Damage detection of civil structures through machine learning approaches
Résumé
Structural Health Monitoring (SHM) has become a potential solution for early damage detection in civil structures, particularly with the deterioration of aging bridges. Vibration-based SHM, among all approaches, is widely adopted due to its ability to capture the dynamic response of structures. Lately, the increasing volume of sensor data and development of machine learning (ML) and deep learning (DL) algorithms triggered an interest in their application to SHM. Many studies focus on the performance demonstrated by these algorithms without systematically comparing their relative strengths, trade-offs, and resource efficiency particularly in the context of real-world SHM and resource constrained edge deployment. This study provides a comprehensive evaluation of ML and DL models for multiclass damage detection using the Z24 benchmark dataset. We analyze classification performance, noise robustness, prediction time, resource usage, and model complexity on acceleration time series. Our findings show that while DL models excel in complex damage classification and noise robustness, well-tuned simpler ML models like STSF and ROCKET can achieve comparable or superior results in specific scenarios with lower resource requirements. Notably, we find that down-sampling the original time series improves resource efficiency without sacrificing performance. In addition, some efficient algorithms that optimize both accuracy and computational cost are presented, offering new perspectives for practical implementation of SHM systems.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |