Outlier ensembles: A robust method for damage detection and unsupervised feature extraction from high-dimensional data

Outlier ensembles are shown to provide a robust method for damage detection and dimension reduction via a wholly unsupervised framework. Most interestingly, when utilised for feature extraction, the proposed heuristic defines features that enable near-equivalent classification performance (95.85%) when compared to the features found (in previous work) through supervised techniques (97.39%) — specifically, a genetic algorithm. This is significant for practical applications of structural health monitoring, where labelled data are rarely available during data mining. Ensemble analysis is applied to practical examples of problematic engineering data; two case studies are presented in this work. Case study I illustrates how outlier ensembles can be used to expose outliers hidden within a dataset. Case study II demonstrates how ensembles can be utilised as a tool for robust outlier analysis and feature extraction in a noisy, high-dimensional feature-space.