Brain-computer interfaces (BCIs) are a fruit of an impressive and long collaboration between fields of neuroscience and signal processing. The purpose of these complex systems is to interpret measured brain activity into useable commands and actions through implementation of different feature extraction, selection, and classification techniques. Depending on the application of a BCI and the exploited type of brain activity a specific set of methods would be implemented. In this thesis, electroencephalography (EEG) signals containing motor imagery (MI) information are analysed using a spatial-temporal technique called dynamic mode decomposition (DMD).
MI-EEG signals can be mainly described through two different types of brain activities: event-related de-/synchronisation (ERD/S) and event-related potentials (ERPs). The studies covered in this thesis focus on the former activity which exhibits strong characteristics in temporal, spectral and spatial domains. Despite being well described in the aforementioned domains, current state-of-the-art feature extraction techniques focus either on spectral (power spectral density (PSD) or bandpower) or spatial (common spatial patterns, CSP) side or on the combination of temporal and spectral domains (spectrograms and scalograms). The introduction of DMD aims to address the lack of more spatial-oriented techniques and three different feature types were explored to extract features based on ERD/S phenomenon. Firstly, standard DMD modes are used to accomplish that task. The measured performance, while being relatively low, still provided valuable information into the correct processing routes for DMD technique. With this knowledge, novel DMD spectrum features were extracted to cover a spatial-spectral domain combination. Despite the literature’s suggestions and links of DMD spectrum to average Fast-Fourier transfroms (FFTs), the perceived performance clearly indicated that DMD spectrum is unfit to extract ERD/S features from MI-EEG signals. Lastly, novel implementation of DMD maps with convolutional neural network (CNN) aimed to fully exploit spatial characteristics of ERD/S phenomenon was not able to successfully do so. Even though
all three proposed hypotheses were rejected based on the evidence seen from classification accuracy and kappa values, the author argues that DMD technique is still at the early stages of development and, given time and enough research, the performance of DMD modes and maps can be greatly improved.
|Date of Award||17 Oct 2022|
- University Of Strathclyde
|Sponsors||EPSRC (Engineering and Physical Sciences Research Council) & University of Strathclyde|
|Supervisor||Bernard A Conway (Supervisor) & Gaetano Di Caterina (Supervisor)|