Browsing by Author "Akagunduz, Erdem"

Now showing 1 - 5 of 5

BERT for Harmonic Time Series Modeling: A Multi-Stage Fine-Tuning Approach
(IEEE, 2025) Hekimoglu, Nevin Sehbal; Tileylioglu, Salih; Akagunduz, Erdem
This study demonstrates the potential of a BERT-based transformer model in harmonic signal modeling using synthetic sinusoidal data. The model was trained through a three-stage fine-tuning process (reconstruction, linear analysis, full tuning) with a masked language modeling approach. In the first stage, the model successfully filled in missing data and learned the basic features, while in subsequent stages, its ability to capture temporal dependencies and sequential patterns was enhanced. Additionally, patch, time, and station embedding strategies effectively represented the harmonic structure of the signal. The results indicate that pre-training with synthetic data can overcome the limited access to real-world data, allowing transformer models to be efficiently used in these types of problems.
Deep Learning-Based Epicenter Localization Using Single-Station Strong Motion Records
(Springer, 2025) Turkmen, Melek; Meral, Sanem; Yilmaz, Baris; Cikis, Melis; Akagunduz, Erdem; Tileylioglu, Salih
This paper explores the application of deep learning (DL) techniques to strong motion records for single-station epicenter localization. Often underutilized in seismology-related studies, strong motion records contain rich information for source parameter inference. We investigate whether DL-based methods can effectively leverage this data for accurate epicenter localization. Our study introduces AFAD-1218, a collection comprising more than 36,000 strong motion records sourced from Turkey. To utilize the strong motion records represented in either the time or the frequency domain, we propose two neural network architectures: deep residual network and temporal convolutional networks. Our findings highlight significant reductions in prediction error achieved through the exclusion of low signal-to-noise ratio records, both in nationwide experiments and regional transfer-learning scenarios. Overall, this research underscores the promise of DL techniques in harnessing strong motion records for improved seismic event characterization and localization. Our codes are available via this repo: https://github.com/melekturkmen/EarthQuakeLocalization
Generation and Analysis of Strong Motion Signals via Diffusion Model
(IEEE, 2025) Dogrusoz, Rem; Yilmaz, Baris; Tileylioglu, Salih; Akagunduz, Erdem
This study examines the potential of Denoising Diffusion Probabilistic Models in earthquake engineering to generate seismic signals and learn deep representations. The complex nature of seismic data and its noise are major obstacles that hinder the extraction of meaningful features. Traditional supervised learning methods are limited in their generalization capacity due to their dependence on labeled data. Diffusion models, however, promise to overcome these limitations by generating conditional seismic signals and enhancing the reliability of early warning systems. This study aims to demonstrate how diffusion-based methods contribute to earthquake engineering and propose an approach for seismic data analysis and detection of the P-wave arrival time. The obtained results show that the model can grasp certain patterns; however, larger-scale datasets are needed for more realistic signal generation and a deeper understanding of seismic features.
Training and Evaluation of a Variational Autoencoder for Seismic Station Data Analysis
(IEEE, 2025) Gurkan, Delfin Bengisu; Tileylioglu, Salih; Akagunduz, Erdem
In this study, a variational autoencoder model is proposed to encode the geological, geophysical, and geographical data of seismic stations in Turkey. The model encodes various station-specific data, such as site frequency, surface lithology, and latitude-longitude, into a low-dimensional latent space, aiming to disentangle the generative factors of these data and improve their representation. This approach effectively represents the regional characteristics surrounding the station, enabling the disentanglement of station-related effects. Evaluations based on disentanglement and completeness scores indicate that the model successfully distinguishes station characteristics, such as the average shear wave velocity in the top 30 meters of the surface. The resulting station data encoder can provide additional information to deep learning models processing acceleration records, contributing to a better understanding and modeling of station effects in seismic analysis.
Variational Autoencoders for P-Wave Detection in Strong Motion Earthquake Records
(IEEE, 2025) Ispak, Turkan Simge; Tileylioglu, Salih; Akagunduz, Erdem
Earthquake early warning systems rely on accurate detection of Primary waves before the destructive Secondary waves arrive. However, identifying P-wave onsets in strongmotion accelerograms is challenging due to high noise, limited labeled data, and complex waveforms. This paper proposes a Variational Autoencoder framework for self-supervised P-wave detection in strong-motion data. A Convolutional VAE is trained to reconstruct P-wave segments while rejecting noise and non-P-wave inputs. We employ a sliding window method, combining reconstruction loss and normalized cross-correlation, to locate P-wave arrivals. Experimental results on 1, 2, and 3 second segments show robust performance with area-under-the-curve up to 0.97, demonstrating improved accuracy for longer segments and reduced computational cost for shorter segments.