Identification and Quantization of Multiple Damages on 1-Dimensional Beam by Minimizing Wasserstein Distance with Sensitive Modal Amplitude Data

Abstract

A novel Predictor-Critic Artificial Neural Network architecture is presented here which minimizes Wasserstein Distance simultaneously with Jensen Shannon Divergence, Mean Squared Error and Mean Absolute Error as regression losses. Modal Amplitude data which is input to Artificial Neural Network, has been made sensitive by elastically striding strip of tiny mass. Damage amount is converted into Probability Mass Function. Proposed Artificial Neural Network not only predicts location of multiple damages but also predicts their amount with great accuracy. The present methodology is verified on a 1-D beam with four different boundary conditions- cantilever, simply supported, overhanging and propped cantilever. Baseline Modal Amplitude Data is not required for the proposed approach.