Time and Frequency Fusion Deep Learning Model for ECG Signal Quality Enhancement
- 주제(키워드) Denoising , ECG , Fourier analysis , Transformer
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Kyung-Ah Sohn
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035466
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Electrocardiogram (ECG) signals are often contaminated by various noise sources such as baseline wander, electrode motion, and muscle artifacts, particularly in wearable or ambulatory monitoring scenarios. Current methods often struggle to restore high-quality signals in these complex noise environments. This thesis introduces a deep learning model that fuses time and frequency domains to improve ECG signal quality by simultaneously learning temporal and spectral features. These combined features are processed using a Transformer encoder with Fourier Analysis Networks, allowing the model to capture multi-scale dependencies while maintaining the morphological features of ECG signals. Experiments conducted on the QT and MIT-BIH Arrhythmia databases show that the proposed model significantly enhances key performance metrics—RMSE, PRD, and SNR—compared to traditional filters and recent deep learning models. Notably, the explicit integration of frequency-domain features enables the model to more effectively recover signal components across a broad frequency range than standard deep learning techniques. In addition to waveform restoration, our model also excels in downstream tasks like R-peak detection and cardiac abnormality classification, highlighting its potential for real-time clinical use. Keywords: Denoising, ECG, Fourier analysis, Transformer
more목차
1. Introduction 1
2. Related Works 4
2.1 ECG Signal and Noise Characteristics 4
2.2 Traditional Approaches for ECG Denoising 5
2.3 Deep Learning-Based ECG Enhancement 6
3. Methodology 8
3.1 Dataset and Noise Simulation 8
3.1.1 Dataset 8
3.1.2 Comprehensive ECG Noise Simulation Dataset 9
3.2 Model Architecture 10
3.2.1 Dual-Branch Feature Extractor 11
3.2.2 Transformer Encoder with Fourier Analysis Network 12
3.2.3 Denoising Decoder 13
4. Experiments 14
4.1 Experimental Settings 14
4.1.1 Training Strategy 14
4.1.2 Evaluation Metrics 15
4.2 Quantitative Results 16
4.2.1 Overall Performance Comparison 16
4.2.2 Performance Comparison Across Noise Intensities 17
4.3 Qualitative Analysis 18
4.4 Downstream Tasks for Denoising Validation 21
4.4.1 R-Peak Detection with Clinical Tolerance 21
4.4.2 ECG Arrhythmia Classification 23
4.5 Ablation Studies 24
5. Conclusion 25
Reference 27
