NeuralGrip: Enhanced Motor Imagery Training and Testing for Assistive Robots Using Deep Learning and Advanced Data Augmentation Techniques

Abstract

About 1.3 billion people in the globe have some sort of motor disability. This amounts to sixteen percent of the entire human population. Disabling in nature, these conditions provide fresh difficulties to those who live with them every day. Assistive robots and other devices that can improve the lives of people with impairments have benefited greatly from the widespread usage of Motor Imagery Electroencephalography (MI-EEG) data in Brain-Computer Interface (BCI) systems. The low performance of brain signal decoding remains a hurdle for the BCI sector and limits its usefulness, despite the extensive research into EEG. In this study, we propose a deep learning-based 2-D convolutional neural network (CNN) for classifying motor imagery (MI). Not only is it possible to efficiently extract EEG-specific features from raw EEG data using a convolutional neural network, but it is also small enough to fit in the palm of your hand. We used BCI competition IV dataset 2a, a publicly accessible MI-EEG dataset, as our standard. The results show that the suggested model outperforms the baseline algorithms in its capacity to adapt to new settings. Additionally, it improves performance even when just a little amount of data is available for training. Furthermore, a data augmentation strategy is proposed to improve training and testing. The dataset is thoroughly examined using a variety of visualization methods before any augmentation is performed. Both un-augmented and augmented datasets are trained and evaluated using the proposed model, and the results are compared. When compared to other state-of-the-art methods, the model's average classification accuracy for four-class classification tasks was 82.13%.