On robust statistical outlier analysis for damage identification

This thesis aims to contribute towards the development of reliable and accurate damage detection monitoring frameworks, applicable for a range of structural health and condition monitoring problems. Central to this purpose, is to be able to detect damage patterns embedded in a system's vibration signal responses sufficiently early. This will enable a condition-based maintenance and inspection to be carried out so as to prevent potentially catastrophic events, as related to each application domain. Firstly, to obviate reliance on data labels, an inclusive outlier analysis study is conducted by means of robust multivariate statistical analysis and a range of other (more common) outlier detection techniques, in both multivariate and time-series settings. Given the parametric nature of robust multivariate statistical techniques, it has also been possible to characterise outliers according to their influence on a method's estimates. Secondly, novelty detection is explored, in which a set of samples representing the nominal state of the system, is assumed to be available. This set includes observations from a system with its dynamics being significantly influenced by environmental and operational variability. Finally, this thesis explored the potential of utilising certain robust techniques as a pre-processing step on damage sensitive features (contaminated with outliers) for novelty detection tasks. Given the large volume of observations, both experimental and computational, different damage sensitive features were extracted, some of which were specific to the range of problems / types of damage being investigated. The performance, in terms of both sensitivity in damage detection and immunity to environmental and operational variability, was assessed for each damage sensitive feature, in combination to the outlier and novelty detection technique used. This thesis has introduced to the condition and structural health monitoring fields a range of methods from robust statistics with attractive properties, such as the effective unmasking of outliers