Primerjava algoritmov nadzorovanega in nenadzorovanega učenja za klasifikacijo zvočnih dogodkov

magistrsko delo magistrskega študijskega programa II. stopnje Strojništvo

Anže Železnik (Author), Jurij Prezelj (Mentor)

Abstract

V avtomobilski industriji je zaradi zvočnega ugodja pomembno zaznavanje zvoka stick-slip efekta, ki se pri počasnem speljevanju lahko pojavi med diskom in ploščicami zavore. Ker je subjektivno ocenjevanje neprijetnega zvoka zavor s pomočjo ocenjevalcev drago in dolgotrajno, bi bilo smiselno ocenjevalce zamenjati z algoritmom strojnega učenja. Da bi subjektivno ocenjevanje zamenjali z najboljšo metodo strojnega učenja, smo najprej preizkusili več metod nadzorovanega in nenadzorovanega učenja na velikem številu značilk, ki so bile pridobljene pri posnetkih ocenjevanja zavor. Nato smo glede na rezultate izbrali manjše število pomembnejših značilk in algoritme učili le na njihovi podlagi. Glede na rezultate smo primerjali algoritme. Za najboljša algoritma se izkažeta samoorganizirajoča mreža in algoritem k-povprečij. Rezultati, pridobljeni z uporabo štirih izbranih značilk pri šestih razredih, so se izkazali za bolj zanesljive in ponovljive kot pridobljene subjektivne ocene.

Keywords

magistrske naloge;strojno učenje;nenadzorovano učenje;klasifikacija zvočnih posnetkov;samoorganizirajoča mreža;k-povprečja;k-najbližji sosedje;digitalno procesiranje signala;

Data

Language:	Slovenian
Year of publishing:	2021
Typology:	2.09 - Master's Thesis
Organization:	UL FS - Faculty of Mechanical Engineering
Publisher:	[A. Železnik]
UDC:	004.85:004.021:534.32(043.2)
COBISS:	61770243
Views:	591
Downloads:	90
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Comparison of algorithms for supervised and unsupervised classification of sound events
Secondary abstract:	In the automotive industry, the perception of the stick-slip effect, which can occur between the brake disc and the brake pads during slow acceleration, is important in terms of noise pleasantness. Since the subjective quantification of unpleasant brake sound is expensive and time-consuming, it would make sense to replace the subjective evaluation with a machine learning algorithm. In order to replace subjective evaluation with a machine learning algorithm, several methods of supervised and unsupervised learning were tested on a large number of features obtained from experimental brake tests. Based on the results, a small number of important features was selected the algorithms were trained only with the selected features. The algorithms were compared based on the results. The self-organizing map and the k-means algorithm proved to be the most appropriate algorithms. The results obtained using four selected features and six classes proved to be more reliable and reproducible than the obtained subjective evaluations.
Secondary keywords:	master thesis;machine learning;unsupervised learning;audio classification;self-organizing map;k-means;k-nearest neighbors;digital signal processing;
Type (COBISS):	Master's thesis/paper
Study programme:	0
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. Ljubljana, Fak. za strojništvo
Pages:	XXIV, 91 str.
ID:	12708623