Abstract

New understanding of the brain function and increasing availability of low-cost-non-invasive
electroencephalograms (EEGs) recording devices have made brain-computer-interface (BCI)
as an alternative option to augmentation of human capabilities by providing a new non-muscular channel for sending commands, which could be used to activate electronic or
mechanical devices based on modulation of thoughts. In this project, our emphasis will be on
how to develop such a BCI using fuzzy rule-based systems (FRBSs), metaheuristics and Neural
Mass Models (NMMs). In particular, we treat the BCI system as an integrated problem
consisting of mathematical modelling, machine learning and classification. Four main steps are
involved in designing a BCI system: 1) data acquisition, 2) feature extraction, 3) classification
and 4) transferring the classification outcome into control commands for extended peripheral
capability. Our focus has been placed on the first three steps.
This research project aims to investigate and develop a novel BCI framework encompassing
classification based on machine learning, optimisation and neural mass modelling. The primary
aim in this project is to bridge the gap of these three different areas in a bid to design a more
reliable and accurate communication path between the brain and external world.
To achieve this goal, the following objectives have been investigated: 1) Steady-State Visual
Evoked Potential (SSVEP) EEG data are collected from human subjects and pre-processed; 2)
Feature extraction procedure is implemented to detect and quantify the characteristics of brain
activities which indicates the intention of the subject.; 3) a classification mechanism called an
Immune Inspired Multi-Objective Fuzzy Modelling Classification algorithm (IMOFM-C), is
adapted as a binary classification approach for classifying binary EEG data. Then, the DDAG-Distance aggregation approach is proposed to aggregate the outcomes of IMOFM-C based
binary classifiers for multi-class classification; 4) building on IMOFM-C, a preference-based
ensemble classification framework known as IMOFM-CP is proposed to enhance the
convergence performance and diversity of each individual component classifier, leading to an
improved overall classification accuracy of multi-class EEG data; and 5) finally a robust
parameterising approach which combines a single-objective GA and a clustering algorithm
with a set of newly devised objective and penalty functions is proposed to obtain robust sets of
synaptic connectivity parameters of a thalamic neural mass model (NMM). The
parametrisation approach aims to cope with nonlinearity nature normally involved in
describing multifarious features of brain signals.