|
Title |
AI-based Real-time Event Classification Framework of Cable-stayed Bridge Utilizing Probability-based Data Augmentation
|
|
Authors |
조연상(Younsang Cho) ; 강만성(Man-Sung Kang) ; 김호진(Hojin Kim) ; 김재환(Jaehwan Kim) ; 안윤규(Yun-Kyu An) |
|
DOI |
https://doi.org/10.11112/jksmi.2025.29.5.84 |
|
Keywords |
구조물 건전성 모니터링; 통계 기반 데이터 증강; 공간 센서 행렬; Gumbel 분포; 이벤트 분류; IoT 센서 Structural health monitoring; Probability-based data augmentation; Spatial sensor matrix; Gumbel distribution; Event classification; IoT sensor |
|
Abstract |
This study proposes an automated structural event (ship collision, typhoons, earthquakes, etc.) classification model that applies a statistical-based probabilistic data augmentation technique to overcome the learning limitations of rare events by utilizing data collected via IoT sensors for real-time safety diagnosis of cable-stayed bridges. Representative data were extracted from actual measurement data to analyze abnormal signals and rare events occurring on real bridges. Adaptive thresholds were selected for each sensor based on the extracted representative data per sensor channel to determine the presence of abnormal signals. Based on the values classified according to the presence of abnormal signals, a Sensor Spatial Matrix (SSM) was constructed, and the patterns of the SSM observed for event were analyzed. For the SSM patterns, a Gumbel distribution-based probabilistic modeling was used to statistically reproduce the active patterns of rare events, generating synthetic data for training. Generating synthetic training data confirmed the ability to address the problem of sparse data for rare events on bridges. Applying this data augmentation technique to train a CNN-based event classification model ensured stability in model training and reduced the tendency for the trained model to overgeneralize. Field experiments conducted on the Jindo Bridge in Jindo, South Korea, achieved a classification accuracy of 97.46%. This research enhances the real-time automatic diagnostic capability of Structural Health Monitoring (SHM) and presents a methodology to compensate for insufficient data on rare events.
|