Damage Detection of Frame Structures Based on Acceleration Using Deep Learning

In this work, a damage detection method for frame structures based on time-history acceleration data using a deep learning (DL) technique is presented. For that aim, a dataset randomly created by finite element analysis (FEA) is employed to build the learning model based on a deep neural network (DNN). In which, inputs are the time-series acceleration at several degrees of freedom (DOFs) of a structure, while outputs are damage ratios of frame members. The accuracy of the trained and tested DL models is continuously updated by eliminating low-risk members which are predicted in a previous DL model via a damage threshold. Accordingly, the proposed methodology can reliably diagnose the location and severity of damaged members. Several examples programmed by Python are exhibited to validate the reliability of the suggested paradigm.