Machine Learning-Based Structural Damage 2 Identification Utilising Frequency Response 3 Function Peak Amplitude Ratios

Abstract

With the rapid advancement of technologies such as machine learning (ML) and edge computing, ML-integrated structural damage identification has gained significant traction. Modal-based methods are frequently preferred for these applications due to the inherent robustness of modal features. Frequency Response Functions (FRFs) obtained via experimental modal analysis are rich in damage-sensitive information; however, the resulting datasets often suffer from high dimensionality. This paper investigates the feasibility of employing dimensionally reduced features extracted from FRFs to train ML models, specifically to minimise computational overhead during the training phase. Experimental validation was conducted using a laboratory-scale plate test rig. The study evaluates the predictive accuracy, model size and training duration for three distinct feature sets: raw FRFs, FRF peak amplitudes, and FRF peak amplitude ratios. While all three feature sets maintained predictive accuracies above 99.5%, the FRF peak amplitude ratio set yielded the most compact model and the fastest training time. Therefore, utilizing the FRF peak amplitude ratio as a primary training feature provides an effective method to reduce computational overhead while preserving high accuracy in structural damage identification.