# Knowledge base: Warsaw University of Technology

Back

## Autopilot system based on artificial neural networks

### Paweł Kicman

#### Abstract

In the following thesis project set of simulation was performed in Simulink program, which is part of computing environment Matlab. Their goal was to develop longitudal autopilot system for 30-seat, two-engine jet aircraft in which pitch angle is controlled by a neural network. First part of the paper contains theoretical introduction, where information about autopilots and neural networks were gathered – history, architectures, operating rules and basic structures that are in practical use. Additionally for neural networks the number of possible applications in aviation were listed, mostly concerned on control issues. After explaining and justifying methods used to solve the problem the development process was described. First of all autopilot based on standard PI and PID controllers was designed (PID for pitch and speed control and PI for height control). This system was point of reference and basis for further experiments with neural networks. Its scheme is presented on rysunek 2. Next, using direct inverse control method (general training version), was performed experiment which aim was to create neural network that can be used as a controller of pitch angle for considered model (PID1 on rysunek 2). Using Levenberg-Marquardt algorithm 420 networks were trained – 4 structures of input signal, 3 values of feedback signal delay and different sizes of networks hidden layer. Unfortunately this experiment gave no satisfying results and the best of created networks was unable to perform the task in proper way. In order to check the methodology that was used additional network was trained – their goal was to model created previously pitch PID controller. In this case much better results were achieved. After analysis of results the following conclusions and possible reasons for failure were listed: • Training set prepared in wrong way; • Δt delay not sufficiently adjusted to controlled object; • Author’s lack of experience in field of neural network; Despite the failure it is possible to successfully complete task and to develop properly working neural controller using presented methods after allowing for suggestions that were made.
Diploma type
Engineer's / Bachelor of Science
Diploma type
Engineer's thesis
Author
Paweł Kicman (FPAE/IAAM) Paweł Kicman,, The Institute of Aeronautics and Applied Mechanics (FPAE/IAAM)Faculty of Power and Aeronautical Engineering (FPAE)
Title in Polish
Układ autopilota oparty na sieciach neuronowych
Supervisor
Robert Głębocki (FPAE/IAAM) Robert Głębocki,, The Institute of Aeronautics and Applied Mechanics (FPAE/IAAM)Faculty of Power and Aeronautical Engineering (FPAE)
Certifying unit
Faculty of Power and Aeronautical Engineering (FPAE)
Affiliation unit
The Institute of Aeronautics and Applied Mechanics (FPAE/IAAM)
Study subject / specialization
, Lotnictwo i Kosmonautyka
Language
(pl) Polish
Status
Finished
Defense Date
15-09-2009
Issue date (year)
2009
Pages
58
Internal identifier
MEL; PD-913
Reviewers
Krzysztof Kędzior (FPAE/IAAM) Krzysztof Kędzior,, The Institute of Aeronautics and Applied Mechanics (FPAE/IAAM)Faculty of Power and Aeronautical Engineering (FPAE) Robert Głębocki (FPAE/IAAM) Robert Głębocki,, The Institute of Aeronautics and Applied Mechanics (FPAE/IAAM)Faculty of Power and Aeronautical Engineering (FPAE)
Keywords in Polish
SIECI NEURONOWE, STEROWANIE, AUTOPILOTY, LOTNICTWO I KOSMONAUTYKA
Keywords in English
xxx
Abstract in Polish
urn:pw-repo:WUTcf988c3d31b64d63aa986cd960efaf5e