논문 상세보기
Rapid, real-time detection of anomalies and locate structural defects during earthquakes is critical for ensuring safety and enabling timely decision-making. Although deep learning-based structural health monitoring (SHM) has shown considerable promise, conventional supervised models are often impractical because labeled damage data from real-world structures are extremely scarce. To address this challenge, this paper proposes a Multi-Class Deep Support Vector Data Description (SVDD) framework for structural defect detection. The proposed Multi-Class Deep SVDD approach learns the boundary of normal data using only normal seismic acceleration responses. When new data are recorded, the system infers both the occurrence and location of defects by evaluating whether the responses fall within or deviate from the learned normal boundary. The framework is validated using the Los Alamos National Laboratory 3-story bookshelf structure benchmark dataset. Experimental results show that the proposed model achieves a peak average accuracy of 87.12% in a 4-dimensional latent space, substantially outperforming traditional baseline methods, including Kernel Density Estimation (KDE), SVDD, and One-Class Deep SVDD. These findings indicate that the Multi-Class Deep SVDD framework provides a robust and objective metric for rapid post-earthquake safety assessment without requiring prior exposure to faulty datasets.
1. 서 론
2. 방법론
2.1 단일 클래스 Deep SVDD
2.2 다중 클래스 Deep SVDD
3. 적용 방법
4. 실험 데이터셋
5. 실험 및 결과
5.1 잠재 공간 차원 수에 따른 정확도 분석
5.2 반지름 분위수 값에 따른 정확도 분석
5.3 저주파 통과 필터 적용에 따른 성능 변화 분석
5.4 타 방법론과의 정확도 비교 분석
6. 결 론
/ 감사의 글 /
/ REFERENCES /
