COMPARATIVE STUDY OF ANN CONFIGURATIONS FOR SAW-CUT DAMAGE IDENTIFICATION IN CANTILEVER STEEL BEAMS USING RFP-CALIBRATED FEM

Abstract

This study presents a hybrid framework for damage identification in cantilever steel beams by combining experimental modal analysis using the Rational Fraction Polynomial (RFP) method, Finite Element Method (FEM) simulation, and two Artificial Neural Network (ANN) models with different design philosophies: a data-driven optimized model (ANN_1) and a classical shallow network (ANN_2). A total of 427 FEM-generated and experimentally calibrated damage scenarios were divided into training, validation, and testing datasets, with the testing set consisting entirely of unseen damage locations to rigorously evaluate spatial generalization capability. Both ANN models were trained to predict crack location, width, and depth from the first ten natural frequencies. Results indicate that ANN_1 consistently outperformed ANN_2 on unseen test data, reducing mean absolute error by 24.6% for location, 35.0% for depth, and 13.0% for width of saw cut. Paired t-tests confirmed that these improvements were statistically significant for all target variables. In addition, Pearson correlation analysis revealed that crack width exhibited substantially weaker correlation with the input frequencies than crack location and depth, explaining its lower prediction accuracy and indicating an inherent limitation of frequency-based features rather than a weakness of the ANN architecture itself. The findings demonstrate that, under modern hyperparameter optimization and regularization strategies, data-driven ANN architectures provide superior predictive accuracy and generalization capability compared with traditional single-hidden-layer designs for vibration-based structural health monitoring applications.