Institutional Repository
Technical University of Crete
EN  |  EL



My Space

A deep machine learning architecture for filtering and classification of motor imagery EEG data

Lytridis Nikolaos

Full record

Year 2020
Type of Item Diploma Work
Bibliographic Citation Nikolaos Lytridis, "A deep machine learning architecture for filtering and classification of motor imagery EEG data", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2020
Appears in Collections


The Brain-Computer Interface (BCI) is a scientific field, which with the development of new technologies is gaining more and more interest. Research in this area concerns both applications of medical interest (e.g. for people with mobility problems) and applications with multiple uses in everyday life (e.g. driving assistance system to avoid obstacles). A group of BCI applications that provide solutions to both of the above categories are those based on Motor Imagery. Each BCI application includes the recording and pre-processing of brain signals, as well as the use of machine learning methods in order to classify signals for certain tasks. The main problem that researchers face is the non-stationary nature of electroencephalogram (EEG) signals, which poses many challenges in the successful pre-processing and classification of signals. This diploma thesis investigates the performance of various electroencephalogram (EEG) signal classification systems, derived from a public data set based on BCI for Motor Imagery. As part of the work, an original integrated system was implemented, in a unified model, trainable from end to end, with two basic components: a spatial filtering method (CSP) and a classifier based on a deep neural network. The neural network parameters are adjusted directly during the training, so that the classification creates highly distinguishing features directly from the EEG signals. The structure of the system is such as to deal with both the spatial and temporal variability that exist naturally in the problem of classifying EEG signals. The proposed structure is compared to other common structures on the same data set. A point of reference is the comparison with the recent OPTICAL architecture that basically uses CSP and LSTM and seems to excel over the rest. The results reinforce the belief that the problem is generally user-dependent. It seems, however, that the proposed methodology, which shows comparable performance, could give excellent results in future research, given its flexibility in structure and configuration, with appropriate per-user optimization.

Available Files