Deep convolution neural network for damage identifications based on time-domain PZT impedance technique

Abstract

Recently, Intelligence-based structural health monitoring (SHM) methods have investigated widely. Most of these methods are for detecting and classifying different structural damages by the means of features extraction from the structural responses signals, for instance different back propagation artificial neural networks SHM based methods. However, automatic features extraction, that eliminates the need for expertise and performing visual inspection to evaluate structures status is still a big challenge. In this study, therefore, a novel convolution neural network-based algorithm along with a hybrid training method has been proposed to detect, quantify and localize structural damage. The proposed method has been evaluated experimentally, many damaged and undamaged structural conditions have been conducted, acquiring samples of time-domain PZT impedance response signals from a beam. As the results show that, the method obtained a significant execution on damage detection, damage size evaluation and damage location recognition with high accuracy and reliability.