본문 바로가기

Data-Driven Estimation of the Critical Structural Strain from Boundary Acceleration Using a 2-Stage Shallow Time-Delay Neural Network

문서 유형
연구출판물 · 학술지 논문
등재
KCISCIEScopus
저널 순위
JCR Q3Scopus Q2KCI Q3
피인용
Web of Science 0Scopus 0
FWCI (Scopus)
0
JCR IF
JCR 2024 1.4
ISSN
2093-274X

저자

  1. Seok, Jin Hyeok제1저자Seoul Natl Univ, Dept Aerosp Engn, Extreme Environm & Impact Lab, Seoul 08826, South Korea
  2. Seo, Ji WanSeoul Natl Univ, Dept Aerosp Engn, Extreme Environm & Impact Lab, Seoul 08826, South Korea
  3. 조영관Pusan Natl Univ, Dept Mech Engn, Intelligent Addit Mfg Lab, Busan 46241, South Korea
  4. 박상후Pusan Natl Univ, Dept Mech Engn, Intelligent Addit Mfg Lab, Busan 46241, South Korea
  5. Lim, Jae HyukKyung Hee Univ, Sch Mech Engn, Space Mobil & Digital Twin Lab, Yongin 17104, South Korea
  6. 김윤호교신저자Seoul Natl Univ, Dept Aerosp Engn, Extreme Environm & Impact Lab, Seoul 08826, South Korea

초록

Estimating internal strain responses from boundary acceleration measurements remains a challenging problem in structural monitoring. This study proposes a two-stage neural network framework based on a shallow time-delay neural network (TDNN) to reconstruct internal strain responses using triaxial acceleration measurements from a single accessible boundary region. The proposed approach separates the acceleration-to-strain transformation into two sequential steps where acceleration is first used to estimate displacement, and the displacement response is then used to infer strain. This staged learning structure simplifies the mapping between structural responses and enables consistent reconstruction using a compact model trained on limited experimental data. The framework was validated through a rotary excitation experiment with a prediction duration of 242 s. When the measured and predicted strain signals were compared using the coefficient of determination (R 2), the resulting values were strongly affected by phase drift that accumulated over time between the signals. After compensating for this phase difference using dynamic time warping (DTW), both models reproduced the overall strain waveform with approximately 93% similarity. In fatigue-related structural assessment, accurate estimation of strain peaks is particularly important because damage accumulation is dominated by peak amplitudes. The proposed 2-stage model improved the reconstruction of peak strain values by approximately 12% compared with the 1-stage model. Further frequency-domain analysis based on Fast Fourier Transform (FFT) revealed that the staged learning structure preserves peak-related high-frequency information, enabling more accurate reconstruction of rapid strain variations. These results demonstrate that localized internal strain behavior can be inferred from minimal boundary sensing and provide a practical pathway for data-efficient structural monitoring for fatigue evaluation and digital-twin-based structural assessment.

키워드

Dynamic response estimationStructural health monitoring (SHM)Time-delay neural network (TDNN)Virtual sensingInternal strain reconstructionDAMAGE IDENTIFICATIONFORCE IDENTIFICATION

식별자

Scopus EID
2-s2.0-105045177101
WoS UID
001820609100001