Structural Damage Detection Using Reduced Free Vibration Data and Deep Learning

This work presents a damage detection method for trusses using reduced free vibration data and multiple deep neural networks (DNNs). For this aim, a dataset randomly created by finite element analysis (FEA) is employed to build the DNN model. Inputs are a reduced free vibration dataset only including eigenvalues at several degrees of freedom (DOFs) of a few first modes, while outputs are damage ratios of truss members. Accordingly, the DNN requires a simpler architecture and less computational cost for the training and testing processes. By eliminating low-risk members via a damage threshold, the subsequently trained and tested DNN models become more accurate in predicting the location and severity of damaged members. A 2D truss programmed by Python is tested with two different damage scenarios to verify the reliability of the suggested approach.