Change Detection and its closely associated sister fields provide fundamental
components for many vital applications such as quality control, data mining, power distribution, network intrusion detection and adaptive classification.
There is a tremendous body of research in statistics, quality control, data
mining and applied areas that has contributed to a diverse arsenal of change detectors.
Whilst there has been a greater focus on the univariate problem, there
are many approaches to the more challenging problem of multivariate change
detection. Novel change detection methods continue to be actively developed.
Supervised change detection methods have a clear pathway to improvement,
by training on labelled data. However, there are a number of problems for which abundant labelled data is scarce or unavailable. For these problems, an unsupervised approach must be taken using incoming data. It is proposed here to develop general, composable modules to improve on the existing methods for unsupervised multivariate change detection. The modules should be composable such that they can all be applied together without interfering with each other.
This thesis proposes three such modules. Firstly, Principal Components Analysis
(PCA) is assessed as a general purpose feature extraction and selection
step. Secondly, it is proposed to chain univariate change detection methods to
multivariate criteria, such that they act as adaptive thresholds. Finally, univariate
change detectors are built into subspace ensembles where each detector monitors a single feature of the input space, allowing them to function as a multivariate change detector. These three modules are jointly assessed against
a challenging problem of unsupervised endogenous eye blink detection.