Pelatihan dan Evaluasi Model EQTransformer Berbasis Dataset Gempa Indonesia untuk Deteksi dan Picking Fase Gelombang Gempa

Achsan, Faiz Maulana (2025) Pelatihan dan Evaluasi Model EQTransformer Berbasis Dataset Gempa Indonesia untuk Deteksi dan Picking Fase Gelombang Gempa. Other thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 5017211014-Undergraduate_Thesis.pdf] Text
5017211014-Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only

Download (4MB) | Request a copy

Abstract

Deteksi event dan picking fase gelombang gempa berperan penting dalam seismologi. Metode konvensional kerap memakan waktu dan rentan terhadap bias terutama pada data berskala besar. Peningkatan volume data dan tuntutan akurasi yang tinggi mendorong penggunaan deep learning. EQTransformer merupakan salah satu model deep learning yang dapat mendeteksi event dan picking fase gelombang gempa secara bersamaan pada stasiun tunggal. Namun, performa EQTransformer sering tidak konsisten akibat pemilihan parameter overlap dan probability threshold yang tidak optimal. Penelitian ini bertujuan untuk melatih dan mengevaluasi EQTransformer menggunakan data gempa Indonesia dengan magnitudo lebih dari tiga pada rentang 2015-2021. Tahapan riset mencakup pembuatan dataset, pelatihan dan validasi model, serta pengujian menggunakan data kontinu. Hasil pelatihan menunjukkan kurva loss yang stagnan, mengindikasikan pelatihan yang belum maksimal. Meskipun demikian, nilai f1-score dari proses validasi mencapai 0,9 untuk kedua fase P dan S. Pengujian dilakukan dengan dua pendekatan preconditioning data kontinu, mekanisme overlap dan manual. Mekanisme overlap memanfaatkan metode sliding window untuk memproses seluruh data, sedangkan mekanisme manual melibatkan filtering, short-time fourier transform, cutoff energi gelombang untuk memperoleh perkiraan timestamp gempa. Hasil pengujian menunjukkan penerapan optimasi parameter dan mekanisme overlap tidak cukup efektif. Sebaliknya, mekanisme manual mampu meningkatkan performa model lokal secara signifikan dari f1-score awal sebesar 0,14 menjadi 0,647. Performa terbaik didapatkan melalui kombinasi optimasi parameter dan mekanisme manual. Nilai f1-score model lokal mencapai 0,718 dan 0,328 untuk fase P dan S, sedikit lebih tinggi dibandingkan model native dengan 0,531 dan 0,298. Hasil ini menunjukkan bahwa model lokal memiliki performa yang hampir sama dengan model native dengan pendekatan preconditioning manual. Penelitian ini memberikan landasan untuk pengembangan lebih lanjut EQTransformer dengan dataset gempa Indonesia, termasuk penerapan transfer learning serta penambahan proses matching dan otomatisasi dalam preconditioning data.
============================================================================================
Event detection and seismic phase picking play a crucial role in seismology. Conventional methods are often time-consuming and prone to bias, especially when applied to large-scale data. The increasing volume of seismic data and demand for high accuracy have encouraged the adoption of deep learning approaches. EQTransformer is one such deep learning model that can simultaneously detect events and pick seismic phases using single-station data. However, its performance tends to be inconsistent due to suboptimal choices of overlap and probability threshold parameters. This study aims to train and evaluate EQTransformer using Indonesian earthquake data from 2015 to 2021, focusing on events with magnitudes above 3. The research includes dataset construction, model training and validation, and testing using continuous data. The training results show a stagnant loss curve, indicating suboptimal learnning process. Nevertheless, the validation process yields high F1-scores of 0.9 for both P and S phases. The model is tested using two data preconditioning approaches: an overlap-based mechanism and an alternative manual approach. The overlap mechanism uses a sliding window to process continuous data, while the manual method applies filtering, short-time Fourier transform, and energy-based cutoffs to estimate earthquake timestamps. Testing results show that optimasi parameter and the overlap mechanism alone are insufficient to improve performance. In contrast, the manual approach significantly boosts the local model's performance, increasing the F1-score from 0.14 to 0.647. The best results are achieved by combining parameter optimization with manual preconditioning, yielding F1-scores of 0.718 and 0.328 for P and S-phase, slightly outperforming the native model, which scored 0.531 and 0.298. These results suggest that the local model performs comparably to the native one when supported by manual preconditioning. This study provides a foundation for further development of EQTransformer using Indonesian earthquake data, including the potential application of transfer learning and the integration of automatic matching and preconditioning techniques.

Item Type: Thesis (Other)
Uncontrolled Keywords: Deep learning, EQTransformer, Fase, Gelombang gempa, Optimasi parameter, Deep learning, EQTransformer, Seismic phase, Earthquake waveform, Parameter optimization
Subjects: Q Science > QE Geology > QE538.8 Earthquakes. Seismology
Divisions: Faculty of Civil, Planning, and Geo Engineering (CIVPLAN) > Geophysics Engineering > 33201-(S1) Undergraduate Thesis
Depositing User: Faiz Maulana Achsan
Date Deposited: 04 Aug 2025 02:39
Last Modified: 04 Aug 2025 02:39
URI: http://repository.its.ac.id/id/eprint/124078

Actions (login required)

View Item View Item