Identification of damage in steel beam by natural frequency using machine learning algorithms

Abstract

In recent times, the efficacy of machine learning (ML) algorithms as tools for forecasting structural damage has become increasingly evident. However, input data in structural health monitoring predominantly comprises normal operational states or states with minor deviations from the initial condition, lacking potentially hazardous states. Consequently, creating a realistic dataset for machine learning models to identify structural damage poses a challenge. If such data were obtainable, it might involve parameters like stress intensity factor range and stress ratio, which are often difficult to measure within real structures. In this paper, ML models, including Artificial Neural Network (ANN), Extreme Gradient Boosting (XGB), and Random Forest (RF), were constructed to predict the locations, widths, and depths of saw-cuts in steel beams. The prognostications were based on fluctuations in natural frequencies. The natural frequencies under various damage scenarios were identified using the Finite Element Method (FEM). The natural frequencies in the absence of saw-cuts, obtained from the two methods, Finite Element Method (FEM) and Frequency Domain Decomposition (FDD), were compared to validate their agreement. Conclusions regarding the selection of appropriate machine learning models, as well as the combination of FEM, FDD, and machine learning methods, will be drawn upon completion.