Damage detection and localization in steel structural components using strain-based deep learning networks

Structural Health Monitoring (SHM) systems have been utilized in recent years for both damage detection and, more recently, for damage localization. Many proposed approaches integrate SHM systems with deep learning techniques. However, most of these approaches, specifically for damage localization, rely on supervised learning algorithms that require datasets of damaged state of the structure to train the deep learning model. Since collecting such data is impractical for real-world civil structures, using supervised learning techniques is unfeasible. While significant progress has been made in using vibration-based methods for detecting global damage in structures, these methods face challenges in accurately localizing the source of damage. This is due to the inherent nature of vibration-based methods, which are effective at identifying global structural characteristics but struggle with pinpointing the location of damage. In this thesis, we present the Zone Discriminator Network (Z-DNet), a new damage detection and localization approach inspired by Generative Adversarial Networks (GANs), particularly PatchGAN. The network consists of multiple discriminators, each associated with a specific zone within the structure (i.e., support zones, connection zones, high-stress zones, and locations of stress concentration). The approach works by penalizing the structure at the predefined zones and classifying whether the strain data at each zone corresponds to a real (undamaged) or fake (damaged) state. Z-DNet is trained in an unsupervised learning environment using strain-based deep learning. Unlike traditional methods, the approach does not require prior knowledge of damage states of the structure for damage detection and localization in steel structural components. This data-driven approach uses raw strain data from the SHM system as inputs in the deep learning network. A baseline deep learning model is trained using only healthy (undamaged) state datasets of the structure. Damage is detected by comparing the results of the unknown state dataset to the baseline model thresholds. The proposed approach is validated through large-scale experimental testing on a steel beam at Bert Cooper Civil Engineering Laboratory at Oklahoma State University (OSU), using three distinct validation datasets.