Detekcija potresa u sjeverozapadnim Dinaridima uz upotrebu neuronskih mreža

master's thesis

• Title (english): Neural network based earthquake detection in the northwest Dinarides
• Author: Dorian Wild
• Mentor: Josip Stipčević (mentor)
• Committee member: Josip Stipčević (predsjednik povjerenstva)
• Committee member: Iva Dasović (član povjerenstva)
• Committee member: Zoran Pasarić (član povjerenstva)
• Committee member: Ivana Herceg Bulić (član povjerenstva)
• Committee member: Igor Mandić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Geophysics); Zagreb
• Defense date and country: 2024-02-27, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Geophysics; Seismology and Physics of the Earth's Interior

Abstract

Dolaskom doba nazvanim Big Data, dogodilo se eksponencijalno povećanje količine seizmičkih podataka zbog povećanja broja seizmičkih postaja i napretka tehnologije, zbog čega tradicionalne metode analize i ručne detekcije faza postaju neodržive. Ususret tom vremenu, testira se novi pristup koji koristi neuronske mreže za automatizaciju detekcije i karakterizaciju seizmičkih događaja, s ciljem povećanja točnosti i učinkovitosti u odnosu na postojeće automatizirane sustave. U ovom radu analiziran je sveobuhvatni skup podataka kojeg čini testni katalog potresa u razdoblju 2009. - 2018. godine u sjeverozapadnim Dinaridima primjenom modela PhaseNet za detekciju seizmičkih faza i GaMMA Associatora za izradu sintetičkih kataloga te procedura koje značajno poboljšavaju detekciju potresa. Korišteno je pet varijanti modela PhaseNet treniranih na pet različitih baza podataka: INSTANCE, STEAD, ETHZ, SCEDC i GEOFON sa svrhom detekcije P- i S-faza, te algoritam na temelju strojnog učenja zvan GaMMA Associator sa svrhom asocijacije i izrade sintektičkog kataloga. Najprije se varijante modela PhaseNet primjenjuju na uzorak iz stvarnog kataloga u svrhu određivanja najefikasnije varijante modela PhaseNet, te se potom ta varijanta primjenjuje na cijeli skup podataka. Kombinacija modela PhaseNet s tehnikom učenja prijenosom i GaMMA Associatora omogućava stvaranje kataloga potresa koji sadrži preko 92 % identičnih događaja iz usporednog stvarnog kataloga, te također 10 puta više novih događaja. Rezultati naglašavaju potencijal strojnog učenja za revoluciju u praćenju potresa pružanjem bržih, većih i pravovremenijih sposobnosti detekcije. Ovo istraživanje ne samo da doprinosi napretku metoda detekcije potresa, već otvara i puteve za buduća istraživanja u seizmologiji i povezanim područjima.

Abstract (english)

With the arrival of the era called Big Data, there has been an exponential increase in the volume of seismic data due to the expansion of seismic stations and the advancement of technology, due to which traditional methods of analysis and manual phase detection are becoming unsustainable. Therefore, a new approach that uses neural networks to automate the detection and characterization of seismic events is being tested, with the goal of increasing accuracy and efficiency compared to existing automated systems. In this thesis, the PhaseNet models and GaMMA Associator were applied in the analysis of a comprehensive data set of the test earthquake catalogue for the period of 2009 – 2018 in the northwest Dinarides to significantly improve earthquake detection and create a synthetic earthquake catalogue. Five variants of the PhaseNet model trained on five different databases were used: INSTANCE, STEAD, ETHZ, SCEDC and GEOFON to detect P and S phases, and GaMMA Associator, an algorithm based on machine learning for association and creation of a synthetic catalog. The models were first applied to a sample from the real catalog to determine the most efficient variant of PhaseNet, and then that variant is applied to the entire data set. The combination of PhaseNet and GaMMA Associator generates an earthquake catalog that contains over 92% identical events from the comparative real catalog, and 10 times more new events. The findings underscore the potential of machine learning to revolutionize earthquake monitoring by providing more reliable and timely detection capabilities. This research not only contributes to the advancement of seismic detection methods but also opens avenues for future investigations in seismology and related fields.

• Keywords: Detekcija potresa; Sjeverozapadni Dinaridi; Neuronske mreže
• Keywords (english): Earthquake detection; Neural networks; Northwest Dinarides
• Language: croatian
• URN:NBN: urn:nbn:hr:217:618108
• Study programme: Title: graduate academic research programme: Master of Physics – geophysics, modules: Seismology and solid Earth physics, Meteorology and physical oceanography; specializations in: Seismology and solid Earth physics, Meteorology and physical oceanography; Course: Seismology and solid Earth physics; Study programme type: university; Study level: graduate; Academic / professional title: sveučilišni magistar fizike – geofizike (sveučilišni magistar fizike – geofizike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-03-19 10:32:34