Towards Control of EEG-Based Robotic Arm Using Deep Learning via Stacked Sparse Autoencoder

摘要

Recently, an EEG-based robotic arm control has offered a key solution to the problem of high level amputation or severe neuromuscular damage. Several attempt to use EMG-based pattern recognition (PR) failed due to insufficient myoelectric signals from the residual limb to perform control functions. The EEG activity recorded from human scalp is used to control the movement of a robotic arm. This can either be achieved when the arm is attached to or separated from the amputee stump by interfacing the brain directly to the robotic arm through the brain-machine interface (BMI). To build an intelligent robotic (or prosthetics) system that will manipulate object seamlessly with multiple degrees of freedom (DoF), it is required that a robust learning algorithm which is able to control the prosthetic arm while interacting with the environment should be implemented. However, the conventional machine learning approach of using handcrafted features to design a robot controller that can perform multiple task is not a feasible option. Hence, we propose a robust learning control which is based on unsupervised learning algorithm of deep autoencoder. We applied stacked autoencoder to generate our features, and softmax layer is then used to classify five different motor imagery tasks. The proposed method produced an overall accuracy of 98.9% across the four amputees recruited for the experiments. Our algorithm shows a better performance when compared with the state-of-the art classifiers. Thus, the proposed results demonstrates the possibility of providing better control performance for EEG-based prosthesis.