Loading...

Development of Motor Imagery Based Real-Time BCI for Lower-Limb Exoskeleton

Koureshi Hazrat, Arya | 2025

88 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58757 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Ghazizdeh Ehsaei, Ali; Shamsollahi, Mohammad Bagher
  7. Abstract:
  8. Severe motor impairments resulting from neurological injuries challenge the quality of life for millions. Brain-Computer Interface (BCI) technology offers a novel solution by creating a direct communication channel between the brain and external devices, thereby bypassing damaged neuro-muscular pathways. This research focuses on the development of a Brain-Computer Interface (BCI) system based on Motor Imagery (MI), utilizing non-invasive Electroencephalogram (EEG) signals. Aimed at controlling lower-limb exoskeleton robots, this system decodes the user's motor intention and rapidly translates it into physical movement. This capability, by facilitating sensory feedback and stimulating neuroplasticity through the robot, offers the potential to accelerate the neurorehabilitation process. The primary challenge of this research is the accurate real-time classification of five knee motor imagery states (right/left knee up/down, and rest). Differentiating these intra-limb movements is particularly difficult due to the deep and overlapping cortical representation of the lower limbs. Furthermore, given the low signal-to-noise ratio and high variability of EEG signals, achieving a system with low processing latency is a key requirement for real-time robot control and creating a natural sense of control for the user. While deep learning models have achieved high accuracies, their heavy computational demand is a barrier to implementation on portable hardware. Therefore, this study focuses on optimizing computationally efficient classical machine learning algorithms. To this end, EEG data were collected from six healthy subjects over three sessions using a 30-channel device with a sampling rate of 250 Hz. The experimental protocol included a five-state cue implemented with the Psychtoolbox in MATLAB. The signal preprocessing pipeline consisted of filtering (4-40 Hz), Common Average Reference spatial filtering, artifact removal, and normalization. A comprehensive set of temporal, frequency, connectivity, nonlinear, wavelet, and Common Spatial Pattern features was extracted, and the best features were selected using the LASSO method. Six classical machine learning algorithms (LDA, KNN, SVC, RandomForest, Logistic Regression, XGBoost) were evaluated using Stratified K-Fold cross-validation (K=10) and metrics of accuracy, Cohen's kappa, and F1-score. To simulate online performance, the analysis was conducted on sliding time windows of varying lengths (0.5, 1, and 2 seconds). The results showed that the Support Vector Classifier (SVC) model achieved the best performance, reaching a mean accuracy of over 70% in 2-second time windows for some subjects. This accuracy is highly competitive compared to previous studies on multi-class and intra-limb classification of lower-limb movements. Analysis of the confusion matrices revealed that most errors occurred in discriminating between similar movements (e.g., up/down), while the distinction between rest and motion states was highly successful. From an online processing perspective, the total processing time for a 2-second window was between 193 and 204 milliseconds, enabling the transmission of approximately 5 control commands per second to the robot. This low latency, achieved with optimized classical algorithms, is entirely acceptable for real-time control applications and represents a significant step towards creating intelligent exoskeletons and improving the quality of life for individuals with motor disabilities
  9. Keywords:
  10. Brain-Computer Interface (BCI) ; Motor Imagery (MI) ; Lower Limb Exoskeleton ; Real-Time Classification ; Nerve Injury ; Electroencephalogram Signal Analysis

 Digital Object List

 Bookmark

No TOC