Khanam, Taslima ORCID: https://orcid.org/0000-0001-5842-930X (2023) Developing artificial intelligence based algorithms for automatic analysis of brain signal data for advanced BCI systems. Research Master thesis, Victoria University.

Abstract

Motor disability is a limitation of a person's physical functioning, movement, ability or stamina to move and maintain balance. To eliminate this suffering from society, motor imagery (MI) based brain computer interface (BCI) performs a significant role. An MI-based BCI converts human intention into control signals to communicate with their external device through brain activity without direct physical movement. This brain activity can be recorded by electroencephalography (EEG). The EEG is a test that measures the electrical activity of the brain. EEG generates a vast amount of non-stationary, non-linear, and non-periodic signal data (called ‘brain signal data’). Undesirable signals, called noise, commonly contaminate these signal data. Therefore, it is necessary to remove the noise in the raw EEG data to acquire helpful information that reflects brain activities and mental states to identify motor disabled people’s intentions. The fast speed and accuracy of BCI systems for recognising MI activity or tasks based on EEG signals is another significant problem. Traditional techniques in BCI have some limitations. These include noise sensitivity, longer iteration, complex method, require enormous amounts of training data, time consuming, applicable only in small dataset which reduces the accuracy, efficiency and robustness of an anticipated method. Therefore, it is necessary to address these issues to develop optimised artificial intelligence (AI) based machine learning technique to identify human intentions of people with motor disabilities for an improved BCI system. Hence, this study aims to introduce AI-based algorithms for identifying human intentions of physical movements through EEG data in BCI’s development and to compare them with existing traditional methods. In this dissertation, we have developed three methods to fulfil our research objective by analysing two publicly available EEG-based BCI datasets: 1) a common spatial pattern (CSP) based Medium K-nearest neighbour machine learning approach 2) a hybrid method: CSP based optimised ensemble (OE) 3) Markov chain-based support vector method. A short detail of the three developed approaches and their contributions are described. Approach 1 (CSP-MKNN): To make a more robust approach with lower execution time, we utilized CSP algorithm for feature extraction technique and a MKNN for classification. For noise free data, we applied Butterworth filter ranges from 0.1 Hz to 4 Hz frequency and 5th- order derivations. Our advanced approach produced the highest score for all subjects more than 90% and achieved 3.4% to 14.52% advancements compared with the other existing approaches with the same dataset. Approach 2 (CSP-OE): To improve the performance, we reformed the CSP- MKNN model by presenting a CSP- based OE algorithm. Our proposed system achieved 99.64% overall accuracy and 1.24% to 26.14% enhancement compared to the earlier machine learning algorithms. Approach 3 (MC-SVM): We developed a Markov chain-based SVM algorithm to improve the classification of the MI tasks more directly and efficiently. Our proposed system produced the highest score for all individual subjects above 99%, compared with the existing prominent approaches. The tentative results proved great achievement for detecting MI task. The algorithms can help clinical diagnosis and restoration of motor- impaired persons with external devices.

Search Google Scholar

Repository staff login