Robust methods for outlier detection and regression for SHM applications

Abstract

In this paper, robust statistical methods are presented for the data-based approach to structural health monitoring (SHM). The discussion initially focuses on the high level removal of the 'masking effect' of inclusive outliers. Multiple outliers commonly occur when novelty detection in the form of unsupervised learning is utilised as a means of damage diagnosis; then benign variations in the operating or environmental conditions of the structure must be handled very carefully, as it is possible that they can lead to false alarms. It is shown that recent developments in the field of robust regression can provide a means of exploring and visualising SHM data as a tool for exploring the different characteristics of outliers, and removing the effects of benign variations. The paper is not, in any sense, a survey; it is an overview and summary of recent work by the authors.