Unsupervised Novelty Detection Techniques for Structural Damage Localization: A Comparative Study

Abstract

In recent decades, novelty detection has attracted considerable attention in the structural health monitoring field. Numerous machine-learning algorithms have been proposed to carry out novelty detection for structural damage detection and localization, owing to their abilities to identify abnormal data in large numbers of datasets. This paper introduces a number of unsupervised novelty detection methods, based on machine learning, and applies them to localize different types of damage in a laboratory scale (lab-scale) structure. The key concept behind unsupervised novelty detection is that the novelty detection model must be trained using only normal data. In this study, the model used to identify abnormal data in the testing datasets has been well trained using normal data. The unsupervised novelty detection methods in this paper include the Gaussian mixture method, one-class support vector machines, and the density peak-based fast clustering method, which was developed recently. To enable these methods to carry out novelty detection and to increase their localization accuracy, a number of improvements have been made to the original algorithms. In this comparative study of structural damage localization, two damage-sensitive features are extracted from the acceleration signals measured by the sensors installed on a complex lab-scale structure. The advantages and disadvantages of these methods are analyzed based on experimental comparative case studies of damage localization.