.scopus.com/inward/record.url?scp=85062465360&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85062465360&partnerID=8YFLogxK

U2 - 10.1016/j.ast.2019.02.021

DO - 10.1016/j.ast.2019.02.021

M3 - Article

VL - 87

SP - 254

EP - 264

JO - Aerospace Science and Technology

JF - Aerospace Science and Technology

SN - 1270-9638

ER -