Capsule Neural Networks for structural damage localization and quantification using transmissibility data

Abstract

One of the current challenges in structural health monitoring (SHM) is to take the most advantage of large amounts of data to deliver accurate damage measurements and predictions. Deep Learning methods tackle these problems by finding complex relations hidden in the data available. Amongst these, Capsule Neural Networks (CapsNets) have recently been developed, achieving promising results in benchmark Deep Learning problems. In this paper, Capsule Networks are expanded to locate and to quantify structural damage. The proposed approach is evaluated in two case studies: a system with springs and masses that simulate a structure, and a beam with different damage scenarios. For both case studies, training and validation sets are created using Finite Element (FE) models and calibrated with experimental data, which is also used for testing. The main contributions of this study are: A novel CapsNets-based method for dual classification–regression task in SHM, analysis of both routing algorithms (dynamic routing and Expectation–Maximization routing) in the context of SHM, and analysis of generalization between FE models and real-life experiments. The results show that the proposed Capsule Networks with dynamic routing achieve better results than Convolutional Neural Networks (CNN), especially when it comes to false positive values.