Basic Experimental Study on the Application of Natural Frequency to Transfer Learning

Abstract

Integrating artificial intelligence (AI) into structural health monitoring systems significantly improves the structural integrity and safety of buildings. This integration necessitates extensive data on various structural damage scenarios; however, acquiring comprehensive damage-state data for real-world buildings is difficult, resulting in data scarcity and imbalances between intact- and damage-state datasets. Transfer learning offers a compelling alternative that enables the utilization of damage-state data acquired from scaled experimental models for real-world building applications. This study explores the use of natural frequency ratios before and after damage to facilitate transfer learning. This study established an equation that describes the relationship between the natural frequency ratio before and after damage, accounting for scaling impacts. The natural frequency ratios for both the prototype and scaled models were experimentally determined and compared, focusing on bolted joints in the steel structure. The results of the experiment were consistent with those of the finite-element analysis. The experimentally obtained natural frequency ratios of the prototype and scaled models under identical damage conditions exhibited high congruence. The experimental and FEA results demonstrated analogous patterns of decreasing natural frequency ratio with increasing damage severity. These results indicate that natural frequency ratio data from various damage conditions in scaled models could mitigate data scarcity issues and train AI models for real-world building applications.