Identification of Exciting Force on Structures Using Vibration Characteristics and Deep Learning Algorithms

Abstract

This study presents a novel method for force identification, as force-location and force-history, in one and two-dimensional structures by utilizing vibration characteristics and the combination of deep learning algorithms. The force-location is first determined by obtaining the vibration responses from the structure and transforms these responses to one image, by proposed Image Mapping process, to be able to fit and test into the trained deep learning algorithm. The training dataset is the image mapped of frequency responses of same structure which are obtained using mathematical models such as the Spectral Element Method (SEM) or Finite Element Method (FEM). By using the relation of equality between dividing two frequency responses are obtained, numerically, from two points and dividing the two vibration responses obtained, experimentally, from the same two points in frequency-domain, we could exploit the effective of the deep learning algorithm to capture the similar pattern among several data for purpose localization the force. Convolutional Neural Network (CNN) learning algorithm was used to identify the force-location in 1D structure, and in the case of 2D structures, Deep Convolutional Generative Adversarial Networks (DCGANs) was also used. The force-history is then identified by using inverse problem with apply Artificial Damping Technique instead of apply filters to avoided ill-conditioned such Truncated Singular Value Decomposition (TSVD) or Tikhonov filter which lead to loss useful information from the signal. The results appear high accuracy expectations in force-location identification and somewhat high prediction in a reconstruction of force-history identification.