علي عابدي

The field of structural health monitoring is rapidly evolving to meet the growing global demand for efficient and proactive maintenance strategies, achieved through the use of innovative technologies that ensure the safety and sustainability of critical infrastructure. These infrastructures are increasingly being equipped with monitoring systems that generate vast amounts of data on structural behavior, damage, and condition, enabling data-driven analyses and assessments. However, sensor errors often introduce anomalies into the data. Therefore, automatic damage and anomaly detection are essential for reliable monitoring. While deep learning methods such as convolutional neural networks have been effective in damage detection and anomaly classification, two key challenges remain: (1) the issue of saturation in traditional models as model complexity increases, and (2) class imbalance. This study introduces a novel vision-based framework to address these challenges, leveraging a vision transformer and the fusion of multi-domain features. Initially, raw time-series data are divided into hourly intervals and visualized in three primary domains: time, frequency, and time-frequency. These domains are then integrated into a single three-channel image. The performance of this framework was eva‎luated using accelerometer data from two structural health monitoring systems of long-span cable-stayed bridges. The results indicate that the proposed approach exhibits excellent performance in both damage detection and anomaly classification, demonstrating its competitiveness with other networks.

پايش سلامت سازه , يادگيري عميق , ويژن‌ترانسفورمر , بينايي كامپيوتر , شناسايي آسيب , روش مبتني بر ارتعاش

Structural Health Monitoring , Deep Learning , Vision Transformer , Computer Vision , Damage Detection , Vibration Base Approaches

Ali Abedi

Dr. Gholamreza Ghodrati Amiri