Recently, machine learning (ML) algorithms have proven to be highly effective tools for predicting structural damage. However, the data used in structural health monitoring often consists primarily of normal operational conditions or slight deviations from the original state, with a scarcity of data representing potentially dangerous conditions. This limitation makes it challenging to create realistic datasets for training ML models to detect structural damage. If such data were available, it would likely involve parameters like the stress intensity factor range and stress ratio, which are difficult to measure in real-world structures. In this study, a random forest (RF) model was developed to predict the locations, widths, and depths of saw-cuts in steel beams based on variations in natural frequencies. These natural frequencies under various damage scenarios were determined using the Finite Element Method (FEM). To ensure accuracy, the natural frequencies in the undamaged state were compared between the FEM and Frequency Domain Decomposition (FDD). After training, the RF model showed an R-squared value of 0.996 for location, 0.876 for width, and 0.880 for depth. The mean squared error (MSE) was found to be 0.0003 for location, 0.0313 for width, and 0.0420 for depth. The results indicate that combining the FEM and FDD with the RF model holds significant potential for applications in structural health monitoring.