The 6th Korea Artificial Intelligence Conforence
한국 통신학회 제 6회 한국 인공지능 학술대회, 2025년 9월, 서귀포 칼 호텔
© 2025 KAIC
This paper presents a systematic comparative study of eight deep learning architectures for seismic signal denoising, addressing a critical challenge in building reliable earthquake prediction systems.
Unlike prior works that focus on improving single-model performance, this research emphasizes cross-architecture evaluation under controlled experimental settings. The comprehensive design and standardized comparison metrics make the findings highly reproducible and practically valuable for future AI-based seismology research.
The main contribution lies in identifying PatchTST, a Transformer-based model, as the most effective structure for denoising seismic waveforms. The paper demonstrates that PatchTST achieves 0.0004 MSE, 99.98% accuracy, and 0% false negatives, outperforming CNN, RNN, and Autoencoder-based models. This result validates the potential of patch-based temporal representation learning for complex time-series data such as seismic signals.
The methodology is rigorous and transparent.
Dataset: The use of 40,000 MSEED-format samples with real earthquake and noise data ensures realism and robustness.
Preprocessing: The authors carefully apply DC removal, cosine tapering, bandpass filtering (0.1–10 Hz), and normalization. This process accurately reflects the physical characteristics of Korean seismic data.
Experimental Control: All models were trained with identical hyperparameters (Adam optimizer, lr=1e-3, batch=16, 20 epochs) on the same hardware, ensuring fairness in performance comparison.
The study also introduces a multi-layer evaluation scheme—using both regression metrics (MSE, MAE) and classification metrics (Accuracy, F1, FPR, FNR)—which gives a holistic view of model performance from both statistical and operational perspectives.
Reproducibility and Fairness: Every model was tested under strictly equal conditions, removing typical confounders in comparative AI studies.
Empirical Clarity: Quantitative results clearly show the dominance of Transformer-based approaches in handling both local and global dependencies.
Applicability: The focus on false-negative reduction aligns directly with the safety-critical nature of earthquake prediction, enhancing societal relevance.
While the paper is methodically sound, several areas could be expanded in future work:
The dataset is limited to a single national seismic source (Korean region). Generalization to global datasets or heterogeneous geological conditions should be tested.
The FFTformer and Denoising Autoencoder results are notably poor, but the analysis of why these models fail could be deepened. A spectral attention visualization or error distribution analysis would strengthen interpretability.
The study could benefit from additional uncertainty quantification and real-time performance evaluation to bridge the gap between laboratory and deployment environments.
This paper represents a well-executed comparative benchmark in the intersection of AI and seismology.
It highlights that recent Transformer models—particularly PatchTST—outperform traditional CNN, RNN, and Autoencoder structures in seismic noise reduction.
The work is both methodologically rigorous and application-oriented, serving as a strong foundation for future research on reliable, real-time earthquake detection systems.
This research effectively bridges data-driven AI approaches and geophysical signal processing.
Its disciplined experimentation, transparent reporting, and clear conclusions make it a credible contribution to applied machine learning in the geoscience domain.
Further studies extending to ensemble architectures or uncertainty-aware prediction would likely enhance both scientific and practical impact.
Sunjun Hwang is an undergraduate researcher at the RAISE Lab, Yonsei University. His academic and research interests are centered on cutting-edge areas within computer science and quantum technology. He is particularly focused on quantum computing, exploring the principles and applications of quantum systems to solve complex computational problems. Additionally, Sunjun is deeply engaged in the study of quantum algorithms, investigating innovative approaches to designing algorithms that leverage quantum mechanics for enhanced computational efficiency. His interest in quantum artificial intelligence reflects his pursuit of integrating quantum computing techniques with AI methodologies to develop advanced intelligent systems. Furthermore, he is involved in artificial intelligence security, focusing on safeguarding AI systems against vulnerabilities. A key aspect of his work includes studying adversarial attacks, where he examines techniques used to manipulate or compromise AI models, aiming to develop robust defenses to enhance the security of intelligent systems.
Sehee Park is an undergraduate researcher at the Viscom Lab, Yonsei University. Her academic and research pursuits are centered on advanced domains within computer science and engineering. She has a keen interest in image processing, where she explores techniques to enhance, analyze, and manipulate visual data for various applications. Additionally, Sehee is focused on communication systems, investigating methods to optimize data transmission and improve the efficiency of networked systems. Her work also extends to artificial intelligence, where she engages in the development and application of intelligent algorithms to solve complex problems. Furthermore, she is deeply involved in signal processing, studying techniques to analyze and interpret signals for improved system performance and data accuracy.
Kangmin Ko is a member of the Software Department at Yonsei University. Her academic and professional interests encompass several specialized domains within computer science and engineering. She is particularly focused on artificial intelligence, where she investigates the development and application of intelligent systems to tackle complex challenges. Additionally, Kangmin has a keen interest in embedded hardware, engaging in the design and optimization of hardware systems that integrate seamlessly with software to achieve efficient performance. Her work also includes information retrieval, where she explores techniques to improve the accuracy and efficiency of accessing and organizing data from large datasets.
Jiyoon Baik is an undergraduate researcher at the ISP Lab, Yonsei University. Her academic and research interests are centered on advanced areas within signal processing and artificial intelligence. She is particularly focused on speech signal processing, exploring techniques to analyze and enhance audio signals for applications such as speech recognition and synthesis. Additionally, Jiyoon is engaged in digital signal processing (DSP), investigating methods to manipulate and optimize digital signals for improved performance in various systems. Her work also encompasses artificial intelligence, where she applies intelligent algorithms to solve complex problems in data processing and system optimization. Furthermore, she has a strong interest in signal processing and signal systems, studying the theoretical and practical aspects of signal analysis, transformation, and system design to enhance the efficiency and accuracy of technological applications.