Akademik Veri Yönetim Sistemi
Neural Computing and Applications, cilt.38, sa.2, 2026 (Scopus)
This study uses a ground motion recorder to measure surface vibrations and analyzes patterns to enhance earthquake prediction. Surface vibrations caused by far-focus earthquakes were recorded, and their responses in the temporal and frequency domains were evaluated. A recurrent unsupervised artificial neural network model, validated through feature extraction studies, was developed to predict pattern changes during earthquakes. A secondary objective was to assess the impact of strong sinusoidal vibrations from turbine and hydro generator-induced rotational motion on the resonance state of bedrock beneath the Gürsöğüt Hydroelectric Dam. Continuous ground motion data collected from July 2023 to October 2024 included 20 far-focus earthquakes with magnitudes between 4.2 and 5.6. A Long Short-Term Memory (LSTM) deep learning model was trained incrementally using spectral and statistical features from 19 earthquakes to predict the excluded one in a repeated evaluation. Using 5-s sliding windows, the model forecasted anomalies approximately 50 s before each earthquake. The deep learning approach successfully predicted seismic events by recognizing patterns in spectral and statistical features before and during earthquakes. This revealed directional and magnitude relationships that generalize seismic behavior. The model correctly identified approximately 64% of actual seismic peaks (true positives) while misclassifying about 18% of non-seismic peaks as seismic (false positives). The Mean Absolute Error (MAE) in non-peak zones was 0.254, indicating effective background noise suppression. These results suggest the model effectively suppresses background noise while focusing on meaningful seismic activity. These results demonstrate that the model effectively suppresses background noise while focusing on meaningful seismic activity. Emphasizing feature identification over predictive tasks, the study provides a framework for labeling seismic event anomalies before and during earthquakes and designing attention mechanisms in machine and deep learning models. The model offers a generalized understanding of seismic behavior by incorporating seismic traces from different fault lines. Additionally, its ability to determine the natural frequency of bedrock supports improved soil–foundation interaction modeling, which is critical for large-scale structures.