A Hybrid CNN-Transformer (HCNNET) Architecture For Real-Time Seismic Signal Noise Reduction

Abduh, M Ulin Nuha (2026) A Hybrid CNN-Transformer (HCNNET) Architecture For Real-Time Seismic Signal Noise Reduction. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5017221007_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB)

Abstract

Metode denoising klasik seperti Butterworth bandpass dan wavelet thresholding memiliki keterbatasan fundamental: keduanya tidak dapat memisahkan sinyal dan noise yang berada dalam passband yang sama, serta merusak polaritas first motion akibat distorsi fase filter. Penelitian ini mengusulkan HCNNeT (Hybrid CNN-Transformer), sebuah arsitektur deep learning untuk denoising sinyal seismik tiga komponen pada frekuensi 100 Hz. HCNNeT mengintegrasikan normalisasi robust berbasis median absolute deviation, front-end STFT terdiferensiasi, encoder-decoder CNN dengan mekanisme axial radial basis function attention, mask-plus-residual reconstruction head, serta tahap refinement berbasis dilated residual convolutional. Model dilatih menggunakan fungsi loss komposit yang menggabungkan event-weighted Charbonnier loss dan multi-resolution STFT loss. Pelatihan dan validasi dilakukan pada dataset STEAD (Stanford Earthquake Dataset) dengan channel Broadband (High-Gain Seismometer). Evaluasi menggunakan 1.000 sampel dengan SNR −10 sampai +8 dB menunjukkan bahwa HCNNeT mencapai median correlation coefficient sebesar 0,908, peak amplitude ratio −0,028, dan SNR onset 11,83 dB, yang secara konsisten lebih baik dari seluruh baseline klasik pada rentang SNR rendah di bawah 0 dB. Dengan pretrained EQTransformer picker, P-pick rate mencapai 93,6% dibandingkan 49–57% pada baseline, dan polaritas first motion meningkat dari 72,6% (noisy) menjadi 79,7% (HCNNeT). Pada rekaman kontinu 24 jam stasiun YS.LEGO, HCNNeT mereduksi rasio RMS terhadap raw menjadi 0,0665 dengan mempertahankan waveform correlation 0,8275. Pada GPU NVIDIA GeForce RTX 2050, pemrosesan data satu hari membutuhkan waktu 258,30 detik atau 69 milidetik per menit data.
==============================================================================================================================
Classical denoising methods such as Butterworth bandpass filtering and wavelet thresholding suffer from a fundamental limitation: they cannot separate signal and noise residing within the same passband and corrupt first-motion polarity due to filter phase distortion. This study proposes HCNNeT (Hybrid CNN-Transformer), a deep learning architecture for denoising three-component broadband seismic waveforms sampled at 100 Hz. HCNNeT integrates pre-event robust normalization based on the median absolute deviation, a differentiable STFT front end, a CNN encoder-decoder with axial radial-basis-function attention, a mask-plus-residual reconstruction head, and a dilated residual convolutional refinement stage. The model is trained with a composite loss combining event-weighted Charbonnier loss and multi-resolution STFT loss. Training and validation were performed on the STEAD (Stanford Earthquake Dataset) using only the Broadband channel (High-Gain Seismometer). Evaluation on 1,000 samples with SNR ranging from −10 to +8 dB demonstrated that HCNNeT achieved a median correlation coefficient of 0.908, a peak amplitude ratio of −0.028, and an SNR onset of 11.83 dB, consistently outperforming all classical baselines in the low-SNR regime below 0 dB. Coupled with a pretrained EQTransformer picker, the P-pick rate reaches 93.6% compared to 49–57% for classical baselines, and first-motion polarity preservation improves from 72.6% (noisy) to 79.7% (HCNNeT). On a 24-hour continuous recording from station YS.LEGO, HCNNeT reduces the RMS ratio to 0.0665 of the raw input while maintaining a waveform correlation of 0.8275. Using an NVIDIA GeForce RTX 2050 GPU on a standard laptop, processing one day of data requires 258.30 seconds, corresponding to 69 milliseconds per minute of seismic data.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Denoising sinyal seismik, Hybrid CNN-Transformer, Axial RBF attention, Normalisasi pra-event robust, Evaluasi multi-axis, Phase picking, STEAD, Seismic signal denoising, Hybrid CNN-Transformer, Axial RBF attention, Pre-event robust normalization, Multi-axis evaluation, Phase picking, STEAD
Subjects:	Q Science > QE Geology > QE538.8 Earthquakes. Seismology T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Civil, Environmental, and Geo Engineering > Geophysics Engineering > 33201-(S1) Undergraduate Theses
Depositing User:	M Ulin Nuha Abduh
Date Deposited:	22 Jul 2026 01:38
Last Modified:	22 Jul 2026 01:38
URI:	http://repository.its.ac.id/id/eprint/136523

Actions (login required)

View Item