Politecnico di Torino (logo)

Neural network data analysis for virtual air data sensors

Angelo Scacciavillani

Neural network data analysis for virtual air data sensors.

Rel. Piero Gili, Alberto Brandl. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Aerospaziale, 2018

PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution.

Download (6MB) | Preview

Over the years it is getting more present the use of machine learning techniques applied to many different disciplines. In aeronautic field they are used as a substitute or as an enrichment of flight systems in order to have a better production of flight data. The main three aspects studied in neural networks, which are one of the many machine learning techniques and the one used in this thesis, are the network architecture, the training phase and the testing phase.There are other very important aspects such as the various methods for carrying out these operations. The knowledge of the latter is very important in order to obtain very precise networks in the purpose that is set. This thesis is in fact aimed at finding a method for the improvement of one of these aspects. The focus of this project is on the analysis of input data, that is the training and, in a small part, which impact they do have on neural network quality, that is the testing phase, which constitutes the output. In this specific case the neural network is used to simulate an angle of attack sensor. The main advantage of a NN is that it is possible to model, even if having an approximation, the complex behavior of a given system, without knowing equations and relations which govern the system. In this case the advantage is quite clear: knowing data as accelerations over the three body axis and the rotation rates around them and some other data, it is possible to find the angle of attack, without modeling the entire flight mechanics of the plane. Characteristics of the input data will be evaluated via statistics and a data reduction technique will be used in order to select only meaningful data points with the usage of k-means clustering algorithm. In previous studies an encouraging method was found to exploit neural networks in order to simulate the angle of attack sensor. However, certain of the complexity of neural networks, it is an improvable method. Therefore this thesis has the role of proposing an improvement of a previously used method. This study has been done in collaboration with Ing. Nando Groppo and flight data utilized in this thesis are from a Groppo G70.

Relators: Piero Gili, Alberto Brandl
Academic year: 2017/18
Publication type: Electronic
Number of Pages: 81
Corso di laurea: Corso di laurea magistrale in Ingegneria Aerospaziale
Classe di laurea: New organization > Master science > LM-20 - AEROSPATIAL AND ASTRONAUTIC ENGINEERING
Aziende collaboratrici: UNSPECIFIED
URI: http://webthesis.biblio.polito.it/id/eprint/7754
Modify record (reserved for operators) Modify record (reserved for operators)