Machine Learning Vibration-Based Damage Detection and Early-Developed Damage Indicators

The vibration-based damage detection of structures has been an active topic among many research fields for, at least, the past four decades. In its early years, great effort has been devoted towards developing damage indicators computed from experimental data. Recently, as this problem is increasingly studied with a data science perspective, efforts are shifting towards finding highly meaningful features in experimental vibrational data, which is key to Machine Learning success. The present work had the objective of analysing how the performance of a Machine Learning approach compares to and may benefit from early-developed damage indicators. A performance comparison is presented between some early-developed damage indicators, based on both frequency and mode shapes, and an Artificial Neural Network (ANN) supplied with vibrational data. Also, the use of the early-developed damage indicators as inputs in the ANN was investigated. A two-span simply supported beam was used as the numerical experiment for the tests. The performance of each approach in different damage scenarios was discussed, as well as insights were drawn by putting the Machine Learning vibration-based damage detection approach into perspective with some of the simplest early-developed damage indicators.