Development of feature extraction algorithms for real-time brain computer interface

Arnin, Jetsada

Thesis

Development of feature extraction algorithms for real-time brain computer interface

Download PDF

Creator

Arnin, Jetsada

Rights statement

Strathclyde Thesis Copyright

Awarding institution

University of Strathclyde

Date of award

2022

Thesis identifier

T16168

Person Identifier (Local)

201766624

Qualification Level

Doctoral (Postgraduate)

Qualification Name

Doctor of Philosophy (PhD)

Department, School or Faculty

Department of Biomedical Engineering

Abstract

Brain-computer interface (BCI) research is a multidisciplinary field of biomedical engineering that incorporates information from neuroscience, computer engineering, and electrical engineering. A BCI attempts to decode the brain’s electrical activity to comprehend the neural system’s function and transform it into outputs that augment, replace, repair, or improve human activities. The primary focus of BCI research is on creating robust, dependable, and user-friendly real-time systems. One of the most challenging aspects of developing real-time BCI is minimising the delay introduced by processing pipelines, which often decreases a system’s computing performance. The thesis focuses on developing a real-time BCI signal processing method, particularly on the acceleration of the feature extraction phase. The purpose of the study is three-fold: a) to evaluate a feature extraction method that could be potential for real-time BCIs b) to modify the existing feature extraction algorithms in order to support real-time processing using different accelerating techniques c) to demonstrate the use of modified algorithms for real-time BCI applications. The first study examined potential methods for real-time BCI feature extraction using a publicly available EEG data set. The findings established a difference between time and frequency domain techniques for a BCI with two classes of motor imaginary. Furthermore, the key findings show that although varying feature extraction methods did not result in a significant increase in classification accuracy, they did result in a considerable decrease in computation time. An additional examination is parallel computing to address the remaining complexity issues in the selected feature extraction methods. As a result, boosting computing time using a parallel processing technique is theoretically possible and flexible enough for real-time processing. However, in order to further reduce system latency, both hardware and software must be optimised when utilised in a real-time environment. Additionally to the assessment findings, the discrete Fourier transform technique offered comparable classification accuracy but was computationally costly. This almost eliminates the possibility of processing in real time. Another study solved this problem by proposing a new technique known as the “enhanced sliding discrete Fourier transform (eSDFT).” The eSDFT method is optimal for processing time varying physiological data in real-time or near-real-time. The computational complexity analysis revealed that the proposed approach outperformed the traditional method. Furthermore, the suggested approach can be utilised for large-scale, real-time signal processing that uses a distributed computing infrastructure comprised of many compute units. Apart from the Fourier transform-based feature extraction method, a straightforward Fourier transform cannot be used effectively for BCI processing because of the high temporal resolution of EEG signals, which often results in the loss of time information. Hence, the discrete wavelet transform (DWT) was selected to study, as it operates in both the time and frequency domains. This study examined the feasibility of creating a parallel computing approach for extracting time-frequency domain characteristics. A modification of the DWT method has been proposed to enable parallel computing. The results show how effective hardware acceleration is in handling enormous computational demands. The suggested processing technique may be utilised to speed up any signal processing methodology. The last study shows how to handle real-time signal processing using recursive approaches and parallel computing technologies. On a sliding window basis, the first and second generations of DWT algorithms were chosen for research. A variation of the conventional DWT method has been suggested to minimise duplication in the computation of coefficients. The findings indicate that the recursive approach has the shortest runtime and the most outstanding speedup ratio for all input lengths. Additionally, the proposed technique may be used for a wide variety of wavelet functions without degrading performance.

Advisor / supervisor