Construction of neural nets in brain-computer interface for robot arm steering

The paper presents an attempt to construct a direct brain-to-machine interface (BCI) that could be used to control movements of a robotic arm. A series of experiments was performed to collect data from subjects, train and validate an effective neural network and search for a generalized solution. The ERD/ERS imagery paradigm was chosen to extract valid data from the EEG signal. Since categorizing between two opposite states (like Left-Right) proved to be the most reliable, a control structure containing multiple neural networks was proposed. New research concerns the method of development and the use of neural networks classifiers. An automated procedure was used to select the best bipolar classifiers from the set of machine-generated neural networks. The chosen classifiers were used in a hierarchical structure responsible for signal interpretation. Adoption of this method was motivated mainly by the complexity of arm movement. The movement consisted of several phases, such as the initiation, continuation, change of direction, change of speed. It was observed that simple bipolar classifiers produced better output than classifiers designed to recognize complex decisions.