Razvrščanje signalov v posnetkih vibracijske krajine

Štrus, Tomaž (author), Marolt, Matija (mentor), Pesek, Matevž (comentor)

Vibracijska krajina je naravno vibracijsko okolje, sestavljeno iz bioloških, geofizikalnih in antropogenih vibracij. V tem še zelo neraziskanem svetu vibracijske krajine nas zanima predvsem vibracijska komunikacija žuželk. Zato smo v okviru magistrskega dela raziskali, kako uspešno detektirati oglašanja žuželk na posnetkih vibracijske krajine z uporabo različnih strojnih modelov. Učinkovitost različnih preprostih modelov strojnega učenja, kot je model SVC, smo primerjali z usmerjeno nevronsko mrežo. Te modele smo kombinirali s tradicionalnimi ročnimi značilkami, kot sta LFCC in MFCC, hibridnimi značilkami, ki združujejo obe prej omenjeni metodi, ter globokimi značilkami openl3. Poleg tega smo uporabili tudi globoka modela CNN in TFNet, ki nimata potrebe po predhodnem izračunu značilk. Z globokim modelom smo preizkusili tudi metode, kot je normalizacija PCEN in metodi obogatitve podatkov mixup in specAugment. Za detekcijo vseh oglašanj smo dobili najboljše rezultate s kombinacijo značilk openl3 in modela SVC, kjer smo dosegli 79,950-odstotno mero F1 na testnih podatkih.

• Language: Slovenian
• Keywords: šibko nadzorovano učenje, klasifikacija zvokov, vibracijska krajina
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-152768
• COBISS.SI-ID: 177616643
• Publication date in RUL: 06.12.2023

ŠTRUS, Tomaž, 2023, Razvrščanje signalov v posnetkih vibracijske krajine [online]. Master’s thesis. [Accessed 17 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=152768

Secondary language

• Language: English
• Title: Signal classification in vibroscape recordings

Abstract

Vibroscape is a naturally occurring vibrational environment composed of biological, geophysical, and anthropogenic vibrations. In this still largely unexplored realm of the vibrational landscape, our primary focus is on the vibrational communication of insects. Therefore, as part of our master's thesis, we examined the ability to detect insect calls in vibrospace recordings using various machine learning models. We compared the performance of manual features such as LFCC, MFCC, and the combination of both as hybrid features to the performance of deep features extracted using openl3. We compared the effectiveness of different simple machine learning models, such as the SVC model, with directed neural networks. We also employed deep CNN and TFNet models without pre-computed features, with PCEN normalization and data augmentation techniques like mixup and specAugment. For the detection of all insect calls, the best results were obtained using a combination of openl3 features and the SVC model, achieving an F1 score of 79.950% on the test data.

• Keywords: weakly-supervised learning, acoustic classification, vibroscape