차원축소, 군집분석 기반 소리 분리 및 CNN을 통한 위험 소리 인식

다국어 초록 (Multilingual Abstract)

This study proposes a hybrid framework for separating hazardous sounds from single-channel mixed audio signals by combining dimensionality reduction based on Non-negative Matrix Factorization (NMF) with K-medoids clustering. The method decomposes the spectrogram of a mixed signal into low-rank basis components and subsequently groups them using a distance-based clustering scheme designed to enhance the separation of hazard-related spectral patterns. After reconstructing cluster-specific signals, the separated audio is evaluated using the Signal-to-Distortion Ratio (SDR) to quantify separation quality. To further verify the practical effectiveness of the proposed pipeline, Mel-Frequency Cepstral Coefficients (MFCCs) extracted from the separated signals are passed to a Convolutional Neural Network (CNN) classifier to identify the underlying hazard sound classes. Experimental results demonstrate that the distance-driven clustering approach improves separation performance and that the post-separation CNN classification achieves meaningful recognition accuracy, confirming the viability of the overall method for real-world hazard sound analysis

목차 (Table of Contents)

• 제1장 서 론 1
• 제2장 본 론 3
• 제1절 분석 자료 설명 3
• 제2절 Mel-Frequency Cepstral Coefficients(MFCC) 5
• 2.1. MFCC 계산 방법 6
• 제3절 분석 방법 10
• 3.1. 소리 분리 11
• 3.1.1. 비음수 행렬 분해(NMF) 11
• 3.1.2. K-medoids 클러스터링 14
• 3.2. 소리 분류 19
• 3.2.1. CNN 20
• 제4절 성능 평가 22
• 4.1. 실험 환경 22
• 4.2. Sound-to-Distortion ratio(SDR) 24
• 4.3. 성능 평가 결과 25
• 4.3.1. 소리 분리 결과 26
• 4.3.2. 소리 분류 결과 28
• 제3장 결 론 31
• 참 고 문 헌 32
• ABSTRACT 34