Large-Scale Structural Anomaly Detection During Seismic Events Using Optical Flow and Transfer Learning from Video Data

Abstract

Civil structures inevitably experience anomalies and damage, especially during disasters like earthquakes, tsunamis, and hurricanes, causing performance degradation or even collapse. Identifying such anomalies plays an extremely critical role in the maintenance and life extension of civil structures. This study proposes a novel approach based on video data due to its accessibility and rich temporal-spatial information for anomaly detection in large-scale civil structures by integrating transfer learning (TL) techniques with optical flow. Given the low importance of structural Region-of-Uninterest (RoU) like windows and doors, TL with BEIT+UPerNet pre-trained models identifies them. The extended node strength network then leverages video data to focus on structural components and detect disturbances in the nonlinearity vector field. The approach was validated using open video data from E-Defense, capturing two large-scale structural shaking-table tests that featured both pronounced shear cracks and tiny cracks. The detection and quantitative analysis results confirmed the method’s effectiveness in detecting structural anomalies and improved computational efficiency by approximately 10%, with a positive correlation observed between this efficiency gain and the proportion of structural RoUs in the video. This study advances anomaly detection in large-scale structures, offering a promising approach to enhancing safety and maintenance practices for critical infrastructure.

Taisei Saida

Ph.D Student (JSPS DC2 Researcher)

Topics of interest will be structural reliability, structural health monitoring, and machine learning.